Methodol

Reijo Sund
Methodological Perspectives for

Register-Based Health System
Performance Assessment
Developing a Hip Fracture Monitoring System in Finland
National Research and Development Centre

for Welfare and Health
Research Report 174

Editorial board
Matti Heikkilä, chairman
Marjatta Bardy
Marko Elovainio
Mika Gissler
Riitta Haverinen
Timo Tuori
Matti Virtanen
The research presented in this series has been approved for publication after
undergoing a formal referee evaluation process.
© Author and STAKES
Translation of the Swedish abstract: Done Information Oy
Cover design: Harri Heikkilä

Layout: Christine Strid
ISBN 978-951-33-2114-7 (nid.)

ISSN 1236-0732
ISBN 978-951-33-2132-1 (PDF)
Gummerus Printing
Vaajakoski 2008
Finland
Dedicated to the memory of my father
Acknowledgements
First of all, I want to praise my beloved wife Anne, the queen of my heart and the
goddess of my soul, for giving a purpose to my life. I would also like to express my
gratitude to my mother Irja and my sister Paula for their unconditional love, and to
my goddaughter Saana-Emilia for bringing so much sunshine that this world has
become a happier place.
The preparation of this thesis would not have been possible without a research
grant for the project ’Register-based data in hospital outcomes research’ from
the Academy of Finland. I am grateful to Professor Pekka Rissanen and Adjunct
Professor Ilmo Keskimäki for guiding me to the world of health services research.
I also thank the collaborators within the ’Performance, effectiveness and cost of
treatment episodes’ -project led by Research Professor Unto Häkkinen. In addition,
I want to acknowledge the role of my other co-authors and the colleagues I have
worked with. Thank you; you know who you are.
I express my sincere appreciation to Professor Emeritus Seppo Mustonen. His
extraordinary creativity, talent, and experience combined with the ability to say
the right things at the right time have offered an invaluable source of inspiration
and support during my whole statistical career. Moreover, the general computing
environment provided by Survo, developed by Professor Mustonen, offered an
excellent platform for register-based data analysis.
I also thank Professor Risto Lehtonen, Professor Hannu Niemi and Professor
Esa Läärä for comments and discussions concerning the manuscript of this thesis.
I am thankful to the official reviewers of the dissertation, Adjunct Professor Timo
Alanko and Adjunct Professor Helka Hytti. Their comments helped me to further
simplify and clarify the message.
Finally, one thing has become clear to me during the preparation of this thesis.
It is extremely time-consuming to conduct and publish multidisciplinary scientific
research that aims to offer alternative solutions to pragmatic problems. Although
I had to make some compromises, I still hope that this thesis is able to reflect my
personal way of thinking as well as my enthusiasm for scientific research.
Research Report 174 Methodological Perspectives for Register-Based

STAKES 2008 Health System Performance Assessment
Abstract
Reijo Sund. Methodological Perspectives for Register-Based Health System

Performance Assessment. Developing a Hip Fracture Monitoring System in Finland.
STAKES, Research Report 174. Helsinki 2008. ISBN 978-951-33-2114-7
The resources of health systems are limited. There is a need for information
concerning the performance of the health system for the purposes of decision-
making. This study is about utilization of administrative registers in the context of
health system performance evaluation.
In order to address this issue, a multidisciplinary methodological framework
for register-based data analysis is defined. Because the fixed structure of register-
based data indirectly determines constraints on the theoretical constructs, it
is essential to elaborate the whole analytic process with respect to the data. The
fundamental methodological concepts and theories are synthesized into a data
sensitive approach which helps to understand and overcome the problems that are
likely to be encountered during a register-based data analyzing process.
A pragmatically useful health system performance monitoring should produce
valid information about the volume of the problems, about the use of services and
about the effectiveness of provided services. A conceptual model for hip fracture
performance assessment is constructed and the validity of Finnish registers as a
data source for the purposes of performance assessment of hip fracture treatment
is confirmed. Solutions to several pragmatic problems related to the development
of a register-based hip fracture incidence surveillance system are proposed. The
monitoring of effectiveness of treatment is shown to be possible in terms of care
episodes. Finally, an example on the justification of a more detailed performance
indicator to be used in the profiling of providers is given.
In conclusion, it is possible to produce useful and valid information on health
system performance by using Finnish register-based data. However, that seems
to be far more complicated than is typically assumed. The perspectives given in
this study introduce a necessary basis for further work and help in the routine
implementation of a hip fracture monitoring system in Finland.
Keywords: administrative registers, methodology, secondary data, hip fractures,

health services, performance, effectiveness, incidence, profiling, care episodes

Abstract in Finnish
Reijo Sund. Methodological Perspectives for Register-Based Health System Perfor-

mance Assessment. Developing a Hip Fracture Monitoring System in Finland [Me-
netelmällisiä näkökulmia rekisteriperusteiseen terveydenhuoltojärjestelmän vai-
kuttavuuden arviointiin. Lonkkamurtumien seurantajärjestelmän kehittäminen
Suomessa]. Stakes, Tutkimuksia 174. Helsinki 2008. ISBN 978-951-33-2114-7
Terveydenhuoltojärjestelmän resurssit ovat rajallisia. Jotta näitä resursseja pystyt-

täisiin hyödyntämään mahdollisimman tarkoituksenmukaisella tavalla, tarvitaan
tietoa terveydenhuoltojärjestelmän vaikuttavuudesta. Tässä tutkimuksessa tarkas-
tellaan kuinka suomalaisia rekisteriaineistoja voidaan käyttää päätöksentekoa tu-
kevan terveydenhuollon vaikuttavuustiedon tuottamiseen.
Tutkimuksessa kehitetään menetelmällinen viitekehys rekisteriperusteisen da-
ta-analyysin tueksi yhdistämällä eri tieteenaloilta periytyviä menetelmällisiä ideoi-
ta. Lähtökohtana on, että rekistereitä hyödyntävän tutkimusprosessin tulee edetä
myös aineiston ehdoilla, koska hallinnollisiin tarkoituksiin kerättyjen rekisteri-
aineistojen tietosisältö rajoittaa käyttömahdollisuuksia. Kehitetyn viitekehyksen
puitteissa näiden välttämättömien rajoitusten syitä ja seurauksia on mahdollista
käsitellä systemaattisella tavalla, joka auttaa ymmärtämään ja ratkaisemaan rekis-
teriperusteisessa tutkimuksessa tyypillisesti kohdattavia ongelmia.
Terveydenhuollon vaikuttavuutta arvioitaessa tarvitaan tietoa niin terveys-
ongelmien ilmaantuvuudesta, palvelujen käytöstä kuin hoidon laadustakin. Näitä
tarkoituksia varten tässä tutkimuksessa muodostetaan käsitteellinen malli lonkka-
murtumien seurantajärjestelmän tietotarpeille ja varmistetaan, että rekisteriaineis-
tojen validiteetti on näiltä osin riittävä. Useaan lonkkamurtumien ilmaantuvuu-
den seurantaan liittyvään käytännölliseen ongelmaan esitetään rekisteriaineistojen
kanssa sopusoinnussa oleva ratkaisu ja osoitetaan kuinka hoidon toteutumista voi-
daan arvioida rekisterien avulla koko hoitoketjun osalta. Lopuksi tarkastellaan
kuinka hoidon laatua on mahdollista vertailla eri sairaaloiden välillä käyttäen yksi-
tyiskohtaisemmin perusteltua vaikuttavuusindikaattoria.
Johtopäätöksenä voidaan todeta, että suomalaiset rekisteriaineistot tarjoa-
vat erinomaiset mahdollisuudet terveydenhuollon vaikuttavuutta koskevan tie-
don tuottamiseen, vaikka hyödyllisen ja tieteellisyyden kriteerit täyttävän tiedon
tuottaminen onkin varsin haasteellista. Tässä tutkimuksessa esitetyt näkökulmat
muodostavat perustellun lähestymistavan rekisteritietojen entistä tehokkaampaan
hyödyntämiseen ja auttavat lonkkamurtuman rutiiniluonteisemman seurantajär-
jestelmän kehittämisessä.
Avainsanat: hallinnolliset rekisterit, metodologia, toisen käden aineistot, lonkka-

murtumat, terveydenhuolto, vaikuttavuus, ilmaantuvuus, laatu, hoitoketjut
Methodological Perspectives for Register-Based Research Report 174

Health System Performance Assessment STAKES 2008
Abstract in Swedish
Reijo Sund. Methodological Perspectives for Register-Based Health System Perfor-

mance Assessment. Developing a Hip Fracture Monitoring System in Finland [Me-
todologiska perspektiv på registerbaserad utvärdering av effektiviteten i hälso- och
sjukvårdssystem. Utveckling av ett system för uppföljning av höftfrakturer i Fin-
land]. Stakes, Forskningsrapport 174. Helsingfors 2008. ISBN 978-951-33-2114-7
Hälso- och sjukvårdssystemets resurser är begränsade. För att dessa resurser skall
kunna utnyttjas på ett så ändamålsenligt sätt som möjligt, behövs information om
hälso- och sjukvårdssystemets effektivitet. I denna undersökning granskas hur
finska registermaterial kan användas för att producera uppgifter om hälso- och
sjukvårdens effektivitet som stöd för beslutsfattande.
I undersökningen utvecklas en metodologisk referensram till stöd för da-
taanalysen genom att kombinera metodologiska idéer från olika vetenskapsgrenar.
Utgångspunkten är, att en undersökningsprocess som utnyttjar register behöver
ske på materialets villkor, eftersom innehållet i registermaterial som insamlats för
administrativa ändamål begränsar materialets användningsmöjligheter. Med hjälp
av den utvecklade referensramen finns det möjlighet att hantera orsakerna till des-
sa oundvikliga begräsningar och deras följder på ett systematiskt sätt, som under-
lättar förståelsen och lösningen av de problem som är typiska för registerbasera-
de undersökningar.
Vid bedömningen av hälso- och sjukvårdens effektivitet krävs uppgifter om
hälsoproblemens incidens, användningen av tjänsterna och vårdens kvalitet. För
dessa ändamål utvecklas i denna undersökning en begreppsmodell för databeho-
vet när det gäller uppföljningssystem för höftfrakturer och bekräftas att validiteten
hos data i registermaterialet till dessa delar är tillräcklig. En metod som är förenlig
med registermaterialen och som löser flera praktiska problem som är förknippade
med uppföljningen av incidensen av höftfrakturer presenteras. Dessutom visas hur
genomförandet av vården kan utvärderas med hjälp av registren för hela vårdked-
jans del. Slutligen granskas hur vårdens kvalitet kan jämföras mellan olika sjukhus
med hjälp av en mer detaljerat underbyggd effektivitetsindikator.
Som slutsats kan man konstatera, att de finska registermaterialen erbjuder ut-
märkta möjligheter att producera uppgifter om hälso- och sjukvårdens effektivi-
tet, även om det innebär en ganska stor utmaning att försöka ta fram uppgifter
som uppfyller kriterierna för användbarhet och vetenskaplighet. De synpunkter
som presenteras i den här undersökningen bildar en underbyggd angreppsmetod
som gör det lättare än tidigare att utnyttja registeruppgifter på ett effektivt sätt och
hjälper vid utvecklandet av ett med rutinmässigt system för uppföljning av höftf-
rakturer.
Nyckelord: administrativa register, metodologi, andrahandsmaterial, höftfrakturer,

hälso- och sjukvård, effektivitet, incidens, kvalitet, vårdkedjor

Contents
Acknowledgements
Abstract
Abstract in Finnish
Abstract in Swedish
List of original publications............................................................................................ 13
1 Introduction......................................................................................................... 15
1.1 Aim of the study.............................................................................................. 16
1.2 Structure of the study...................................................................................... 17
2 Register-based data analysis............................................................................. 19

2.1 From technical data processing to a knowledge discovery process............... 19
2.2 Secondary data as a methodological problem............................................... 20
2.3 Prerequisites in the register-based data analysis............................................ 21
2.3.1 Principles of measurement.................................................................. 22
2.3.2 Information science............................................................................. 23
2.3.3 Statistical computing........................................................................... 25
2.3.4 Statistics................................................................................................ 26
2.3.5 Theory.................................................................................................. 27
3 Register-based health system performance assessment............................ 28

3.1 Finnish health system...................................................................................... 28
3.2 Information production in the Finnish context............................................ 29
3.3 Measurement of the performance of a health system................................... 29
3.4 Assumptions for the production of concrete health system
performance information .............................................................................. 30
3.4.1 Register-based health system performance monitoring.................... 31
4 Register-based data on hip fractures – sources and validity.................. 33

4.1 Hip fracture..................................................................................................... 33
4.1.1 Hip fracture treatment in Finland...................................................... 33
4.1.2 Register data on hip fractures in Finland........................................... 34
4.1.3 Previous register-based hip fracture studies in Finland.................... 35
4.1.4 Register data for the current study..................................................... 35
4.2 Validity of register data in the case of hip fracture........................................ 36
4.2.1 Completeness of registration.............................................................. 37
4.2.2 Conceptual model for hip fracture performance monitoring.......... 37
4.2.3 Consistency between prospective and register-based data................ 40
4.2.4 Conclusions on validity....................................................................... 41
5 Hip fracture incidence....................................................................................... 43

5.1 Aging-related hip fractures............................................................................. 44
5.1.1 A method for identifying a first aging-related hip fracture............... 45
5.2 Risk factor extraction...................................................................................... 47
5.3 Risk population data....................................................................................... 49
5.4 Hip fracture incidence between 1998 and 2002 in Finland........................... 51
5.5 Conclusions on hip fracture incidence monitoring...................................... 54
Research Report 174 Methodological Perspectives for Register-Based 11

6 Hip fracture treatment..................................................................................... 55
6.1 State space for a hip fracture care episode..................................................... 55
6.2 Effectiveness in terms of care process............................................................. 57
6.3 Multivariate responses in outcomes research................................................ 57
6.3.1 Traceplots for care episode visualization............................................ 58
6.4 Comparing the care episode profiles of subpopulations............................... 61
6.4.1 Risk adjustment................................................................................... 61
6.4.2 Summarization of the follow-up data................................................ 62
6.5 Conclusions on monitoring hip fracture treatment...................................... 64
7 Operative delay as a performance indicator................................................ 66

7.1 Effects of different operative delays on mortality.......................................... 66
7.2 Profiling of providers...................................................................................... 67
7.3 Adjusted effect of operative delay on mortality............................................. 72
7.4 Provider-level hypotheses............................................................................... 73
7.5 Conclusions on operative delay as a performance indicator......................... 76
8 Discussion.............................................................................................................. 77
References......................................................................................................................... 80
Original publications
12 Methodological Perspectives for Register-Based Research Report 174

List of original publications
[1] Sund R: Utilisation of administrative registers using scientific knowledge

discovery. Intelligent Data Analysis 2003, 7(6):501–519.
[2] Sund R, Nurmi-Lüthje I, Lüthje P, Tanninen S, Narinen A, Keskimäki I:

Comparing properties of audit data and routinely collected register data in
case of performance assessment of hip fracture treatment in Finland. Methods
of Information in Medicine 2007, 46(5):558–566.
[3] Sund R: Utilization of routinely collected administrative data in monitoring

of aging dependent hip fracture incidence. Epidemiologic Perspectives &
Innovations 2007, 4(2).
[4] Sund R, Kauppinen S: Kuinka laskea ikääntyneiden pitkäaikaisasiakkaiden

määriä rekisteritietojen perusteella? Sosiaalilääketieteellinen Aikakauslehti
2005, 42(2):137–144.
[5] Sund R: Lonkkamurtumien ilmaantuvuus Suomessa 1998–2002. Duodecim

2006, 122(9):1085–1091.
[6] Sund R, Liski A: Quality effects of operative delay on mortality in hip fracture
treatment. Quality and Safety in Health Care 2005, 14(5):371–377.
Articles are included in the thesis with permissions from the publishers.

1 Introduction
1 Introduction
The resources available to health systems are limited. The equitable, efficient, and
effective use of these resources requires administrative planning and political
willingness. Information is needed concerning the performance of the health
system for the purposes of decision-making [10].
Recently, a continuous evidence-based quality improvement aiming to “do
the right thing right” has become very popular [11]. Essentially, this approach can
be seen as an application of organizational learning and knowledge management
ideas to the health care context [12]. To fulfill the requirements of this paradigm,
information should be evidence-based, well-organized, pragmatic, task-specific,
and available when and where it is needed [13].
It is obvious that information systems have a central role in this kind of
information production. A comprehensive health information system should be
suitable for clinical work as well as nationwide health policy purposes [14, 15]. The
idea of such population data-based health information systems is not new [16], and
some good examples do exist [17]. Finland has had a long tradition of maintaining
administrative information systems, and extensive nationwide registration—using
an individual’s single personal identification number for all systems—is exceptional
also from the international point of view [18]. However, even though the potential
advantages of Finnish information systems are well recognized, the utilization of
these data sources has been rather limited.
This controversy is not specific to Finnish health information systems. In fact,
advances in information technology have made it possible to produce and store all
kinds of data effectively, but the emphasis has been on technical aspects and not
on the information itself [19]. For as long as there have been such information
systems, the problem of giving too much data in an unusable form has been a
common complain [20]. In fact, the predominant belief that secondary data stored
in the health information system consist of autonomous, atom-like building
blocks is fundamentally erroneous. Therefore, it is also erroneous to believe that
the production of more detailed data would solve the basic problems. On the
contrary, the more detailed and complex are the variables to be recorded, the more
background information and tacit knowledge is required for secondary utilization.
Such a fact is known as the law of medical information [21]. In this sense, it can be
expected that as the amount of data increases with the introduction of electronic
patient records and correspondingly complex information systems, it will be even
more difficult to transform data into useful information.
Traditionally, the problems in analyzing data have been considered as
statistical issues, but statistical research focusing on probabilistic inference based
on mathematics has not been able to offer enough concrete help in the analysis

1 Introduction
of growing amounts of data [22] . The practical need for information has led to
the development of alternative ways of analyzing data, such as data mining [23].
However, the development of new tools for data analysis has not solved the actual
problem of transforming raw data into useful information.
As a matter of fact, it is naïve to believe that there would be some magic trick to
overcome problems arising from the philosophical and methodological limitations
of empirical research. Rather than trying to squeeze the data into a predefined
model or saying too much on what can and cannot be done, data analysis should
work to achieve an appropriate compromise between the practical problems and
the data [24]. This kind of activity has been characterized as “greater statistics”,
which tends to be inclusive, and eclectic with respect to methodology, while
being closely associated with other disciplines and also practiced by many non-
statisticians [25]. Because formal statistical expertise provides an excellent basis
for the understanding and evaluation of methodological ideas, statisticians ought
to take advantage of the situation, get involved in interdisciplinary activities, learn
from the experience, expand their own minds—and thereby also their field—and
act as catalysts for the dissemination of insights and methodologies [26].
1.1 Aim of the study

This thesis is about scientific research methodology in the context of register-based
health system performance assessment. The application field is health services
research. Health services research differs from most areas of applied research, since
it does not exist in isolation from the decision-making procedures needed in health
policy. A multidisciplinary approach and knowledge is required in health services
research, including perspectives from at least biological, medical, social, clinical,
management, economic, statistical and information sciences [27].
Another special aspect here is the use of register-based data. Since register-
based data have been originally produced for other purposes than for this specific
research, the traditional methodological assumptions concerning the nature of
data are not valid.
In addition, information that allows for the monitoring of health system
performance should be available if a health information system is to be adequate.
This means that the perspectives of routine statistics production and the technical
possibilities of information systems must also be considered simultaneously by
means of a more research-oriented approach. Unfortunately, there are no widely
accepted principles for the production of register-based statistics, which makes the
development of register-statistical methodology challenging [28].
These issues make things rather complex. It is not possible to rely solely on a
particularly well-developed paradigm and so the only reasonable way is to accept
the diversity of the approaches and learn to deal with them simultaneously. This

1 Introduction
can be done by “raising” the ideas from the application level to the philosophical–
methodological level. However, the abstract ideas may become real only through
the applications. In other words, the innovative results can be achieved only by
rethinking the methodology separately for each specific application.
In short, the aim of this thesis is to give multidisciplinary methodological
perspectives for scientific register-based information production for the purposes
of health system performance assessment. More specifically, the goals are 1) to
develop a methodological framework that helps to utilize register-based data
effectively, 2) to demonstrate how the register-based data can be used as a data
source for the performance monitoring of a health system. The specific application
area is hip fracture from the public health point of view.
1.2 Structure of the study

This thesis consists of this summary and the following six original contributions.
[1] Sund R: Utilisation of administrative registers using scientific knowledge
discovery. Intelligent Data Analysis 2003, 7(6):501–519.
[2] Sund R, Nurmi-Lüthje I, Lüthje P, Tanninen S, Narinen A, Keskimäki I:
Comparing properties of audit data and routinely collected register data in
case of performance assessment of hip fracture treatment in Finland. Methods
of Information in Medicine 2007, 46(5):558–566.
[3] Sund R: Utilization of routinely collected administrative data in monitoring
of aging dependent hip fracture incidence. Epidemiologic Perspectives &
Innovations 2007, 4(2).
[4] Sund R, Kauppinen S: Kuinka laskea ikääntyneiden pitkäaikaisasiakkaiden
määriä rekisteritietojen perusteella? Sosiaalilääketieteellinen Aikakauslehti
2005, 42(2):137–144.
[5] Sund R: Lonkkamurtumien ilmaantuvuus Suomessa 1998-2002. Duodecim
2006, 122(9):1085–1091.
[6] Sund R, Liski A: Quality effects of operative delay on mortality in hip fracture
treatment. Quality and Safety in Health Care 2005, 14(5):371–377.
This summary synthesizes the methodological ideas from these contributions.

The second chapter is about register-based data analysis in general. It begins with
the database-oriented approach to data analysis and then extends the perspective
to a more comprehensive research-process approach which is able to deal with
secondary data, which is the topic of article 1. The latter part of chapter two gives
a brief review of certain multidisciplinary ideas that help in understanding the
problems likely to be encountered during a register-based data analyzing process.
The general methodological ideas that I have presented mainly in articles 1 and
Articles are included in the thesis with permissions from the publishers.

1 Introduction
3 and developed further elsewhere [7, 8], are in this chapter reflected against the
existing literature. As a whole, chapter two defines a methodological framework for
register-based data analysis and describes fundamental assumptions that have been
applied in the later chapters.
In chapter three, some fundamental background information on health system
performance assessment in Finland is given and a perspective for the applied
research problem is fixed with certain focusing assumptions. This chapter extends
and focuses the introduction given in article 2 on the basis of my other publications
[7–9]. The fourth chapter includes a brief introduction to hip fracture, register-
based sources of information on hip fractures, and a review on previous register-
based research on hip fractures. The latter part of the fourth chapter summarizes
the contents of article 2, which defined a conceptual model for the hip fracture
performance assessment and examined the validity of Finnish registers as a data
source for the purposes of a performance assessment of hip fracture treatment.
Chapter five, based on articles 3–5, deals with the pragmatic problems related to
the development of a surveillance system for register-based hip fracture incidence.
The sixth chapter updates the conceptual model on the hip fracture treatment
process presented in article 1 so as to be compatible with the conceptual model on
hip fracture performance monitoring reported in article 2 and gives a very basic
example of an application of the conceptual model. I have already presented the
ideas summarizing the approach elsewhere [9], and prepared several manuscripts
that deal with the application in more detail. Chapter seven gives a more detailed
example on the justification of a performance indicator, and elaborates on the
methodological ideas presented in article 6, including a basic model for the profiling
of providers first described in article 1.
Finally, in the eighth chapter, the motives for this study are once more discussed,
the idea of the methodological approach is briefly summarized, and whether the
aims of the study have been fulfilled is evaluated.

2 Register-based data analysis
A register is an information system which continuously records subject-based data

for a particular complete set of subjects. A register contains a logically coherent
collection of related data with some inherent meaning, typically reflecting some
events occurring in the observational reality. A register is implemented for a
specific purpose and has an intended group of users. These properties also fulfill the
definition of a database and hence a register can be considered as a database [29].
Typically a register is a large collection of data and is maintained using computers.
As long as there have been (computerized) databases, there has been an increasing
interest to provide information at the correct level of detail to support decision-
making [30, 31].
2.1 From technical data processing to a

knowledge discovery process
In principle, it is quite straightforward to implement procedures that take register
data as input and produce a sensible summarization of them as an output. For
example, intensive cross-tabulations and basic visualization tools can be used for
the comprehensive summarization of stored data. In addition, implementations of
many traditional statistical techniques are suitable for such processing. However,
the massive size of a database may impose certain computational difficulties in
applying methods initially developed for smaller data sets [32]. Therefore, more
recent and computationally efficient methods and algorithms offer an invaluable
enhancement to data analysis [33, 34].
In practice, the description of data analysis as a finite series of precisely encoded
rules needed to transform raw data in a database into interesting information gives
an overly simplistic impression of data analysis, because it remains unclear what
should actually be done in order to achieve information in data analysis. In this
sense, it is convenient to consider data analysis as a collection of tasks, such as
exploratory data analysis, descriptive and predictive modeling, pattern discovery,
and identification [35]. In fact, it has been suggested that reasonable results can
be achieved by considering the whole series of actions required to transform data
into information as a knowledge discovery process [36]. A descriptive model of
a standard process of data mining—which gives detailed pragmatic guidelines
for performing a knowledge discovery process—has been developed [37]. In
this business-oriented tradition, the goals of the process are evaluated in terms
of its ability to produce interesting information. Suggested criteria for interesting

Context
Problem
Debate
Question
Analysis
Data
Perspective
Idea
Answer
Theory
Figure 1. Research process schema
information are, for instance, evidence, non-redundancy, simplicity, novelty, and

usefulness [38].
The problem with such a pragmatically useful description of a business-
oriented knowledge discovery process is that certain important aspects of
traditional empirical research and scientific method have been disregarded.
Therefore I have complemented the process model to give a more comprehensive
and detailed description of the phases commonly encountered during an empirical
research process that incorporates register-based data [1]. The improved schema
for a knowledge discovery process is presented in Figure 1. The main phases in
such process are: understanding the phenomenon, understanding the problem,
understanding data, data preprocessing, modeling, evaluation and reporting [1].
The case study of this thesis represents an application of this schema.
2.2 Secondary data as a methodological

problem
If the schema presented in Figure 1 is compared to a standard scientific inquiry,
the most important difference is the connection between the problem and the
data. In fact, typically a theory of a given phenomenon of interest should drive
the primary data collection. This makes it possible to concentrate on just those
parts of observable reality that are considered most relevant for the current
theoretical purpose. A similar approach is not possible with register-data, as the
formulation of a problem in terms of register-based data is opportunistic, given
that the measurement can only be based on existing secondary data originally
produced for some other purposes than the research problem at hand [39]. In fact,
the limitations of secondary data are determined by the choices made in relation
to the production of the data, such as a decision to collect only easily available data
and using fixed categorizations in the data production, with consequent variation
in the production practices and categorizations used, and a lack of flexibility in the
information systems to take into account changes in phenomena and society [40].

Bias/Noise
Phenomenon Coding Register Decoding Data
Figure 2. Schematic diagram of information communication via administrative registers
I have proposed that the methodological challenges related to the utilization

of register-based data can be illustrated using ideas from communication theory
[1]. A schematic diagram illustrating information communication via registers is
presented in Figure 2. First, it is assumed that some phenomenon exists that can be
observed. Since it is impossible to completely observe all details or perform exact
measurements, some kind of coding is used to describe things. This coded signal is
then stored in a database. The noise and bias can be interpreted as an explanation for
measurement compromises, possible inconsistencies and coding errors, and coding
practices existing in the stored signal. When this signal is then utilized, it must be
decoded into a suitable form. This phase is also subject to noise and bias caused by
incompatibility of choices and interpretations made by the data producer and the
data user. Even the decoded signal (data) is not a final phase in the research process,
because further analysis and processing is needed in order to transform the data
into information. Even though this is a very simple and technical representation of
communication, it seems to contain the essential elements needed in the common-
sense understanding of secondary data.
2.3 Prerequisites in the register-based data

analysis
The schemas in Figure 1 and 2 give a basic overview of the process of register-
based data analysis, but leave many pragmatic details unanswered. Even though
the actual realization of the process is determined by the research problem and
the available register data, it has been pointed out that the effective use of register-
data presumes skills in at least four areas: in computer science, in statistics, in the
principles of measurement, and in the theory of the subject matter [41]. In the
following, I briefly review and synthesize the most important ideas, assumptions
and approaches that were prerequisites for the case study of this thesis.

2.3.1 Principles of measurement

Typically it is assumed that reality can be confronted by recording observations that
reflect the phenomenon of interest. Measurement aims to create data as symbolic
representations of the observations. The two main aspects for measurement
are the representational and the practical approach [42]. Representational
measurements are quantifications of attributes of objects existing in reality, while
the practical measurements are operationalizations of the phenomena of interest.
All representational measurements resulting in concrete data can be formulated in
terms of a practical approach.
The operationalization used in the practical approach for measurement
determines how the phenomenon P that becomes visible via observations O
is mapped to data D. In the pragmatic sense, operationalization can be taken as
successful if it becomes possible to make valid interpretations I of symbolic data
D in regard to the phenomenon P, i.e. that a satisfactory saturation between the
phenomenon, conceptualization and data is achieved [1]. The key point is that three
recognized levels of contextual dependencies accompanying the empirical research
are explicitly separated: a theory driven observation (O), the operationalized data
(D), and a theoretical interpretation of the data (I) [43].
It is particularly important to notice that the actual information is not assumed
to be contained in the data, but is something that has to be produced from the data
and the pre-knowledge. Such an idea has been presented more formally in terms of
an infological equation
I = i(D,S,t),
which states that the information I is produced from the data D and the pre-
knowledge S by the interpretation process i during time t [44]. From this point
of view it is obvious that any sharing of data can only be a proxy for the process
of sharing of information, because the unbiased sharing of information would
require the background knowledge S to be identical with the producer and the
users of data. In this sense it is not surprising that data as such are of no value and
become interesting only if there is also a meaning and a context for them [45].
The range of possible interpretations of the data can be reduced by increasing the
common background knowledge of the potential data users by offering descriptive
data about data, that is, metadata [46]. While it is true that the increased amount
of metadata increases the proportion of shared information, the more diverse
contexts the data have to be usable in, the more work is required to disentangle the
data from the context of its production [21].
These are also starting points in register-based data analysis, and the problem is
to find some shared perspective between the original and intended data utilization
purposes in terms of available data and (more or less tacit) metadata [3]. Following

the ideas of a cognitive fit approach [47], there could be a problem-solving task
related to phenomenon P, but only data Dp’ are available that tell about some
other phenomenon P’. A “true” task solution Ip can only be approximated with a
conditional task solution Ip|p’ . The conditional task solution is based on conditional
data Dp|p’,which are a result of some intelligent transformation of available data
Dp’ and are based on the cognitive fit between the internal representation of
hypothetical “true” data Dp and available data Dp’ . In practice this means that an
extra interpretation-operationalization phase is required in the research process if
secondary data are to be utilized [1], i.e. that data in a register need to be decoded
to become suitable research data (Figure 2: register → decoding → data). This
important idea has been applied throughout this dissertation.
The problem with the approach is that phenomena are different in terms of
measurement possibilities. As the law of medical information states, it is more
difficult to utilize more complex variables for purposes other than those originally
intended [21]. In this sense, it is in practice fruitful to evaluate the amount of
subjectivity in the variables. One criterion is the separation between representational
and practical measurement: the measurement is representational only if it is
reasonable to assume that the data D are direct reflections of observations O that
refer to some entity existing in reality [42], i.e. that the phenomenon of interest is
directly observable. Another useful criterion is related to the amount of additional
background knowledge S+ required in the interpretation of data D. I have suggested
that if there is no need for knowledge beyond the shared human common sense, the
measurement can be characterized as stable [8]. In fact, it is known that space and
time provide a shared biological-development basis for ordering human common
sense as our world of sensory observations is arranged to observe things changing
in space over time [48]. If measurement is representational and stable, it is factual.
These ideas have an important role in this thesis, because the purpose has been
to minimize unnecessary subjectivity by maximizing the use of factual data. For
example, the hip fracture care episodes are reconstructed in terms of factual data.
2.3.2 Information science

A concrete data management requires some kind of data structures. In fact, the raw
data are nothing but a sequence of bits stored in a database. There must be some
a priori fixed rules for the interpretation and handling of bit sequences, which
define the basic data objects, such as integers, floating point numbers and character
strings. A data model specifies which kinds of data and data manipulations are
permissible [49]. New types of data objects and associated operations can be easily
constructed by combining the existing types and operations [50]. This idea makes
it possible to provide a conceptual representation of data which hides storage and
implementation details which are of no interest to most database users [29].

Phenomenon Knowledge component

Taxonomy ⇓
Concept
Partonomy
⇓
Object
Logical component
Host Attributes
Theoretical measurement properties
Data component
Time Place Realized observation
Figure 3. An overview of the extended pyramid framework
More generally, data modeling can be used to construct (computer-based)

symbol structures which capture the meaning of data and organize it in ways that
make it understandable and useful [51]. Data modeling is restricted in the sense
that only what is (or can be) represented is considered to exist [52]. A recently
introduced unified meta-information architecture of statistics (UMAS) provides a
comprehensive conceptual framework for data modeling [53]. The aspects of the
complicated framework that are useful in pragmatic register-based data analysis can
be illustrated with the help of a pyramid framework that is based on the cognitive
principles of how humans store everyday knowledge [48]. A slightly extended
pyramid framework is presented in Figure 3. Following the ideas that I have
presented independently [7], but which the UMAS also captures, the framework
had to be complemented with an additional logical component.
A semantic object (conceptual entity), reflecting some phenomenon in
reality, is a basic element in the framework. Taxonomy and partonomy reflect
the principles of a cognitive categorization of the object and represent also the
schemas needed to interpret the observational data. These parts form a knowledge
component of the framework, i.e. reflect the required pre-knowledge that cannot
be extracted from the data. The object must be logically defined by identifying
an observable host for a concept and by determining which attributes related to
the concept are to be observed. The logical component also includes theoretical
measurement properties of variables given in terms of random variables that
represent the (assumed) operationalization of the variables. The data component
of the framework represents the actual data by means of three distinct data
perspectives: place (where), time (when), and realized observation (what). The
whole path from a knowledge component to the realized observation is needed to

describe what is actually observed. The (extended) pyramid framework provides a

human–computer environment for (register-based) data analysis that incorporates
models of human cognition so that the utilization of large databases becomes more
effective than without the framework [54]. A generalized event-sequence approach
that I have presented in article 1 is obviously a special case of this more general
framework. The framework also had an important role in the development of the
conceptual model in article 2.
2.3.3 Statistical computing

Information sciences also deal with more pragmatic tasks concerning the acquisition
and preparation of data, which can also be classified as statistical tasks. The
acquisition of relevant data in practice may be a difficult and time-consuming task,
because the sources of data need first to be identified and the use of register-based
individual-level data for scientific research requires permissions to be applied for
[55]. Technically the problem is similar to the one of extracting and integrating data
from multiple sources into a new data view (data warehouse) [56]. An important
special case aims to identify the instances which include data about the same real-
world entity [57]. Such duty is better known as record linkage [58]. The record
linkage becomes problematic, however, if the entity identifiers in the data sources
are not identical and probabilistic techniques are needed to predict the equivalence
[59]. More generally, data integration may require schema matching, which aims
to find compatible interpretations between multiple database structures [60]. For
example, typical mortality data have a different structure to hospital discharge
data, but in practice it is useful to define a data structure that allows incorporation
of data from both sources.
There are also some commonly encountered data preparation tasks [1]. Data
cleaning task involves detecting and removing errors and inconsistencies from data
in order to improve the quality of data [61]. In statistics literature, data cleaning
is known as editing and imputation [62], and in applied mathematics as error
correction [63]. Data cleaning deals with logical errors, such as violations against
integrity constraints or duplicate observations. The nature of data cleaning is
technical, meaning that such corrective manipulation of data could be done no
matter what the actual research problem under investigation is. Data reduction
aims to produce a reduced representation of data which is much smaller in
volume than the original data set, yet produces (almost) the same answers to the
research problem [64]. This kind of data preparation can consist of anything from
simple dropping of unimportant variables, combination of several variables or
observations into single one, or more radical changes in aggregate levels to even
more complicated analyses. For example, in some cases it may be reasonable to
assume a multivariate normal distribution for variables of interest, which means

that only the covariance matrix and means (sufficient statistics) are needed for
subsequent analyses. Data abstraction tries to embed an intelligent interpretation
(enrichment) of “raw” data into analyses so that the resulting derived data set
would be at the level of abstraction corresponding to the current problem [65]. In
practice this kind of abstraction may be feasible by processing data with explicit
algorithms so that important a priori schemas that are not directly available in data
become formalized in terms of the data. For example, register-based data are usually
patient-specific, while medical knowledge is patient-independent and consists of
generalizations that apply across patients. A complication after a surgical operation
is a medical concept, but from individual-based hospital discharge data it must be
abstracted by using some rules, such as a list of particular diagnosis codes with
appropriate time stamps recorded in the data. Data abstraction always results in a
problem-specific derived data set.
In summary, the data preparation has been characterized as a process that
requires human participation, which means that as much art as science is needed in
good data preparation [66]. Whatever preprocessing tasks are applied, it is obvious
that incorporated explanatory analyses offer insights and realistic perspectives into
the data. It can be further stated that a sophisticated preprocessing—which is full
of ideologically dependent qualitative choices—in order to scale matters down to a
size more suitable for specific analyses is the most important and time-consuming
part of register-based data analysis [1]. Data sensitive preprocessing had a key role
in the empirical part of this thesis.
2.3.4 Statistics
Statistics is a science devoted to the production, analysis, modeling, and presentation
of data. Statistical models help to distinguish systematic patterns from random
fluctuations, measurement errors, and confounding biases occurring in the data.
However, the traditional mathematically oriented statistical paradigm has had
some difficulties in adapting to a new situation that requires concrete analyses of
massive (register-based) datasets [22]. One reason is that the basic assumptions
about the independent observations and the sampling error as the main source of
uncertainty are often violated in complex register-based data sets which possibly
include total populations [67]. Further, the traditional statistical significance has
become as an issue, because even the practically unimportant differences easily
become statistically significant within the large data sets. In addition, with massive
datasets it can be expected that there are some distortions and errors in the data,
and typically it is unfeasible to manually check all of them. On the other hand,
the reasonable use of data typically requires the use of problem-specific data
abstraction rules which transform the data to a useful form. This kind of data
analysis is typical for “soft” qualitative analysis, but with massive register-based

datasets, strong technical skills in quantitative data management are also required.
In summary, a flexible approach for statistical inference is typically beneficial with
register-based data. Moreover, statistics offers not only a set of tools for problem-
solving, but also a formal way of thinking about the modeling of the actual problem
[68]. In this sense, statistical thinking also has a key role in the process of register-
based data analysis.
In fact, one useful perspective for register-based data analysis is to consider the
whole process of information communication (Figure 2), and try to explicate the
sources of noise and bias that may distort the communication [69]. The basic idea
is—in addition to a traditional modeling of the phenomenon of interest—to take
into account suspected sources of bias, possibly with the help of a priori information
[70]. Such is one way to deal with the particular restrictions of secondary data by
using theory driven statistical modeling. Chapter 7 in this thesis also incorporates
this kind of statistical thinking.
Another option for statistical modeling is to apply statistical algorithms
directly to data and let the data speak for themselves [71]. Also this approach
is often useful with massive secondary data sets, because data exploration and
description plays a more general role than it does in the case of small data sets. The
screening of data for significant associations without having specific hypotheses
may be questionable [72], but it may also be erroneous to assume that data are
automatically in concordance with the theory. Within this approach it becomes
possible to be data sensitive so that empirical justification is given to the proposed
operationalizations. Most analyses in this thesis are of this type.
2.3.5 Theory
Theory of the subject matter is needed for the formulation of the research problem.
As shown in Figure 1, theory is also the driving force in the generation of the
question and in the choice of the perspective, as well as in the interpretation of the
data and analyses. In other words, the theory determines the framework within
which justified data analysis becomes possible.
The special feature in register-based data analysis is that more than one theory
must be simultaneously dealt with [3]. For meaningful results there is a need to
find some communality in terms of realized data between problem-driven subject
matter theory and the (more or less unknown) theories that have been used during
the production of the register-based data. The extended pyramid framework helps
to structure the data in a concrete and methodologically sound way that makes the
finding of suitable compromises much easier. For example, the components of data
that have the most stable measurement properties are obviously the ones that can
be most easily reused in various contexts and the basics of many useful theoretical
approaches can be built on these factual main elements.

3 Register-based health system performance assessment
3 Register-based health system

performance assessment
In the previous chapter it was argued that the register-based data analysis should be
considered as a research process. Obviously that is true also in the case of scientific
health system performance assessment. In this chapter some essential background
information needed for the understanding of the phenomena of interest are given.
The applied research problem is also introduced and the perspective is fixed with
certain focusing assumptions.
3.1 Finnish health system

A (public) health system is a creation of the human community, and the health system
of any society can only be understood in the light of its societal operating principles
and policies [73]. The Finnish societal system is typically characterized as a Nordic
welfare state, which is internationally rather exceptional. The Finnish health care
system is very decentralized, and the country’s numerous (>400) local authorities
(municipalities) are responsible for arranging services [74]. Each municipality is a
member of one of the 21 hospital district joint authorities, which are responsible
for organizing specialized medical services and coordinating hospital treatment in
its own district. Secondary and tertiary level medical care is provided by a hierarchy
of regional, central and university teaching hospitals. Services for older people are
provided in both social and health care, both being incorporated into the same
national planning and financing system [75]. The organization and financing of
health care has long been considered a public responsibility [76]. There are also
many recent reforms to the Finnish health system, such as the setting of thresholds
for admission onto waiting lists for elective surgical procedures, the introduction
of a set of maximum waiting-time targets for non-urgent examinations and
treatments, a national electronic patient record and a project aiming to restructure
municipalities and services [77].

3.2 Information production in the Finnish

context
One important tool in the maintenance of a Nordic welfare state has been to
produce information which supports the (nationwide) decision-making. The
main emphasis has been on producing official statistics and indicators that reflect
important aspects of health and social welfare, and therefore offer information
for decision-making and controlling purposes. However, the national-level
control mechanisms were decomposed in 1993 due to the change from direct
authorative control (imperatives, rules, norms, earmarked funds) into indirect
control based on information (instructions, guidelines, hopes, total funds) [78].
This new nationwide policy on “steering by information” has aimed to produce
information for meso-level decision-makers (such as municipalities and hospital
districts) [79]. The problem is that now there should be information to suit the
purposes of hundreds of local policies instead of one global policy as in the past
[80]. Moreover, the potential users of information may easily refuse to act on the
given information if they consider it to be irrelevant for their purposes [81]. In this
sense, there is an obvious need for methodological studies that aim to make policy
relevant information production more efficient.
3.3 Measurement of the performance of a health

system
Even though there does not exist a universal value base to all health care systems,
most health systems in developed countries aim to promote, restore and maintain
health [82]. The common goal is the optimization of the health of individual
patients and populations in an equitable, efficient, and effective manner that is
acceptable to patients, providers and administrators [83]. The growing need for
appropriate services with limited resources and the concern about the continuing
inequities in health and in access to health care means that there is much interest
among decision-makers in improving the performance of health systems [84].
However, performance measurement has proved to be very difficult in practice
[85]. The main problem in performance measurement seems to be that there is
little agreement on the philosophy of measurement and on what to measure [86].
For example, at least 15 dimensions for health care performance can be identified
from the existing frameworks [87].
In fact, the problem seems to be related to a more general change in societies
and in information production [88]. The evaluation of performance has become
a more powerful organizing concept in activities of societies and there is a real
need for information about performance. However, it has been claimed that the

abstract rhetoric of administration blurs the actual meaning of performance: there

are no more norms about what and how things should be done, but the political
responsibility is implicitly included by using vague formulations of strategic
objectives which delegate the actual implementation to “actors” [89]. In this sense,
performance assessment as such is an empty concept [88]. Only if the goals are
fixed and operationalized to a measurable form can the concept of performance
become interesting and potentially useful.
Concrete quantitative measurement of performance has mainly been
conducted using various performance indicators [90–92]. Recently the whole idea
of performance measurement seems to be focused on the sophisticated development
and reporting of indicators [93-95]. Furthermore, methods for evidence synthesis
and assessment in the context of multiple indicators have been suggested [96, 97]. As
the indicators certainly reflect something that can be measured, it is typically useful
to compare indicators between the appropriate subpopulations, because descriptive
data on relative performance may help to identify “best practices” or “what works”.
However, the causal mechanism underlying differences between the indicators
typically remains unknown in such an approach [98], which “invalidates” the
appealing analogy to the experimental research design [99]. In this sense, register-
based indicators without detailed justification of their theoretical-methodological
basis do not represent scientifically valid information.
3.4 Assumptions for the production of concrete

health system performance information
In this thesis, I focus on the methodology of producing health system performance
information. The key assumption is that the routinely collected register data are to
be used for performance assessment purposes. The use of register-data in health
research is known to be prone to several problems, such as the perceived lack of value
of administrative data, privacy and confidentiality, data availability, population
coverage, registration period, record linkage possibilities, lack of clinical data, data
format, coding systems, coding practices, completeness of registration, accuracy of
registered data, data processing, size of data, and discovery of chance occurrences
[100–104]. I do not claim that the problems are not real, but adopt the view that
proponents of certain approaches have been more interested in advocating their
ready-made mechanical procedures than in understanding alternative logics
of interpretation [105]. In fact, many of the mentioned problems are related to
particular study designs or properties of existing data sets and therefore relevant
only for certain types of research questions or data. For the most fruitful results it
seems to be essential to elaborate the whole analytic process with respect to the data,
because the fixed structure of secondary data indirectly determines constraints on

the theoretical constructs [106]. In this sense, the methodological approach for
register-based data analysis presented in Chapter 2 also provides a basic framework
for the register-based performance assessment of the health system. I focus on the
data sensitive analyzing approach, and the aim is a step-by-step transformation of
data into as objective information as possible by maximizing the use of the most
factual register data.
I assume that the interesting dimensions of health system performance
assessment can be examined by using the framework which conceptualizes health
system performance assessment in terms of structure, process and outcomes [107].
In short, the structure comprises resources that are devoted to producing actions
whose primary purpose is to improve health, while the process means the realized
utilization of these resources, and the outcomes reflect the effects of resource
utilization on health. More recent formulations of the framework emphasize also
the importance of a context and health policy goals that have led to certain health
system implementations [10, 108].
The definition of what kind of expectations the commonly accepted abstract
health policy goals truly reflect is difficult [109]. However, it can be defined that
effectiveness in this context measures how the health system (or resources spent on
it) affects the health of the target population [10]. In this thesis I concentrate on
the effectiveness dimension of health system performance, because effectiveness
seems to be a key dimension in several performance frameworks [87]. It is also
known that the production of health benefits (effectiveness) plays a central role in
assessing the cost of producing health benefits (efficiency) as well as the distribution
of these benefits and costs across groups (equity) [10]. In addition, there is an acute
need for information on effectiveness for the purposes of information steering in
Finland [110].
3.4.1 Register-based health system performance

monitoring
The Finnish registers offer an exceptional coverage of health and social welfare data
[18]. In principle, a separate scientific study utilizing these data could be conducted
to answer some specific question. However, from the information steering point
of view, the available register-based data should be routinely converted to useful
information. As it is unfeasible to repeat separate studies manually over and
over again, a better alternative is to incorporate new information production
pipelines to the health information system that can be used in the production of
routine statistics [14]. In fact, the implementation of a nationwide health system
performance monitoring system is one of the main development goals of for the
information systems of the health and social welfare services in the near future
[111].

In the Finnish proposals for the health information system, register data (raw
data) and evidence-based decision support (know-how of transforming the raw
data into information about issues that should be known) are essential ingredients
for the continuously improving epidemiological and quality components that
routinely produce new scientifically valid information [14, 15, 112]. Following
these proposals, health system performance monitoring should consist of two
components. The epidemiological component is needed for monitoring the
volume and incidence of health problems, which are essential information for the
evaluation of prevention strategies and which also help to prepare for changes
in the need for care caused by changes in the population structure. The quality
component concentrates on producing information on the effectiveness and quality
of treatment, which are essential information for the evaluation of health system
performance and for the purposes of finding ways to improve the health system.
For the purposes of this thesis, I assume that pragmatically useful health system
performance monitoring should be able to produce valid information about the
volume and incidence of the problems, use (and costs) of services as well as about
the effectiveness (and quality) of provided services.
It is well known that the performance assessment is practically feasible in
a reasonable way only if it is carefully developed and tailored for each specific
health problem separately [106, 113]. Therefore, I concentrate only on one health
problem, hip fracture.

4 Register-based data on hips fractures – sources and validity
4 Register-based data on hip

fractures – sources and validity
The assumptions and theoretical considerations in the chapters above create

adequate starting points for more concrete health system performance assessment.
The case study in this dissertation is intended to offer perspectives and methods
for the implementation of a nationwide system for monitoring health system
performance. Hip fracture is used as a concrete example. A brief review of essential
background knowledge is given first. Thereafter, the feasibility of using Finnish
register data in the case of hip fracture treatment performance assessment is
discussed based on article 2.
4.1 Hip fracture

Hip fractures are common injuries among older people, and associated with
substantial morbidity and mortality [114]. The term hip fracture refers to a fracture
of the upper end of the thigh bone (femur). Most hip fractures in persons aged
50 years and over result from moderate low-energy trauma, usually a fall from a
standing height or lower [115]. For younger persons it is more likely that a case of
hip fracture results from a high-energy trauma, such as traffic accidents or a fall
from a height [116]. Prevention has focused on minimizing the risk of falls and
on reducing the injury potential of those falls [117]. About 7000 (of which more
than 95% occur for patients aged 50 years and over) hip fractures per year occur in
Finland currently [5].
Ageing among populations is increasing hip fracture patients’ mean age and
the number and severity of their pre-existing co-morbidities, which is likely to cause
additional problems in patients’ treatment and rehabilitation in the future [118].
Sometimes the hip fracture can be interpreted as an indication of the “beginning
of the end” (patients were doing well until they broke a hip and went downhill
quickly) and sometimes as an “end of the beginning” (hip fracture signals that the
cumulative effect of small declines has reached a critical level) [119].
4.1.1 Hip fracture treatment in Finland

Virtually all suspected hip fracture patients are first referred for examination and
treatment to the nearest hospital with orthopedic services. The main objective in
hip fracture treatment is to return the patient to his or her level of function before

the fracture [115]. The diagnosis of fracture of the hip is straightforward, using x-
ray examination. A surgical operation is performed on the majority of patients.
The main methods used in treatment are reduction of the fracture using internal
fixation and hip replacement arthoplasty. The care pathway for a hip fracture
patient is rather complex with several phases such as surgical management and
rehabilitation [120], and is known to result in diverse episode profiles in Finland
[121]. Typically a patient is transferred for rehabilitation to the health center serving
the patient’s resident municipality after a short postoperative hospital treatment
[122]. Finnish health centers are local primary health care units, which also contain
inpatient wards. Other institutional environments of care include residential homes
and service housing with 24-hour assistance, which both correspond to the nursing
home type of care. Non-institutional services utilized by hip fracture patients
include outpatient health services, home nursing, ordinary service housing, home-
help services, and support for informal care [75].
For six Finnish hospitals, patients aged 50 and over had an average mortality
at 30 days after the fracture of about 7%, 17% at four months, 26% at one year, and
about 50% at four years [123]. At four months, about 40% of patients lived at home,
about 15% were unable to walk, and about 8% had a lot of pain in the injured hip
[123]. The functional capacity of the patients does not typically revert to the level
prior to the fracture [124]. Hip fractures are also costly to the society. The average
patient-specific costs during the first post-fracture year in Finland were estimated
to be around €14 410 and more than €35 000 in case of a previously home-dwelling
individual who becomes a long-term care patient following the fracture [125].
Treatment processes as well as the outcomes vary considerably between
areas and hospitals, and improved auditing of hip fracture treatment has been
suggested [126, 127]. A recent Finnish current care guideline on the management
of hip fracture patients proposes that a nationwide hip fracture register allowing
continuous auditing should be established in Finland [128]. In this sense, there
is a pragmatic justification for the methodological studies aimed at transforming
routinely collected register data into relevant hip-fracture-specific information
about the performance of the health system. In addition, hip fracture is a good
choice for a pilot study on performance assessment, because it can also be viewed
as a tracer condition in health systems, testing how well health and social services
are integrated in the provision of acute care, rehabilitation, and continuing support
for a large and vulnerable group of patients [129].
4.1.2 Register data on hip fractures in Finland

Finland has a long history of collecting data on health and social services. At the
end of 20th century, there were about twenty different routinely collected national
administrative health registers [18, 55]. A unique national personal identification

number is used in all Finnish registers which can be utilized as a linkage key in
order to combine data from multiple sources. A particularly important register
for the purposes of health services research is the Finnish Health and Social
Welfare Care Registers (including the Finnish hospital discharge register). These
registers contain data on all inpatient care periods in hospitals, in health centers,
in residential homes and in service housing with 24-hour assistance. The data
warehouse of the Finnish Hospital Benchmarking Project (nowadays a part of
the Finnish Health Care Register) is the corresponding register for hospitals, but
also incorporates data on outpatient visits. In general, the complete registration
combined with easily linkable registers makes large, longitudinal population-based
studies feasible in Finland.
4.1.3 Previous register-based hip fracture studies in

Finland
The first nationwide register data based on hospital discharges in Finland are
available for the year 1960, although the continuous hospital discharge data
collection began no earlier than 1967. Hip fractures were not reported separately,
but were combined with other fractures of limbs [130]. The register data from
1968 included hip fractures as a separate group [131]. The hospital discharge
data from 1968 is also a data source for the first hip-fracture-specific register-
based study in Finland [132]. Most of the register-based studies concerning hip
fractures have had an epidemiological perspective [133–137]. In the late 1980s and
early 1990s, a project aimed at improving the reporting of treatment and costs
data based on registers had hip fractures as a separate group [138]. At the same
time, the first register-based small-area analyses examining treatment practices in
terms of surgical operations reported data on hip fractures treated with arthoplasty
[139]. The next step in the utilization of register-based data on the description of
hip fracture treatment practices was the reconstruction of the care episodes of the
patients [121, 140]. There followed a pilot study examining the effectiveness of hip
fracture treatment using register-based data [141, 142].
4.1.4 Register data for the current study

The data used in this study were also based on Finnish registers. As virtually all
hip fracture patients are treated at the hospital inpatient ward and given that hip
fracture is easy to diagnose, it is very likely that the patient population can be easily
identified from the hospital discharge data by using a simple database query with
a list of diagnosis codes. To make sure that all hip fracture patients were included
in the study population, a total population of fracture of femur (corresponding to

ICD-10: S72) in the period 1998–2002 was identified in the Finnish Health Care
register.
In order to capture the medical histories of these patients, data on all inpatient
(1987–2002) and outpatient hospital care (1998–2002), residential home care and
care in service houses with 24-hour assistance (1997–2002), and deaths for this
population were obtained from the Finnish hospital discharge register, the Finnish
Health and Social Welfare Care Register, the data warehouse of the Finnish Hospital
Benchmarking Project and the National Causes of Death statistics using the unique
national identification numbers of the patient population. It was first time that
the hospital outpatient visits from the benchmarking project data and the data
on inpatient care in social institutions from the Social Welfare Care Register were
included in a hip fracture specific register-study.
The results of simple database queries were integrated into a new data set
containing 988 762 records for 39 041 patients. Each record in this data set
corresponded to one care period in inpatient institutional care or outpatient visit
in hospital (or death), and included variables such as patient and provider ID-
numbers, age, sex, area codes, and diagnosis and operation codes, as well as dates
of admission, operation and discharge (or death).
4.2 Validity of register data in the case of hip

fracture
Valid data is a prerequisite for any information system aiming to produce practically
useful information on health system performance. Several studies have reported
a high completeness of registration and a high validity for most variables in the
Finnish registers [143–146]. However, there have been some suspicions about
the quality of register data in the case of hip fracture [147], and recent a Finnish
current care guideline has requested more validity analyses [128]. Moreover, there
have not been studies where the validity of data is judged against the health system
performance assessment purposes.
Therefore, an investigation aimed at clarifying these issues was conducted
[2]. Prospectively collected data for 106 consecutive hip fracture cases in the
Kuusankoski Regional Hospital between January 1, 1999 and January 31, 2000 and
the register data for patients treated during the same period in the surgical ward
of the Kuusankoski Regional Hospital were compared and the observed differences
were checked from the medical records.

4.2.1 Completeness of registration

It was rather straightforward to check whether the patients in the register data
having a recorded hip fracture diagnosis (ICD-10: S72.0, S72.1, S72.2) in a
particular hospital period were found in the prospective data and vice versa. As
it was very likely that all hip fracture patients were identified from the register or
prospective data, and as the hip fracture diagnosis could be confirmed from the
medical records, it was possible to evaluate whether all true hip fracture cases were
identifiable from the data sources. It turned out that there were in total 105 patients
with a confirmed hip fracture diagnosis. The audit data missed one patient and the
register data two patients. That indicated very good completeness of registration
in both data sources. Three extra hip fracture candidates clearly had a false positive
diagnosis in the register, and a further five candidates had been excluded from
the prospective data on the basis of additional information, i.e. register data may
slightly overestimate the number of fresh hip fractures unless appropriate data
abstraction rules are used for the exclusion of extra cases. In summary, this analysis
confirmed that the validity of the data seems to be acceptable for simple incidence
monitoring purposes.
4.2.2 Conceptual model for hip fracture performance

monitoring
It was more difficult to evaluate the quality of register data for performance
assessment purposes, because data validity must be judged against the intended
utilization purposes [148]. However, there is no obvious golden-standard that
would have told which aspects and issues of performance monitoring should have
been evaluated between the data sources [149]. In addition, it was obvious that
the quality criteria could change with time. For instance, the detailed diagnosis
of hip fracture is more important in the operating room than during the final
stages of rehabilitation, i.e. the requirements for correct diagnosis in data may be
more important in a surgical ward than in a nursing home. Therefore, in order to
make any statements concerning the quality of data, it was essential to first outline
the properties of required measures carefully under a fixed conceptual model for
performance assessment (determine how data result from the theory), and then
evaluate the observed data against these requirements (examine how well the
theory can be reconstructed using the properties of actual observed data). The
same idea was expressed more explicitly in chapter 2.3.1.
In the case of hip fracture treatment, the importance of practical and reliable
information on evidence-based performance assessment has been recognized [126],
and good experiences of the Swedish Rikshöft-registry have led to the formalization
of data production procedures for the purposes of a standardized audit of hip

fractures in Europe (SAHFE) [127]. The suggested prospective collection of clinical

data represents the practical “state-of-the-art” consensus for data requirements for
assessing the treatment of hip fracture. This work has produced a list of SAHFE-
variables which reflect the operationalizations of patient-level concepts that have
been considered important in the case of hip fracture. However, the more or less
randomly ordered list of variables is not a useful conceptual model.
In order to define a conceptual model, the SAHFE-variables were chosen to
represent an adequate base for data requirements in terms of patient level concepts.
The extended pyramid framework presented in Chapter 2 was used in the systematic
characterization of these concepts. Each phenomenon represented by a concept was
assumed to be attributable to an individual on a continuous time scale. It was also
assumed that each concept can be described using a systems approach with a (limited
or unlimited) number of states so that the system is always in some of these states
at any time point. All concepts were classified on the basis of a subjective evaluation
of potentially suitable measurement scales and types of statistical distributions
(theoretical measurement properties). The second classification dimension
was the temporal stability of the concept, i.e. it was considered, for example, at
what time intervals the concept makes sense and how often the concept should
be measured. In addition to these statistical classification principles, a contextual
interpretation was linked to each concept in order to make sure that the resulting
operationalization offers desirable insights on the phenomenon of interest. The
linkage was mainly based on the conceptualization used in SAHFE-metadata, but
it was complemented using theoretical models of performance assessment [10,
150, 151]. After this initial classification phase, the concepts were further classified
into more general groups in terms of their contextual interpretation, temporal
stability and theoretical measurement properties. In the third phase selected
measures from the SAHFE-list, from the prospective Kuusankoski data, and from
the register data were mapped to the conceptual model and the goodness of the fit
in terms of comprehensibility and simplicity of the model was evaluated. These
three qualitative phases were repeated until the resulting conceptual model was
considered satisfactory.
The final conceptual model presented in Table 1 gave a pragmatic mapping
between concepts from a performance assessment theory, SAHFE metadata and
structured data entry in prospective data and secondary register-based data in
terms of contextual interpretation, temporal stability and theoretical measurement
properties. Of the model dimensions, the biological constants do not change in
time and therefore one measurement is generalizable to all times. Biological events,
accident/fall history, and hip fracture event represent dimensions for which actual
values of measures are recorded in the proximity of some observable event. All
other dimensions relate to phenomena that potentially change in time and should
be continuously monitored. In practice continuous monitoring is impossible, and
actual measurements of continuously changing phenomena can be made only

Table 1. Conceptual model dimensions and their realizations in different data sources
SAHFE measure Prospective data measure Register data measure

Biological facts
Biological events date of birth, death, year of date of birth and death date of birth and death
menopause, menarche
Biological constants sex sex sex
(dna)
Biological measures
Physical examinations height, weight
Laboratory tests hemoglobin, creatine, albumin,
bone density
Demographic history
Household composition living alone marital status, living alone
Place of living residential status residential status, area of living area of living, level of care1
Socioeconomic history
Education, occupation,
economic resources
Health related behavior
Risk behavior smoking, alcohol intake alcohol usage alcohol- related inpatient care1
Diet, activities
Subjective quality of life
Overall well-being pain, psychological state, fear pain
of fall
Objective need for care
Diseases/symptoms comorbidity, complications complications recorded diagnoses, use of care1
Physical functioning walking walking, activities of daily living need for care
Cognitive functioning abbreviated mental test score
Use of care
Technical aid walking aids use of assistive technology,
changes in living environment
Medication use of painkillers
Health care utilization provider, type of stay, length type of care, rehabilitation, provider, type of care,
of stay dates for hospital visits admission and discharge dates,
diagnoses, operations
Accident/fall history
Properties of fall date of fall, place of fall date of fall, place of fall place of and reason for
hospitalized injury1
Hip fracture event
Properties of fracture time of fracture, occurrence day of fracture, reason for place of and reason for
place, fracture type, fracture fracture, occurrence place, hospitalized injury
side fracture type, fracture side
Initial treatment process hospital ID, admission time, admission day, admission hospital ID, admission day,
admission source, start of source, day of operation, reason admission source, diagnoses,
operation, reason for operative for operative delay, operative day of operation, operation
delay, type of surgeon and method, complications in codes, discharge day, discharge
anesthetist, type of anesthetic, operation, reoperations, destination
operative method, length of discharge day, discharge
surgery, type of prophylaxis, destination
complications in operation,
reoperations, time to
mobilization, discharge time,
discharge destination
Clinical stability ASA grade, comorbidities, ASA grade, reason for operative recorded diagnoses, use of care1
coexisting fractures, reason for delay in medically unfit patients
operative delay in medically
unfit patients
1) Indirect measurement requiring record linkage.

at selected time points. In the prospective data, the measurement was done in
connection with a hip fracture event and for certain measures also at two weeks, at
four months, and one year following the fracture. In the register data, the recording
took place at each discharge.
The constructed conceptual model explicitly revealed commonalities between
the prospective and register data. As expected, the content of prospective data was
richer in clinical detail than the register data, even after using indirect measures
of certain concepts. The conceptual model also showed which dimensions were
comparable between the available data sources. In the actual empirical analyses
these comparable dimensions were examined by using the available complete
prospective data for 106 hip fractures and corresponding register-based data.
4.2.3 Consistency between prospective and register-

based data
Three types of comparisons between the data sources were used depending on the
theoretical measurement properties of the variables. In the type I comparisons,
the same rather easily measurable variable was available in both data sources,
and detected differences could be further checked from the medical records
of the patients. The percentage of cases with a recorded value for the variable
(completeness) and proportion of correctly recorded values for cases with some
recorded value (accuracy) were calculated for such variables. Completeness and
accuracy were very good for age, sex, area of living, admission and discharge
information as well as for hip fracture diagnosis and operation day and type. Side
of fracture was seldom recorded and accuracy of information on place of accident
was poor in the register.
In the type II comparisons, conceptually similar measures which had
substantially different operationalizations in both data sources were examined.
From this category, follow-up information was a particularly important issue
for performance assessment purposes. Using the register-based data it was
straightforward to extract information on the use of inpatient health services and
mortality on a daily basis. The derivation of corresponding information from the
prospective data was much more complicated and required extra assumptions and
interpolation. The overall follow-up patterns were very similar indicating that the
register-based data suits better for the follow-up of use of inpatient health services
than prospective data. However, some extra work needed for the prospective
collection of follow-up data may be useful in order to produce data on issues that
are not available in the register-based data, but offer clinically interesting follow-
up information, such as the use of technical aids, pain, and validated measures of
functional disability and quality of life.

Type III comparisons were actually a mixture of type I and type II comparisons.
This approach was used to evaluate the more detailed diagnosis of hip fracture. Using
a detailed hip fracture classification of femoral neck fracture, trochanteric fracture
or subtrochanteric fracture, the overall agreement between prospective and register
data—calculated as a sum of equivalent classifications in data sources divided by
sample size—was 86.3%. The 95% confidence interval for overall agreement—
calculated with a nonparametric bootstrapping technique [152]—was 79.4–92.2%.
The proportion of agreements that were not due to chance was measured using a
delta coefficient, which is based on the model of multiple-choice tests and avoids
certain known deficiencies of the traditionally used kappa coefficient [153]. Kappa
may be considered as an approximation of delta, as both yield values that are very
similar when at least one marginal distribution is balanced or when both marginal
distributions are moderately unbalanced in the same direction [154]. According
to the estimated delta coefficient a high proportion 81.3% (95% CI: 71.3–91.3%)
of agreements were not due to chance. The delta coefficient was somewhat higher
than the kappa coefficient of 73.5% (95% CI: 61.2–85.7%) indicating that the
assumptions of kappa were not satisfied. In addition, the accuracy of diagnoses
was addressed by considering bone anatomy as a biological entity identifying true
fracture status [155]. This made it possible to assume that the trait underlying
different fracture types was continuous, and a simple latent trait model in the form
of a polychoric correlation coefficient was estimated [156], yielding a value of
0.92 (95% CI: 0.86–0.98) which also reflected a very high consistency between the
register and prospective data.
4.2.4 Conclusions on validity

In order to analyze the quality of register data for the purposes of performance
assessment, a conceptual model had to be constructed. There were several novel
aspects in the model. The statistical properties of the concepts had a very important
role in the classifications and the special features of different data sources were
simultaneously taken into account during the construction process. As a result, the
presented conceptual model revealed the most important assumptions affecting the
actual results of analyses, and showed how it was possible to produce a pragmatically
useful mapping between problem-oriented and data-driven theories.
As expected, the register data lacked clinical detail, but had a data structure
that allowed the complete observation of inpatient care history, and therefore
outperformed prospective data in this respect. The empirical validity analyses
revealed that the completeness of registration was very good, but appropriate data
abstraction rules should be used in order to avoid overestimation of hip fracture
cases. Also the accuracy of the most easily measurable variables and the quality
of follow-up information in the register was found to be very good. In addition,

detailed hip fracture classification showed a high consistency between the register
and prospective data. As a conclusion, register-based data seemed to be a valid data
source for the purposes of the performance assessment of hip fracture treatment.
However, there was room for improvements in the register. An inflexible data
structure for the performed operations causes problems if a patient has several
operations during a single hospitalization. There was also a need for additional
data, while especially the validated functional disability or quality of life measure
would have been useful. The content of these register data could also be improved
by giving standardized instructions that guide practitioners to record the relevant
additional operation codes available in the classification, such as on which side the
fracture was. Also a voluntary input of additional hip fracture event data to the
centralized register should be made possible, because the best option for practical
performance assessment purposes would be to combine SAHFE-type of data with
the register data.

5 Hip fracture incidence
The epidemiological component of the hip fracture surveillance system has an

important role, because without continuous monitoring of hip fracture incidence
it would be difficult to assess the effectiveness of prevention interventions and also
to prepare for changes in the need of care caused by changes in the population
structure. This chapter is based on articles 3–5.
In principle, the Finnish register data are well suited to the hip fracture
incidence calculation. Hospital admission and discharge days are easily observable
facts (i.e. have stable measurement properties in the terminology presented
in chapter 2), which are correctly and completely recorded in the Finnish data.
However, in practice it is not straightforward to determine the true number of hip
fracture events from the register data. In fact, there are several recognized potential
pitfalls in using hospital discharge data for calculating the number of injuries [157,
158].
Even though all hip fractures are identifiable from the hospital discharge data
(virtually all hip fracture patients require inpatient hospital treatment and hip
fracture is easy to diagnose), there are several discussable details, such as the exact
diagnosis codes to be used in the identification, the use of secondary diagnoses
in the hip fracture identification, and exclusion of certain nonstandard cases.
However, from the methodological perspective a much more important source of
bias (which is independent of the use of diagnosis codes) is related to the detection
of multiple hospitalizations of a single patient. At the conceptual level this means
that all events related to the same underlying disorder should be recognized. A
so-called care episode approach provides a sound methodological framework for
dealing with multiple hospitalizations in terms of discharge data [159]. In the case
of hip fracture, it is known that a typical treatment process starts with an emergency
admission to a hospital after the hip fracture event followed by an operation at the
surgical ward and continues with rehabilitation at other inpatient wards.
With a careful development of data abstractions rules to exclude multiple
hospitalizations it was possible to demonstrate that several previous Finnish
studies on hip fracture incidence had overestimated the numbers of hip fractures
[5]. More specifically, the decision to use the calendar year boundaries to combine
multiple hospitalizations in these studies was artificial and had no epidemiological
justification, because the underlying disorder had nothing to do with the beginning
or end of the year. In fact, it was likely that many patients having their fracture
during the final months of each year were erroneously counted as separate cases
for two consecutive years [3].
However, these considerations, combined with the problem that even the
clinical criterion of a fresh hip fracture may vary, point out that it is impossible to

obtain a unique estimate for the number of hip fractures using register-based data.
This ambiguity reflects the undesirable incompatibility between the theoretical
model and data. As it was not possible to change data, there was a need for an
alternative theoretical interpretation that would be more compatible with the
restrictions of data.
5.1 Aging-related hip fractures

It is known that hip fracture incidence increases exponentially with age [160].
This kind of population level functional dependency means that the condition
is related to aging [161]. Biological aging can be seen as a complex process
occurring stochastically in organs and tissues, which results in irreversible damage
accumulation and vulnerability to the failures in maintaining the integrity of
tissues and organs [162], and the components of a biological system are prone
to aging-related failures in the system [163]. Most aging-related conditions are
closely connected to alterations in certain tissue or organ, although a well-known
exception is death, which is a result of fatal failure(s) in any vital components of
human body.
Following this line of though, it became obvious that a hip fracture was also
a special case in aging-related conditions. In fact, hip fracture is not just an aging-
related failure in some tissue or organ, but typically a result of an accidental event
(such as fall). Cumulative damage in some tissues (such as bone and muscle)
increases the probability of (accidental or pathological) fracture, but it is also likely
that some aging-related conditions increase the risk of accident. In this sense, hip
fracture seemed to capture the effects of non-fatal failures in components of the
human body, which most likely appear in the form of some kind of disabilities.
Therefore, by drawing a methodological analogy to the event of death, it could be
hypothesized that the event of hip fracture actually represents a multidimensional
disability event (“disability death”).
From this point of view, it was clear that the occurrence of first aging-related
(low energetic) hip fracture would be the most interesting one, because it gives
the upper limit for the time of development of critical hip fracture risk factors
(for a hip-fracture-related chronic disability condition). Therefore, the actual
problem of identifying the interesting hip fracture episodes could be transformed
to the identification of the first aging-related hip fractures of the patients. This
theoretically justifiable approach turned out to be in better concordance with the
register data than the identification of all hip fracture events [3]. In terms of chapter
2, an extra interpretation–operationalization phase was needed.

5.1.1 A method for identifying a first aging-related hip

fracture
The pragmatic problem in identifying the first hip fractures was that only a limited
backward follow-up was possible with the register data, i.e. some fractures could
have occurred outside the observation window. A straightforward solution would
have been to use all available data for the determination of first hip fractures, but
it was considered more convenient to use the same “hip fracture free” clearance
period, that is shorter than the minimum backward follow-up time in the data, for
all individuals.
In order to justify the length of clearance period, it was assumed that there
may be a cut-off point after which the probability of having a new hip fracture
is reduced to the level it would be even without a preceding hip fracture, i.e. that
the preceding hip fracture would be an “uninformative” predictor of the new hip
fracture [3]. In the concrete analyses, the first admissions (index points) with
the diagnosis of hip fracture in 1998–2002 were identified for each patient. For
each index point, backward time to the previous hip fracture admission of the
same patient was calculated. Time was measured in months. There were follow-
up data from the period 1987–2002, so the minimum (backward) follow-up time
was 11 years. If there was no previous hip fracture admission, the observation was
considered to be censored at the beginning of 1987.
The discrete hazard probabilities of the form

P(T = t | T ≥ t),
reflecting the conditional probability for an event occurrence at time t on condition

that no event had occurred until t, were estimated for each follow-up time 0 < t ≤ 160
as
c(t) ,
h(t) =
n(t)
where c(t) was the number of events at time t and n(t) was the number of persons
at risk at time t. To diminish the degree of random variation, estimated hazard
rates were smoothed using a simple moving average technique with polynomial
weights [164]. The smoothed hazard rates were referred as observed probabilities.
For the calculation of corresponding probabilities in the case that no preceding
hip fractures had occurred (expected probabilities), there was first a need for
reasonable estimates for the incidence of first hip fractures. It was assumed that
clearance periods between one and ten years are reasonable a priori, and that the
average incidence between the years 1998 and 2002 can be interpolated to the
whole observation period. After the preparation of risk population data (described

below in chapter 5.3) it became possible to calculate age- and sex-group specific
incidence rates by using standard methods [165]. Sexes were analyzed separately
and clearance periods of one and ten years were used. A logarithmic transformation
was made for the estimated age-group specific incidence rates, which linearized the
exponential increase of rates with age. After that, a linear regression model was
fitted to the logarithmic incidence rates so that an interpolated incidence rate could
be predicted for any age by using the estimated model [166]. For ages below 40
years, a constant incidence rate was used in the prediction, because an assumption
of the exponential increase in incidence with age was not satisfied for young ages
[167].
As age and sex were known for each index period identified from the data, it
was possible to subtract one month from each age and then predict the expected
probability of a hip fracture after a one-month (backward) follow-up for each
case in the risk population by using the estimated models. As the incidence rates
were person-year based, they had to be divided by 12 so that they corresponded
to a time period of approximately one month. An average expected probability
of hip fracture could then be calculated by summing the expected probabilities
across the risk population and finally dividing by the size of the risk population. By
subtracting one month more from each age and making the predictions again, an
average expected probability of hip fracture after a two-months (backward) follow-
up could be derived. Following the same procedure, expected probabilities up till a
160-month (backward) follow-up were calculated.
The observed (solid line) and expected probabilities (dotted lines) are presented
in Figure 4. The curve of observed probabilities showed that there was a non-
stabilized risk for a hip fracture admission up to about 120 months. The difference
between expected probabilities based on one- and ten-year clearance periods was
quite small, and the observed risk also remained higher than expected during the
first 7–10 years. Based on these analyses, a conservative ten-year clearance period
was selected to make sure that the risk was reduced to the same level as that without
a preceding hip fracture. The ten-year criterion has also been used in other studies
[168].

Figure 4. Conditional (hazard) probabilities of having a preceding hip fracture as a function

of backward time in months from the first fracture in 1998–2002. The smaller picture is a
tenfold magnification of the final months. Dotted curves represent the expected probabilities
of having a hip fracture with one year (upper) and ten year (lower) clearance periods.
5.2 Risk factor extraction

The use of a clearance period in the determination of first hip fractures identified
the persons with first hip fracture in addition to the determination of the number of
first hip fractures. This made it possible to extract the status of certain hip fracture
risk factors available in the data in relation to these patients. Three different risk
factor extraction techniques, internal, external, and empirical extraction, were used
for this purpose.
Internal extraction corresponded to the use of data abstraction rules within the
register data. For example, using the Finnish data it was not enough to determine
the index hip fracture period, but the actual day of hip fracture had to be inferred
using appropriate rules. The basic idea in the rule construction was to find one
simple rule at a time that was applicable to the majority of the population, and
each additional rule was then developed and applied only for the remaining
population. The index period could be a very long stay in residential care, while
the diagnosis referred to an accident that happened near the discharge day, or the
index period could be a surgical period, while the actual accident had happened

before the admission. Therefore, the first rule aimed to identify hospitalizations
with an operation code for a hip fracture procedure in the proximity of the
candidate period, and the second rule identified hospitalizations at the surgical
ward. The third rule was developed to identify patients who had died shortly after
the first diagnosis. With these rules a reasonable index period could be found for
about 96.9% of patients, while the last three percent required more complicated
identification strategies.
After the detection of the most reasonable index period, the basic covariates
(such as the patient’s age, sex, area of living) could be extracted from that period.
Type of fracture was determined on the basis of diagnoses from the operative
surgical period if possible, because those were most likely the first ones based on x-
ray examination. Further abstraction rules were required for the definition of more
complex covariates. For example, a person was classified as a long-term care patient
if there were at least 90 days of inpatient care or a recorded administrative decision
for long-term care immediately before the fracture.
In external extraction, variables linking individuals to aggregate levels (such
as area codes) were first internally extracted and then external data describing
aggregate units were linked back to each individual with the help of these variables.
For example, the area code (municipality) at the index period was used to classify
each patient as rural or urban (including semi-urban areas) using the official
grouping defined by Statistics Finland (rural municipalities are those municipalities
in which less than 60 per cent of the population lives in urban settlements).
Figure 5. Daily numbers of the hip fractures in Finland 1998–2002

In empirical extraction, the observed data were used to justify the definitions.
For example, the Finnish register data allowed accurate calculation of the daily
numbers of hip fractures, also separately for long-term care patients. After smoothing
with a moving average technique, a small but clear seasonal variation attributable
to non-institutionalized persons was detected (Figure 5). Without these kinds of
preliminary analyses it would have been more difficult to justify that winter season
should last from November to April and correspondingly the summer season from
May to October. The method developed above for the definition of the length of
clearance period was another example of empirical extraction.
5.3 Risk population data

For incidence calculations, data on risk population were also needed, and official
population figures (taken on the last day of the years 1997–2002) in 5-year groups
were obtained for each municipality. Municipality worked as an aggregate unit in
external extraction, and also allowed easy determination of population figures for
any combination of these basic units.
It was more difficult to determine the risk population for internally extracted
risk factors. There were no data available in the official statistics on the numbers
of people in long-term institutional care with the required, rather detailed
stratification. Therefore, data on long-term institutional care for the period 1997–
2002 were extracted from the Finnish Social and Welfare Care register using the
recorded decision for long-term care or more than 90 days of continuous care in
a single institution as the extraction criteria (n = 532 169). In principle, these data
allowed the calculation of the total numbers of clients in long-term institutional
care with the required stratification, as the register included individual-level data
on all inpatient hospital and nursing home care in Finland.
However, it turned out that the calculation technique used in the derivation
of official statistics utilized only census data (collected at the final day of each
year), while the number of long-term care patients seemed to be much higher if
discharge data were also taken into account [4]. This was examined more closely
by cross-tabulating both numbers with the stratification of sex, year, age, and type
of institution. The proportion p of census data clients (O) in combined data clients
(E) corresponded to the prevalence proportion p = O / E. In order to describe the
differences in the proportions and to estimate the missing E for the latest year, it
was assumed that the counts O were Poisson distributed and a Poisson regression
model was formulated of the type
log(O / E) = b0 + b1x1 + b2x2 +···+ bkxk,

where xi are the covariates and the model parameters βi are log relative risks [169].
The final model estimated with the years 1997–2001 included main effects for
age, sex, and type of institution as well as the year-effect for each age-institution
combination. The fit of the model was very good in terms of χ2-test, and no
oddities were detected in residual plots. As an assumption of binomial distribution
instead of Poisson distribution was theoretically more elegant, the model with
a log-binomial formulation was also estimated [170]. The point estimates were
practically identical, but the variances of the coefficients were somewhat smaller,
and turned out to be almost equivalent to the robust variance estimates in Poisson
regression. As a conclusion, an easily estimable Poisson regression model resulted
in good point estimates, but gave more conservative (wider) confidence intervals.
The proportions increased with age and also with year in social institutions,
but were smaller among men and in social institutions in general. Reasons for
most differences could be traced back to the definition of long-term care, but
also to the properties of the register data especially in the Social Welfare Care
register. It was concluded that the approach where both census and discharge data
had been taken into account was better suited for epidemiological purposes [4].
The practical problem from the incidence calculation point of view was then to
calculate the numbers of clients on the final days of the years 1997–2002. That was
straightforward for all but the most recent year which required prediction using
the estimated model, because no discharge data concerning the year to follow were
available.
The procedures described above resulted finally in simultaneously observed
risk populations on the final day of the years 1997-2002 with stratification by sex,
age, urbanity, and institutionalization. A common method has been to interpolate
average population at risk by using the mean of observed census data on two
consecutive years. In order to generalize the interpolation so as to also incorporate
the seasons, the population figures were plotted against the years in each group
with stratification by sex, age, institutionalization, and urbanity. Visual inspections
suggested that the associations between the years and population figures were rather
simple, and that simple curve-linear models could be used to approximate the
relationship. Therefore, linear regression models with constant, year, and squared
year as exploratory variables were used to summarize the curve-linear trend in the
population figures. These models allowed an interpolation of population figures
to any day during 1997-2002. These models were then used in the calculation of
approximate follow-up times with the required stratification including seasons and
years.
These denominators were desirable in terms of accuracy, but not perfect,
because an assumption of a stable population (no short-term fluctuations in
migration or mortality) was needed for interpolating census day population
figures to appropriate mean follow-up times [171]. In addition, it was not possible
to exclude persons who had had hip fracture during the preceding ten years. Bias

resulting from keeping the prevalent pool in the risk population was estimated to be
small because of low incidence and high mortality, and to be towards conservative
estimates (i.e. to underestimation rather than overestimation of group specific
incidences).
5.4 Hip fracture incidence between 1998 and

2002 in Finland
After the adequate data were finally assembled, it was straightforward to examine
incidence rates with standard methods [165]. Standardized incidence was higher
among women and long-term care patients, and a long increase in standardized
incidence in Finland was found to be stabilized [5].
From the hip fracture surveillance system point of view, it was particularly
important to be able to monitor the small-area variation in hip fracture incidence.
The estimation of the hospital-district-specific standardized incidence rates was
performed by using a hierarchical estimator, because the small sex- and age-
group-specific numbers of events were prone to extensive random variation. More
specifically, it was assumed that better estimates (in terms of mean squared error)
could be obtained by a weighted sum of area specific estimates and an appropriate
guess [172]. In an empirical Bayes-approach the guesses are chosen to be the average
estimates obtained without area stratification [173]. The hospital-district-specific
weights and group-specific rates were estimated for each year by using an iterative
algorithm [172]. According to the results [5], the incidence had been somewhat
higher in Helsinki, but otherwise the differences were rather small (Figure 6).
The constructed population-based data also allowed the examination of the
simultaneous impact of several risk factors on hip fracture incidence [3]. Year was
not considered as an interesting factor, so the data from years 1999–2002 (during
which the average incidence had remained constant) were combined. A Poisson
regression model was used in the analysis [169]. There were a lot of significant
interactions between risk factors, and it seemed to be more fruitful to model the
cross-tabulated data in the style of log-linear analysis [174] instead of aiming at
determining the independent impacts of the risk factors on the incidence by means
of epidemiological modeling. In this sense, the modeling goal was to compress
the representation of the cross-tabulated data, and all main effects and statistically
significant interactions up to the third degree were included in the model. The fit
of the final model was reasonable in terms of χ2-test, and residuals did not reveal
any exceptional observations. The relaxation of the Poisson assumption by using
a negative binomial regression model [175] turned out to be unnecessary as it was
after estimation almost equal to the Poisson model, indicating that no significant
overdispersion was apparent.

225 275 325 375 425
1998
Helsinki
1999
2000
2001
2002
1998
Keski-Suomi
1999
2000
2001
2002
1998
Varsinais-Suomi
1999
2000
2001
2002
1998
Hyvinkää-Porvoo
1999
2000
2001
2002
1998
Kymenlaakso
1999
2000
2001
2002
1998
Satakunta
1999
2000
2001
2002
1998
L-Pohja
1999
2000
2001
2002
1998
Jorvi-Peijas
1999
2000
2001
2002
1998
Lohja-L-Uusimaa
1999
2000
2001
2002
1998
Pirkanmaa
1999
2000
2001
2002
1998
Päijät-Häme
1999
2000
2001
2002
1998
Vaasa
1999
2000
2001
2002
1998
P-Karjala
1999
2000
2001
2002
1998
Kanta-Häme
1999
2000
2001
2002
1998
P-Pohjanmaa
1999
2000
2001
2002
1998
I-Savo
1999
2000
2001
2002
1998
Lappi
1999
2000
2001
2002
1998
P-Savo
1999
2000
2001
2002
1998
Keski-Pohjanmaa
1999
2000
2001
2002
1998
E-Savo
1999
2000
2001
2002
1998
E-Karjala
1999
2000
2001
2002
1998
Kainuu
1999
2000
2001
2002
1998
E-Pohjanmaa
1999
2000
2001
2002
200 250 300 350 400 450
Figure Figure
6. Age- 6.and sex-adjusted
Age- hip fracture
and sex-adjusted incidence
hip fracture in Finland
incidence 1998-2002
in Finland by hospital
1998–2002 district,
by hospital
district, rates per 100 000 persons years. Year-specific rates with confidence
rates per 100 000 persons years. Year-specific rates with confidence intervals are plotted forintervals areeach
plotted for each hospital district, from 1998 (uppermost) to 2002 (lowest). The short vertical
hospitalline
district,
at eachfrom 1998district
hospital (uppermost) to 2002
corresponds (lowest).
to the The
average short
rate vertical line
in 1998–2002 in at each
that hospital
hospital
districtdistrict. The dashed
corresponds to theline gives rate
average the overall mean incidence
in 1998-2002 during 1998–2002
in that hospital district. Theindashed
Finland.line gives
the overall mean incidence during 1998-2002 in Finland.

As there were quite a lot of parameter estimates, it was assumed that the simplified
cross-tabulation based on the model predictions had captured the systematic
characteristics of the original data and removed the random noise, and the
actual effects were investigated in the terms of group-specific predicted rates. The
main results illustrated in Figure 7 indicated that there was a higher hip fracture
incidence among older and institutionalized persons. The incidence among women
managing at home increased more rapidly that among men, but in institutions
the risks were rather similar. The incidence was higher during the winter time for
non-institutionalized persons. The effect of urbanization was more complicated
suggesting that the effect was different between men and women, and that there
was a tendency for higher incidence among institutionalized persons, the youngest
of the men and oldest of the women in urban areas.
The results could be interpreted from the population aging point of view,
and seemed to reflect aging-related disability [3, 5]: Institutionalization is a sign
of reduced coping with daily activities [176], while women tend to have more
disabilities than men as age increases [177], and the additional hip fractures during
winter occur mostly among persons with a good enough functional status to allow
walking outdoors [178]. As reduced functional capacity and coping with physical
activities of daily living are known to be important risk factors for hip fracture
[179], it was hypothesized that hip fracture incidence would be proportional to the
prevalence of disability related hip fracture risk factors in the underlying population.
The hypothesis was not directly falsified and requires further examination, as a
Men Men Women Women
Men Women
Incidence Incidence
women(logarithmic scale), women
Incidence (logarithmic scale), men (logarithmic scale), men Incidence (logarithmic scale),
2500 2500 Institutionalized Urban Urban

2500 Institutionalized 2500 Institutionalized
Institutionalized Rural
Winter Rural
Winter Summer
Urban Summer
Urban
1000 1000 1000
1000
Rural Rural
400 400 400

400
Non-institutionalized Non-institutionalized
Non-institutionalized Non-institutionalized
150 150 150
150
Urban Urban
50 50 50 Winter
50 Winter
Rural Summer Summer
Rural
60 70 80 90 60 70 60
80 70
90 80 90
60 70 80 90
Age Age Age
Age
Figure 7. Hip fracture incidence in Finland in 1999-2001 by sex, age, institutionalization, season,
and urbanity. Rates are per 100 000 person years, logarithmic scale. Dotted lines refer to rates
Figure 7.urban
Figure 7. Hip fracture
for Hip areas.
fracture
incidence incidence
in Finland inin Finland inby1999-2001
1999-2001 by sex, age, institutionalization,
sex, age, institutionalization, season, season,
and urbanity.and urbanity.
Rates are perRates are person
100 000 per 100years,
000 person years,scale.
logarithmic logarithmic
Dottedscale. Dotted
lines refer to lines
rates refer
for to rates for
urban areas. urban areas.
high correlation between the prevalence of outdoor walking ability and hip fracture
incidence was detected [3].
5.5 Conclusions on hip fracture incidence

monitoring
The feasibility of producing information for the purposes of monitoring hip
fracture incidence based on Finnish administrative register data was demonstrated
in the current chapter. It was pointed out that unwary use of Finnish register data
had resulted in biased estimates of the numbers of hip fractures. As there were
problems in the reliable identification of all hip fractures, the definition of hip
fracture was linked to population aging. Within the renewed theory only the first
aging-related hip fracture of each individual had to be found from the data, and
a new method was developed to support this purpose. A systematic account of
pragmatic problems in the extraction of risk factor and risk population data and in
the data analyses was given with proposed solutions. The empirical results suggested
that hip fracture incidence would be linked to the prevalence of population level
disability. This connection may be important from the monitoring point of view,
because under this hypothesis, hip fracture incidence trends would also tell about
more general disability trends of the population.

6 Hip fracture treatment
Not all hip fractures can be prevented so it is important that the effectiveness
and quality of treatment following hip fracture are continuously evaluated.
The basic idea in this kind of outcomes research is to elucidate cause-and-effect
relationships between health actions (treatments, interventions) and health gains
(outcomes) [150]. Usually large scale empirical investigations that aim to do that
are observational studies [180], as it is not possible to assign subjects at random
to treatment or control groups, as would be done in a controlled experiment. It
is quite common that administrative register data are used in outcomes research
[103]. Good examples exist, but the actual assumptions that would justify the use
of register data in outcomes research are seldom reported. Such practice makes it
impossible to evaluate whether the key assumptions have been uncritically taken
as given or whether the assumptions have been carefully tailored to the case under
study. Therefore, it was considered important to provide detailed justification for
the register-based monitoring of hip fracture treatment, i.e. to aim at finding an
adequate compromise between problem-oriented and data-driven theories as
proposed in chapter 2.
6.1 State space for a hip fracture care episode

The available data consisted of individual-level data on the use of health services for
hip fracture patients. Such data reported directly observable events that determine
whether an individual had used certain health service at a particular moment in
time. The events describing the use of health services connected data to an idea
of record linkage, where such observable events are assembled to the “book of
life” of an individual [181]. In this sense, data provided descriptions of patient-
specific paths through the observable components of the health system. As the
aim was not to analyze complete health service usage, but to examine hip fracture
related treatment, an episode-of-care approach was adopted for operationalization
purposes [182]. A care episode refers to a series of health-related events linked to
a particular health problem [183]. That is, the episodes-of-care approach helps to
define health-problem-specific boundaries for the health system structure and to
identify interesting elements from the care processes [159].
The first task in the reconstruction of hip fracture care episodes was to
identify the beginning of the episode. A natural choice for an index event was the
surgical hospitalization after the first aging-dependent hip fracture as described
in subchapter 5.1. In addition to being compatible with incidence monitoring, a

restriction to first aging-dependent hip fracture homogenized the study population

appropriately.
The next task was to categorize the usage data so that the observed events
would have a relevant interpretation from the hip fracture point of view. The idea
was to construct a state space that could be used to describe possible transitions
across a set of care regions in time [184, 185]. Many different ad hoc classifications
of health services had been utilized in the analyses of hip fracture treatment [120,
186, 187]. In addition, theories and models describing the processes of care for
the elderly were also applicable to the case of hip fracture [188]. More specifically,
it has been suggested that care and services intended for older persons could be
categorized in terms of the level of care and the need for care [189]. The level
of care is a description of the intensity and type of service, while the need for
care is attributable to the patient’s characteristics (such as health status, severity of
medical condition and functional disabilities).
Following this idea, the actual care regions were considered to consist of
four different levels of care that were identifiable from the data (in increasing
order of intensity): 1) home (including home care, ordinary service houses and
outpatient care), 2) nursing home (service houses with 24-hour assistance and
residential homes), 3) health center (inpatient ward of local primary care unit),
4) hospital. In addition, an absorbing death state was included to the state space.
These levels of care are common to the whole country in the sense that all the
more specialized components of local structures can be classified into these main
categories. Validation analyses reported in article 2 showed that this categorization
was compatible with the one used in a standardized hip fracture audit, and that the
register data had excellent properties in this respect. The constructed state space is
presented graphically in Figure 8.
Home
Nursing home
Initial hip fracture

Health center
hospitalization
Hospital
Death
Figure 8. State space for the hip fracture care episode

6.2 Effectiveness in terms of care process

The constructed state space was well suited to the description of care processes. In
fact, by using only data that were based on directly observable events, it became
possible to produce hard facts about the care episodes. The problem was the lack of
theoretical links between the care episode data and the outcomes of care. However,
by assuming that the outcome of interest was effectiveness—operationalized as a
change in health of the target population—and that the patients were not treated
ambiguously, it was possible to hypothesize that the decision to treat a patient
in a given care region is associated with the patient’s health status [190-192]. In
other words, it was assumed that the transitions between care regions reflected
the appraised need for care, which is also a sign of a qualitative change in health
status. For instance, an acute admission to a hospital typically reflects a sudden
drop in health status of a patient, while a discharge to lower levels of care would
be expected to reflect a somewhat improved health status. This assumption was
supported by empirical evidence [75, 176] showing that at the population level
there had been a strong positive correlation between the intensity-ordered levels of
care and the need for care in Finland.
Using the notation from chapter 2.3.1, there was a problem-solving task
related to effectiveness, where a mental representation for the task solution
Iuser(effectiveness) | producer(health service usage), was based on the data Duser(effectiveness) | producer(health service usage),
that had been derived from the externally available data Dproducer(health service usage), and
that had an acceptable cognitive fit with the internal representation of suitable data
Duser(effectiveness). As a conclusion, an appropriate compromise between the data-driven
and the problem-driven theories could be reached in the terms of available data.
6.3 Multivariate responses in outcomes research

By following the episodes-of-care approach it became clear that the whole
realization of the care process was important in performance measurement. This
was somewhat conflicting with the traditional outcomes research models that
typically deal with univariate response variables or try to combine information
from several indicators [96]. In the methodological sense it was obviously more
proper to deal with a multivariate response variable that described the whole care
process. This posed certain challenges, because no well-established methods were
available for concrete analyses.
More specifically, the available register data allowed complete observation of
realized levels of care on a daily basis for each individual. The care process of an
individual j consisted of a sequence of multinomial observations, i.e. corresponded
to the repeated categorical response profile

Yj = Y1j ,..., Ykj ,
where k is the length of a sequence. One option for data analysis would have been to
use a multivariate probability distribution for the responses [193]. However, from a
pragmatic point of view it was considered more important to be able to summarize
the multivariate responses in a comprehensive, yet intuitively meaningful way. This
kind of exploratory perspective for the data analysis of categorical response profiles
has been developed within the sequence methods framework, where the basic idea
is to analyze whether useful patterns can be found among the sequences [194].
However, no methods matching the targeted purpose were found in the literature.
The most relevant studies concentrated on the clustering of sequences to appropriate
groups [195, 196], but did not offer much help on the visual summarization of
sequences within the groups.
6.3.1 Traceplots for care episode visualization

I show here that it was possible to achieve the required comprehensive summarization
by simply plotting the actual realizations of patient-specific care episodes. First, it
was assumed that some symbol is linked to each possible category. It was obvious
that each care episode could then be expressed with a sequence of these symbols.
For example with the hip fracture case, if 3 stands for hospital use, 0 for being
home, and 4 for death, the sequence
333333300000000000033444444444
tells that the individual was followed for 30 days (one symbol for each day); first
she was seven days at the hospital, and then after twelve days at home she had a
readmission to hospital where she died.
Matters became more complicated when there were sequences for more than
one individual. Instead of using numerical symbols, it was beneficial to use colors
for telling the realized state. With the hip fractures it was natural to use a gray
scale where a darker color reflected the more intensive level of care. By using the
colors to indicate the level of care, it was easy to plot diagrams in which each row
corresponded to a care episode of one individual. This idea is illustrated in the
Figure 9, where the register-based hip fracture care episodes with one-year follow-
up were plotted for the randomly ordered cohort of patients living in the town of
Espoo during 1998–2001 (n = 516).

Figure 9. Unsorted traceplot for the hip fracture care episodes in Espoo 1998-2001. Each row
Figure 9. Unsorted traceplot for the hip fracture care episodes in Espoo 1998–2001. Each
corresponds to row
a care episodetoof
corresponds oneepisode
a care patient, which
of one consists
patient, of 365 ofdaily
which consists realizations
365 daily realizations(one year
(one year follow-up) of the level of care, where the darker color reflects more intensive level
follow-up) of the level
of care of care,
(white where
= home, the=darker
light gray nursingcolor reflects
home, middle more
gray intensive
= health level
center, dark of=care
gray
hospital,
(white=home, light black = death).home, middle gray=health center, dark gray=hospital,
gray=nursing
black=death).
The presentation of care episodes in a random order resulted in a diagram
without any simple interpretation. Intuitively, it was clear that the sorting of
episodes could reveal some patterns among the sequences. As there were no
known solutions to this kind of problem, a heuristic algorithm for the sorting was
developed. The basic idea in the algorithm was to measure similarity between the
care episodes and to find an optimal location for each episode in the figure. As the
aim was to find a visually simple presentation, common sense logic suggested that
the episodes with a lot of common realized states had to be similar. This idea was
formalized by measuring similarity between two episodes in terms of the number
of equal daily states, i.e. as a sum of the daily specific Hamming-distances (zero if
the realized states were the same, otherwise one).
In the actual algorithm two linked lists were used. The first singly linked list
(unsorted list) was initialized to contain all episodes in a random order, and the
second doubly circular linked list (sorted list) with a single seed episode consisting
of only death statuses. The fundamental loop in the algorithm aimed to move
one episode from the unsorted list to an optimal place in the sorted list. More
specifically, the unsorted list was looped over one episode at a time, and distances

Death without discharge to home
Death after discharge to home
Long-term care at health centre
Discharge to home
Long-term care at nursing home

Mixed care episodes
Figure 10. Sorted traceplot for the hip fracture care episodes in Espoo 1998-2001. Each row
Figure 10. Sorted traceplot for the hip fracture care episodes in Espoo 1998–2001. Each
corresponds to thecorresponds
row care episode
to theofcare
oneepisode
patient, which
of one consists
patient, whichofconsists
365 daily
of 365realizations (one year
daily realizations
follow-up) of(one
the year
levelfollow-up)
of care, of the level
where theofdarker
care, where
colorthe darker more
reflects color reflects morelevel
intensive intensive level
of care
of care (white = home, light gray = nursing home, middle gray = health center, dark gray =
(white=home,hospital,
light gray=nursing
black = death).home, middle gray=health center, dark gray=hospital,
black=death).
between this current candidate episode and the preceding and following episodes at
each possible place in the sorted list were calculated. The episode with the smallest
distance in the unsorted list was then moved to the place in the sorted list that had
given this smallest distance. The fundamental loop was repeated until all episodes
from the unsorted list had been moved to the sorted list.
A sorted traceplot for the Espoo cohort is presented in the Figure 10. The sorted
episodes had a simple structure which identified several qualitatively different
types of episodes. In this sense, the face validity of the developed algorithm was
pretty good. From the methodological point of view the aim of achieving an easier
interpretation from the sorting reminds one about the use of rotation in factor
analysis: there was no unique solution and the interpretability of the resulting simple
structure was the criterion that finally determined the utility of the solution.

6.4 Comparing the care episode profiles of

subpopulations
Even though the illustrations of the realized care processes of total populations
based on the factual data were useful as such, there was also a need for comparisons
of multivariate responses in different subpopulations as especially the comparisons
may tell something about the effects that are of interest. In order to illustrate such
comparisons with the hip fracture data, patients who were institutionalized at the
time of fracture were selected as a base group. This was an interesting group from
the health services point of view, because it could be interpreted as representing
the upper boundary of the expected use of resources for hip fracture patients.
In addition, the recent Finnish clinical guideline for hip fracture treatment had
requested further research on the outcomes of those patients who already were in
long-term care at the time of fracture
In order to examine how institutionalization changes the treatment to be
received, the patients managing at home (at the time of fracture) were considered
an interesting group to be compared with the institutionalized patients. More
specifically, the patient was defined as home-dwelling at the time of hip fracture if
there had not been care during the 30 days before the fracture, and no long-term
care in the year before the fracture. This population had a low initial need for care,
and it represented the lower boundary for the expected resource usage.
After the exclusion of patients aged less than 65 at the time of facture, there
were 4325 patients identified as institutionalized at the time of hip fracture in
1998–2001 in Finland. This was about 21.2% of all patients aged 65 or more with
the fracture of a proximal femur. The population of home-dwelling patients was
larger and consisted of 10 347 patients.
6.4.1 Risk adjustment

An observational study is considered biased, if the populations to be compared
have differences in observed or unobserved background factors in ways that
confound the outcomes of interest [180]. In the strict sense, the comparison
between institutionalized and home-dwelling patients was not an observational
study, because it was not reasonable to conceive of the groups as interventions and
the interest was not on the causal effects of any intervention but on the observed
differences between the groups. Anyhow, there were background factors, such as
age and sex, whose impact had to be controlled, and an actual method that could
be used to control for these confounding variables was needed [197].
Regression models have been widely used for the adjustment [151]. However,
in the case of current multivariate responses, that was not a realistic approach. A

more attractive and pragmatic choice for adjustment was the matching framework
[198]. In addition to the fact that matching works with any response (even if the
response was still unobserved at the time of matching), it can ensure that the
characteristics of the compared populations truly overlap, and that each patient
remains intact as a single patient and observed differences directly reflect disparities
between real patients and not between any artificial predictions based on estimated
model parameters [199].
As the number of covariates to be adjusted was small, a pair directly matched
according to sex, fracture type, and age (in years) was chosen for each patient in the
institutionalized population (n = 4 325) from the randomly ordered set of home-
dwelling patients (n = 10 347) without replacement. Perfect matches were found for
98.7% of the patients, while the rest had minimal differences in age. If the number
of factors to be matched would have been larger, other matching techniques (such
as the propensity score approach) could have been utilized [198].
6.4.2 Summarization of the follow-up data

In order to get a comprehensive explorative view of the data, the traceplots describing
the care episodes were considered to be potentially useful. As the subpopulation to
be compared had a different initial need for care, the main interest was on the
summarization of the need for care after the hip fracture. It was also known that the
Institutionalized patients Home-dwelling patients
Figure 11. Traceplots for the hip fracture care episodes of institutionalized patients in Finland in
Figure 11. Traceplots for the hip fracture care episodes of institutionalized patients in
1998-2001 and for home-dwelling patients matched for age, sex, and fracture type. Each row
Finland in 1998–2001 and for home-dwelling patients matched for age, sex, and fracture
corresponds
type. Each to
rowa care episode of to
corresponds oneapatient, which consists
care episode of one ofpatient,
120 daily realizations
which consists(four-month
of 120 daily
realizations
follow-up) (four-month
of the follow-up)
level of care, where theof the level
darker of care,a where
color reflects the darker
more intensive levelcolor
of carereflects a
more intensive
(white=home, level
light of care (white
gray=nursing home,=middle
home,gray=health
light gray =center,
nursing home,
dark middle gray = health
gray=hospital,
center, dark
black=death). gray = hospital, black = death).

levels of care described the intensity of care in an ordinal fashion, and that one step
in the levels of care corresponds to approximately a constant step in the functional
dependency measures [75, 176]. Therefore, it was assumed that an average level
of care over the follow-up period would work as a proxy variable for the need for
care of an individual. These averages were calculated and the traceplots were sorted
according to them. As Figure 11 reveals, even the straightforward mean sorting
resulted in rather simple and interpretable structures. For both subpopulations the
sorted order could be interpreted as an approximate continuum from good health
(top) to bad health (bottom). As expected, there were huge differences between the
institutionalized and home-dwelling patients.
Another way to summarize follow-up data in the case of hip fracture
treatment was developed in Sweden in the middle-1980s [200]. In these state
diagrams, the proportion of patients in each state is calculated for each day, and
the daily proportions are then plotted. A logarithmic scale on the time dimension
is typically used in order to emphasize the beginning of the follow-up time in
relation to its importance in the hip fracture case. It turned out that the state
diagrams corresponded to a kind of margin of the traceplot, i.e. could be derived by
summarizing the vertical dimension of the traceplot with proportions of patients
in different states.
The state diagrams in Figure 12 clearly demonstrate the most obvious
differences between the institutionalized and home-dwelling patients, such as the
higher mortality among the institutionalized patients and the greater proportions
of lower levels of care for the home-dwelling patients. In fact, the state diagrams
gave an easily interpretable yet comprehensive view of the multivariate response
by reporting simultaneously a huge number of measures that could be used as
traditional performance indicators, such as day-specific mortalities.
B. Home-dwelling patients (matched pairs of

A. Institutionalized patients institutionalized patients)
% %
100 100
80 80
60 60
Death Death
40 40
Hospital Hospital
Health center Health center
20 Nursing home 20 Nursing home
Home Home
0 0
Before 10 15 30 60 90 120 180 365 10 15 30 60 90 120 180 365
Before
fracture Days after fracture (logarithmic scale) fracture Days after fracture (logarithmic scale)
Figure 12. State diagrams for hip fracture patients aged 65 or more
Figure 12. State diagrams for hip fracture patients aged 65 or more

The information in the state diagram could be further compressed. An

obvious way was to compress the multivariate response to the state indicator that
reported the mean proportions of different levels of care during the follow-up after
the fracture for both subpopulations. Such state indicators are reported in Table 2.
These percentages would correspond to areas in the Figure 12, if a standard time
scale (instead of a logarithmic one) had been used in the figure.
Table 2. Percentages of days spent at different levels of care during the one-year follow-
up of hip fracture patients
Nursing Health
Home home centre Hospital Death
Institutionalized patients 6.3 32.3 26 3.7 31.7
Home-dwelling patients* 54.0 6.6 17.1 4.8 17.6
* Age, sex, and fracture type matched pairs of institutionalized patients.
6.5 Conclusions on monitoring hip fracture

treatment
In summary, this chapter aimed to provide a justification for the register-based
monitoring of hip fracture treatment. First a state-space model for hip fracture
treatment that was compatible with the register-based data was developed.
Then assumptions which were used to link the developed data-driven model to
the theory of effectiveness measurement were discussed. It was pointed out that
the combination of theory and data naturally leads to the use of multivariate
responses, and an illustration of summarization of multivariate response with
novel methodology were given. In addition, an example of adjusted comparisons
between multivariable responses was given. Finally it was demonstrated how
complicated responses could be reduced to simpler summarizations and further to
more traditional indicators.
In conclusion, this chapter demonstrated how it was possible to find an
adequate compromise between problem-oriented and data-driven theories. This
idea can be illustrated within the general framework for register-based health
services research that I have proposed [9]. Figure 13 visualizes the general structure
of the conceptual framework. The figure should be read from left to right. The
components that are closer to each other in the figure are more directly connected
in the framework. The two central paths from left to right in the figure correspond
to theory-oriented modeling (solid line) and to data-driven modeling (dashed
line). It must also be remembered that the selected health problem gives a third and
most concrete theory to be utilized in the actual application of the framework.

Public Health Equity

Quality
Health system Goals Effectiveness

Performance
Outcomes assessment
Structure
Processes model
Health
Register- Care episodes
information
based data
Event history
systems Preprocessing
analysis
Figure 13. Visualization of the conceptual framework for register-based health services
research. A solid line reflects the theory-driven approach, and the dashed line the data-
driven approach. The approaches are most compatible under the oval.

7 Operative delay as a performance indicator
7 Operative delay as a
performance indicator
In the previous chapter it was demonstrated what kind of assumptions were needed
to transform register-based data into factual information about hip fracture
treatment. In this chapter, it is illustrated what kind of issues may be related to the
justification of the use of a performance indicator. This chapter is mainly based on
article 6, but the presentation of a model for the provider profiling originates from
article 1.
A surgical operation is performed for most patients during the acute
management of hip fracture. A typical hip fracture patient is confined to hospital
bed-rest before the surgical operation. A fairly short delay between hospital
admission and surgical operation has been advised in the clinical guidelines [128,
129], but considerable variation in the operative delays between the providers has
been reported [201–203]. This discrepancy, which points out a potential area for
improvements, makes the operative delay an attractive performance indicator
[204].
However, the evidence concerning the timing of surgery in relation to the
overall outcomes of hip fracture treatment has been quite mixed. This is unfortunate,
because the operative delay should have confirmed effects on outcomes, otherwise
improvements may possibly be targeted at issues which do not improve health
[205]. Therefore, it was considered important to examine the association between
operative delay and mortality more carefully. The data were restricted to patients
aged 65 or more and operated on using an internal fixation, a prosthesis, or a total
hip replacement during 1998-2001 in Finland. The study population consisted of
16 881 patients, which was 83.3% of all first-time hip fracture patients aged 65 or
older in Finland 1998–2001.
7.1 Effects of different operative delays on

mortality
The first open question was the definition of the time period for delayed surgery
[204]. The register data included admission and operation days, so the operative
delay could be calculated only with limited accuracy, as the number of nights spent
in a hospital before the operation. Time from admission to death was measured
in days. The cumulative probabilities of death for different operative delays
were calculated using Kaplan-Meier estimators. As can be seen from Figure 14,

Proportion of death patients, %

40
5+ nights
3 4
30
0 1 2
20
10
0
10 30 60 90 120 180 240 300 365
Days
Figure 14. Cumulative probabilities of unadjusted mortality following hip fracture for
different
Figure 14. Cumulative operative delays
probabilities of unadjusted mortality following hip fracture for different
operative delays
waiting times from 0–2 nights had the same effect on mortality, but there was a
clear increase in mortality with longer waiting times. Therefore, early surgery was
defined as consisting of patients with a 0–2 nights waiting time and late surgery as
consisting of patients with at least 3 nights waiting time.
7.2 Profiling of providers

Next, it was interesting to compare whether the potential performance indicator
(proportion of late surgery patients) varied by hospital. This kind of activity has
been commonly referred to as the profiling of providers [206]. It has been pointed
out that the results of provider profiling have been commonly interpreted as a
ranking of providers, even though the extent of variation attributable to the quality
of provider may have been negligible [207], and an appropriate statistical modeling
has been considered as a necessary prerequisite for justified conclusions in provider
profiling [208]. In fact, it has been suggested that there exist variations between
providers for at least three reasons: 1) differences may be attributable to a random
variation related to the size of the provider, 2) the patient case-mix varies from
provider to provider, and 3) providers may differ in the performance of their care
[209].

There were several commonly applied risk-adjustment methods available for

the case of a binary response variable [210]. Following the standard approach,
a model for the probability of late surgery was constructed by using a logistic
regression model
¥ p ´
log ¦ µ = B 0 + B1x1 + B2x2 + ··· + Bkxk ,
§1–p¶
where xi are the predictor variables and the model parameters bi are log odds
ratios. Age, sex, fracture type, comorbidities, pre-admission residence, inpatient
care preceding fracture, admission weekday, and year were used as predictors. The
estimated probability of late surgery was calculated for each individual j as
pj = logit–1(Bˆ0 + Bˆ 1x1j + B̂2x2j + ··· + B̂kxkj),
where Bî are estimated parameters, and

exp(z)
logit–1 (z) = ..
1 + exp(z)
The fit of the model was acceptable yielding a c-statistics value 0.65, indicating a
reasonable ability for discrimination.
Since the focus in profiling was on providers and not on individuals, the
observed (Y) and expected outcomes (p) were then aggregated to the provider level
[207], i.e.
Oi = Yj 3
3p,
and
Ei = j
where Yj refers to the individual’s observed status of late surgery and the sums are
over patients treated by provider i. The ratio of observed to expected numbers of
late surgery O / E was then used to quantify the risk-adjusted differences between
providers. This quantity can be interpreted as the relative risk of a given provider
to have the same share of late surgery patients as would have been expected on
average for the similar patient population [210].
However, as the adjustment is practically always only partial, there remains
residual variability between the providers that may result in overestimated
differences in profiling analyses [211], and it has been recommended to use of
hierarchical (random effect) models in the profiling [208]. The advantages of
hierarchical modeling include the automatic removal of regression to the mean

problem by shrinkage, an improved precision for estimates among small providers,

an adequate adjustment for multiple comparisons, and a more realistic framework
for comparison judgments [212].
The common approach in the modeling of the O / E-ratio has been to assume
that the observed counts Oi being non-negative integers describing frequencies of
rare events have a Poisson distribution with mean γi , that is
Oi ~ Poisson (gi).
Two types of hierarchical models based on the Poisson assumption were constructed,
a gamma-Poisson model [213] and a Poisson-lognormal model [214].
The gamma-Poisson case was modeled as

log gi = log Ei + log(λi),
where the expected outcomes Ei adjust the patient-level variation and

λi ~ Gamma (a, b)
describe the provider-level variation of interest. A vague exponential prior was used
for a and a gamma prior for b. The posterior distributions of λi corresponded to
the shrunken O / E-ratios and were of primary interest. In the Poisson-lognormal
case the model took the form
log γi = log Ei + θi ,
where
θi ~ N(µ,σ2),
which could have been easily extended to incorporate provider-level covariates.

A normal prior was used for µ and a gamma prior for precision 1/σ2 . Here the
posterior distributions for exp(θi) were the interesting ones.
In principle, the conditions justifying the Poisson approximation for the
distribution of Oi were not very well satisfied, because there was a fixed number of
patients at each provider and the proportions of late surgery patients were quite
large. Therefore, a beta-binomial model [213] and a logistic model [207] were also
formulated. For both models it was assumed that
Oi ~ Binomial (pi ,ni),
where ni stands for the number of patients at provider i. The proportion pi was
modeled in the beta-binomial case as
pi ~ Beta (a, b),
and in the logistic case as

logit(pi ) = θi ,
where
θi ~ N (µ,σ2).
The gamma priors were used for α and β, and the priors for µ and 1/ σ2 were the
same as in the Poisson-lognormal model. The posterior distributions for (ni / Ei)pi
gave the shrunken O / E-ratios in both cases.
The actual estimation was carried out by using Markov chain Monte Carlo
(MCMC) simulation [215]. The mixing of the estimation procedures was examined
by using two chains in each estimation, and the convergence was evaluated on
the basis of Gelman-Rubin convergence plots [216]. After an initial 1000 burn-
in iterations, a further 10 000 iterations were used in the actual estimation of the
parameters. The complexity and relative fit of the models were assessed with the
deviance information criterion (DIC) [217]. The obtained shrunken O / E-ratios
were further multiplied by the average share of late surgery patients so that the
reporting could be done in terms of easily interpretable risk-adjusted proportions
for providers [207].
The actual results were rather similar for all models. In terms of DIC, the fit
was somewhat better with binomial assumption than with the Poisson assumption.
With the binomial assumption variance tend to be smaller if the provider was
small or the proportion of late surgery patients was large. The shrinkage was a bit
stronger for models with non-normal random effects distribution. As the purpose
was not to rank the providers or identify unusual performance, but to indicate the
potentiality for improvements, the most conservative gamma-Poisson model was
selected in order to avoid too strong interpretations.
The potentiality for improvements was assessed by assuming that the
percentage of late surgery patients can be reduced to some potentially achievable
level [218]. The level was determined empirically from the provider-specific
shrunken estimates by assuming that if 20% of providers have achieved that level
or better, the others should also be able to do it. In addition, the 20th centile is about
one standard deviation from the mean and can be interpreted to reflect the best
practice proportion [213]. As can be seen from the Figure 15, there was extensive
variation in the proportions of late surgery between the providers, and the 95%
credible intervals crossed the potentially achievable share of 92.7% of early surgery
patients only for less than half of the providers.

0 10 20 30 40 50 60
%
Figure 15. Provider-specific
Figure risk-adjusted
15. Provider-specific percentages
risk-adjusted of patients
percentages with latewith
of patients hip fracture
late hip surgery.
fracture
surgery. Thin parts of lines correspond to 95% credible intervals and thick parts correspond
Thin parts of lines correspond to 95% credible intervals and thick parts correspond to 50% credible to
50% credible intervals. The diamonds indicate the mean risk-adjusted shrunken proportions
intervals. The diamonds indicate the mean risk-adjusted shrunken proportions and the crosses the
and the crosses the originally observed proportions. The dotted line refers to the potentially
originallyachievable
observedpercentage
proportions. The dotted
(7.7%) line refers
of patients to late
with the the potentially
surgery. achievable percentage
(7.7%) of patients with the late surgery.
As the difference between unadjusted mortalities in late and early surgery groups
was considerable, it would have been straightforward to infer that the reduction of
the operative delay would also diminish the mortality. However, there were no such
huge variations in provider-specific mortalities, and the linkage between operative
delay and mortality had to be examined more carefully.

7.3 Adjusted effect of operative delay on

mortality
In fact, because of the potential biasing effects of different patient characteristics
between the early and late surgery group, there was first a need to adjust for these
factors. As the data included times from admission to death or end of follow-up,
a Cox proportional hazards model were used to control for the observed patient
characteristics.
A proportional hazard model specifies the hazard for individual j as

λj(t) = λ0(t)exp(b1x1j + b2x2j + ··· + bkxkj),
where λ0(t) is a nonnegative function of time t known as baseline hazard, xi are

the predictor variables and the model parameters bi are log hazard ratios [219].
In the Cox model, the baseline hazard function λ0(t) is considered as a nuisance
parameter, and parameters bi are estimated semi-parametrically without fixing the
form of λ0(t). The estimated hazard ratio between the late surgery and early surgery
was 1.24 (95% CI: 1.15 to 1.34) as no other covariates were included. The hazard
ratio for the late surgery reduced to 1.18 (95% CI: 1.09 to 1.28) after the inclusion
of age, sex, fracture type, comorbidities, pre-admission residence, inpatient care
preceding fracture, admission weekday, year, and hospital dummies as covariates.
Since there were no detailed clinical data available, it was likely that certain
relevant covariates were missing from the model. In addition, the patients were
known to be nested to hospitals. In the case of survival data, the concept of frailty
offers a suitable way to incorporate random effects in the proportional hazards
model to account for association and unobserved heterogeneity. A frailty may be
considered as an unobserved random factor Z that has a multiplicative effect on
the hazard function of an individual or a group of individuals. More specifically,
a constant shared frailty model is a conditional independence model having the
form
λj(t | Zi) = Ziµi(t)exp(b1x1j + b2x2j + ··· + bkxkj),
which corresponds to the hazard function of the individual j belonging to group i

given the group specific frailty Zi, where µi(t) is a baseline hazard rate that may or
may not depend on i, and b and x are as above [220]. The frailty Z was assumed to
have a form of gamma distribution and to vary between hospitals i, but a common
baseline hazard λ0(t) was assumed. This model resulted in a slightly smaller hazard
ratio 1.17 (95% CI: 1.08 to 1.27) than with the fixed effects model.
However, the fixed and random effect models failed to fulfill the assumption
of proportional hazards [221]. Therefore, a stratified Cox model, which allowed a

different baseline hazards for groups of interest but common values for coefficients
bi , was estimated with robust sandwich variance estimators which took the
clustering by hospitals into account [219]. Stratification was done in terms of
age, sex, cancer, dementia, and hospital. The resulting model seemed to fulfill the
proportional hazards assumption and gave a hazard ratio estimate of 1.15 (95% CI:
1.03 to 1.29), still indicating a higher mortality among late surgery patients.
The plotting of smoothed hazard functions revealed that there was clearly
increased risk of death during the four months following the hip fracture in early
and late surgery groups. The plotting of excess cumulative mortality (late surgery
mortality – early surgery mortality) showed that the difference between the groups
first rapidly increased to 3% and then continued to increase until a maximum of
about 5% at 1–3 years. Based on these considerations, a binary variable indicating
one-year mortality was selected to summarize information in the actual time-to-
death variable. It was further hypothesized that logistic regression analyses could
approximately replicate the actual results obtained with the Cox model by using a
simpler binary response variable. The concrete analyses confirmed this hypothesis
to be true.
With the binary response it was straightforward to apply a pseudo-
randomizing instrumental variable approach to the current problem. The basic
idea with the approach is to define an instrument variable that causes variation
in the treatment variable of interest (early or late surgery in this case), but has no
direct effect on the response variable (one-year mortality), because it then becomes
possible to estimate how much the variation in the treatment variable induced
by the instrument affects the response variable [222]. The pseudo-randomized
assignment of patients into the early and late surgery groups was done by using
the admission day of the week as an instrumental variable [223, 224]. This variable
had an effect on operative delay, but no direct effect on mortality in the case of hip
fracture [225]. In practice, the approach took a simple form: to test whether the
risk-adjusted one-year mortality was equal for each admission day of the week. It
turned out that, according to a χ2-test with null hypothesis of equal mortalities for
each admission day of the week, the difference in adjusted one-year mortality was
not attributable to the operative delay. This result was in contrast with the results
from the Cox models above.
7.4 Provider-level hypotheses

The clear difference of the estimates of the operative delay effect on mortality
between the methods indicated that there was a need for additional methods to
explain such differences. Fortunately, the results pointing out the huge variation
between providers in the shares of late surgery patients made it possible to
formulate novel hypotheses to address the problem. First, it was assumed that

an acceptable delay in hip fracture surgery corresponds to the clinical decision

to postpone the operation and that the clinical decisions remain nearly constant
within providers. Under this assumption, the potentially achievable lower limit in
the proportion of late surgery patients could also be interpreted as the upper limit
for the proportion of acceptable delayed patients. Correspondingly, the expected
proportion of unacceptable delayed patients was the proportion of late surgery
patients exceeding this upper limit. This inference resulted in a hypothesis stating
that the overall mortality of hip fracture patients should increase with the rising
share of the late surgery patients given that the longer operative delay would have
an adverse effect on mortality. Another novel provider-level hypothesis was that
the long-term mortality of the late surgery patients is higher if only the patients
who are unfit for surgery are delayed, since the unfit condition for the surgery is
also a risk factor for one-year mortality.
In order to test these hypotheses, the provider-level heterogeneity was further
examined in terms of simultaneous provider specific shares of the late surgery
patients and one-year mortality. The provider-specific overall mortalities were
decomposed into mortalities of early and late surgery patients by recognizing the
fact that the weighted average of these mortalities gives the overall mortality, i.e.
that
moverall (p) = [p · mlate (p)] + [(1 – p) · mearly (p)],
where p corresponds to the proportion of late surgery patients and m(p) to

mortality with proportion p of late surgery patients.
It was straightforward to calculate these three (adjusted) mortalities for
each provider, and use a scatter plot between the share of late surgery patients
and mortalities to describe the associations. In order to extract trends from these
associations, it was, after preliminary examinations, assumed that the trends
between the share of late surgery patients and overall as well as early surgery
mortalities are linear, i.e. that they have the forms
moverall (p) = a + bp,
and
mearly (p) = c + dp,
where p is the proportion of late surgery patients, and a, b, c and d are regression
coefficients to be estimated from the observed data. Late surgery mortality could
then be calculated deterministically from the weighted average formula given
above.
Due to the extensive random variation in mortality figures among small
providers, there was a need for hierarchical modeling giving shrunken estimates.
More specifically, it was assumed that better estimates (in terms of mean squared
error) could be obtained by a weighted sum of observed mortalities and appropriate
guesses, i.e. that

si = wioi + (1 – wi)ei,
where si is the shrunken mortality estimate for provider i, oi corresponds to observed

and ei to guessed mortality, and wi is the (probability) weight to be estimated from
the data [173]. The guesses were chosen to be the trends extracted from the data
by using the formulas given above, and the weights and shrunken estimates were
calculated by using Stein estimation in the form of an iterative algorithm [172].
This simultaneous examination of the provider-level proportions of the late
surgery patients and one-year mortality illustrated in Figure 16 revealed that there
was an almost flat (positive) trend between the greater share of late surgery patients
and one-year overall mortality. This confirmed that the actual independent effect
of operative delay on mortality is very small. Another result indicated that the
smaller proportion of late surgery patients was clearly (nonlinearly) associated
with a higher mortality for these patients. In other words, the providers with low
proportions of late surgery patients had been able to select the patients with such
severity that immediate operation would not have been clinically justified.
One year mortality, %
70
60
50
40
30
20
10
0
0 10 20 30 40 50
Share of patients with late surgery, %
Figure 16. A scatter plot of adjusted mortality for the operated hip fracture patients and the
Figure 16. of
share A hip
scatter plotpatients
fracture of adjusted mortality
with late surgeryfor
for the operated
different hip fracture
providers. patients
(The triangles and the share of
describe
the overall mortality of the operated hip fracture patients for each provider, and the crosses
hip fracture patients with late surgery for different providers. (The triangles describe the overall
the mortality of the late surgery patients for each provider. The line with a positive slope is
mortality of the
the trend of operated hip fracture
overall mortality, patients
and the dottedfor each
curve provider,
describes theand the crosses
association the mortality
between the of the
share ofpatients
late surgery late surgery patients
for each and mortality
provider. across
The line withthea positive
providers.)slope is the trend of overall
mortality, and the dotted curve describes the association between the share of late surgery patients
and mortality across the providers.)

The provider-level associations were further analyzed for five severity groups
of patients by using the same methodology. The severity was measured as predicted
one-year mortality for observed patient characteristics, and groups were formed
by dividing the patients into five equivalently sized classes on the basis of sorted
severity score. Intuitively speaking, the least severe patients were young, coming
from home and were without severe medical conditions. Correspondingly the most
severe patients were older, with much comorbidity and coming from residential
care. For the patients in the least severe group, mortality was higher for the late
surgery patients for all providers. Probably this indicated that in this group, the hip-
fracture-related mortality was caused by the medical problems which necessitated
late surgery and the prolonged surgical delay did not itself increase the mortality.
For the patients in the most severe group, mortality was also higher for the late
surgery patients for all providers. However, for this group the late surgery mortality
was lower in the providers with a small share of late surgery patients and there were
no differences in early surgery mortalities between the providers. It seems that in
this group, it was essential to perform early surgery for all patients who could bear
it, since the significantly prolonged surgical delay makes the patient’s condition
worse and increases mortality. For groups 2–4 the interpretation was more difficult
than for extreme patients. Since the mortality of early surgery patients increased
significantly and was even higher than the mortality of the late surgery patients
for the providers with a large share of late surgery patients, it seems that in these
groups an operation too early may have caused more harm than the prolonged
waiting time. In summary, these results actually gave reasonable explanations for
the mixed results in the existing literature.
7.5 Conclusions on operative delay as a

performance indicator
The analyses presented above demonstrated that methods that have been
traditionally used may result in biased results if used uncritically in the case of
register-based data, with such a bias being difficult to detect. The problem was
that the register data did not include detailed clinical data about the individual-
specific factors that may have postponed the operation, and that the actual
selection mechanism related to becoming a late surgery patient varied between the
providers. In this case the formulation of novel provider-level hypotheses helped
to reveal these associations indirectly from the register-based data. In conclusion,
the average effect of operative delay on mortality was small, but the provider-level
association between the share of late surgery patients and non-optimal treatment
was obvious, and the proportion of patients with a prolonged waiting time for a
hip fracture operation seems to be an effective quality indicator, since it clearly
indicates an area for improvement.

8 Discussion
8 Discussion
As a statistician and a scientist I was quite enthusiastic when I first participated

in a study that used Finnish administrative registers as a data source. The massive
amount of data had been shown to be of good quality and with an internationally
exceptional coverage as a result of unique individual identification numbers
allowing easy linkages, offering excellent opportunities for valuable health services
research. There were many good examples of such research, and it seemed that the
administrative data could be used effectively as long as the potential pitfalls were
avoided. However, practical data analyses quickly revealed that the meaningful use
of data was very challenging, and that the most fundamental parts of the analysis
corresponded more closely to a qualitative than to a quantitative methodological
approach. This troubled me, because—even though it was possible to successfully
use administrative data in health services research—the links between scientific
theories and secondary data were mostly extremely fuzzy. It was also clear that the
data sources and therefore the methodological problems abroad were not identical
to Finnish ones, and the optimal solution for the problems in a Finnish context
obviously requires that the methodology for a solution is developed from the
Finnish context point of view. In this sense, the main question in the utilization of
register-based data turned out to be: How the use of secondary register data could
be justified by using strong methodological arguments tailored to the Finnish
context?
In this thesis I have tried to give an answer to that question. The first aim
of this study was to develop a methodological framework that helps in the
effective utilization of register-based data. The problem was approached from a
multidisciplinary methodological perspective. Chapter 2 gave a brief review to the
methodological approaches, issues, and assumptions that had a fundamental role
in the more applied work of later chapters. In this sense, Chapter 2 could also be
interpreted as representing summarizations of those philosophic-methodological
ideas that are likely to be fruitful in the characterization and solving of the concrete
problems in register-based data analysis in general, and thereby sketching the core
of the framework.
Two main ideas of the framework should be noted. First is the aim to find
suitable compromises between the problem-oriented and data-driven theories,
and the second is the empirical justification of theory-driven operationalized
definitions, i.e. the confirmation that the theoretical constructs are in concordance
with the register-based data. Both of these ideas were used throughout this study,
combined with the idea of minimizing the overall bias via the maximization of the
use of the most factual data, and reflected the need for data sensitive methodological

8 Discussion
thinking. In this sense, the framework is not just about the methods, but about the
use of the right methods in the right situations.
There are two potential problems with the framework: 1) the framework is not
a sufficient condition for the register-based data analysis, because a strong and more
or less tacit background knowledge on multidisciplinary methodological issues
(including application experience) is a prerequisite for a full scale understanding
and evaluation of the approach, and 2) the framework is not a necessary condition
for a register-based data analysis, because it may be possible to end up with similar
results without any knowledge of the framework. However, as the presented
framework included a multidisciplinary collection of extensions of well-grounded
methodological approaches, it is likely that a skilled methodologist would find at
least some ideas helpful in the efficient utilization of register-based data analysis.
The second aim was to demonstrate how register-based data can be used as a
data source for the performance monitoring of a health system, i.e. to connect the
methodological developments to the practical needs of information and further to
concrete health policy goals, such as the “steering by information” policy regime.
The actual demonstration included four components. The first was the validity
of register data, the second concentrated on the epidemiological perspective of
the monitoring system, the third on the monitoring of treatment in terms of care
episodes, and the fourth gave an example on the justification of a more detailed
performance indicator. At the end of each chapter the main conclusions were
summarized. From the methodological point of view, each component aimed to
find an acceptable compromise between the theory and the data. Especially the
first three components also included the tailoring of conceptual definitions and
assumptions. Data sensitive empirical analyses had a central role in each component,
and throughout these chapters the empirical results were novel contributions to
the medical literature.
The development of a comprehensive hip fracture monitoring system was
such a huge task, that it was not possible to include everything into this thesis. In
fact, several manuscripts developing issues further are already under preparation.
These include a review article about prerequisites in register-based research, several
papers that extend the care episode ideas presented in Chapter 6 or develop more
detailed risk-adjustment methods, and working papers that compare the costs
of treatment [226] and model the lengths of stay in treatment processes [227].
In this sense, the practical work is still in progress, but the perspectives given in
this thesis introduce a necessary basis for further work and help in the routine
implementation of a hip fracture monitoring system [228].
It should be noted that this study concentrated on methodological issues in
the wide sense. The main idea in the analyses was to solve a problem, not to use
some particular model or technique. Statistical thinking was the key that made
it possible to deal with multidisciplinary methodological ideas at a sufficiently
abstract level. After an extraction of the “wireframe model”, the actual search for

8 Discussion
a suitable method or analytical approach was in some cases straightforward, but

occasionally there was also a need to extend the available methods or even create
new ones. In this sense, the methodological problems considered important were
not on the (mathematical) details of any single model, but on finding appropriate
methodological solutions to various interactively encountered problems. This kind
of approach naturally required a lot of preliminary and sensitivity analyses, but it
was feasible to report only a small selection of those, and mostly the ones that were
required to justify the use of the selected models. The drawback in this kind of
underreporting is that the actual process of transforming register-based data into
useful information may seem to be easier than it really was [3].
All in all, the methodological approach demonstrated in this thesis represents
an alternative style of reasoning for scientific research utilizing register-based data.
This style of reasoning could be characterized as methodological constructionism
or—in the sense of desire for factual data for data analyses—as hermeneutic
positivism. But that is another story.

References
References
1. Sund R: Utilisation of administrative 13. Tierney WM: Improving clinical decisions

registers using scientific knowledge and outcomes with information: a review.
discovery. Intelligent Data Analysis 2003, Int J Med Inform 2001, 62(1):1–9.
7(6):501–519. 14. Härö AS (ed.): Planning Information
2. Sund R, Nurmi-Lüthje I, Lüthje P, Services for Health. Decision-Simulation-
Tanninen S, Narinen A, Keskimäki I: Approach. Health Services Research by
Comparing properties of audit data the National Board of Health in Finland
and routinely collected register data in 22. Helsinki: National Board of Health in
case of performance assessment of hip Finland; 1981.
fracture treatment in Finland. Methods 15. Nenonen M, Nylander O, Rintanen
Inf Med 2007, 46(5):558–566. H, Suominen T: What might a health
3. Sund R: Utilization of routinely collected information system look like? Inform
administrative data in monitoring of Prim Care 2002, 10:135–138.
aging dependent hip fracture incidence. 16. White KL: The ecology of medical care:
Epidemiol Perspect Innov 2007, 4(2). origins and implications for population-
4. Sund R, Kauppinen S: Kuinka laskea based healthcare research. Health Serv
ikääntyneiden pitkäaikaisasiakkaiden Res 1997, 32(1):11–21.
määriä rekisteritietojen perusteella? Sos 17. Roos NP: Establishing a population
Laaketiet Aikak 2005, 42(2):137–144. data-based policy unit. Med Care 1999,
5. Sund R: Lonkkamurtumien ilmaantu- 37(6 Suppl):JS15–26.
vuus Suomessa 1998-2002. Duodecim 18. Gissler M: Routinely collected registers
2006, 122(9):1085–1091. in Finnish health research. In: Statistics,
6. Sund R, Liski A: Quality effects of Registries, and Science Experiences
operative delay on mortality in hip from Finland. Edited by Alho J. Keuruu:
fracture treatment. Qual Saf Health Care Statistics Finland; 1999.
2005, 14(5):371–377. 19. Shani M: The impact of information on
7. Sund R: Huono-osaisuus tiedollisena medical thinking and health care policy.
haasteena. In: Toinen Tieto. Kirjoituksia Int J Med Inform 2000, 58-59:3–10.
Huono-osaisuuden Tunnistamisesta. 20. Bevan JM: Some general remarks on
Edited by Hänninen S, Karjalainen J, the nature and purpose of medical
Lahti T. Helsinki: Sosiaali- ja terveysalan information systems. Statistician 1972,
tutkimus- ja kehittämiskeskus; 2005. 21(4):237–243.
8. Sund R, Nylander O, Palonen T: Raa’asta 21. Berg M, Goorman E: The contextual
rekisteriaineistosta terveyspoliittisesti nature of medical information. Int J
relevanttiin informaatioon. Yhteiskunta- Med Inform 1999, 56(1-3):51–60.
politiikka 2004, 69(4):372–379. 22. Friedman HP, Goldberg JD: Knowledge
9. Sund R: Terveydenhuoltojärjestelmän discovery from databases and data
vaikuttavuuden arvioinnista. Yhteiskun- mining: New paradigms for statistics
tapolitiikka 2005, 70(5):507–513. and data analysis. Biopharmaceutical
10. Aday LA, Begley CE, Lairson DR, Report 2000, 8(2):1–12.
Slater CH: Evaluating the Healthcare 23. Han J, Kamber M: Data Mining: Concept
System: Effectiveness, Efficiency, and and Techniques. San Fransisco: Morgan
Equity. Second Edition. Chicago: Health Kaufman; 2001.
Administration Press; 1998. 24. Tukey JW: The future of data analysis.
11. Gray JAM: Evidence-Based Healthcare. Ann Math Statist 1962, 33(1):1–67.
How to Make Health Policy and 25. Chambers JM: Greater or lesser statistics:
Management Decisions? New York: A choice for future research. Stat Comput
Churchill Livingstone; 1997. 1993, 3:182–184.
12. Brailer DJ: Management of knowledge in 26. Huber PJ: Massive datasets workshop:
the modern health care delivery system. four years after. J Comput Graph Stat
Jt Comm J Qual Improv 1999, 25(2):55– 1999, 8:635–652.
67.

References
27. St Leger AS, Schnieden H, Walsworth- 44. Langefors B: Essays on Infology. Summing
Bell JP: Evaluating Health Services’ Up and Planning for the Future.
Effectiveness. Milton Keynes: Open Gothenburg: Department of Information
University Press; 1992. Systems, University of Göteborg; 1993.
28. Wallgren A, Wallgren B: Register-Based 45. Hand DJ: Data, metadata and information.
Statistics: Administrative Data for Stat J UN Econ Comm Eur 1993, 10:143–
Statistical Purposes. Chichester: John 151.
Wiley & Sons; 2007. 46. Sundgren B: Making statistical data more
29. Elmasri R, Navathe SB: Fundamentals available. Int Stat Rev 1996, 64(1):23–
of Database Systems. Third Edition. 38.
Reading: Addison-Wesley; 2000. 47. Khatri V, Vessey I, Ram S, Ramesh V:
30. Marshall JT: Canada’s national vital Cognitive fit between conceptual schemas
statistics index. Popul Stud 1947, 1(2): and internal problem representations: The
204–211. case of geospatio-temporal conceptual
31. Nordbotten S: A statistical file system. schema comprehension. IEEE T Prof
Stat Tidskr 1966, 4(2):99–103. Commun 2006, 49(2):109–127.
32. Wegman EJ: Huge data sets and the 48. Peuquet DJ: Representations of Space
frontiers of computational feasibility. J and Time. New York: The Guilford Press;
Comput Graph Stat 1995, 4(4):281–295. 2002.
33. Hastie T, Tibshirani R, Friedman J: The 49. Codd EF: Data models in database
Elements of Statistical Learning; Data management. ACM SIGMOD Record
Mining, Inference, and Prediction. New 1981, 11(2):112–114.
York: Springer-Verlag; 2001. 50. Armstrong DJ: The quarks of object-
34. Bishop CM: Pattern Recognition and oriented development. Commun ACM
Machine Learning. New York: Springer; 2006, 49(2):123–128.
2006. 51. Mylopaolos J: Information modeling in
35. Hand DJ, Mannila H, Smyth P: Principles the time of the revolution. Inform Syst
of Data Mining. Cambridge: The MIT 1998, 23(3/4):127–155.
Press; 2001. 52. Chandrasekaran B, Josephson JR,
36. Fayyad U, Piatetsky-Shapiro G, Smyth Benjamins VR: What are ontologies, and
P: The KDD process for extracting why do we need them? IEEE Intelligent
useful knowledge from volumes of data. Systems 1999, 14(1):20–26.
Commun ACM 1996, 39(11):27–34. 53. Froeschl KA, Grossmann W, Del Vecchio V
37. Shearer C: The CRISP-DM model: the (eds.): The Concept of Statistical Metadata.
new blueprint for data mining. Journal Edinburgh: A Network of Excellence
of Data Warehousing 2000, 5(4):13–22. for Harmonising and Synthesising the
38. Silberschatz A, Tuzhilin A: What makes Development of Statistical Metadata;
patterns interesting in knowledge 2003.
discovery systems. IEEE T Knowl Data 54. Mennis JL, Peuquet DJ, Qian L: A
En 1996, 8(6):970–974. conceptual framework for incorporating
39. Cartwright DW, Armknecht PA: cognitive principles into geographical
Statistical uses of administrative records. database representation. Int J Geogr Inf
Rev Public Data Use 1980, 8(1):13–27. Sci 2000, 14(6):501–520.
40. Irvine J, Miles I, Evans J (eds.): 55. Gissler M, Haukka J: Finnish health and
Demystifying Social Statistics. London: social welfare registers in epidemiological
Pluto Press; 1979. research. Norsk Epidemiologi 2004,
41. David M: Access to data: the frustration 14(1):113–120.
and utopia of the researches. Rev Public 56. Lenzerini M: Data integration: A
Data Use 1980, 8:327–337. theoretical perspective. In: Proceedings
42. Hand DJ: Measurement Theory of 21th ACM SIGMOD-SIGACT-
and Practice: The World Through SIGART Symposium on Principles of
Quantification. London: Hodder Arnold; Database Systems. New York: ACM
2004. Press; 2002: 233–246.
43. Faber J, Scheper WJ: Social scientific 57. Hernández MA, Stolfo SJ: Real-world
explanations? On Quines’s legacy and data is dirty: Data cleansing and the
contextual fallacies. Qual Quant 2003, merge/purge problem. Data Min Knowl
37:135–150. Disc 1998, 2(1):1–31.

References
58. Gu L, Baxter R, Vickers D, Rainsford European Observatory on Health Care

C: Record Linkage: Current Practice Systems; 2002.
and Future Directions. CMIS Technical 75. Kauppinen S, Forss A, Säkkinen S, Noro
Report 03/83. Canberra: Commonwealth A, Voutilainen P: Care and Services for
Scientific and Industrial Research Older People 2002. Official Statistics of
Organication (CSIRO); 2003. Finland, Social Security 2003:1. Helsinki:
59. Fellegi IP, Sunter AB: A Theory for National Research and Development
record linkage. J Am Stat Assoc 1969, Centre for Welfare and Health; 2003.
64(328):1183–1210. 76. Häkkinen U, Lehto J: Reform, change,
60. Rahm E, Bernstein PA: A survey of and continuity in Finnish health care.
approaches to automatic schema J Health Polit Policy Law 2005, 30(1–
matching. VLDB J 2001, 10:334–350. 2):79–96.
61. Rahm E, Do HH: Data cleaning: 77. OECD: OECD Reviews of Health
problems and current approaches. IEEE Systems. Finland. Paris: Organisation
Data Eng Bull 2000, 23(4):3–13. for economic co-operation and
62. Fellegi IP, Holt D: A systematic approach development; 2005.
to automatic edit and imputation. J Am 78. Laaksovirta TH: Tiedonvälittymisen
Stat Assoc 1976, 71(353):17–35. trendit ja informaatio-ohjaus. Yhteis-
63. Bruni R: Error correction for massive kuntapolitiikka 2000, 65(4):355–358.
datasets. Optim Method Softw 2005, 79. Smith PC, Häkkinen U: Information
20(2–3):291–310. strategies for decentralization. In:
64. Barbará D, DuMouchel W, Faloutsos Decentralization in Health Care:
C, Haas PJ, Hellerstein JM, Ioannidis Y, Strategies and Outcomes. Edited by
Jagadish HV, Johnson T, Ng R, Poosala Saltman R, Bankauskaite V, Vrangbaek
V et al: The New Jersey data reduction K. Milton Keynes: Open University
report. IEEE Data Eng Bull 1997, Press; 2006.
20(4):3–42. 80. Rauhala P-L, Simpura J, Uusitalo H:
65. Shahar Y: A framework for knowledge- Hyvinvoinnin tutkimusperinteet, hyvin-
based temporal abstraction. Artif Intell vointipolitiikka ja 1990-luvus uudet
1997, 90:79–133. avaukset. Yhteiskuntapolitiikka 2000,
66. Pyle D: Data Preparation for Data Mining. 65(3):191–207.
San Fransisco: Morgan Kaufmann; 1999. 81. Nylander O, Ståhle P, Nenonen M:
67. Hand DJ: Data mining: New challenges Informaatio-ohjauksesta tietointen-
for statisticians. Soc Sci Comput Rev siiviseen vuorovaikutukseen ter-
2000, 18(4):442–449. veydenhuollon kehittämisessä. Yhteis-
68. Box GEP: Science and statistics. J Am kuntapolitiikka 2003, 68(1):3–18.
Stat Assoc 1976, 71(356):791–799. 82. WHO: The World Health Raport 2000,
69. Maclure M, Schneeweiss S: Causation of Health Systems: Improving Performance.
bias: the episcope. Epidemiology 2001, Geneva: World Health Organization;
12(1):114–122. 2000.
70. Greenland S: Multiple-bias modelling 83. Naylor CD, Iron K, Handa K: Measuring
for analysis of observational data (with health system performance: problems
discussion). J R Stat Soc Ser A Stat Soc and opportunies in the era of assessment
2005, 168:267–306. and accountability. In: Measuring Up
71. Breiman L: Statistical modeling: the two Improving Health System Performance
cultures (with discussion). Stat Sci 2001, in OECD Countries. Edited by Smith
16(3):199–231. PC. Paris: Organization for Economic
72. Jensen D: Data snooping, dredging and Co-operation and Development; 2002:
fishing: the dark side of data mining. 13–34.
SIGKDD Explorations 2000, 1(2):52– 84. Hurst J: Performance measurement and
54. improvement in OECD health systems:
73. Thomas RK: Society and Health. overview of issues and challenges. In:
Sociology for Health Professionals. Measuring Up Improving Health System
New York: Kluwer Academic / Plenum Performance in OECD Countries.
Publishers; 2003. Edited by Smith PC. Paris: Organization
74. Järvelin J: Health Care Systems in for Economic Co-operation and
Transition: Finland. Copenhagen: Development; 2002:35–54.

References
85. Eddy DM: Performance measurement: and medical decision-making: current

problems and solutions (with approaches. J R Stat Soc Ser A Stat Soc
discussion). Health Aff 1998, 17(4):7– 2006, 169(1):5–35.
41. 97. Hauck K, Street A: Performance
86. Loeb JM: The current state of assessment in the context of multiple
performance measurement in health objectives: a multivariate multilevel
care. Int J Qual Health Care 2004, 16 analysis. J Health Econ 2006, 25(6):1029–
Suppl 1:i5–9. 1048.
87. Arah OA, Westert GP, Hurst J, Klazinga 98. Lipsey MW: Theory as method:
NS: A conceptual framework for the small theories of treatments (with
OECD Health Care Quality Indicators discussion). In: Research Methodology:
Project. Int J Qual Health Care 2006, 18 Strengthening Causal Interpretations
Suppl 1:5–13. of Nonexperimental Data. Edited
88. Rajavaara M: Vaikuttavuusyhteiskunta. by Sechrest L, Perrin E, Bunker J.
Sosiaalisten Olojen Arvostelusta Vai- Washington DC: U.S. Department of
kutusten Todentamiseen. Sosiaali- ja Health and Human Resources; 1990: 33–
Terveysturvan Tutkimuksia 84. Helsinki: 60.
Kansaneläkelaitos; 2007. 99. Walshe K: Understanding what works-
89. Sulkunen P: The welfare state and -and why--in quality improvement: the
the state of knowledge: The Finnish need for theory-driven evaluation. Int J
case. In: Research on the Study of the Qual Health Care 2007, 19(2):57–59.
Nordic Welfare State Renvall Institute 100. Potvin L, Champagne F: Utilization of
Publications 16. Edited by Marjanen administrative files in health research.
J, Stenius H, Vauhkonen J. Helsinki: Soc Indic Res 1986, 18:409–423.
Renvall Institute; 2004:58–74. 101. Connell FA, Diehr P, Hart LG: The use
90. Klazinga N, Stronks K, Delnoij D, of large data bases in health care studies.
Verhoeff A: Indicators without a cause. Annu Rev Public Health 1987, 8:51–74.
Reflections on the development and use 102. Romano PS, Luft HS: Getting the
of indicators in health care from a public most out of messy data: Problems
health perspective. Int J Qual Health and approaches for dealing with large
Care 2001, 13(6):433–438. administrative data sets. In: Medical
91. Freeman T: Using performance Effectiveness Research Data Methods.
indicators to improve health care quality Edited by Grady ML, Schwartz HA.
in the public sector: a review of the Rockville: Agency for Health Care Policy
literature. Health Serv Manage Res 2002, and Research; 1992: 57–75.
15(2):126–137. 103. Powell AE, Davies HT, Thomson RG:
92. Mattke S, Epstein AM, Leatherman S: The Using routine comparative data to assess
OECD Health Care Quality Indicators the quality of health care: understanding
Project: history and background. Int J and avoiding common pitfalls. Qual Saf
Qual Health Care 2006, 18 Suppl 1:1–4. Health Care 2003, 12(2):122–128.
93. Mainz J: Developing evidence-based 104. Motheral B, Brooks J, Clark MA, Crown
clinical indicators: a state of the art WH, Davey P, Hutchins D, Martin BC,
methods primer. Int J Qual Health Care Stang P: A checklist for retrospective
2003, 15 Suppl 1:i5–11. database studies—report of the ISPOR
94. Bird SM, Cox D, Farewell VT, Goldstein Task Force on Retrospective Databases.
H, Holt T, Smith PC: Performance Value Health 2003, 6(2):90–97.
indicators: good, bad, and ugly. J R Stat 105. Schwartz HA: Medical effectiveness
Soc Ser A Stat Soc 2005, 168(1):1–27. research methods and data. In: Medical
95. Berg M, Meijerink Y, Gras M, Goossensen Effectiveness Research Data Methods.
A, Schellekens W, Haeck J, Kallewaard M, Edited by Grady ML, Schwartz HA.
Kingma H: Feasibility first: developing Rockville: Agency for Health Care Policy
public performance indicators on and Research (AHCPR); 1992: 1–3.
patient safety and clinical effectiveness 106. Wray NP, Ashton CM, Kuykendall DH,
for Dutch hospitals. Health Policy 2005, Hollingsworth JC: Using administrative
75(1):59–73. databases to evaluate the quality of
96. Ades AE, Sutton AJ: Multiparameter medical care: a conceptual framework.
evidence synthesis in epidemiology Soc Sci Med 1995, 40(12):1707–1715.

References
107. Donabedian A: Evaluating the quality 120. Currie CT, Hoy D, Tierney AJ, Bryan-
of medical care. Milbank Mem Fund Q Jones J, Lapsley I: HipMod: Development
1966, 44(3):Suppl:166–206. of a multi-agent audit-based computer
108. Handler A, Issel M, Turnock B: A simulation of hip fracture care. Health
conceptual framework to measure Inf J 2003, 9(3):183–197.
performance of the public health system. 121. Noro A, Rissanen P, Laitinen O, Hammar
Am J Public Health 2001, 91(8):1235– T, Pöyry P: Lonkkamurtumapotilaiden
1239. hoito- ja kotiuttamiskäytännöt ja
109. STAKES: From Values to Choices. kotihoidossa selviäminen. Gerontologia
Report of the Working Group on Health 2000, 14(2):83–88.
Care Prioritisation. Helsinki: National 122. Huusko T, Karppi P, Avikainen V,
Research and Development Centre for Kautiainen H, Sulkava R: Significant
Welfare and Health; 1995. changes in the surgical methods and
110. Stenvall J, Syväjärvi A: Onks Tietoo? length of hospital stay of hip fracture
Valtion Informaatio-Ohjaus Kuntien patients occurring over 10 years in
Hyvinvointitehtävissä. Helsinki: Valtio- Central Finland. Ann Chir Gynaecol
varanministeriö; 2006. 1999, 88(1):55–60.
111. Hämäläinen H, Gissler M, Muuri 123. Heikkinen T, Willig R, Hänninen A,
A: Sosiaali- ja Terveydenhuollon Koskinen K, Mannismäki P, Alavaikko
Tietouudistus 2005. Palvelujärjestelmä A, von und zu Fraunberg M, Jalovaara P:
ja Sen Toimivuus -ryhmän raportti. Hip fractures in Finland—a comparison
Helsinki: Sosiaali- ja terveysministeriö; of patient characteristics and outcomes
2004. in six hospitals. Scand J Surg 2004,
112. Nenonen M, Nylander O: A theoretical 93(3):234–240.
framework for health information 124. Nurmi I, Narinen A, Lüthje P, Tanninen
systems. Int J Healthcare Technology S: Functional outcome and survival after
and Management 2002, 4(6):467–477. hip fracture in elderly: a propective study
113. OECD: A Disease-Based Comparison of of 106 consecutive patients. J Orthopaed
Health Systems. What is Best and at What Traumatol 2004, 5(1):7–14.
Cost? Paris: Organisation for Economic 125. Nurmi I, Narinen A, Lüthje P, Tanninen
Co-operation and Development; 2003. S: Cost analysis of hip fracture treatment
114. Liporace FA, Egol KA, Tejwani N, among the elderly for the public health
Zuckerman JD, Koval KJ: What’s new in services: a 1-year prospective study in
hip fractures? Current concepts. Am J 106 consecutive patients. Arch Orthop
Orthop 2005, 34(2):66–74. Trauma Surg 2003, 123(10):551–554.
115. Zuckerman JD: Hip fracture. N Engl J 126. Treml J, Kroker PB: Hip fracture audit:
Med 1996, 334(23):1519–1525. time for standards to be evidence-based.
116. Robinson CM, Court-Brown CM, Int Orthop 2000, 24(4):181–183.
McQueen MM, Christie J: Hip fractures 127. Parker MJ, Currie CT, Mountain JA,
in adults younger than 50 years of age. Thorngren KG: Standardised audit of
Epidemiology and results. Clin Orthop hip fracture in Europe (SAHFE). Hip Int
Relat Res 1995(312):238–246. 1998, 8(1):10–15.
117. Kannus P, Sievänen H, Palvanen M, 128. Duodecim: Käypä hoito suositus. Lonk-
Järvinen T, Parkkari J: Prevention of kamurtumapotilaiden hoito. Duodecim
falls and consequent injuries in elderly 2006, 122(3):358–379.
people. Lancet 2005, 366(9500):1885– 129. SIGN: Prevention and Management
1893. of Hip Fracture in Older People.
118. Marks R, Allegrante JP, Ronald Edinburgh: Scottish Intercollegiate
MacKenzie C, Lane JM: Hip fractures Guidelines Network (SIGN); 2002.
among the elderly: causes, consequences 130. Lääkintöhallitus: Yleissairaaloiden
and control. Ageing Res Rev 2003, käyttö Suomessa vuonna 1960. Lää-
2(1):57–93. kintöhallituksen Tutkimuksia 1. Helsinki:
119. Heithoff KA, Lohr KN (eds.): Hip Lääkintohallitus; 1966.
Fracture: Setting Priorities for 131. Slätis P: Tapaturmapotilaiden Sairaan-
Effectiveness Research. Washington, DC: sijatarve. Yleissairaaloiden Käytön
National Academy Press; 1990. Analyysi Tapaturmapotilaiden Osalta
Vuonna 1968 ja Ennuste Vuodeksi 1980.

References
Lääkintöhallituksen Tutkimuksia 10. 143. Keskimäki I, Aro S: Accuracy of data on

Helsinki: Lääkintöhallitus; 1972. diagnoses, procedures and accidents in
132. Alhava EM, Puittinen J: Fractures of the the Finnish Hospital Discharge Register.
upper end of the femur as an index of Int J Health Sci 1991, 2(1):15–21.
senile osteoporosis in Finland. Ann Clin 144. Lüthje P, Nurmi I, Kataja M, Heliövaara
Res 1973, 5(6):398–403. M, Santavirta S: Incidence of pelvic
133. Mattila K: Reisiluun yläosan murtuman fractures in Finland in 1988. Acta Orthop
ilmaantuvuus Suomessa. Sos Laaketiet Scand 1995, 66(3):245–248.
Aikak 1983, 20:53–59. 145. Pajunen P, Koukkunen H, Ketonen M,
134. Lüthje P: Incidence of hip fracture in Jerkkola T, Immonen-Räihä P, Kärjä-
Finland. A forecast for 1990. Acta Orthop Koskenkari P, Mähönen M, Niemelä
Scand 1985, 56:223–225. M, Kuulasmaa K, Palomäki P et al:
135. Simonen O: Incidence of femoral neck The validity of the Finnish Hospital
fractures: Senile osteoporosis in Finland Discharge Register and Causes of Death
in the years 1970-1985. Calcif Tissue Int Register data on coronary heart disease.
1991, 49:S8–S10. Eur J Cardiovasc Prev Rehabil 2005,
136. Kurttio P, Gustavsson N, Vartiainen T, 12(2):132–137.
Pekkanen J: Exposure to natural fluoride 146. Tolonen H, Salomaa V, Torppa J, Sivenius
in well water and hip fracture: a cohort J, Immonen-Raiha P, Lehtonen A: The
analysis in Finland. Am J Epidemiol validation of the Finnish Hospital
1999, 150(8):817–824. Discharge Register and Causes of Death
137. Kannus P, Niemi S, Parkkari J, Palvanen Register data on stroke diagnoses.
M, Vuori I, Järvinen M: Hip fractures Eur J Cardiovasc Prev Rehabil 2007,
in Finland between 1970 and 1997 and 14(3):380–385.
predictions for the future. Lancet 1999, 147. Huusko TM, Karppi P, Avikainen V,
353(9155):802–805. Kautiainen H, Sulkava R: The changing
138. Hokkanen E, Vauramo E, Mikkola picture of hip fractures: dramatic change
P, Sippo-Tujunen I: Poistoprojektin in age distribution and no change in age-
Loppuraportti. Aiheita 45/1993. Helsinki: adjusted incidence within 10 years in
Sosiaali- ja terveysalan tutkimus- ja Central Finland. Bone 1999, 24(3):257–
kehittämiskeskus; 1993. 259.
139. Keskimäki I, Teperi J, Aro S, Pohjanlahti 148. Sorensen HT, Sabroe S, Olsen J: A
J-P: Leikkaustoimenpiteiden Yleisyyden framework for evaluation of secondary
Alueellinen Vaihtelu Suomessa. data sources for epidemiological research.
Stakes, Raportteja 83. Helsinki: So- Int J Epidemiol 1996, 25(2):435–442.
siaali- ja terveysalan tutkimus- ja 149. Chin MH, Muramatsu N: What is
kehittämiskeskus; 1993. the quality of quality of medical care
140. Rissanen P,NoroA: Ikääntyvien Potilaiden measures? Rashomon-like relativism
Hoito- ja Kotiuttamiskäytännöt. Re- and real-world applications (with
kisteripohjainen Analyysi Aivohalvaus- discussion). Perspect Biol Med 2003,
ja Lonkkamurtumapotilaista. Aiheita 46(1):5–23.
44/1999. Helsinki: Sosiaali- ja terveysalan 150. Kane RL (ed.): Understanding Health
tutkimus- ja kehittämiskeskus; 1999. Care Outcomes Research. Gaithersburg:
141. Rissanen P, Sund R, Nordback I, Rousi T, Aspen Publishers; 1997.
Idänpään-Heikkilä U: Lonkkamurtuman 151. Iezzoni LI: Risk Adjustment for
Hoidon Vaikuttavuuden Rekisteriperus- Measuring Health Care Outcomes. Third
teinen Mittaaminen ja Vertailu. Edition. Chigago: Health Administration
Aiheita 21/2002. In. Helsinki: Press; 2003.
Sosiaali- ja terveysalan tutkimus- ja 152. Graham P, Bull B: Approximate standard
kehittämiskeskus; 2002. errors and confidence intervals for indices
142. Rissanen P, Sund R, Linna M, Idänpään- of positive and negative agreement. J
Heikkilä U, Rousi T, Nordback I: Vai- Clin Epidemiol 1998, 51(9):763–771.
kuttaako sairaalassa suoritettujen leik- 153. Andres AM, Marzo PF: Delta: A new
kausten määrä lonkkamurtuman hoidon measure of agreement between two
tuloksiin? Suomen Lääkärilehti 2003, raters. Br J Math Stat Psychol 2004,
58(12):1419–1423. 57(1):1–19.

References
154. Andres AM, Marzo PF: Chance-corrected register-based study among older people
measures of reliability and validity in K in Viborg County, Denmark, 1987–1997.
x K tables. Stat Methods Med Res 2005, Osteoporos Int 2006, 17(4):552–564.
14(5):473–492. 169. Frome EL, Checkoway H: Use of Poisson
155. Audige L, Bhandari M, Hanson B, Kellam regression models in estimating incidence
J: A concept for the validation of fracture rates and ratios. Am J Epidemiol 1985,
classifications. J Orthop Trauma 2005, 121(2):309–323.
19(6):401–406. 170. Barros AJ, Hirakata VN: Alternatives
156. Olsson U: Maximum likelihood for logistic regression in cross-sectional
estimation of the polychoric correlation studies: an empirical comparison
coefficient. Psychometrika 1979, 44(4): of models that directly estimate the
443–460. prevalence ratio. BMC Med Res
157. Langley J, Stephenson S, Cryer C, Borman Methodol 2003, 3:21.
B: Traps for the unwary in estimating 171. Schlaud M, Brenner MH, Hoopmann
person based injury incidence using M, Schwartz FW: Approaches to
hospital discharge data. Inj Prev 2002, the denominator in practice-based
8(4):332–337. epidemiology: a critical overview. J
158. Boufous S, Williamson A: Reporting of Epidemiol Community Health 1998, 52
the incidence of hospitalised injuries: Suppl 1:13S–19S.
numerator issues. Inj Prev 2003, 9(4): 172. Greenland S: Principles of multilevel
370–375. modelling. Int J Epidemiol 2000, 29(1):
159. Hornbrook MC, Hurtado AV, Johnson 158–167.
RE: Health care episodes: definition, 173. Efron B, Morris C: Stein’s estimation
measurement and use. Med Care Rev rule and its competitors--An empirical
1985, 42(2):163–218. Bayes approach. J Am Stat Assoc 1973,
160. Melton LJ, 3rd: Epidemiology of hip 68(341):117–130.
fractures: implications of the exponential 174. Agresti A: Categorical Data Analysis.
increase with age. Bone 1996, 18(3 New York: Wiley; 2002.
Suppl):121S–125S. 175. Hilbe JM: Negative Binomial Regression.
161. Brody JA, Schneider EL: Diseases and Cambridge: Cambridge University Press;
disorders of aging: an hypothesis. J 2007.
Chronic Dis 1986, 39(11):871–876. 176. Laukkanen P, Karppi P, Heikkinen E,
162. Holliday R: The multiple and irreversible Kauppinen M: Coping with activities of
causes of aging. J Gerontol A Biol Sci daily living in different care settings. Age
Med Sci 2004, 59(6):B568–572. Ageing 2001, 30(6):489–494.
163. Gavrilov LA, Gavrilova NS: The 177. Murtagh KN, Hubert HB: Gender
reliability theory of aging and longevity. differences in physical disability among
J Theor Biol 2001, 213(4):527–545. an elderly cohort. Am J Public Health
164. Wang J-L: Smoothing hazard rate. In: 2004, 94(8):1406–1411.
Encyclopedia of Biostatistics Second 178. Finsen V: Improvements in general
Edition. Edited by Armitage P, Colton health among the elderly: a factor in
T, vol. 7. Chichester: John Wiley & Sons; the rising incidence of hip fractures? J
2005: 4986–4997. Epidemiol Community Health 1988,
165. Kirkwood BR, Sterne JAC: Essential 42(2):200–203.
Medical Statistics. Second Edition. 179. Walter LC, Lui LY, Eng C, Covinsky
Malden: Blackwell Science; 2003. KE: Risk of hip fracture in disabled
166. Johnell O, Gullberg B, Allander E, Kanis community-living older adults. J Am
JA: The apparent incidence of hip Geriatr Soc 2003, 51(1):50–55.
fracture in Europe: a study of national 180. Rosenbaum PR: Observational Studies.
register sources. MEDOS Study Group. Second Edition. New York: Springer-
Osteoporos Int 1992, 2(6):298–302. Verlag; 2002.
167. Brody JA, Grant MD: Age-associated 181. Dunn HL: Record linkage. Am J Public
diseases and conditions: implications for Health 1946, 36:1412–1416.
decreasing late life morbidity. Aging Clin 182. Rosen AK, Mayer-Oakes A: Episodes
Exp Res 2001, 13(2):64–67. of care: theoretical frameworks versus
168. Giversen IM: Time trends of age-adjusted current operational realities. Jt Comm J
incidence rates of first hip fractures: a Qual Improv 1999, 25(3):111–128.

References
183. Solon JA, Feeney JJ, Jones SH, Rigg 197. Greenland S, Morgenstern H:
RD, Sheps CG: Delineating episodes Confounding in health research. Annu
of medical care. Am J Public Health Rev Public Health 2001, 22:189–212.
Nations Health 1967, 57(3):401–408. 198. Rubin DB: Matched Sampling for
184. Navarro V: A systems approach to health Causal Effects. Cambridge: Cambridge
planning. Health Serv Res 1969, 4(2):96– University Press; 2006.
111. 199. Rosenbaum PR, Silber JH: Matching and
185. Fetter RB, Thompson JD: A decision thick description in an observational
model for the design and operation of study of mortality after surgery.
a progressive patient care hospital. Med Biostatistics 2001, 2(2):217–232.
Care 1969, 7(6):450–462. 200. Holmberg S, Thorngren KG:
186. Intrator O, Berg K: Benefits of home Rehabilitation after femoral neck
health care after inpatient rehabilitation fracture. 3053 patients followed for
for hip fracture: health service use by 6 years. Acta Orthop Scand 1985,
Medicare beneficiaries, 1987–1992. Arch 56(4):305–308.
Phys Med Rehabil 1998, 79(10):1195– 201. Freeman C, Todd C, Camilleri-Ferrante
1199. C, Laxton C, Murrell P, Palmer CR,
187. Braithwaite RS, Col NF, Wong JB: Parker M, Payne B, Rushton N: Quality
Estimating hip fracture morbidity, improvement for patients with hip
mortality and costs. J Am Geriatr Soc fracture: experience from a multi-
2003, 51(3):364–370. site audit. Qual Saf Health Care 2002,
188. Chaussalet TJ, El-Darzi E: Modelling the 11(3):239–245.
process of care. Health Care Manag Sci 202. Heikkinen T, Willig R, Hänninen A,
2001, 4(1):5–66. Koskinen K, Mannismäki P, Alavaikko
189. Lagergren M: ASIM: a system for A, von und zu Fraunberg M, Jalovaara P:
monitoring and evaluating the long- Hip fractures in Finland--a comparison
term care of the elderly and disabled. of patient characteristics and outcomes
Health Serv Res 1993, 28(1):27–44. in six hospitals. Scand J Surg 2004,
190. Smallwood RD, Murray GR, Silva DD, 93(3):234–240.
Sondik EJ, Klainer LM: A medical service 203. Bottle A, Jarman B, Aylin P, Taylor
requirements model for health system R: Some way to go for consistent
design. Proc IEEE Inst Electr Electron implementation of guidance on hip
Eng 1969, 57(11):1880–1887. fracture. BMJ 2004, 328(7448):1097.
191. Smallwood RD, Sondik EJ, Offensend 204. Fairbank J, Goldacre MJ, Mason A,
FL: Toward an integrated methodology Wilkinson E, Fletcher J, Amess M,
for the analysis of health-care systems. Eastwood A, Cleary R: Health Outcome
Oper Res 1971, 19(6):1300–1322. Indicators: Fractured Proximal
192. Tilquin C: Modeling health services Femur. Report of a Working Group
systems. Med Care 1976, 14(3):223–240. to the Department of Health. Oxford:
193. Jokinen J: Fast estimation algorithm for National Centre for Health Outcomes
likelihood-based analysis of repeated Development; 1999.
categorial responses. Comput Stat Data 205. Brook RH, McGlynn EA, Cleary PD:
Anal 2006, 51:1509–1522. Quality of health care. Part 2: measuring
194. Abbot A: Sequence analysis: New quality of care. N Engl J Med 1996,
methods for old ideas. Annu Rev Sociol 335(13):966–970.
1995, 21:93–113. 206. Brand DA, Quam L, Leatherman S:
195. McVicar D, Anyadike-Danes M: Medical practice profiling: concepts
Predicting successful and unsuccessful and caveats. Med Care Res Rev 1995,
transitions from school to work by using 52(2):223–251.
sequence methods. J R Stat Soc Ser A 207. Goldstein H, Spiegelhalter D: League
Stat Soc 2002, 165:317–334. tables and their limitations: Statistical
196. Cadez I, Heckermann D, Meek C, Smyth issues in comparisons of institutional
P, White S: Model-based clustering and performance (with discussion). J R Stat
visualization of navigation patterns on Soc Ser A Stat Soc 1996, 159(3):385–
a web site. Data Min Knowl Disc 2003, 443.
7:399–424. 208. Christiansen CL, Morris CN: Improving
the statistical approach to health care

References
provider profiling. Ann Intern Med health care. Int J Qual Health Care 2004,
1997, 127(8 Pt 2):764–768. 16(Suppl 1):i37–43.
209. Normand S-LT, Glickman ME, Gatsonis 219. Therneau TM, Grambsch PM: Modeling
CA: Statistical methods for profiling Survival Data. Extending the Cox Model.
providers of medical care: issues and New York: Springer; 2000.
applications. J Am Stat Assoc 1997, 220. Hougaard P: Analysis of Multivariate
92(439):803–814. Survival Data. New York: Springer;
210. DeLong ER, Peterson ED, DeLong DM, 2000.
Muhlbaier LH, Hackett S, Mark DB: 221. Grambsch PM, Therneau TM:
Comparing risk-adjustment methods Proportional hazards tests and diagnostics
for provider profiling. Stat Med 1997, based on weighted residuals. Biometrika
16(23):2645–2664. 1994, 81(3):515–526.
211. Burgess JF, Jr., Christiansen CL, 222. Newhouse JP, McClellan M:
Michalak SE, Morris CN: Medical Econometrics in outcomes research: the
profiling: improving standards and risk use of instrumental variables. Annu Rev
adjustments using hierarchical models. J Public Health 1998, 19:17–34.
Health Econ 2000, 19(3):291–309. 223. Ho V, Hamilton BH, Roos LL: Multiple
212. Ohlssen DI, Sharples LD, Spiegelhalter approaches to assessing the effects of
DJ: A hierarchical modelling framework delays for hip fracture patients in the
for identifying unusual performance in United States and Canada. Health Serv
health care providers. J R Stat Soc Ser A Res 2000, 34(7):1499–1518.
Stat Soc 2007, 170(4):865–890. 224. McGuire KJ, Bernstein J, Polsky D, Silber
213. Howley PP, Gibberd R: Using hierarchical JH: The 2004 Marshall Urist award:
models to analyse clinical indicators: a delays until surgery after hip fracture
comparison of the gamma-Poisson and increases mortality. Clin Orthop Relat
beta-binomial models. Int J Qual Health Res 2004, 428(428):294–301.
Care 2003, 15(4):319–329. 225. Bell CM, Redelmeier DA: Mortality
214. Marshall EC, Spiegelhalter DJ: among patients admitted to hospitals on
Institutional performance. In: Multilevel weekends as compared with weekdays. N
Modelling of Health Statistics. Edited by Engl J Med 2001, 345(9):663–668.
Leyland AH, Goldstein H. Chichester: 226. Liski A, Sund R: A new register based
John Wiley & Sons; 2001: 127–142. method for comparing the costs of
215. Gilks WR, Richardson S, Spiegelhalter D: health care episodes. In: Nordic meeting
Markov Chain Monte Carlo in Practice. on register-based health research:
Boca Raton: Chapman & Hall; 1996. September 22-23 2005; Helsinki,
216. Brooks SP, Gelman A: Alternative Finland; 2005.
methods for monitoring convergence of 227. Riihimäki J, Vehtari A, Sund R, Lampinen
iterative simulations. J Comput Graph J: A Bayesian nonparametric approach
Stat 1998, 7:434–455. for modelling the lengths of stay in a
217. Spiegelhalter D, Best NG, Carlin BP, van treatment process. In: The international
der Linde A: Bayesian measures of model conference of the Royal Statistical Society
complexity and fit (with discussion). J R (RSS 2006): 10–14 September 2006;
Stat Soc Ser B 2002, 64:583–639. Queen’s University Belfast, Northern
218. Gibberd R, Hancock S, Howley P, Ireland, Great Britain; 2006.
Richards K: Using indicators to quantify 228. Sund R: Lonkkamurtuman hoitoketjun
the potential to improve the quality of arviointi kehittyy. CHESS Online 2005
(3):9.

I
Intelligent Data Analysis 7 (2003) 501–519 501
IOS Press
Utilisation of administrative registers using

scientific knowledge discovery
Reijo Sund
National Research and Development Centre for Welfare and Health (STAKES), P.O.Box 220, 00531
Helsinki, Finland
Tel.: +358 9 3967 2469; Fax: +358 9 3967 2485; E-mail: reijo.sund@stakes.fi
Received 1 November 2002

Revised 11 December 2002
Accepted 25 February 2003
Abstract. The volume of data being produced for administrative purposes is increasing rapidly. Data must be analysed in order
to extract useful information to support decision making. The demand for evidence-based information means that the analysis
must be conducted according to the principles of scientific research. Unfortunately, the massive second-hand data sets seem not
to fit very well into the traditional methodological paradigm. A secondary data source imposes limitations on the formulation
of a problem and concepts, because the measurement can only be based on existing data. The aim of this paper is to present a
methodological framework for the utilisation of administrative registers in the creation of scientifically valid information. This
is done by discussing fruitful methodological aspects encountered in the practical knowledge-discovery process. The ideas
presented originate from many different fields, such as statistics, data mining and sociology. The emphasis lies on understanding
connections between problem, data and analysis in the case of massive secondary administrative data sources.
Keywords: Scientific research process, secondary data, administrative registers, event history analysis, statistical data mining,
health care provider profiling
1. Introduction
Data have been produced for hundreds of years. The reasons for such production were originally
administrative in nature. There was a need for systematically collected numerical facts on a particular
subject. Later, a belief in the advantages of quantitative information accompanied by advances in
statistical techniques led to the birth of a widely used methodological paradigm for scientific research.
Even more recently, advances in information technology have made it possible to more effectively collect
and store larger and larger data sets. Administrative information systems have been at the forefront of
data gathering, since there have been a growing demand for evidence-based information to support
decision making and other administrative purposes. In many cases, however, raw data as such are of little
value. Data must be analysed in order to convert them to useful information. Unfortunately data-driven
analysis and the massive size of data sets seem not to fit very well into the traditional methodological
paradigm.
A traditional way of creating scientifically valid information is to design and carry out a scientific
study, in which data production, analysis and reporting are all strictly problem-specific. Data represent
tailored measurements of observational reality that are needed in problem solving. Analysis is then used
1088-467X/03/$8.00  2003 – IOS Press. All rights reserved

502 R. Sund / Utilisation of administrative registers using scientific knowledge discovery
to reveal ‘how things work’ in empirical data. By interpreting these empirical results in the theoretical
context and discussing relevant limitations, it is hopefully possible to create a sophisticated answer to a
problem in such a way that it can be evaluated by the scientific community.
The most important challenge for the traditional paradigm is the nature of data. If it is not possible to
produce tailored data according to the problem in question, the whole problem-solving process becomes
more restricted. The secondary data sources utilised impose limitations on possibilities for technical
analysis. Moreover, the formulation of a problem and the associated concepts inevitably becomes
opportunistic, because the measurement can only be based on existing data [6]. For some researchers,
even the validity of data is suspect if they were not produced for the specific purpose for which they
are used. Such criticism is not necessarily based on evidence concerning the quality of the information
contained by the data but rather on philosophical beliefs regarding the connection between data origin
and data quality [35].
In addition, the growing size of data sets has raised some issues that must be dealt with in a different
way: data exploration and description plays a more general role than it does in the case of small data
sets; efficient algorithms (and computers) are needed in analyses; the unfeasibility of manual analysis
results in separation between the data and the analyst; the manual management of all errors in the data
is not possible; and the chance occurrence of patterns is more likely in large data sets. In addition, some
databases include the whole population and not just a sample, which changes the nature of the whole
statistical inference [13].
In other words, in the analysis of secondary data, the problem must be defined under preconditions
arising from the data. Moreover, massive data sets are usually so overwhelming that their processing and
analysis is difficult (‘data overload syndrome’). It may also be tempting to screen the data for significant
associations without having an adequate prior hypothesis (‘snooping/dredging/fishing’). Rather than
tackling too much on what can and can not be done, analysis should achieve an appropriate compromise
between the practical problem and the data. This kind of activity has been characterised as ‘greater
statistics’, which tends to be inclusive, eclectic with respect to methodology, closely associated with
other disciplines and also practised by many non-statisticians [3]. Moreover, the larger the data sets are,
the more important the general-science and collaboration aspects of the analysis process seem to become
relative to the ‘statistical’ aspects.
Since statistical research has traditionally focused on probabilistic inference based on mathematics, it
has not been able to offer very much concrete help in these different circumstances. The practical need
for information has led to the development of alternative ways of analysing data, such as data mining [9].
However, these strategies for analysing data have much in common, and formal statistical expertise also
provides an excellent basis for the understanding and evaluation of these ‘new’ ideas. Statisticians ought
to take advantage of the situation, get involved in interdisciplinary activities, learn from the experience,
expand their own minds – and thereby their field – and act as catalysts for the dissemination of insights
and methodologies [15].
1.1. The aim of the paper
The aim of this paper is to present a methodological framework for the utilisation of administrative
registers in the creation of scientifically valid information. This is done by discussing the fundamental
methodological issues encountered in the practical knowledge discovery process. The ideas presented
originate from many different fields, such as statistics, data mining and sociology. The paper emphasises
a broader understanding of connections between problem, data and analysis in the case of massive
secondary data sources.

R. Sund / Utilisation of administrative registers using scientific knowledge discovery 503
Debate
Problem
Context
Question Method
Perspective Data
Analysis
Idea Answer
Theory
Fig. 1. Research process schema.
The development of this paper is heavily influenced by the context of health services research, a field
in which the most valuable official data sets are individual-based. In these circumstances, the histories
of individuals recorded in administrative data are the most general starting point for any analysis – and
a fruitful starting point for this paper, too. A practical real-world example – used to demonstrate the
research made by following the guidelines presented in the paper – is also taken from the context of
health services research.
2. The scientific knowledge-discovery process
The ultimate goal in all research is – or at least should be – to search for appropriate answers to
relevant problems. To find sensible and scientifically valid solutions requires not only a knowledge of
the phenomena being investigated but also the capacity to think originally about the nature of inference
and the process of conceptualisation. The whole problem-solving process must be conducted in a
systematic way, using appropriate methods. Even though it is not in general possible to split research
process completely into distinct phases, certain common phases are encountered repeatedly and even
simultaneously during the research process. The process does not proceed in such a way that one
phase is completed first, after which one moves on to the next. Findings made during the research
process inevitably have effects on the final conclusions, since these preliminary findings provide a
deeper understanding of the phenomenon and data in question.
One sophisticated characterisation of complex problem-solving, is known as a knowledge-discovery
process (see e.g. [2,4,8]). However, despite the practical usefulness of a process description of this kind,
some critical issues concerning the scientific nature of the research process are not mentioned at all.
To see what is actually happening in the research process, it is helpful to place it in a wider context. For
this purpose, an alternative schema for a research process is presented in Fig. 1. The interactive phases
related to this schema are: understanding the phenomenon, understanding the problem, understanding
data, data preprocessing, modelling, evaluation and reporting.
Research is conducted in order to obtain an answer to some problem. The researcher has to choose the
perspective from which solving the problem is approached, since research requires rigorous communi-
cation and careful definitions of concepts related to a phenomenon of interest. However, all researchers
have some more or less latent foreknowledge, which provides directions and criteria for understanding
the world and also guides understanding gained and interpretations made during the research process.
The researcher must be aware that these choices and interpretations are his or her ideologically and
historically dependent decisions. The choice of perspective and conceptions may even prevent one from
seeing (or enable one to see) something important or interesting.
In order to move from the theoretical level to the empirical level, the researcher must operationalise
the research frame in such a way that it is possible to produce useful information from the phenomenon

and thus create some answers to the problem. This kind of operationalisation is fundamental to the
plausibility of research. This is the stage at which the research defines the point of view from which the
phenomenon and concepts are examined.
Data represent the empirical world in research. Observed information can not cover all details which
are relevant to the phenomenon of interest. The observation and measurement must also be dealt with
in a conceptual context. This means that some compromises must be made between the exactness of
measurement and the ambiguity of observation. Definitions can not replace the empirical information, but
definitions affect the organisation of information and the decisions based on such information. From this
point of view, it is easy to argue that data should be produced strictly according to the research problem.
That is not the case with secondary data sources, such as administrative registers, since secondary data
are originally produced for some other purposes. The available data and the conceptual definitions used
in their production are not necessarily compatible with the research problem at hand. As an extreme
example, imagine that your observational world is all that you see, hear, smell and feel. Would you
trade your observational world for an old black-and-white photo, if you could trade it for a high-quality
three-dimensional digital image with full colour scale, including infrared?
All in all, the researcher must decide what to observe, how to conceptualise variables, what kind of
operationalisation to make and how to acquire appropriate data. The same freedom also applies to the
choice of analysing strategy, interpretation of the results and discussion of alternative interpretations.
Therefore it is important to reflect upon decisions made during the research process. Every scientific
explanation is valid only in so far as the investigator can provide a rationale for what has been done. At
each stage of the research, decisions and arguments must not rest on the uncritical or automatic use of
established techniques but on a set of logical decisions which must be taken on theoretical grounds with
full consciousness of what is being done. Results must be justified. In the example above, one obvious
question to be answered is: Why was that particular black-and-white photo chosen? In this example,
ideology may be viewed as the position of and lighting for the camera, and historical dependency as the
moment of taking the picture in question.
Since in the case of administrative registers the data typically have certain common properties, it is
possible to make some suggestions of which it is useful to be aware if data originating from such a source
are to be utilised. The next section briefly describes a few ideas that have been found very worthwhile
in practice. In the research process these ideas may be seen as aspects of the following three phases:
understanding the problem, understanding data, and data preprocessing.
2.1. Understanding the problem
Time is an essential factor in many problem domains. For example, disease processes evolve in time,
and patient records give the history of patients. In research on such dynamic phenomena, interest very
often focuses on the sequences of events which occur in time and reflect changes in research objects. A
good example is an employment career, which can be regarded as consisting of the sequences and timing
of the various jobs and of their association with other events, such as changes in marital status and place
of residence. In fact, the importance of the longitudinal event history approach has long been recognised
in many areas, especially in the social sciences, econometrics and medical research (see e.g. [5,19,34]).
2.1.1. Operationalisation
For the purposes of problem solving, the problem has to be ‘matched’ against the corresponding domain
knowledge and data. This can be done by characterising the important properties of the phenomenon
carefully, i.e. by operationalising it as a system. A system is defined as a ‘group of things or parts

B: Outpatient
A: No care care D: Death
C: Hospital
care
Fig. 2. A system shown as a directed graph.
working together or connected in some way so as to form a whole’. This definition of a system requires
that the concepts, objects and restrictions related to the system be adequately described.
In the case of event histories, a common approach to characterising the phenomenon is to categorise
events in such a way that they describe transitions across a set of discrete states in time. In the simplest
case only one qualitative change is possible – for example, the transition from the state ‘alive’ to the
state ‘dead’. However, in many cases one transition can not describe the dynamics of a phenomenon in
a realistic way. For example, the slow development of a disease which finally results in the state ‘death’,
can not be modelled accurately using only the time elapsing between the first diagnosis and death.
If the system is constructed in a reasonable way, the individuals’ event histories can be regarded as
‘paths’ through the system. Pictorially the system can be presented as a directed graph whose nodes and
edges have characteristics which describe the properties of possible states and transitions in a system
(Fig. 2).
This kind of presentation can be criticised as simplistic or even confusing, since it may seem that
nothing happens to an individual as long as he or she remains in the same state. On the other hand, it
has been found that this kind of formalisation of a problem is easily understandable even for technical
outsiders and that it thus provides a common language for the experts from different fields which is
required for effective co-operation between members of a research group.
As a matter of fact, the presentation technique described above does not restrict the number of possible
states. If it seems too unrealistic to assume that nothing happens during a stay in a state or that stays
in the same state are not comparable, it is not a problem to split a state into a greater number of more
specific states. In general there is no limit to this kind of hierarchical splitting. The choice of an
appropriate degree of hierarchical structure is a good example of a fundamental decision on how to reach
a compromise between exactness of measurement, ambiguity of observation and – most importantly –
the practical needs of the problem-solving process in question.
2.2. Understanding data
In register-based analyses, the data themselves impose restrictions on the whole problem-solving task.
Such data are known as secondary data, in contrast to the primary data obtained by the investigator through
direct observation or interviewing. The use of secondary data raises particular problems associated with
the need to take account of the process by which the data were compiled. In general, data must have a
form appropriate for storage and analysis purposes, and in addition for intuitive interpretation.

Noise
Phenomenon Coding Register Decoding Data
Fig. 3. Schematic diagram of information communication via administrative registers.
2.2.1. The nature of data

If we apply the traditional Shannon’s communication model in the context of administrative registers,
the transmission of information in register utilisation can be regarded as having a structure such as that
presented in Fig. 3 (cf. [31]).
The main idea is that some parts of the phenomenon of interest are observed and coded to signal, which
is then transmitted using a possibly noisy channel into a database and then decoded in order to obtain
proper data concerning the phenomenon. Even though this is a very simple and technical representation
of communication, it seems to contain the essential elements needed in the common-sense understanding
of secondary data. In other words, it is assumed that there exists some phenomenon which can be
observed. Since it is impossible to completely observe all details or make exact measurements, some
kind of coding is used to describe things. This coded signal is then stored in a database. The noise can be
interpreted as an explanation for measurement compromises, possible inconsistencies and coding errors,
and coding practices existing in the stored signal. When this signal is to be utilised, it must be decoded
into understandable data, i.e. into a form which tells us something about the phenomenon. This phase is
also subject to noise caused by incompatibility of choices and interpretations made by the data producer
and the data user.
In other words, there are many problematic preconditions in the case of administrative data. Most
importantly, the limitations of data are determined by the choices made by the producer of the data:
easily available data are also easier to collect; data are produced using fixed categorisations and may be
dependent on producers’ interpretations; there may be many data producers, with consequent variation
in production practices and categorisations used; and information systems may not be flexible enough to
take changes in phenomena and society into account. In practice, secondary data are interpreted in order
to utilise them for purposes other than those for which they were originally intended. The most typical
problems encountered in such a situation are the lack of data on important background phenomena and
discrepancies between the users and producers of data as regards their beliefs and knowledge concerning
the nature of data. It is possible to grasp the real nature of the data only if one understands (a) what
factors have led to the production of certain data and a particular information system, (b) what are the
sources and consequences of the type of organisation and data structures used, and (c) whether the data
have an effect on the understanding of their domain [24,32].
2.2.2. Data structures

Event-history data consist of observations of the form (τ k ,Dk ), where τk is an ‘occurrence time’ and
Dk is an ‘explanation’ for the event (and n is the number of observations and τ 1 � τ2 � . . . τn and
τi < τj for at least one observation i �= j and i, j, k = 1, 2, . . . , n).
Usually Dk consists of a set of attributes (variables). However, not all the attributes necessarily contain
important or interesting information. Some attributes can be irrelevant to the solving of particular problem
or can be easily derived from other attributes. It is often reasonable to divide up the relevant information
into two subsets of attributes. The first subset defines an event type E k and the other includes important

A
tk Ek B
0 A C
4 C D
8 B
time
0 2 4 6 8 10 12 14 16 18
10 C
13 D
A C B C D
0 2 4 6 8 10 12 14 16 18 time
S = < {(0,A)}, {(4,C)}, {(8,B)}, {(10,C)}, {(13,D)} >

Fig. 4. An example of event-history data and the corresponding event sequence.
‘covariate’ information ik . For the sake of generality it is useful to allow a transformation f k for the
occurrence time τk . If not stated otherwise, fk (τk ) = τk (k = 1, 2, . . . , n).
Let m be the number of distinct occurrence times, t i be the ith distinct occurrence time (i = 1, 2, . . . , m)
and the event set A i be the set of relevant information on observations occurring at the same time, i.e.
Ai = {(ti , Ek , ik )}, where i = 1, 2, . . . , m and for every i, k is over the observations for which
ti = fk (τk ).
The generalised event sequence S is defined to be a queue of event sets sorted by the (transformed)
occurrence time, i.e. S =< A1 , A2 , . . . , Am >. In addition, let the conditional event sequence, to
be called an event subsequence, be a sequence S θ =< Ai |Ai ∈ S and condition θ is true >, where
i = 1, . . . , m.
The definition of the generalised event sequence given above is not very rigorous and can be used in
the absence of any knowledge about the event-history framework. However, if one uses this framework,
generalised event sequences acquire a constructive and intuitively clear interpretation. Moreover, trans-
formations allowed in generalised event sequences make it possible to use the same data structure in the
implementation of different statistical and data mining methods.
2.2.3. An example of a generalised event sequence

Figure 4 shows an example of event-history data and the corresponding event sequence, which could be
produced by system – say system P, presented in Fig. 2. Two graphical representations of this particular
event history are also shown. The first is an event-history description of transitions in a system where an
individual actually stays in a current state until there is a transition to another state. In the second, only
the events occurring are marked on the figure. Since the time between two consecutive events is also
the length of stay in a particular state, it is often very useful to include a ‘length of stay’ attribute in the
covariate attributes ik , even though it can be easily calculated from the corresponding occurrence times.
In the definition of a generalised event sequence, parallel occurrences of events are allowed. In
principle, the systems approach can be rendered valid in this case by defining each combination of event

A C B C D G F
A C B C D
a) 0 2 4 6 8 10 12 14 16 18 time b)
G F
0 2 4 6 8 10 12 14 16 18 time
0 2 4 6 8 10 12 14 16 18 time
S = < {(0,A), (0,G)}, {(4,C)}, {(8,B), (8,F)}, {(10,C)}, {(13,D)} >

Fig. 5. An example of a situation with parallel occurrence times.
sets as a ‘new’ event type. In other words, one can think of the explanation for an event type as a
combined description of events occurring at the same time. However, in many cases it is reasonable to
assign the parallel events to different event types, since it is possible to find ‘natural’ interpretations for
these parallel events.
The data shown in Fig. 4 constitute a description of movements of individual X in a system P. Let Q be
a system with two states (state F: ‘married’ and state G: ‘not married’). As a result, data corresponding
to the movements of individual X in the system Q have the same form as the data in Fig. 4. Figure 5a
shows an example of this kind of situation. Now there are two event histories for individual X, one
corresponding to the path through the system P and the other to that through the system Q. In this case
it is known that the systems which ‘generate’ the data are parallel, but the two event histories can be
combined into one event sequence (Fig. 5b).
In the case of two parallel systems, the information concerning the ‘source system’ of an observation
is a very valuable covariate. In general, there can, of course, be more than two data-generating systems.
If there is a need to restrict analyses to the observations obtained from some particular system, this can
be done easily by using an event subsequence conditional to the corresponding system.
In practice there are always event histories for more than one individual. It is thus trivial to include a
covariate which identifies an individual, for example his or her social security number, while the whole
data set still has the form of a generalised event sequence. Again it is possible to restrict the analyses to
a pretermined set of observations by using event subsequences with suitable conditions.
2.2.4. Censoring
An additional point to consider is the fact that in practice, data constitute only a narrow window on the
dynamics of a phenomenon. In other words, the observations contained in the data set fulfil the condition
a < ti < b, where a and b are finite constants and t i is the occurrence time of an event. The problems
caused by a limited observation window are illustrated in Fig. 6, which shows examples of different
types of ‘censoring’. Each case corresponds to an individual’s length of stay in some particular state, i.e.
each case is combination of two events: transition to the state of interest and away from it. Dotted parts
of the lines are unobserved; this corresponds to the case where the ‘true’ transition to or from the state
is not observed. Censoring in fact guides the possibilities for analysis, and it must be taken into account
in such analyses.
In the cases (a) and (g), there are no observed transitions to or from the state. This kind of censoring
can be very problematic if the very first (or last) occurrence of some event type is considered more
important than other occurrences (for example the first diagnosis of schizophrenia or the first back-
surgery operation). In the case (b), the true transition to the state is not observed, but the individual was

a
b
c
d
e
f
g
h
t1 t2 time
Fig. 6. Examples of censoring.
in that state when the observation began, and the transition time from the state is known. In the case (c)
both transitions are observed, but there is a potential problem, because the data outside the observation
window are not complete for all individuals (typical situation for hospital discharge data). The case
(d) corresponds to the uncensored observation. In the case (e), there has been an unknown or ‘wrong’
transition from the state (drop-outs or ‘competing risk’). It is also possible that the follow-up has ended
before the occurrence of the event of interest (case f) or the transitions to and from the state are outside
of the observation window (case h).
2.2.5. Statistical interpretations of a generalised event sequence

A generalised event sequence can be interpreted as a sample path of a marked point process, if the
occurrence times for all events are distinct. This leads to a very general family of statistical hazard-rate
models suitable for censored data (see e.g. [1]). In practice, all methods in the event-history framework
are special cases of this general interpretation.
For example, there are many situations in which the ‘calendar time’ of event occurrence is not
important, since the ‘real’ information is the time between two consecutive events. In other words, the
‘starting time’ of a follow-up can vary between individuals. A traditional solution, very common in
survival analysis, is to transform the time axis from calendar time to ‘failure time’. This can be done
using a transformation function f k (τk ) = τk − bk , where bk is the occurrence time of ‘starting event’ of
a corresponding individual and k = 1, 2, . . . , n (see e.g. [7,18]).
Assuming that the probability of the next event type (state) depends only on time spent in the current
state, an appropriate choice for a model is a semi-Markov-model (see e.g. [29]). By restricting the
probabilities to change only at discrete time points, the semi-Markov model can be formulated using
the Markov chain, where the state space is expanded in a proper way (see e.g. [14]). In the traditional
Markov chain, the probability of the next event type (state) depends only on the current state (first-order
Markov property), and the probabilities are time-homogeneous. In fact, the Markov chain interpretation
corresponds to a situation in which the exact occurrence time is not important and only the order of
observations matters. A generalised event sequence can be transformed to an event type sequence of this
kind by using the transformation f k (τk ) = k(k = 1, 2, . . . , n).
Moreover, by choosing the transformation function f k so that it is of the form fk (τk ) = c, where c is
an arbitrary constant and k = 1, 2, . . . , n, the time dimension of the event sequence can be eliminated
and the sequence reduces to the data miner’s classical market basket model (see e.g. [12], Chapter 6) .

2.3. Data preprocessing
Usually massive second-hand data sets contain so much information and so many domain-specific
features, inaccuracies and problems that raw data as such are not usable. In order to use data in problem
solving, there must be understanding of the connections between the problem and the data. In register-
based analyses, the problem, domain knowledge and data determine the most suitable model for the final
problem solving. Typically it is possible to construct data sets of the event-history type using register
data. Preprocessing can also be interpreted as a kind of technical operationalising phase in the research
process.
2.3.1. Data abstraction

Often the connections between highly specific raw data and the highly abstract domain knowledge are
so complicated that it is not possible to find any direct links between data and knowledge. An intelligent
interpretation of raw data must be embedded into analyses, so that the resulting derived data set is at the
level of abstraction corresponding to the current problem. Since noise is an unavoidable phenomenon,
some kind of data validation and verification which makes use of knowledge should also be performed.
This kind of task, performed in order to abstract higher-level concepts from possibly time-stamped data,
is called data abstraction (see e.g. [20], Chapter 2; [30]). In the discovery of medical knowledge, data are
usually patient-specific, while medical knowledge is patient-independent and consists of generalisations
that apply across patients. For example, a complication after surgical operation is a medical concept, but
from individual-based data it must be abstracted by using some ‘rules’, such as a list of some particular
diagnosis codes recorded in data[RF6].
2.3.2. Data cleaning

Real-world data are very often more or less incomplete, noisy and inconsistent. Data cleaning involves
detecting and removing errors and inconsistencies from data in order to improve the quality of data (see
e.g. [25]). Compared to data abstraction, cleaning is more data-driven and technically oriented. In other
words, corrections of erroneous and inconsistent codes as well as missing values can be usually made to
the whole database, but data abstraction always results in problem-specific derived data sets.
2.3.3. Data integration and reduction

Two other common types of preprocessing are known as data integration and data reduction. The idea
in data integration is to include data from multiple sources in analyses; the process is also known as
record linkage (see e.g. [26,36]).
Data reduction results in a reduced representation of a data set which is much smaller in volume than
the original data set, yet produces the same (or almost the same) analytical results (see e.g. [12], Chapter
3). Data reduction can in fact consist of anything from simple database queries to very complicated
analyses.
2.3.4. Preprocessing tasks in the event-history framework

Sometimes it can in fact be difficult to distinguish what is operationalising, what is preprocessing and
what is modelling. All these phases need interpretation and evaluation of the results. Since these are also
highly domain-specific and problem-specific, it is difficult to give any general suggestions concerning
tasks and methods for analyses, other than that the problem in question implicitly defines the most
suitable technique and that the assumptions underlying techniques applied must be valid.

However, some non-trivial preprocessing tasks in the event-history framework could include: ‘defin-
ing’ the state space for some system (What happens to a patient after a surgical operation?); finding
interesting and frequent combinations of patterns (What combinations of diagnosis and operation codes
are frequent?); assigning patterns to adequate hierarchies (What diagnoses relate to complications of a
surgical operation?); and confirming the expert’s ‘hypotheses’ about the phenomenon from data.
Most analysing techniques are feasible only in the case of moderately small data sets, since these
typically need access to the whole data set, and the processing time will be directly proportional to
the physical file size. In fact, many data mining methods are very well suited in such circumstances
and are clearly something that is needed in practical data analysis. It can be stated that a sophisticated
preprocessing operation incorporating non-technical domain knowledge in order to scale things down
to a size fit for more specific statistical analyses is the most important and time-consuming part of
register-based data analysis.
One very straightforward but extremely useful preprocessing technique is ‘pattern remapping’. The
idea is to first ‘forget’ the time dimension in the generalised event sequence and to use a levelwise search
(see e.g. [21]) to extract the frequent patterns from data, regardless of occurrence times. These patterns
(such as a list of medical diagnoses) are then given to a domain expert, who can identify and make
hierarchical groupings for the relevant patterns. In the remapping phase, interesting patterns are then
assigned to appropriate event types simply by ‘renaming’ interesting patterns according to suggestions
made by the domain expert. Finally, the time dimension is restored and records lacking interesting
event types are removed. This abstraction usually results in a considerably reduced data set with an
interpretation corresponding to the current problem.
3. Practical example: Hip-fracture surgery
This example is a simplified extract from a study, which aimed 1) to develop and implement register-
based performance indicators to measure the effectiveness of surgical treatment of hip fracture and 2)
to evaluate and compare the effectiveness of health-care providers. The complete results are reported
elsewhere [27,33]. The study is a part of a larger project which aims to develop register-based methods
for the measurement of effectiveness in specialised health care.
3.1. Defining the problem
The first task of the project was to build up a research group consisting of experts from different fields.
The group defined the actual problem more meticulously: the idea was to identify all hip-fracture patients
from the Finnish Health Care Register and follow the life events that they encountered after hip-fracture
surgery according to register data.
A simplified system related to this particular problem is shown in Fig. 7. As can be seen, the first
hip-fracture operation performed on an individual patient has a key role in the characterisation of this
phenomenon. Actually, the state preceding the first hip-fracture operation also matters from the clinical
point of view, since the patients coming from home are usually in better condition than patients who are
already in residential or hospital care. However, the major interest is in the events and pathways of care
following the first hip-fracture operation. In this case, these events are classified into four categories.
If everything goes well, the patient should return home. Hip fracture is a serious condition for the
elderly, and it can also be a starting point or catalyst for other problems which may result in the need for
residential or hospital care. A hip-fracture operation may even be followed by fatal complications. This

Home
Home Residential/
hospital care
Hip-fracture
operation
Residential/ Complication
hospital care
Death
Fig. 7. System characteristics of significant life events related to hip-fracture surgery.
particular categorisation was chosen because the event types comprised the events of interest from the
viewpoint of the original problem; and even more importantly, these life events are recorded in various
registers.
For the sake of generality, the system’s formulation also allows there to be multiple events after hip-
fracture surgery. As a matter of fact, the pathways of care are more interesting than single events in
situations such as a cost-effectiveness evaluation. In the formulation of the system, only death is an
absorbing state; all other states can be followed by any other state, i.e. only the death state ends a path in
the system. Moreover, even though these four states are distinct, they are not necessarily independent.
For example, death may be more probable after complications, or it may not be very likely that the
patient will return home if long-term residential care is approved. The absorbing death state causes
even more problems, because it is not possible to say of any individual that there is an increased risk
of complications after death. However, it is possible to speculate what would have happened if death
had not occurred. All in all, this kind of quite simple system definition seems to result in an extremely
complicated competing risks model, and more simplifications are needed in the actual modelling.
3.2. Understanding and preprocessing data
A cohort of patients with hip fracture in 1998 or 1999 was identified in the Finnish Hospital Care
Register using a simple diagnosis-group abstraction (all patients with at least one ICD-10: S72 diagnosis
in 1998 or 1999). Using the unique personal identity codes of the patient cohort, data on all inpatient
and outpatient hospital care and deaths for this cohort were obtained from the Finnish Health Care
Register, the data warehouse of the Finnish Hospital Benchmarking Project and the National Causes of
Death Register. The results of these straightforward database queries were integrated into a new data set
containing 167 952 records for 17 099 patients.
Each record in this data set corresponds to one care episode in hospital (or death), not any actual event
of the system, i.e. each observation includes information, such as patient and hospital ID-numbers, age,
sex, area codes and diagnosis and operation codes, as well as dates of admission and discharge (or death).
Data cleaning was performed in order to correct impossible simultaneous hospital episodes, systematic

errors in the use of symptom vs. cause diagnoses, and some missing or erroneous attribute values in area
codes.
Many types of censoring occur in this data set: the first hip-fracture operation (or other important
event) can be outside the observation window; some hospital episodes may have begun before 1998;
follow-up finishes at the end of 1999 (there are census data available for the last day of every year in the
Finnish Health Care Register); and follow-up may be terminated by the death of a patient.
Operation codes corresponding to hip-fracture surgery were abstracted into two different operation
types. Using this and the diagnosis-group abstraction of hip fracture, hip-fracture operations were
identified from the data. Since the state preceding a hip-fracture operation was also important, the
histories of the patients were traced backwards and the preceding state was classified as home or
residential/hospital care by using data abstraction of a more complicated nature.
The forward-direction abstractions were even more complicated, and all event types needed special
abstractions and techniques. For example, the acute complication events were identified using the pattern
remapping technique, in which all clinically relevant complication diagnoses were remapped to one event
type.
3.3. Modelling, evaluation and reporting
For statistical modelling purposes it was assumed that any acute complication event occurring after
a hip-fracture operation is an outcome which reflects the effectiveness of the surgical treatment. In
addition, deaths and the upper limit of the observation window were assumed to cause censoring of
the event of interest. With these assumptions, the modelling reduced to standard survival analysis (see
e.g. [7]) where the variables of interest are the time between a hip-fracture operation and a complication
or censoring event, and the censoring indicator. These variables were calculated for all of those patients
who were (a) aged over 60 years and (b) had been living at home before surgery. The final preprocessed
data set had 8824 records, each containing relevant variables for one patient.
In survival analysis, the distribution function of failure time random variable T is F (t) = P (T � t),
where t � 0, and S(t) = P (T > t) = 1− F (t), t � 0, is the corresponding survival function. Assuming
F to be absolute continuous and f to be the corresponding density, the hazard function is defined by
r(t) = f (t)/S(t), t � 0. The differential r(t)dt = P (T ∈ dt|T � t) has the intuitive interpretation of
‘the conditional failure probability at time t, given survival to at least t’. Moreover, the hazard function
determines uniquely the distribution function, and hazard-based models are often a convenient way to
handle censored observations.
In spite of the fact that the time dimension includes a lot of information, the actual effectiveness
indicators should be as easy as possible to interpret. In this study the continuous time scale was reduced
to a simple dichotomous scale which merely indicates whether or not the event of interest has occurred
in the case of an individual during some particular ‘limit’ time period. However, it is not obvious how
such ‘limit’ times should be chosen. In addition to using clinical knowledge, it is also possible to utilise
data empirically in this task. Since the outcome was an acute complication, the hazard function of
acute complication occurrences was estimated. According to the hazard function presented in Fig. 8, the
probability of acute complications was higher during the first 30-day period after a surgical operation.
This finding based on the data corresponded to the domain knowledge and provided some evidence that
the data abstraction was done in a proper way.
Since there are censored observations in the data, the proportion of the risk population for whom an
event of interest occurred within the ‘limit’ time period does not necessarily correspond to the ‘true’

Hazard function for acute

P
0.01
0.009
0.008
0.007
0.006
0.005
0.004
0.003
0.002
0.001
0
0 30 60 100 150 200 250 300 350 400 450 500 550
Time (days)
Fig. 8. Hazard function for acute complication.
occurrence rate. In other words, rates must be estimated using more sophisticated methods. In this
study the censoring was assumed to be independent (survival experience in the future is not statistically
altered by censoring and survival experience in the past). The ‘limit’ time rates were estimated using
product-limit (Kaplan-Meier) estimators. In addition, the counts of events were needed, and the corrected
‘observed’ counts were calculated using the relation O i = Ni F̂i (t), where i indexes the risk population,
Ni is the size of risk population i and F̂i (t) is the estimated cumulative probability of event occurrence
at time t (‘limit’ time rate) in risk population i.
3.4. Profiling the providers
The most useful information is obtained when rates are evaluated for health-care providers, such as
hospitals or hospital districts. Profiling analyses of this kind also allow comparisons of effectiveness
between providers. However, the profiling can be quite complicated, since there is variation between
providers for at least three reasons: 1) differences may be attributable to random variation caused by the
size of the provider, 2) the patient case-mix may vary from provider to provider, and 3) providers may
differ in the effectiveness of their care. For these reasons, a statistical model for provider profiling in
which provider differences are modelled explicitly was constructed in the study.
For an individual patient j , an observed outcome has the form (Y j |xj , zj ), where xj is a vector of
patient characteristics and z j is a provider-specific effectiveness component. The expected outcome for
a patient is E(Yj |xj ), i.e. a constant degree of effectiveness is assumed. Obviously, if the expected
outcome is subtracted from that observed, the remaining residual reflects the effectiveness of care of a
provider.
In the case of binary outcomes, a logistic regression is a suitable tool for the calculation of the expected
outcomes. The idea is to construct and estimate a model in which the observed outcome is a dependent
variable and patient characteristics are independent variables. Using this kind of model, it is possible
to calculate predicted values for all individuals using patient characteristics and estimated values of
parameters with the inverse logit transformation.

Since the focus in profiling is on providers and not�on individuals, �the observed and expected outcomes
must be aggregated to the provider level, i.e. O i = Yj and Ei = logit−1 (xj β), where the sums are
over patients treated by provider i and β is an estimated parameter vector.
Traditionally the ratio of observed to expected outcomes multiplied by the mean rate is used as a
risk-adjusted rate for a provider. This quantity forms the basis for comparisons between providers and
can be interpreted as the estimated event rate for a given provider if the population of patients treated
were identical to the nationwide case-mix.
Since the observed outcomes O i are non-negative integers describing frequencies of events, they can
be assumed to have a Poisson distribution with unknown mean µ i . That is, Oi ∼ Poisson(µi), where
log µi = log Ei + θi and i is the provider index. In other words, it is assumed that the expected outcomes
Ei adjust the patient characteristics, and θ i describes the variation caused by provider. The use of
logarithms guarantees that θ i remains positive in the model.
Case-mix adjustment does not necessarily eliminate the variation in performance indicators in such
a way that the indicators reflect a provider‘s effectiveness of care. For example, in data sets with
a hierarchical structure there often exists correlations between observations, and this may result in
overestimated differences in profiling analyses. Differences in the sizes of providers may also cause
problems. For example, rates can not be estimated accurately for small providers.
Assuming the exchangeability of providers (i.e. that the results for all providers are equal if there is
infinite number of (similar) patients), a two-level hierarchical model can be used to solve the problems
mentioned above. A simple solution is to assume that variation caused by providers is normally
distributed, i.e. θi ∼ N (α, σ 2 ), where exp(α) is the ‘general’ case-mix-adjusted risk ratio and σ 2
describes the variance between providers (in logarithmic scale). This kind of hierarchical Bayes model
needs appropriate prior distributions for the hyperparameters α and σ 2 , such as α ∼ N (0, 106 ) and
σ −2 ∼ Γ(0.001, 0.001). The estimation of posterior distributions for parameters of interest (in this case
θi s for all providers i) can be done using Markov chain Monte Carlo (MCMC) simulation [22].
Multiplying the obtained adjusted risk ratios by the mean rate results in easily interpretable risk-
adjusted rates for providers. The use of a hierarchical multilevel model eliminates many drawbacks
of traditional profiling analyses. Multilevel models are well suited to the simultaneous calculation of
many confidence intervals (multiple comparisons problem), and they give more conservative estimates
for differences between providers than do traditional methods [11].
3.4.1. Example drawn from the results

Figure 9 shows the hospital-district-specific risk-adjusted rates for 30-day complication rates. The
mean 30-day acute complication rate was 13.0%, and there were no significant differences between
districts according to the 95% confidence intervals of rates. However, according to the 50% confidence
intervals, hospital districts 9 and 15 seem to have slightly higher rates than others. For more results,
see [27] and [33].
4. Conclusions
This paper has presented a methodological framework for the utilisation of administrative registers in
the creation of scientifically valid information. This has been done by discussing essential methodological
criteria encountered in the practical research process and by combining fruitful methodological ideas
from different fields, such as statistics, data mining and sociology. The emphasis has been on the
understanding of connections between problem, data and analysis in the case of secondary data sources.

Hyvinkaa-Porvoo 33
Lohja-L-Uusimaa 34
Helsinki 31
Jorvi-Peijas 32
V-Suomi 3
Satakunta 4
K-Hame 5
Pirkanmaa 6
P-Hame 7
Kymenlaakso 8
E-Karjala 9
E-Savo 10
I-Savo 11
P-Karjala 12
P-Savo 13
K-Suomi 14
E-Pohjanmaa (dottedline).
15
Vaasa 16
K-Pohjanmaa 17
P-Pohjanmaa 18
Kainuu 19
L-Pohja 20
Lappi 21
10 11 12 13 14 15 16 17
Fig. 9. Hospital district-specific risk-adjusted 30-day complication rates in 1998 and 1999. Rates as complications per 100
patients. Thin parts of lines correspond to 95% confidence intervals and thick parts correspond to 50% confidence intervals.
Mean rate is 13 complications per 100 patients (dotted line).
In the introduction, the problem caused by a changing paradigm of data analysis was discussed.
Different perspectives are needed in analyses. Bringing carefully defined ideas originating from some
field into the body of knowledge concerning other fields may create new possibilities or solve some
problems – if the scholars working in the ‘object field’ are open-minded enough. Moreover, some pieces
of knowledge, such as the criteria for scientific information and the phases of research processes, may be
‘self-evident’ issues for many researchers and data analysts, but the listing of these principles is useful
in any case, since it helps to anticipate and avoid the most crucial pitfalls.
The paper contains some key points concerning the research process when it is based on massive
secondary data sources. The secondary nature and massive size of data sets highlight the importance of
scientific issues as compared to the technical questions. Effective research utilising massive secondary
data sources requires broad expertise and active collaboration, since it is waste of time to reinvent things
which are trivial to some other scholars. This can easily lead to the problem that there is no common
language shared by experts in different fields. The ‘slaves of ideologies’ think differently and have
different latent assumptions, so it can be difficult to find suitable compromises needed in order to achieve
reasonable results. An even bigger mistake is to forget the possibly major impact that decisions made
during the execution of a study can have on its outcomes. In addition, beliefs concerning the nature of
secondary data are not the same for producers and users, and this generates unknown amounts of ‘noise’
for the results if such issues are not considered carefully. Sophisticated preprocessing – incorporating
knowledge from the non-technical domain in order to scale things down to a size fit for more detailed

statistical analyses – is the most important and time-consuming element in the register-based data
analysis. All in all, it can be stated that ‘data are not collected, but produced; research results are not
findings, but creations’ [17].
There is a great deal of literature concerning the utilisation of register-based data in research (see
e.g. [6,10,16,23,28,35,37]). However, the perspective is usually very closely connected to the domain
field or to some particular problem, which makes it difficult to extract more generally applicable ideas.
Moreover, a purely problem-oriented approach can easily lead to a situation in which the contents of the
data are taken too literally. For example, one pitfall is to think that medical diagnoses are recorded in
administrative registers in just as detailed a manner as the diagnosis classification allows.
There are also two very common erroneous beliefs concerning the utilisation of register data in research.
It has been stated that the formation of research data is often technically easy, and that register data are
‘hard’ data representing the truth and providing evidence-based quantitative information. In spite of the
fact that time-consuming and expensive production of new data is avoided, according to the principles
presented above, the data preprocessing, which is full of ideologically dependent qualitative choices,
is the most fundamental and the most difficult part of register-based research. Data analysis is never
‘easy record linking’. It is at least as important a part of the research as the domain knowledge. All
data analysts should remember that they are not just assistants to ‘real researchers’ but full and equal
members of the research group. Collaboration is the key to obtaining the results. Nobody can be an
expert on everything.
The principles of the research process that were presented above are the basis for the actual utilisation
framework. Since the ideas were somewhat abstract in nature, some suggestions for concrete and
practically useful interpretations were considered more carefully. The event-history framework was
suggested as a well-developed option in outlining the characteristics of dynamic phenomena. A systems
approach to the operationalisation of dynamic phenomenon was described as a helpful tool in generating
a common language for experts from different fields. The structure of event-history data was shown to
have the form of a generalised event sequence and to be a suitable data structure for the handling of
second-hand register data. We drew attention to aspects of censoring which, in connection with event-
history data, implicitly define possible model types for problem solving and provide a glimpse of the
data-generating process, which has an intuitive interpretation in the event-history framework. Moreover,
a wide variety of suitable data structures for traditional models can be produced from the generalised
event sequence as special cases using the time-transformation property presented. In addition, common
preprocessing types were reviewed, some examples of possible preprocessing tasks were given and a
pattern remapping technique was formulated. Finally the ideas were illustrated by means of a practical
example. It was seen that register-based data analysis becomes very complicated and challenging even
in seemingly simple situations.
All in all, this paper is simply a monologue presented by a statistician who is specialised in com-
putational statistics and has a background in the social sciences, research interests in problems related
to health services research, and experience in the production of official statistics, the maintenance of
information systems, the development of statistical software, the teaching of statistics, and consultation
with other researchers. In other words, the issues presented in this article are just an ‘insider’s’ collection
of pieces of preliminary knowledge which are essential to the conduct of scientific research based on
administrative registers.
Acknowledgements
This study is supported in part by research grant from the Academy of Finland.

References
[1] E. Arjas, Survival Models and Martingale Dynamics, Scandinavian Journal of Statistics 16 (1989), 177–225.
[2] R.J. Brachman and T. Anand, The Process of Knowledge Discovery in Databases, in: Advances in Knowledge Discovery
and Data Mining, U.M. Fayyad, G. Piatetsky-Shapiro, P. Smyth and R. Uthurusamy, eds, AAAI Press, Menlo Park, 1996,
pp. 37–58.
[3] J.M. Chambers, Greater or Lesser Statistics: A Choice for Future Research, Statistics and Computing 3 (1993), 182–184.
[4] P. Chapman, J. Clinton, R. Kerber, T. Khabaza, T. Reinartz, C. Shearer and R. Wirth, CRISP-DM 1.0 : Step-by-step data
mining guide, The CRISP-DM consortium, 2000.
[5] D.G. Clayton, The Analysis of Event History Data: A Review of Progress and Outstanding Problems, Statistics in
Medicine 7 (1988), 819–841.
[6] F.A. Connel, P. Diehr and L.G. Hart, The use of large data bases in health care studies, Annual Reviews of Public Health
8 (1987), 51–74.
[7] D.R. Cox and D. Oakes, Analysis of Survival Data, Chapman and Hall, London, 1984.
[8] U. Fayyad, G. Piatetsky-Shapiro and P. Smyth, The KDD Process for Extracting Useful Knowledge from Volumes of
Data, Communications of the ACM 39 (1996), 27–34.
[9] J.H. Friedman, The Role of Statistics in the Data Revolution, Bulletin of the International Statistical Institute, 52nd
Session, Proceedings, Book 1, 1995, pp. 121-124.
[10] M. Gissler, Routinely Collected Registers in Finnish Health Research, in: Statistics, Registries, and Research – Experi-
ences from Finland, J. Alho, ed., Statistics Finland, Keuruu, 1999, pp. 241–254.
[11] H. Goldstein and D.J. Spiegelhalter, League Tables and Their Limitations: Statistical Issues in Comparisons of Institu-
tional Performance (with discussion), Journal of the Royal Statistical Society, series A 159 (1996), 385–443.
[12] J. Han and M. Kamber, Data Mining: Concepts and Techniques, Morgan Kaufmann Publishers, San Francisco, 2001.
[13] D.J. Hand, Data Mining: New Challenges for Statisticians, Social Science Computer Review 18 (2000), 442–449.
[14] R.A. Howard, Dynamic Probabilistic Systems, Volume I: Markov Models, Volume II: Semi-Markov and Decision
Processes, John Wiley & Sons, New York, 1971.
[15] P.J. Huber, Massive Datasets Workshop: Four Years After, Journal of Computational and Graphical Statistics 8 (1999),
635–652
[16] P. Huston and D. Naylor, Health services research: reporting on studies using secondary data sources, Canadian Medical
Association Journal 155 (1996), 1697–1702.
[17] J. Irvine, I. Miles and J. Evans, Introduction: Demystifying Social Statistics, in: Demystifying Social Statistics, J. Irvine,
I. Miles and J. Evans, eds, Pluto Press, London, 1979, pp. 1–7.
[18] J.D. Kalbfleisch and R.L. Prentice, The Statistical Analysis of Failure Time Data, John Wiley & Sons, New York, 1980.
[19] T. Lancaster, The Econometric Analysis of Transition Data, Econometric Society Monographs 17, Cambridge University
Press, Cambridge, 1990.
[20] N. Lavrač, E. Keravnou and B. Zupan, Intelligent data analysis in medicine, in: Encyclopedia of computer science and
technology, (Vol. 42)(Supp. 27), Marcel Dekker, New York, 2000, pp. 113–157.
[21] H. Mannila and H. Toivonen, Levelwise Search and Border of Theories in Knowledge Discovery, Data Mining and
Knowledge Discovery 1 (1997), 241–258.
[22] E.C. Marshall and D.J. Spiegelhalter, Institutional Performance, in: Multilevel Modelling of Health Statistics, A.H.
Leyland and H. Goldstein, eds, Wiley Series in Probability and Statistics, John Wiley & Sons, Chichester, 2001,
pp. 127–142.
[23] C.J. McDonald and S.L. Hui, The analysis of Humongous Databases: Problems and Promises, Statistics in Medicine 10
(1991), 511–520.
[24] I. Miles and J. Irvine, The Critique of Official Statistics, in: Demystifying Social Statistics, J. Irvine, I. Miles and J.
Evans, eds, Pluto Press, London, 1979, pp. 113–129.
[25] E. Rahm and H.H. Do, Data Cleaning: Problems and Current Approaches, Bulletin of the IEEE Computer Society
Technical Committee on Data Engineering 23 (2000), 3–13.
[26] Record Linkage Techniques – 1997, in: Proceedings of an International Workshop and Exposition (March 20–21, 1997,
Arlington, VA), W. Alvey and B. Jamerson, eds, Federal Committee on Statistical Methodology, Office of Management
and Budget, Washington DC, 1997.
[27] P. Rissanen, R. Sund, I. Nordback, T. Rousi and U. Idänpään-Heikkilä, Lonkkamurtuman hoidon vaikuttavuuden rek-
isteriperusteinen mittaaminen ja vertailu (Register-based measurement of effectiveness of surgical treatment of hip
fracture, in Finnish), Aiheita 21/2002, National Research and Development Centre for Welfare and Health (STAKES),
Helsinki, 2002.
[28] L.L. Roos and J.P. Nicol, A Research Registry: Uses, Development, and Accuracy, Journal of Clinical Epidemiology 52
(1999), 39–47.
[29] Semi-Markov Models: Theory and Applications, J. Janssen, ed., Plenum Press, New York and London, 1986.

[30] Y. Shahar, A framework for knowledge-based temporal abstraction, Artificial Intelligence 90 (1997), 79–133.
[31] C.E. Shannon, A mathematical theory of communication, Bell System Technical Journal 28 (1948), 379–423.
[32] P. Starr, The Sociology of Official Statistics, in: The Politics of Numbers, W. Alonso and P. Starr, eds, Russel Sage
Foundation, New York, 1987, pp. 7–56.
[33] R. Sund, Health Care Provider Profiling: Effectiveness of Surgical Treatment of Hip Fracture, Nordic Health Econometric
Workshop, August 22, Helsinki, Finland, 2002.
[34] N.B. Tuma and M.T. Hannan, Social Dynamics: Models and Methods, Academic Press, Orlando, 1984.
[35] B.A. Virnig and M. McBean, Administrative Data for Public Health Surveillance and Planning, Annual Reviews of Public
Health 22 (2001), 213–230.
[36] W.E. Winkler, The State of Record Linkage and Current Research Problems, US Bureau of the Census, Technical Report
4/1999, 1999.
[37] N.P. Wray, C.M. Ashton, D.H. Kuykendall and J.C. Hollingsworth, Using Administrative Databases to Evaluate the
Quality of Medical Care: A Conceptual Framework, Social Science and Medicine 40 (1995), 1707–1715.

II
xxn
^ ÓääÇ -V>ÌÌ>ÕiÀ L
«>À} *À«iÀÌiÃ v Õ`Ì >Ì>

>` ,ÕÌiÞ
iVÌi` ,i}ÃÌiÀ >Ì>

>Ãi v *iÀvÀ>Vi ÃÃiÃÃiÌ
v « À>VÌÕÀi /Ài>ÌiÌ >`
,° -Õ`£] ° ÕÀØÌiÓ] *° ØÌiÎ] -° />iÎ] ° >ÀiÓ] ° iÃB£
£
>Ì> ,iÃi>ÀV >` iÛi«iÌ
iÌÀi vÀ 7iv>Ài >` i>Ì] iÃ] >`
Ó
i>Ì
iÌÀi v ÕÛ> ,i}] ÕÛ>] >`
Î
ÕÕÃ>Ã ,i}> Ã«Ì>] ÕÕÃ>Ã] >`
-Õ>ÀÞ ÌÀ`ÕVÌ QÓ£ÓÎR° /Ã ` v «ÀÃ«iVÌÛi ViVÌ

v VV> `>Ì> LÛÕÃÞ Ài«ÀiÃiÌÃ Ìi
!HaR»~Êa³W / V«>Ài «ÀÃ«iVÌÛi >Õ`Ì `>Ì> >`
ÃiV`>ÀÞ >`ÃÌÀ>ÌÛi Ài}ÃÌiÀ `>Ì> Ìi «À`ÕVÌ /i ÌiÀiÌV> «ÀV«iÃ v Ìi «iÀvÀ «À>VÌV> ºÃÌ>ÌivÌi>ÀÌ» VÃiÃÕÃ vÀ
v «iÀvÀ>Vi >ÃÃiÃÃiÌ vÀ>Ì Ìi V>Ãi v >Vi >ÃÃiÃÃiÌ v i>Ì ÃÞÃÌiÃ >Ûi `>Ì> ÀiµÕÀiiÌÃ vÀ >ÃÃiÃÃ} Ìi ÌÀi>Ì
« vÀ>VÌÕÀi ÌÀi>ÌiÌ] >` Ì VÀÃÃÛ>`>Ìi Ìi µÕ>ÌÞ Lii ÜìÞ ÃÌÕì` Q£] ÓR° *iÀvÀ>Vi >Ã iÌ v « vÀ>VÌÕÀiÃ° 1vÀÌÕ>ÌiÞ] Ìi
v vÀ>Ì° ÃiÃÃiÌ ÀiµÕÀiÃ > ÕÌ`ÃV«>ÀÞ >« Ãi«>À>Ìi `>Ì> «À`ÕVÌ ÀiµÕÀiÃ iÝÌÀ>
a»|[³W ÀÃÌ] > VVi«ÌÕ> ì vÀ Ìi «iÀvÀ «À>V >` Ã VÃiÞ ViVÌi` Ì ÃÃÕiÃ ÜÀ >` ÀiÃÕÀViÃ] >` ÌiÀivÀi Ì Ã Õ
>Vi >ÃÃiÃÃiÌ v « vÀ>VÌÕÀi ÌÀi>ÌiÌ Ü>Ã ìvi`° ÃÕV >Ã µÕ>ÌÞ v i>Ì V>Ài QÎ] {R] ÃÞÃÌiÃ iÞ Ì>Ì Ìi iÝÌiÃÛi ÛÕÌ>ÀÞ `>Ì> V
/Ã ì Ü>Ã Ìi ÕÌâi` V«>ÀÃÃ LiÌÜii >>ÞÃÃ QxR] Üi`}i >>}iiÌ QÈR] iVÌ ÀiµÕÀi` LÞ - ÌÞ«i >Õ`Ì `>Ì>
Ìi «ÀÃ«iVÌÛi >Õ`Ì `>Ì> VViÀ} £äÈ VÃiVÕÌÛi >` ìVÃ ÃÕ««ÀÌ ÃÞÃÌiÃ QÇR°
ÌÕ ÜÕ` Li vi>ÃLi > Ã«Ì>Ã° Ì ÜÕ`
« vÀ>VÌÕÀi «>ÌiÌÃ vÀ Ìi ÕÕÃ>Ã ,i}> ÕÃ «ÀÛiiÌ ÌiVµÕiÃ >Ài ÀiµÕÀi` Li «À>VÌV> >` VÃÌivviVÌÛi v ÀÕÌiÞ
Ã«Ì> >` VÀÀiÃ«`} Ài}ÃÌiÀ `>Ì> vÀ Ìi Ìi ìÃ} >` >>}iiÌ v vÀ>Ì ViVÌi` >`ÃÌÀ>ÌÛi `>Ì> VÕ` Li ÕÃi`
Ã i>Ì
>Ài ,i}ÃÌiÀ >` Ìi
>ÕÃiÃ v i>Ì
ÃÞÃÌiÃ ÃÕÌ>Li vÀ «iÀvÀ>Vi >ÃÃiÃÃ vÀ «iÀvÀ>Vi >ÃÃiÃÃiÌ «ÕÀ«ÃiÃ QÓ{R°
,i}ÃÌiÀ° 7i iÝ>i` Ìi V«iÌiiÃÃ v Ài}ÃÌÀ>Ì
v «>ÌiÌÃ >` >Ã Ìi >VVÕÀ>VÞ >` ì}Àii v V
iÌ Qn] R° /iÀi iÝÃÌ ÃiÛiÀ> iÌ` v>VÌ] >` >`ÃÌÀ>ÌÛi Ài}ÃÌiÀÃ
«iÌiiÃÃ v Ìi Ài}ÃÌiÀi` `>Ì>° "LÃiÀÛi` `vviÀiViÃ }V> V>i}iÃ\ V«V>Ìi` ÌiÀ>VÌÃ >Ûi Lii ÕÌâi` «iÀvÀ>Vi >ÃÃiÃÃ
ÜiÀi ViVi` >}>ÃÌ Ìi i`V> ÀiVÀ`Ã° LiÌÜii Ìi «ÞÃ}V> ÃiÌ v `Ãi>Ãi] iÌ Ìi V>Ãi v « vÀ>VÌÕÀiÃ QÓxÓÇR°
+a³Ä»³W ,i}ÃÌiÀ `>Ì> >V VV> ìÌ>] LÕÌ ÕÌ«iÀ i>Ì ÃÞÃÌi] ÌÀi>ÌiÌ ìVÃÃ] Õ> «ÀV«i] Ìi Ã i>Ì Ài}ÃÌiÀÃ
vÀ «ÀÃ«iVÌÛi `>Ì> Ìi ÀiVÀ`} v «>ÌiÌ V>Ài Li>ÛÀ] ÃViÌÞ >` iÝ«i`ÌÕÀiÃ ii` Ì vviÀ > ÛiÀÞ >ÌÌÀ>VÌÛi >` viÝLi iÛÀ
ÃÌÀÞ°
«iÌiiÃÃ v Ìi Ài}ÃÌiÀ `>Ì> Ã ÛiÀÞ }`° Li Ìi}À>Ìi` ÕìÀ Ìi Ã>i ÌiÀiÌV> iÌ vÀ ÀiÃi>ÀV «ÕÀ«ÃiÃ] LiV>ÕÃi Ìi
/i >VVÕÀ>VÞ v i>ÃÞ i>ÃÕÀ>Li Û>À>LiÃ Ìi Ài} ì Q£ä] ££R° «À>VÌVi] Ìi ÃÌ ÕÃivÕ ÕÛiÀÃ> «iÀÃ> ìÌvV>Ì ÕLiÀÃ
ÃÌiÀ Ã >Ì i>ÃÌ x¯° /i >}ÀiiiÌ LiÌÜii Ài}ÃÌiÀ >««À>V Ì «iÀvÀ>Vi >ÃÃiÃÃiÌ ÃiiÃ >Ài ÕÃi` > Ài}ÃÌiÀÃ QÓnR° iÌiÀÃÌV
>` >Õ`Ì `>Ì> Ü>Ã nÈ°Î¯ vÀ ìÌ>i` « vÀ>VÌÕÀi ` Ì Li > `Ãi>ÃiL>Ãi` V«>ÀÃ v i« ÀiVÀ` >}i V> Li ÕÃi` >` Ìi «ÌiÌ>
>}ÃÃ° *ÞÃiÀ> VÀÀi>Ì LiÌÜii Ìi vÕVÌ> ÃìÃ v V>Ài Q£Ó] £ÎR° ,iiÛ>Ì «iÀvÀ `vvVÕÌiÃ ÜÌ V«iÝ «ÀL>LÃÌV
ì«iìVÞ Û>À>LiÃ Ü>Ã ä°Èn° >Vi VÀÌiÀ> V> Li iÝÌÀ>VÌi` vÀ Ìi >}iÃ >Ài ÌÕÃ >Ûì` QÓÎ£R° Ã Ìi >V
RÄ³~³W ,i}ÃÌiÀ >` >Õ`Ì `>Ì> >Ûi ViÀÌ>

Ì>ÌÃ >` «ÀLiÃ] LÕÌ LÌ Ãii Ì Li ÃÕÌ>Li
iÛìViL>Ãi` ÌÀi>ÌiÌ }ÕìiÃ Q£{R° VÕÀ>VÞ v Ìi ÃÌ «ÀÌ>Ì Û>À>LiÃ Ìi
vÀ Ìi «iÀvÀ>Vi >ÃÃiÃÃiÌ v « vÀ>VÌÕÀi ÌÀi>Ì /Þ«V>Þ V«>ÀÃÃ >Ài >ì LiÌÜii Ã Ài}ÃÌiÀÃ Ã Ü Ì Li }`
iÌ° / «ÀÛi Ìi vi>ÃLÌÞ v Ìi >`ÃÌÀ>ÌÛi Ìi «ÀÛìÀÃ v V>Ài Q£xR] ÌÕ} Ì ÀiµÕÀiÃ QÎÓÎxR° 1vÀÌÕ>ÌiÞ] Ìi >Û>>Li iÛ
Ài}ÃÌiÀ `>Ì>] Ìi ÛÕÌ>ÀÞ «ÕÌ v >``Ì> « Ìi >`ÕÃÌiÌ v Ü ÀÃ v>VÌÀÃ Ì>Ì ìVi VViÀ} Ìi Û>`ÌÞ v Ài}ÃÌiÀ
vÀ>VÌÕÀi iÛiÌ `>Ì> Ì Ìi Ài}ÃÌiÀ ÃÕ` Li >ì «Ã >Þ `ÃÌÕÀL Ìi V«>ÀÃÃ Q£È£nR° `>Ì> V> Ì Li `ÀiVÌÞ }iiÀ>âi` Ì Ìi
ÃLi° -Ì>`>À`âi` ÃÌÀÕVÌÃ Ì>Ì }Õì Ìi ÕÃi v *À>VÌV> >` Ài>Li vÀ>Ì iÛ V>Ãi v « vÀ>VÌÕÀi° "i ÃÌÕ`Þ v >VÕÌi «i
>Û>>Li Ài}ÃÌiÀ V>ÃÃvV>ÌÃ > ÃiÃLi Ü>Þ ìViL>Ãi` «iÀvÀ>Vi >ÃÃiÃÃiÌ Ìi ÛV vÀ>VÌÕÀiÃ >` `ÕÀ} £nn >Ã
ÜÕ` «ÀÛi Ìi µÕ>ÌÞ v `>Ì>° V>Ãi v « vÀ>VÌÕÀi ÌÀi>ÌiÌ Ã LiV} ÛiÃÌ}>Ìi` Ìi VÀÀiÃ«ìVi LiÌÜii Ìi
VÀi>Ã}Þ iViÃÃ>ÀÞ Q£R° ` iÝ«iÀ Ài}ÃÌiÀ `>Ì> >` Ìi i`V> ÀiVÀ`Ã v ££{
iÞÜÀ`Ã
6>`ÌÞ] V«iÌiiÃÃ] >VVÕÀ>VÞ] vÀ>Ì ÃÞÃÌiÃ]
iViÃ v Ìi -Üi`Ã ,ÃvÌÀi}ÃÌÀÞ >Ûi «>ÌiÌÃ £ä¯ v £Ó£Ó «>ÌiÌÃ® QÎÈR° /i
VVi«ÌÕ> ì i` Ì Ìi vÀ>â>Ì v `>Ì> «À`ÕVÌ >VVÕÀ>VÞ v Ìi `>Ì> v Ìi Ài}ÃÌiÀ Ü>Ã
vÀ Ìi ÃÌ>`>À`âi` >Õ`Ì v « vÀ>VÌÕÀiÃ vÕ` Ì Li >Ì i>ÃÌ x¯ vÀ ÃÌ «ÀÌ>Ì
iÌ`Ã v i` ÓääÇÆ {È\ xxnqxÈÈ ÕÀ«i - ® QÓäR° >Ì> V«Þ} Û>À>LiÃ iÝVÕ`} ÃiV`>ÀÞ `>}ÃÃ
`\£ä°££ÈäÉ äÎnÓ ÜÌ Ìi - ÃÌ>`>À`Ã >Ûi >Ã Lii nä¯® >` «>Vi v ÕÀÞ Çx¯®° >ÌiÀ
«À`ÕVi` Ãi Ã«Ì>Ã >` ÃÌÕ`Þ ÛiÃÌ}>Ì} « vÀ>VÌÕÀi VìVi
iÌ`Ã v i` xÉÓääÇ ,iViÛi`\ ÛiLiÀ ££] ÓääxÆ >VVi«Ìi`\ -i«ÌiLiÀ n] ÓääÈ

xx
*iÀvÀ>Vi ÃÃiÃÃiÌ v « À>VÌÕÀi /Ài>ÌiÌ
i i>Ì V>Ài `ÃÌÀVÌ vÀ £nÓÎ >` / >i >Þ ÃÌ>ÌiiÌÃ VViÀ} Ìi >Ì>âi` i>ÃÕÀiÃ] >` Ìi V«i
vÀ £ÓÎ >`] Ìi `>Ì> « vÀ>V µÕ>ÌÞ v ÃiV`>ÀÞ® `>Ì>] Ì Ã iÃÃiÌ> iÌi` ÕÃ} ÌiÀiÌV> ìÃ v «iÀ
ÌÕÀi «>ÌiÌÃ r ÈÈn® ìÌvi` vÀ Ã Ì vÀÃÌ ÕÌi Ìi «À«iÀÌiÃ v ÀiµÕÀi` vÀ>Vi >ÃÃiÃÃiÌ Q£] £n] {x] {ÈR° >V
«Ì> viÃ Ü>Ã VÀÀiVÌi` ÕÃ} vÀ>Ì i>ÃÕÀiÃ V>ÀivÕÞ ÕìÀ > vÝi` VVi«ÌÕ> «ii Ài«ÀiÃiÌi` LÞ > VVi«Ì Ü>Ã
vÀ i`V> ÀiVÀ`Ã QÎÇR° /i Ã«Ì> `Ã ì ìÌiÀi Ü `>Ì> ÀiÃÕÌ vÀ Ìi >ÃÃÕi` Ì Li >ÌÌÀLÕÌ>Li Ì > `Û`Õ>
V>À}i Ài}ÃÌiÀ ÃÃi` È°È¯ v «>ÌiÌÃ ÌiÀÞ®] >` Ìi iÛ>Õ>Ìi Ìi LÃiÀÛi` `>Ì> > VÌÕÕÃ Ìi ÃV>i° Ì Ü>Ã >Ã >Ã
£nÓÎ >` Ó°È¯ £ÓÎ] >` > } «iÀ >}>ÃÌ ÌiÃi ÀiµÕÀiiÌÃ iÝ>i Ü ÃÕi` Ì>Ì i>V VVi«Ì V> Li ìÃVÀLi`
ViÌ>}i v v>Ãi À ÕÌ«i `>Ì> ÜiÀi vÕ` Üi Ìi ÌiÀÞ V> Li ÀiVÃÌÀÕVÌi` ÕÃ} ÕÃ} > ÃÞÃÌiÃ >««À>V ÜÌ > Ìi` À
{£°n¯ £nÓÎ >` Ó£°Ç¯ £ÓÎ®° Ìi «À«iÀÌiÃ v >VÌÕ> LÃiÀÛi` `>Ì>®° ÕÌi`® ÕLiÀ v ÃÌ>ÌiÃ Ã Ì>Ì Ìi ÃÞÃ
iÃìÃ ÌiÃi] Üi ` Ì Ü v >Þ ÌiÀ ÌiÀ ÜÀ`Ã] Ì >>Þâi Ìi µÕ>ÌÞ v >Û> Ìi Ã >Ü>ÞÃ Ãi v ÌiÃi ÃÌ>ÌiÃ >Ì i>V
Ã ÃÌÕìÃ Ì>Ì >>Þâi Ìi >VVÕÀ>VÞ À >Li `>Ì> vÀ «iÀvÀ>Vi >ÃÃiÃÃiÌ > Ìi «Ì° /i VVi«ÌÃ ÜiÀi Ìi V>ÃÃ
V«iÌiiÃÃ v Ìi >Ì> Ã«Ì> `Ã «À>}>ÌV>Þ ÕÃivÕ V«ÀÃi LiÌÜii vi` Ì Ài }iiÀ> }ÀÕ«Ã ÌiÀÃ v
V>À}i Ài}ÃÌiÀ Ài}>À` Ì vÀ>VÌÕÀi «>ÌiÌÃ° «ÀLiÀiÌi` >` `>Ì>`ÀÛi ÌiÀiÃ ÌiÀ VÌiÝÌÕ> ÌiÀ«ÀiÌ>Ì] Ìi«À>
À > }iiÀ> «iÀÃ«iVÌÛi] >VVÕÀ>VÞ Ã ÀiµÕÀi`° ÌÃ ÃiÃi] Ìi ìvÌ vÀ > ÃÌ>LÌÞ >` ÌiÀiÌV> i>ÃÕÀiiÌ «À«
>` V«iÌiiÃÃ V> Ì Li ìvi` }ì ÃÌ>`>À` v i>ÃÕÀiiÌ LiViÃ iÀÌiÃ° /Ã µÕ>Ì>ÌÛi V>ÃÃvV>Ì «À
ÕµÕiÞ ÜÌÕÌ > >ÃÃÕ«Ì v > }ì iÀiiÕÌV\ Ìi iÞ ÃÃÕi Ã Ì ÕìÀÃÌ>` Vi`ÕÀi Ü>Ã Ài«i>Ìi` ÕÌ Ìi ÀiÃÕÌ} V
ÃÌ>`>À` v i>ÃÕÀiiÌ] >` `>Ì> Û>`ÌÞ ÜÞ Ìi `>Ì> >Ài i ÌiÞ >Ài ÜÌÕÌ vÝ} Vi«ÌÕ> ì Ü>Ã VÃìÀi` Ã>ÌÃv>VÌÀÞ
ÕÃÌ >Ã Li Õ`}i` >}>ÃÌ Ìi Ìiì` Ìi Ài>ÌÞ ÕÃ} >Þ Ã}i `>Ì> ÃÕÀVi° Ìi ÃiÃi v "VV>¿Ã À>âÀ «ÀV«i°
ÕÌâ>Ì «ÕÀ«ÃiÃ QÎn] ÎR° ÃÌÕ`Þ} ÜiÛiÀ] Ì V> Li >ÃÃÕi` Ì>Ì «À>ÀÞ Õ`Ì `>Ì> ÜiÀi ViVÌi` «ÀÃ«iVÌÛiÞ
VìVi] Ìi ìvÌ v > }ì ÃÌ> `>Ì> ViVÌi` vÀ « vÀ>VÌÕÀi >Õ`Ì «ÕÀ vÀ £äÈ VÃiVÕÌÛi « vÀ>VÌÕÀi «>ÌiÌÃ
`>À` Ã ÃÌÀ>}ÌvÀÜ>À`] LÕÌ ÌÃ Ã Ì Ìi «ÃiÃ VÀÀiÃ«` Ài VÃiÞ Ì >Þ Ài> Ìi ÕÕÃ>Ã ,i}> Ã«Ì> Li
V>Ãi «iÀvÀ>Vi >ÃÃiÃÃiÌ° ÀÌÕ Ã>Li }ì ÃÌ>`>À` Ì> ÃiV`>ÀÞ ÌÜii >Õ>ÀÞ £] £ >` >Õ>ÀÞ Î£] Óäää
>ÌiÞ] Ìi `>Ì> ÀiµÕÀiiÌÃ vÀ VìVi Ài}ÃÌiÀ `>Ì> Ì>Ì ÜiÀi À}>Þ ViVÌi` QÓÓ] ÓÎR° *>ÌiÌÃ ÌÀi>Ìi` Ìi ÃÕÀ}V>
V>VÕ>Ì >` vÀ ìÌiÀ} Ìi ÃÌ>ÀÌ} vÀ ÌiÀ «ÕÀ«ÃiÃ° Ü>À` >Ì Ìi ÕÕÃ>Ã ,i}> Ã«Ì>
«ÌÃ v V>Ài i«ÃìÃ «iÀvÀ>Vi >Ã `ÕÀ} Ìi Ã>i «iÀ` ÜiÀi ìÌvi` vÀ
ÃiÃÃiÌ >Ài ÛiÀÞ Ã>À Q{äR° /Ã ÀiÃi Ìi Ã i>Ì
>Ài ,i}ÃÌiÀ] >` >
L>Vi LiÌÜii ÌÜ >««À>ViÃ «VÌÞ ÀiVÀ`Ã v ÌiÃi «>ÌiÌÃ vÀ £nÇÓääÓ
ÃÜÃ Ì>Ì >ÃÃÕ«ÌÃ v > }ì ÃÌ>`>À` "LiVÌÛiÃ ÜiÀi iÝÌÀ>VÌi` vÀ Ìi Ã Ã«Ì>
>Ài ÛiÀÞ Ã«ÃÌV VìVi ÃÌÕìÃ\ ÃV>À}i ,i}ÃÌiÀ] Ìi Ã i>Ì
>Ài
ViÀÌ> LÃiÀÛ>Li v>VÌÃ >Ài i>ÃÕÀi` >Ì > /i >Ã v ÌÃ ÃÌÕ`Þ ÜiÀi Ì ìvi > V ,i}ÃÌiÀ] >` Ìi
>ÕÃiÃ v i>Ì ,i}ÃÌiÀ
vÝi` Ìi «Ì° «iÀvÀ>Vi >ÃÃiÃÃ Vi«ÌÕ> ì vÀ «À`ÕV} `>Ì> vÀ «iÀ ÕÃ} Ìi ÕµÕi «iÀÃ> ìÌvV>Ì
iÌ Ìi > ÌiÀiÃÌ Ã Ìi `Þ>V vÀ>Vi >ÃÃiÃÃiÌ Ìi V>Ãi v « vÀ>V ÕLiÀÃ v Ìi «>ÌiÌ ««Õ>Ì° iÌiVÌi`
«ÀViÃÃiÃ v V>Ài] ÜV iÃÃiÌ>Þ i>Ã ÌÕÀi ÌÀi>ÌiÌ] Ì V«>Ài ÌÜ `vviÀiÌ `vviÀiViÃ LiÌÜii `>Ì> ÃÕÀViÃ ÜiÀi
ì>} ÜÌ `vviÀiÌ «iÀÃ«iVÌÛiÃ v ÃÕÀViÃ v `>Ì> q «ÀÃ«iVÌÛi VV> >Õ`Ì vÕÀÌiÀ ViVi` >Õ>Þ vÀ Ìi i`V>
i>ÃÕÀiiÌ V>}} Ìi Q{£R° À `>Ì> ìÃ}i` vÀ ÌÀ} ÌÀi>ÌiÌ ÀiVÀ`Ã v ÌiÃi «>ÌiÌÃ *®° /i iÌV>
ÃÌ>Vi] Ì Ã LÛÕÃ Ì>Ì Ìi ìÌ>i` µÕ>ÌÞ® >` ÃiV`>ÀÞ >`ÃÌÀ>ÌÛi Ài} VÌÌii v Ìi Þi>>Ã Ã«Ì>
`>}ÃÃ v « vÀ>VÌÕÀi Ã Ài «ÀÌ>Ì ÃÌiÀ `>Ì> ÕÃi` >Þ V«} ÃÌ>ÌÃ ÃÌÀVÌ >««ÀÛi` Ìi ÃÌÕ`Þ° *iÀÃÃ Ì
Ìi «iÀ>Ì} À Ì> `ÕÀ} Ìi v> ÌVÃ® q Ìi «À`ÕVÌ v «iÀvÀ>Vi ÕÃi Ìi `>Ì> Ü>Ã LÌ>i` vÀ Ìi ÕÕÃ>
ÃÌ>}iÃ v Ài>LÌ>Ì] >` > >ÃÃÕ«Ì >ÃÃiÃÃiÌ vÀ>Ì] >` >Ã Ì VÀÃÃ Ã ,i}> Ã«Ì>] Ìi >Ì> ,i
v VÃÌ>Ì >VVÕÀ>VÞ Ã «À>VÌV>Þ ÕÀi> Û>`>Ìi Ìi µÕ>ÌÞ v ÌiÃi `>Ì> ÃÕÀViÃ Ãi>ÀV >` iÛi«iÌ
iÌÀi vÀ 7iv>Ài
ÃÌV° ÀiÛiÀ] Ìi >VÌÕ> `>}ÃÃ v « >` i>Ì] >` >Ã vÀ -Ì>ÌÃÌVÃ >`°
vÀ>VÌÕÀi Ã Þ > ÌiÀ«ÀiÌ>Ì iÛi >vÌiÀ /i ÃiiVÌi` i>ÃÕÀiÃ v `vviÀiÌ `>Ì>
Ãii} ÝÀ>ÞÃ] ÜV ÃÕ}}iÃÌÃ Ì>Ì Ìi >Ã ÃÕÀViÃ ÜiÀi vÀÃÌ >ÌVi` Ì Ìi ìvi`
ÃÕ«Ì v > }ì ÃÌ>`>À` >Þ Li iÀ iÌ`Ã VVi«ÌÕ> ì° /i V«iÌiiÃÃ v Ài}
ÀiÕÃ Q{ÓR° /Ã «ÌiÌ> >L}ÕÌÞ Ã ÃÌÀ>Ì v « vÀ>VÌÕÀi «>ÌiÌÃ >Ã Üi >Ã
Vi>ÀiÀ ÌiÀ `>}ÃiÃ ÃÕV >Ã ìiÌ> À Ìi ìvÌ v Ìi VVi«ÌÕ> ì] Ìi >VVÕÀ>VÞ >` ì}Àii v V«iÌiiÃÃ v
À ÃVâ«Ài>] >` iÛÌ>Li Ìi V>ÃiÃ Ìi ÃÌ v - Û>À>LiÃ Ü>Ã VÃi Ì Ìi Ài}ÃÌiÀi` `>Ì> ÜiÀi iÝ>i` LÌ
v Ài V«iÝ VVi«ÌÃ ÃÕV >Ã «ÃÌ Ài«ÀiÃiÌ > >ìµÕ>Ìi L>Ãi vÀ `>Ì> ÀiµÕÀi Ìi >Õ`Ì >` Ài}ÃÌiÀ `>Ì> ÕÃ} ÃÌ>`>À`
«iÀ>ÌÛi V«V>Ì À i>Ì ÃÌ>ÌÕÃ Q{ÎR° iÌÃ ÌÌ«\ÉÉÜÜÜ°Ã>vi°ÀÌ°Õ°ÃiÉ}Õì° iÌ`}Þ QÎnR° À Ãi Û>À>LiÃ Ì Ü>Ã
/iÃi «ÀLiÃ LiVi iÛi Ài V Ì®° /iÃi >Ài >VÌÕ> «iÀ>Ì>â>ÌÃ Ài Ài>Ã>Li Ì i>ÃÕÀi >}ÀiiiÌ Ì>
VÀiÌi >` VÀÌV>] v ÃiV`>ÀÞ `>Ì> q v ÌÃi «>ÌiÌ iÛi VVi«ÌÃ VÃìÀi` >VVÕÀ>VÞ° /Ã ÀiµÕÀi` >««À>ViÃ L>Ãi`
ÌiÀ ÜÀ`Ã `>Ì> ViVÌi` À}>Þ vÀ «ÀÌ>Ì Ìi V>Ãi v « vÀ>VÌÕÀi° ÌÃ ì} Ìi ìVÃ>} «ÀViÃÃ]
Ãi ÌiÀ «ÕÀ«Ãi q >Ài Ì Li ÕÃi` QÓ{] ÃÌÕ`Þ] Ìi ìÃÃ v Ìi VVi«ÌÕ> ÃÕV >Ã > «ÞVÀV VÀÀi>Ì ì >
ÓÈ] {{R° ì ÜiÀi >LÃÌÀ>VÌi` vÀ Ìi >VÌÕ> «iÀ vÀ v >ÌiÌ ÌÀ>Ì ì® Q{ÇR° /i «À«À
iÌ`Ã v i` xÉÓääÇ

xÈä
-Õ` iÌ >°
Ì v >}ÀiiiÌÃ Ì>Ì ÜiÀi Ì `Õi Ì ÃÌÕ`Þ «iÀ` ÜÌ > « vÀ>VÌÕÀi `>}ÃÃ° Ì `>Ì> ÃÌÀÕVÌÕÀi v Ìi Ã i>Ì
>Ài
V>Vi Ü>Ã i>ÃÕÀi` ÕÃ} > ìÌ> Viv ÌÕÀi` ÕÌ Ì>Ì ÌiÀi ÜiÀi ÌÌ> £äx «> ,i}ÃÌiÀ >ÜÃ Þ i «iÀ>Ì `>Þ Ì Li
vViÌ] ÜV Ã L>Ãi` Ìi ì v ÌiÌÃ ÜÌ > VvÀi` « vÀ>VÌÕÀi `>} ÀiVÀì` `ÕÀ} i «iÀ`] >` ÀiVÀ`}
ÕÌ«iVVi ÌiÃÌÃ >` >Û`Ã ViÀÌ> ÃÃ />Li Ó®° /i >Õ`Ì `>Ì> ÃÃi` i v ÌÜ Ã>À «iÀ>Ì VìÃ vÀ Ìi Ã>i
Ü ìvViViÃ v Ìi >««> VivvViÌ V>Ãi >` Ìi Ài}ÃÌiÀ `>Ì> ÌÜ V>ÃiÃ `V>Ì «iÀ` >Þ Ãii Ì Li > iÀÀÀ° LÌ
Q{nR° >««> >Þ Li VÃìÀi` >Ã > >« } ÛiÀÞ }` V«iÌiiÃÃ° /i «>ÌiÌ V>ÃiÃ] Ìi `>Ì> Ìi Ài}ÃÌiÀ VÀÀiÃ«ì`
«ÀÝ>Ì v ìÌ>] >Ã LÌ Þi` Û>ÕiÃ ÃÃ Ì Ìi >ÌÌiÀ «iÀ>Ì°
Ì>Ì >Ài ÛiÀÞ Ã>À Üi >Ì i>ÃÌ i >À } vÀ Ìi >Õ`Ì `>Ì> >` > «>VÌi` « /Àii «>ÌiÌÃ ÜÌ > iÜ « vÀ>VÌÕÀi
}> `ÃÌÀLÕÌ Ã L>>Vi` À Üi LÌ vÀ>VÌÕÀi viÀ> iV® >` Ü>Ã iÀÀiÕÃÞ `ÕÀ} Ìi vÀÃÌ Þi>À v vÜÕ« ÜiÀi ì
>À}> `ÃÌÀLÕÌÃ >Ài ìÀ>ÌiÞ Õ iÝVÕì` vÀ Ìi «ÀÃ«iVÌÛi ÃÌÕ`Þ° Ì ÌiVÌi` vÀ Ìi Ài}ÃÌiÀ° /Ü ÜiÀi >Ã
L>>Vi` Ìi Ã>i `ÀiVÌ Q{R° V>ÃiÃ ÃÃ} vÀ Ìi Ài}ÃÌiÀ `>Ì> >` > vÕ` vÀ Ìi >Õ`Ì `>Ì> >` Ìi ÌÀ` Ü>Ã
`>}ÃÃ v `ÃÌ> viÕÀ vÀ>VÌÕÀi ÀiVÀì` VvÀi` ÕÃ} Ìi vÀ>Ì Ìi
Ìi Ài}ÃÌiÀ° /Àii iÝÌÀ> « vÀ>VÌÕÀi «> i`V> ÀiVÀ`° /i VvÀi`® Ài}ÃÌiÀ
ÌiÌÃ ÜÌ Vi>ÀÞ v>Ãi « vÀ>VÌÕÀi `>} `>Ì> >Ã ÀiÛi>i` Ì>Ì i v ÌiÃi ÌÀii
,iÃÕÌÃ ÃiÃ ÜiÀi vÕ` Ìi Ài}ÃÌiÀ° À Ìi vÀÃÌ «>ÌiÌÃ >` > ÌÀ` «iÀ>Ì LiV>ÕÃi v
«>ÌiÌ] ÌiÀi Ü>Ã > V`} iÀÀÀ Ìi i` «ÀLiÃ ÜÌ Ìi « «ÀÃÌiÃÃ `ÕÀ} Ìi
/i ìÃÃ v Ìi VVi«ÌÕ> ì V> ÀiVÀ` `>}ÃÃ -{Ó°ä - -ÇÓ°ä® >` vÀÃÌ Þi>À v vÜÕ«°
>` ÃiiVÌi` i>ÃÕÀiÃ >ÌÌÀLÕÌ>Li Ì ÌiÃi Ìi ÃiV` «>ÌiÌ >` > `ÃÌ> viÕÀ vÀ>V
ìÃÃ Û>ÀÕÃ `>Ì> ÃÕÀViÃ >Ài Ài ÌÕÀi] LÕÌ vÀ Ìi ÌÀ` «>ÌiÌ >Þ iÝ«>>Ì
«ÀÌi` />Li £° "v Ìi ìÃÃ] Ìi vÀ > LÛÕÃ iÀÀÀ Ü>Ã Ì vÕ`° "i «> VVÕÀ>VÞ >`
«iÌiiÃÃ
L}V> VÃÌ>ÌÃ ` Ì V>}i Ìi ÌiÌ ÜÌ « vÀ>VÌÕÀi `>}ÃÃ Ìi Ài}
>` ÌiÀivÀi i i>ÃÕÀiiÌ Ã }iiÀ> ÃÌiÀ >` > «iÀ«ÀÃÌiÌV « vÀ>VÌÕÀi >`
v >ÃÞ i>ÃÕÀ>Li 6>À>LiÃ
â>Li Ì > ÌiÃ° }V> iÛiÌÃ] >VV ÌiÀivÀi ÌÃ «>ÌiÌ Ü>Ã Ì > « vÀ>VÌÕÀi /i V«iÌiiÃÃ >` >VVÕÀ>VÞ v Ìi Ài}
ìÌÉ «>ÌiÌ° ÌiÀ «>ÌiÌ Ü>Ã >`ÌÌi` Li ÃÌiÀ `>Ì> ÜiÀi iÝ>i` ÕÃ} £äÈ «
v> ÃÌÀÞ] >` « vÀ>VÌÕÀi iÛiÌ Ài«ÀiÃiÌ V>ÕÃi v ÃÕÃ«iVÌi` « vÀ>VÌÕÀi] LÕÌ vÀ>V vÀ>VÌÕÀi iÛiÌÃ ÜÌ V«iÌi >Õ`Ì `>Ì>°
ìÃÃ vÀ ÜV >VÌÕ> Û>ÕiÃ v ÌÕÀi Ü>Ã vÕ`° /Àii «>ÌiÌÃ iÝVÕì`
«iÌiiÃÃ `V>ÌiÃ Ìi «iÀViÌ>}i v
i>ÃÕÀiÃ >Ài ÀiVÀì` Ìi «ÀÝÌÞ v vÀ Ìi >Õ`Ì `>Ì> ÜiÀi >Û} }} V>ÃiÃ ÜÌ > ÀiVÀì` Û>Õi vÀ Ìi Û>À>Li
Ãi LÃiÀÛ>Li iÛiÌ° ÌiÀ ì ÌÀi>ÌiÌ vÀ > ÀiViÌ i>ÀiÀ « vÀ>VÌÕÀi q µÕiÃÌ° VVÕÀ>VÞ ÌiÃ Ìi «À«ÀÌ
ÃÃ Ài>Ìi Ì «ii> Ì>Ì «ÌiÌ>Þ i «iÀ>Ìi` >Ì Ìi ÕÕÃ>Ã ,i v VÀÀiVÌÞ ÀiVÀì` Û>ÕiÃ vÀ V>ÃiÃ >Û
V>}i Ìi >` ÃÕ` Li VÌÕÕÃÞ }> Ã«Ì> >` ÌÜ >Ì Ãi ÌiÀ Ã }>Ì i>ÃÌ Ãi ÀiVÀì` Û>ÕiÃ° /i VÀ
ÌÀi`° «À>VÌVi] Ìi ÕÃÌ Li vÝi` «Ì>° Ì >Ã ÌÕÀi` ÕÌ Ì>Ì vÕÀ «>ÌiÌÃ ÀiVÌiÃÃ q Ìi} Ìi ÛiÀ> ÕÌÌÞ
vÀ >VÌÕ> i>ÃÕÀiiÌ° Ìi >Õ`Ì `>Ì>] VÕì` Ìi >Õ`Ì `>Ì> >` >` q >VVÀ`} i>ÃÕÀi` >Ã Ìi «À`ÕVÌ v V«iÌiiÃÃ
Ìi i>ÃÕÀiiÌ Ã ì ViVÌ ÜÌ Ì Ìi Ài}ÃÌiÀ q > i>ÀiÀ « vÀ>VÌÕÀi `ÕÀ} >` >VVÕÀ>VÞ® v Ìi Û>À>Li q Ü>Ã ÛiÀÞ
> « vÀ>VÌÕÀi iÛiÌ >` vÀ ViÀÌ> i>ÃÕÀiÃ Ìi «ÀiVi`} Ìi Þi>ÀÃ° /i Ãì v Ìi vÀ>V }` vÀ ÃÌ v Ìi Û>À>LiÃ />Li Î®°
>Ã ÌÜ ÜiiÃ] vÕÀ ÌÃ] >` i Þi>À ÌÕÀi Ã ÀiVÀì` ÛiÀÞ Ãi` Ìi Ài}ÃÌiÀ] /i >VVÕÀ>VÞ v >`ÃÃ ÃÕÀVi >` `Ã
>vÌiÀ Ìi vÀ>VÌÕÀi° Ìi Ài}ÃÌiÀ `>Ì>] Ìi Ài LÕÌ Ìi i`V> ÀiVÀ`Ã ÀiÛi>i` Ì>Ì Þ V>À}i ìÃÌ>Ì Ü>Ã «ÀÛi` LÞ ÕÃ}
VÀ`} Ì>iÃ «>Vi >Ì i>V `ÃV>À}i°Ã iÝ i v ÌiÃi «>ÌiÌÃ >` Ìi iÜ vÀ>VÌÕÀi ÀiVÀ` >}i ÃÌi>` v Û>À>LiÃ Ìi
«iVÌi`] Ìi VÌiÌ v >Õ`Ì `>Ì> Ã ÀViÀ Ìi Ã>i Ãì >Ã Ìi i>ÀiÀ vÀ>VÌÕÀi° /i ìÝ >`ÃÃ «iÀ`° ÃV>À}i `>Þ
Ì> Ìi Ài}ÃÌiÀ `>Ì>] iÛi >vÌiÀ ÕÃ} «ÃÌÛi >}ÀiiiÌ LiÌÜii Ài}ÃÌiÀ >` >Ã ÀiµÕÀi` ÀiVÀ` >}i] LiV>ÕÃi i>V
`ÀiVÌ i>ÃÕÀiÃ v ViÀÌ> VVi«ÌÃ° Ü >Õ`Ì `>Ì> Ü>Ã ÛiÀÞ }` {°¯®] >` ÌÀ>ÃviÀ vÀ i Ü>À` Ì >ÌiÀ iÛi
iÛiÀ] Ìi i>ÃÕÀiiÌ vÀiµÕiVÞ Ài}ÃÌiÀ ÜÕ` Li iÛi }iÀ v >««À«À>Ìi `>Ì> Ìi Ã>i Ã«Ì> ÀiÃÕÌÃ > iÜ `Ã
`>Ì> Ã ÃÕ«iÀÀ Ì >Õ`Ì `>Ì>] >Ü} V >LÃÌÀ>VÌ ÀÕiÃ vÀ ìÌvV>Ì v v>Ãi V>À}i Ìi Ài}ÃÌiÀ° ÃÃ} «iÀ>Ì
«iÌi LÃiÀÛ>ÌÃ v iÛiÌ ÃÌÀiÃ vÀ «ÃÌÛiÃ >Ài ìvi`° v `>Ì> >LÃÌÀ>VÌ VìÃ >Ài > ÕÃ>Vi Ìi V>Ãi v « vÀ>V
«>ÌiÌ V>Ài° ÀÕiÃ >Ài Ì ÕÃi`] Ìi Ài}ÃÌiÀ `>Ì> Ü ÛiÀ ÌÕÀiÃ] LiV>ÕÃi ÌiÞ >Þ >Ã `V>Ìi V
iÃÌ>Ìi Ìi ÕLiÀ v iÜ « vÀ>VÌÕÀiÃ° ÃiÀÛ>ÌÛi ÌÀi>ÌiÌ° /i `>Ì> ÃÌÀÕVÌÕÀi v
Ìi Ài}ÃÌiÀ ìÃ Ì ÃÕÌ Üi Ì ÀiVÀ`}
>VÕÌi Ài«iÀ>ÌÃ `ÕÀ} Ìi Ã>i Ã«
«iÌiiÃÃ v ,i}ÃÌÀ>Ì

«iÌiiÃÃ v ,i}ÃÌÀ>Ì vÀ Ì> «iÀ`] ÀiÃÕÌ} Ìi ÀiVÀ`} v
/iÀi ÜiÀi £ä{ «>ÌiÌÃ Ìi «ÀÃ«iVÌÛi ÜÀ} VìÃ >` `>ÞÃ vÀ «À>ÀÞ «iÀ
`>Ì> ÌÜ >` > iÜ vÀ>VÌÕÀi >Ì Ìi ÌiÀ Ãì
,i«iÀ>ÌÃ >` iÜ À>VÌÕÀiÃ >Ì° /i iÝÌÀ> «iÀ>Ì Vì `V>Ì}
v Ìi « `ÕÀ} Ìi vÜÕ«®° Ìi Ài} /Ü v Ìi £äx « vÀ>VÌÕÀi «>ÌiÌÃ iiì` Ìi Ãì v vÀ>VÌÕÀi Ü>Ã ÕÃi` ÛiÀÞ À>ÀiÞ°
ÃÌiÀ ÌiÀi ÜiÀi £££ «>ÌiÌÃ Ü ÜiÀi >` > >VÕÌi Ài«iÀ>Ì `ÕÀ} Ìi vÀÃÌ >`Ã VVÕÀ>VÞ vÀ Ìi «>Vi v >VVìÌ Ü>Ã
ÌÌi` Ì Ìi ÃÕÀ}V> «>ÌiÌ Ü>À` >Ì Ìi Ã° iÌiÀ v ÌiÃi Ài«iÀ>ÌÃ Ü>Ã ì «À° /Ã Ü>Ã >Þ LiV>ÕÃi v
ÕÕÃ>Ã ,i}> Ã«Ì> `ÕÀ} Ìi ÌiVÌ>Li vÀ Ìi Ài}ÃÌiÀ `>Ì>° v>VÌ] Ìi V>«>LÌÞ v Ìi
£ä V>ÃÃvV>Ì Ì

xÈ£
/>Li £
Vi«ÌÕ> ì ìÃÃ >` ÌiÀ Ài>â>ÌÃ `vviÀiÌ `>Ì> ÃÕÀViÃ
- i>ÃÕÀi Õ`Ì `>Ì> i>ÃÕÀi ,i}ÃÌiÀ `>Ì> i>ÃÕÀi

}V> v>VÌÃ
}V> iÛiÌÃ `>Ìi v LÀÌ] ì>Ì] Þi>À v i«>ÕÃi] i>ÀVi `>Ìi v LÀÌ >` ì>Ì `>Ìi v LÀÌ >` ì>Ì
}V> VÃÌ>ÌÃ `>® ÃiÝ ÃiÝ ÃiÝ
}V> i>ÃÕÀiÃ
*ÞÃV> iÝ>>ÌÃ i}Ì] Üi}Ì
>LÀ>ÌÀÞ ÌiÃÌÃ i}L] VÀi>Ìi] >LÕ] Li ìÃÌÞ
i}À>«V ÃÌÀÞ
ÕÃi` V«ÃÌ Û} >i >ÀÌ> ÃÌ>ÌÕÃ] Û} >i
*>Vi v Û} ÀiÃìÌ> ÃÌ>ÌÕÃ ÀiÃìÌ> ÃÌ>ÌÕÃ] >Ài> v Û} >Ài> v Û}] aÊa n R:°a
-ViVV ÃÌÀÞ
`ÕV>Ì] VVÕ«>Ì]
iVV ÀiÃÕÀViÃ
i>Ì Ài>Ìi` Li>ÛÀ
,Ã Li>ÛÀ Ã}] >V Ì>i >V ÕÃ>}i :R| °a:»a[ ~:»~a» R:°a
iÌ] >VÌÛÌiÃ
-ÕLiVÌÛi µÕ>ÌÞ v vi
"ÛiÀ> ÜiLi} «>] «ÃÞV}V> ÃÌ>Ìi] vi>À v v> *>
"LiVÌÛi ii` vÀ V>Ài
Ãi>ÃiÃÉÃÞ«ÌÃ VÀL`ÌÞ] V«V>ÌÃ
«V>ÌÃ °aR°[a[ [~:t³a³X Ä³a n R:°a
*ÞÃV> vÕVÌ} Ü>} Ü>}] >VÌÛÌiÃ v `>Þ Û} ii` vÀ V>Ài

}ÌÛi vÕVÌ} LLÀiÛ>Ìi` iÌ> ÌiÃÌ ÃVÀi
1Ãi v V>Ài
/iVV> >` Ü>} >`Ã ÕÃi v >ÃÃÃÌÛi ÌiV}Þ] V>}iÃ Û}
iÛÀiÌ
i`V>Ì ÕÃi v «>iÀÃ
i>Ì V>Ài ÕÌâ>Ì «ÀÛìÀ] ÌÞ«i v ÃÌ>Þ] i}Ì v ÃÌ>Þ ÌÞ«i v V>Ài] Ài>LÌ>Ì] `>ÌiÃ vÀ Ã«Ì> ÛÃÌÃ «ÀÛìÀ] ÌÞ«i v V>Ài] >`ÃÃ >` `ÃV>À}i
`>ÌiÃ] `>}ÃiÃ] «iÀ>ÌÃ
VVìÌÉv> ÃÌÀÞ
*À«iÀÌiÃ v v> `>Ìi v v>] «>Vi v v> `>Ìi v v>] «>Vi v v> :Ra n :[ °a:³ n° |³~»:~Ña[ ~Ä°Í
« vÀ>VÌÕÀi iÛiÌ
*À«iÀÌiÃ v vÀ>VÌÕÀi Ìi v vÀ>VÌÕÀi] VVÕÀÀiVi «>Vi] vÀ>VÌÕÀi ÌÞ«i] `>Þ v vÀ>VÌÕÀi] Ài>Ã vÀ vÀ>VÌÕÀi] VVÕÀÀiVi «>Vi] «>Vi v >` Ài>Ã vÀ Ã«Ì>âi` ÕÀÞ
vÀ>VÌÕÀi Ãì vÀ>VÌÕÀi ÌÞ«i] vÀ>VÌÕÀi Ãì
Ì> ÌÀi>ÌiÌ «ÀViÃÃ Ã«Ì> `] >`ÃÃ Ìi] >`ÃÃ ÃÕÀVi] ÃÌ>ÀÌ >`ÃÃ `>Þ] >`ÃÃ ÃÕÀVi] `>Þ v «iÀ>Ì] Ã«Ì> `] >`ÃÃ `>Þ] >`ÃÃ ÃÕÀVi]
v «iÀ>Ì] Ài>Ã vÀ «iÀ>ÌÛi ì>Þ] ÌÞ«i v Ài>Ã vÀ «iÀ>ÌÛi ì>Þ] «iÀ>ÌÛi iÌ`] `>}ÃiÃ] `>Þ v «iÀ>Ì] «iÀ>Ì VìÃ]
ÃÕÀ}i >` >iÃÌiÌÃÌ] ÌÞ«i v >iÃÌiÌV] V«V>ÌÃ «iÀ>Ì] Ài«iÀ>ÌÃ] `ÃV>À}i `ÃV>À}i `>Þ] `ÃV>À}i ìÃÌ>Ì
«iÀ>ÌÛi iÌ`] i}Ì v ÃÕÀ}iÀÞ] ÌÞ«i v `>Þ] `ÃV>À}i ìÃÌ>Ì
«À«Þ>ÝÃ] V«V>ÌÃ «iÀ>Ì]
Ài«iÀ>ÌÃ] Ìi Ì Lâ>Ì] `ÃV>À}i Ìi]
`ÃV>À}i ìÃÌ>Ì

V> ÃÌ>LÌÞ - }À>ì] VÀL`ÌiÃ] ViÝÃÌ} vÀ>VÌÕÀiÃ] - }À>ì] Ài>Ã vÀ «iÀ>ÌÛi ì>Þ i`V>Þ °aR°[a[ [~:t³a³X Ä³a n R:°a
Ài>Ã vÀ «iÀ>ÌÛi ì>Þ i`V>Þ ÕvÌ «>ÌiÌÃ ÕvÌ «>ÌiÌÃ
£® `ÀiVÌ i>ÃÕÀiiÌ ÀiµÕÀ} ÀiVÀ` >}i

xÈÓ
-Õ` iÌ >°
/>Li Ó n°£¯ x¯

\ Î°{°n¯® v V>ÃiÃ
LÌ `>Ì> ÃiÌÃ\ ÃÃ} vÀ Ài}ÃÌiÀ\
«iÌiiÃÃ v `>Ì> ÃiÌÃ />Li Î®° ÜiÛiÀ] Ìi iÛ>Õ>Ì v Ìi
n iÜ « vÀ>VÌÕÀi «>ÌiÌÃ Ó v>Ãii}>ÌÛi Ài}ÃÌiÀ `>}ÃiÃ
iÝ>VÌ >VVÕÀ>VÞ v Ìi > `>}ÃÃ Ü>Ã
s :»~a»³ Ë~»| a:°~a° n°:R»Ä°a
Ì ÃÌÀ>}ÌvÀÜ>À`] >Ã `vviÀiÌ V>ÃÃv
ÃÃ} vÀ >Õ`Ì `>Ì>\ *ÃÌÛi >}ÀiiiÌ\ V>ÌÃ v « vÀ>VÌÕÀiÃ ÜiÀi ÕÃi` Ìi
£ ÛiÀvi` iÜ «>ÌiÌ {°¯ x¯
\ ä°ÎÈ°n¯® Ài}ÃÌiÀ >` Ìi >Õ`Ì `>Ì>° Ü} Ìi
Â a°:»a[ ~ »|a° |³~»:
"ÛiÀiÃÌ>Ì Ìi Ài}ÃÌiÀ\ iÌ`}V> ÃÕ}}iÃÌÃ] Ìi >VVÕÀ>VÞ
°aRa» a:°~a° n°:R»Ä°a
Ç°Ó¯ x¯
\ Î°Ó£Î°Ç¯® v `>}ÃiÃ Ü>Ã >``ÀiÃÃi` LÞ VÃìÀ}
a:°~a° »»: °³»|a³~³
³Ä³aR»a[ |~ n°:R»Ä°a 1ìÀiÃÌ>Ì Ìi Ài}ÃÌiÀ\
Ìi Li >>ÌÞ >Ã > L}V> iÌÌÞ
Î v>Ãi«ÃÌÛi Ài}ÃÌiÀ `>}ÃiÃ £°¯ x¯
\ ä°ÓÈ°Ç¯® ìÌvÞ} ÌÀÕi vÀ>VÌÕÀi ÃÌ>ÌÕÃ QxäR°
« vÀ>VÌÕÀiÃ >Ài VÞ V>ÃÃvi`
Ì ÌÀ> >` iÝÌÀ>V>«ÃÕ>À >VVÀ`} Ì
ÌiÀ Ài>ÌÃ« Ì Ìi V>«ÃÕ>À >ÌÌ>ViÌ
/>Li Î
6>À>Li
«iÌiiÃÃ£ VVÕÀ>VÞÓ
ÀÀiVÌiÃÃÎ
«iÌiiÃÃ >` >VVÕ v Ìi «° ÌiÀÃ v VÕÀÀiÌ `>Ì>] vÀ>VÌÕÀiÃ
*iÀÃ> ìÌvV>Ì ÕLiÀ{ £ää¯ £ää¯ £ää¯ À>VÞ v i>ÃÞ i>ÃÕÀ>Li v Ìi iV v Ìi viÕÀ >Ài ÌÀ>V>«ÃÕ>À]
Û>À>LiÃ Ìi Ài}ÃÌiÀ >` ÌiÀ « vÀ>VÌÕÀiÃ >Ài iÝÌÀ>V>«ÃÕ>À°
i ÕV«>ÌÞ £ää¯ £ää¯ £ää¯
/i Li >>ÌÞ >Þ Li ÕÃi` Ì ÃÕL`Ûì
`ÃÃ `>Þ £ää¯ £ää¯ £ää¯ iÝÌÀ>V>«ÃÕ>À vÀ>VÌÕÀiÃ vÕÀÌiÀ Ì L>Ã
`ÌÌi` vÀ £ää¯ nÎ°ä¯ nÎ°ä¯ ViÀÛV>] ÌÀV>ÌiÀV >` ÃÕLÌÀV>ÌiÀV
`ÌÌi` vÀx £ää¯ x°Î¯ x°Î¯ vÀ>VÌÕÀiÃ° 1vÀÌÕ>ÌiÞ] ÌiÀi Ã Ãi«>À>Ìi
ÃV>À}i `>Þ £ää¯ £ää¯x £ää¯ `>}ÃÃ Vì
£ä vÀ L>ÃViÀÛV>
vÀ>VÌÕÀi° 1Ã} > ìÌ>i` vÀ>VÌÕÀi V>ÃÃv
ÃV>À}i Ì £ää¯ È°Ó¯x È°Ó¯
x
V>Ì v viÀ> iV vÀ>VÌÕÀi] ÌÀV>ÌiÀV
ÃV>À}i Ì £ää¯ £ää¯ £ää¯ vÀ>VÌÕÀi À ÃÕLÌÀV>ÌiÀV vÀ>VÌÕÀi] Ìi
"«iÀ>Ì `>Þ °£¯ Ç°£¯ È°Ó¯ >}ÀiiiÌ LiÌÜii >Õ`Ì `>Ì> >` Ài}Ã
*À>ÀÞ «iÀ>Ì °£¯ n°£¯ Ç°Ó¯ ÌiÀ Ü>Ã nÈ°Î¯ x¯
\ Ç°{Ó°Ó¯®° *>À
-ì v vÀ>VÌÕÀi Ó°n¯ £ää¯ Ó°n¯ ÜÃi >}ÀiiiÌ >>ÞÃiÃ ÀiÛi>i` Ì>Ì vi
*>Vi v >VVìÌÈ Î°{¯ È£°Î¯ xÇ°Î¯
À> iV >` ÃÕLÌÀV>ÌiÀV vÀ>VÌÕÀiÃ
ÜiÀi VÀÀiVÌÞ ìÌvi` Ìi Ài}ÃÌiÀ
> `>}ÃÃ « vÀ>VÌÕÀi® £ää¯ n°£¯ n°£¯
} ÃiÃÌÛÌÞ®] LÕÌ ìÌvV>Ì v ÌÀ
£® *iÀViÌ>}i v V>ÃiÃ ÜÌ ÀiVÀì` Û>Õi vÀ Ìi Û>À>Li
Ó® *À«ÀÌ v VÀÀiVÌÞ ÀiVÀì` Û>ÕiÃ vÀ V>ÃiÃ ÜÌ Ãi ÀiVÀì` Û>Õi
V>ÌiÀV vÀ>VÌÕÀiÃ Ü>Ã Ì >Ã >VVÕÀ>Ìi
Î® /i ÛiÀ> ÕÌÌÞ v Ìi Û>À>Li «À`ÕVÌ v V«iÌiiÃÃ >` >VVÕÀ>VÞ® />Li {®° /Ã Ã ÃÌ «ÀL>LÞ LiV>ÕÃi Ìi
{®
Ì>Ã `>Ìi v LÀÌ >` ÃiÝ iÝÌÀiÌiÃ >Ài i>ÃiÀ Ì ìÌvÞ VÀÀiVÌÞ°
x® ìÌvV>Ì LÞ ÕÌâ} ÀiVÀ` >}i
È® V«>ÌLi V>ÃÃvV>Ì Ìi Ài}ÃÌiÀ /Ã ÌiÀ«ÀiÌ>Ì Ã ÃÕ««ÀÌi` LÞ Ìi v>VÌ
Ì>Ì Ìi Ài}ÃÌiÀ `>}ÃiÃ v ÌÀV>ÌiÀV
vÀ>VÌÕÀi ÜiÀi «À>VÌV>Þ >Ü>ÞÃ ÌÀÕi ÌÀ
V>ÌiÀV vÀ>VÌÕÀiÃ } «ÃÌÛi «Ài`VÌÛi
/>Li { VVÕÀ>VÞ v « vÀ>VÌÕÀi `>}ÃÃ Ìi Ài}ÃÌiÀ
Û>Õi®° /Ã Ü>Ã Ì Ìi V>Ãi vÀ viÀ>
iV vÀ>VÌÕÀiÃ À ÃÕLÌÀV>ÌiÀV Ài}ÃÌiÀ
-iÃÌÛÌÞ -«iVvVÌÞ *ÃÌÛi i}>ÌÛi
«Ài`VÌÛi Û>Õi «Ài`VÌÛi Û>Õi `>}ÃiÃ ÜV ÜiÀi VV>Ã>Þ ÕÃi`
vÀ ÌÀÕi ÌÀV>ÌiÀV vÀ>VÌÕÀiÃ Ü «ÃÌÛi
iÀ> iV vÀ>VÌÕÀi È°Ç¯ nä°x¯ nn°£¯ {°Î¯
£°x£ää¯® Èn°Ó£°Ç¯® Ç°Çx°Î¯® nx°Ç£ää¯® «Ài`VÌÛi Û>Õi®° ÃV>ÃÃvV>Ì >Ì Ìi
LÀìÀÃ v Ìi ÌÀV>ÌiÀV Ài} v Ìi
/ÀV>ÌiÀV vÀ>VÌÕÀi Èn°È¯ n°x¯ È°ä¯ nx°Ç¯
xÓ°nÎ°Ç¯® x°££ää¯® nÇ°ä£ää¯® ÇÇ°xÎ°£¯® viÕÀ >Þ Li i Ài>Ã° ÌiÀ V>ÕÃi Ã
Ìi
£ä V>ÃÃvV>Ì ÕÌâi`
-ÕLÌÀV>ÌiÀV vÀ>VÌÕÀi nÎ°Î¯ È°¯ ÈÓ°x¯ n°¯
{Ó°£ää¯® Ó°n£ää¯® Óx°ä£ää¯® È°Ç£ää¯® >`] ÜV ÃÕ}}iÃÌÃ Ìi ÕÃi v viÀ> iV
`>}ÃÃ vÀ ÕÃ«iVvi` « vÀ>VÌÕÀiÃ° /i
«À«ÀÌ v >}ÀiiiÌÃ Ì>Ì ÜiÀi Ì `Õi
Ì V>Vi Ü>Ã ÛiÀÞ }` >Ã i>ÃÕÀi`
Ãi«>À>Ìi i >VVìÌÃ vÀ Ìi >VV VVÕÀ>VÞ >` ,i>LÌÞ ÜÌ Ìi ìÌ> VivvViÌ n£°Î¯Æ x¯
\
ìÌÃ VVÕÀÀ} ÀiÃìÌ> V>Ài >` v>Ã Ç£°Î£°Î¯® À Ìi >««> VivvViÌ
VVÕÀÀ} ÕÌ`ÀÃ vÀ Ìi iÃ VVÕÀ v « À>VÌÕÀi >}ÃiÃ ÇÎ°x¯Æ x¯
\ È£°Ónx°Ç¯®° >``Ì]
À} `ÀÃ° ,i}ÃÌiÀ `>}ÃÃ Ü>Ã ÀiVÀì` >Ã « vÀ>V ÃVi Ì Ü>Ã Ài>Ã>Li Ì >ÃÃÕi Ì>Ì Ìi
ÌÕÀi
£ä\ -ÇÓ°ä] -ÇÓ°£] -ÇÓ°Ó® ÕìÀÞ} ÌÀ>Ì Ã VÌÕÕÃ] Ìi «Þ

xÈÎ
VÀV VÀÀi>Ì VivvViÌ Ü>Ã V>VÕ

>Ìi` `V>Ì} > ÛiÀÞ } VÃÃÌiVÞ
LiÌÜii Ìi >Õ`Ì `>Ì> >` Ài}ÃÌiÀ ä°ÓÆ
x¯
\ ä°nÈä°n®°
}ÀiiiÌ LiÌÜii ÕVÌ>

i«iìVÞ i>ÃÕÀiÃ
Ìi >Õ`Ì `>Ì>] vÕVÌ> ì«iìVÞ Ü>Ã
i>ÃÕÀi` ÕÃ} > >VÌÛÌiÃ v `>Þ Û}
® ÌÞ«i v ÃV>i QÓÎR° Ìi Ã
i>Ì
>Ài ,i}ÃÌiÀ] ÛiÀ> ì«iìVÞ
>` ii` vÀ V>Ài® Ã ÀiVÀì` ÕÃ} > ÃV>i
v vÛi V>Ìi}ÀiÃ Qx£R° /i V«iÌiiÃÃ v
ì«iìVÞ Û>À>Li Ìi Ài}ÃÌiÀ Ü>Ã
£ää¯° /i LÝ«ÌÃ ìÃVÀL} ì«iìVÞ
i>ÃÕÀiÃ >Ì `ÃV>À}i® >Ài ÃÜ }
ÕÀi £° /i Õ««iÀ V>Ìi}ÀiÃ v Ìi Ài}ÃÌiÀ
L>Ãi` i>ÃÕÀi >Ài VLi` `Õi Ì Ìi
Ã> ÕLiÀ v LÃiÀÛ>ÌÃ° /i «ÞÃiÀ
> VÀÀi>Ì LiÌÜii Ìi Û>À>LiÃ Ã ä°Èn
x¯
\ ä°xxä°n£®] ÃÜ} À>ÌiÀ }`
VÃÃÌiVÞ° i>Ã >` i`>Ã Li>Ûi
Ài>Ã>LÞ] LÕÌ }ÕÀi £ `V>ÌiÃ Ì>Ì Ìi
ÃÌ VÞ ÕÃi` V>Ìi}ÀÞ Ìi Ài}
ÃÌiÀ `>Ì> >Ã > ÛiÀÞ Üì À>}i° Ì ÃiiÃ
Ì>Ì } Ài}ÃÌiÀ V>Ìi}Àâ>ÌÃ ÌÀÕÞ Ài }° £ ÕVÌ> ì«iìVÞ i>ÃÕÀiÃ Ài}ÃÌiÀ >` >Õ`Ì `>Ì>° Õ`Ì `>Ì> >Ã > >VÌÛÌiÃ v `>Þ Û} ® ÌÞ«i
v ÃÕ ÃV>i QÓÎR° ,i}ÃÌiÀ `>Ì> iÛiÃ >Ài\ >® V«iÌiÞ À >ÃÌ ì«iìÌ] L® VV>Ã> ii` vÀ V>Ài] V® ÀiVÕÀÀiÌ ii`
viVÌ L>` vÕVÌ>ÌÞ] LÕÌ >LÃÕÌi ÌiÀ vÀ V>Ài] `³® >ÃÌ À ÌÌ>Þ® VÌÕÕÃ ii` vÀ V>Ài Qx£R°
ÀÃÃ ÀiviÀÃ Ì i> LÝ«Ì°
«ÀiÌ>Ì v Ìi Ài}ÃÌiÀ ÃV>i >Ì > `Û`Õ
> iÛi >Þ Li iÀÀiÕÃ°
ÃVÕÃÃ ÃÕÌ>iÕÃÞ QxÎxÈR° /Ã Ã «>ÀÌVÕ>ÀÞ

V>i}} v `>Ì> >Ài ÕÃi` vÀ ÌiÀ «ÕÀ
ÜÕ« vÀ>Ì «ÃiÃ Ì> À}>Þ Ìiì` QxÇR° ÌÃ
-Ì>Ìi `>}À>Ã vÀ Ìi Ài}ÃÌiÀ >` >Õ`Ì ÌÃ ÃÌÕ`Þ ÌÜ vÀÃ v `>Ì> «À`ÕVÌ ÃÌÕ`Þ] > «À>}>ÌV >««} LiÌÜii V
`>Ì> ÃÕ>Àâ} Ìi vÜÕ« vÀ> vÀ «iÀvÀ>Vi >ÃÃiÃÃiÌ «ÕÀ«ÃiÃ ÜiÀi Vi«ÌÃ vÀ > «iÀvÀ>Vi >ÃÃiÃÃiÌ Ìi
Ì >Ài ÃÜ }ÕÀi Ó° /i ÛiÀ> V«>Ài` Ìi V>Ãi v « vÀ>VÌÕÀi ÌÀi>Ì ÀÞ] - iÌ>`>Ì>] >` ÃÌÀÕVÌÕÀi` `>Ì>
Ã>«iÃ >Ài «ÀiÌÌÞ Ã>À] LÕÌ Ìi Ài}ÃÌiÀ iÌ° ÕÃÌv>Li V«>ÀÃ ÀiµÕÀi` > iÌÀÞ «ÀÃ«iVÌÛi >Õ`Ì `>Ì> >` ÃiV`
`>Ì> >Ûi Ài >VVÕÀ>VÞ° /i `>ÌiÃ v ìvÌ v Ìi VVi«ÌÕ> ì] ÜV >ÀÞ Ài}ÃÌiÀ `>Ì> Ü>Ã VÃÌÀÕVÌi`° Ã«Ìi
ì>ÌÃ] i}ÌÃ v Ì> Ã«Ì>â>ÌÃ] >Üi` vÀ > ÃÞÃÌi>ÌV ÃÌÀÕVÌÕÀ} v Ã v Ìi }iiÀ>ÌÞ v Ìi VÕÀÀiÌ ì] Ì Ã
>` ÀiÃìÌ> ÃÌ>ÌÕÃiÃ >Ì vÕÀ ÌÃ >` LiÌÜii LÃiÀÛ>Li i>ÃÕÀiÃ >` ÌiÀiÌ Ìi` Ìi ÃiÃi Ì>Ì Ì ì>Ã ÜÌ `
>Ì i Þi>À >vÌiÀ Ìi vÀ>VÌÕÀi Ìi >Õ`Ì `>Ì> V> VVi«ÌÃ° -ÕV ìÃ >Ài Þ À>ÀiÞ Û`Õ> iÛi VVi«ÌÃ Þ° À LiV>À
>Ài >VVÕÀ>Ìi] LÕÌ Ìi vÜÕ« vÀ>Ì Ài«ÀÌi` Q{ÈR] iÛi ÌÕ} Ìi iÝ«VÌ } «ÕÀ«ÃiÃ] «ÀÛìÀ iÛi «ii> Ì>Ì
VViÀ} >VÕÌi V>Ài ÃÕLÃiµÕiÌ Ì Ìi ÌÀ`ÕVÌ v > VVi«ÌÕ> ì LÛÕÃÞ }Ûi iÝ«>>ÌÃ v LÃiÀÛi` `vviÀiViÃ
Ì> Ã«Ì>â>Ì Ìi >Õ`Ì `>Ì> Ã «>ÀÌÞ Ã}vV>ÌÞ «ÀÛiÃ Ìi ÕìÀÃÌ>`} v LiÌÜii «ÀÛìÀÃ v V>Ài ÃÕ` >Ã Li
V«iÌi >` ÌiÀivÀi ÌiVV>Þ `v Ìi «ÀLi >` >iÃ Ìi ÃÌ «ÀÌ>Ì VÃìÀi`° "i ÛiÀÞ «ÀÌ>Ì `Û`Õ>
vVÕÌ Ì >ì° /Ã V> Li ÃÌ Vi>ÀÞ >ÃÃÕ«ÌÃ >vviVÌ} Ìi >VÌÕ> ÀiÃÕÌÃ v iÛi i>ÃÕÀi q >iÞ VÃÌÃ q Ã >Ã
Ãii vÀ Ìi «À«ÀÌ v «>ÌiÌÃ ÕÀÃ >>ÞÃiÃ ÛÃLi° ÃÃ} vÀ Ìi ì° Ì ÜÕ` >Ûi
} iÃ] ÜV Ã vÀÃÌ ÛiÀiÃÌ>Ìi` >` -ÞÃÌi>ÌV >««À>ViÃ V> Li ÕÃi` vÀ Ã«iV> i>ÃÕÀiiÌ «À«iÀÌiÃ] LiV>ÕÃi Ì
>ÌiÀ ÕìÀiÃÌ>Ìi` }° Ó®° Ìi ìÛi«iÌ v > VVi«ÌÕ> ì Ã VÕÕ>ÌÛi Ìi°
QxÓR° +ÕÌi vÌi VVi«ÌÕ> ìÃ >` Ì Ã iÝ«iVÌi`] Ìi Ài}ÃÌiÀ `>Ì> >Vi`
Li Ài«ÀiÃiÌi` Ã Ì>Ì Ì LiViÃ «ÃÃLi VV> ìÌ>° ÜiÛiÀ] ÃÌ v Ìi Ài
Ì VÃìÀ ÃiÛiÀ> `vviÀiÌ «iÀÃ«iVÌÛiÃ ìÌ>i` i>ÃÕÀiÃ Ìi >Õ`Ì `>Ì> ÜiÀi >V

xÈ{
-Õ` iÌ >°
}° Ó -Ì>Ìi `>}À>Ã vÀ Ài}ÃÌiÀ >` >Õ`Ì `>Ì>
ÌÕ>Þ Ài>Ìi` Ì «ii> Ì>Ì VÌÕ ÌÀ>ÃviÀÀi` Ì ÌiÀ Ü V> Ã«Ì>] VÕ` Ã«iVvV ÕLiÀÃ v VÃiVÕÌÛi « vÀ>V
ÕÃÞ V>}i Ìi° -ÕV `>Ì> ÀiµÕÀi Ì>VÌ >Ûi Lii ÃÃi` vÀ Ìi `>Ì>° ÜiÛiÀ] ÌÕÀiÃ `ÕÀ} «>ÀÌVÕ>À Ìi «iÀ`Ã q >Û>
Üi`}i Ì Li ÕÃivÕ > Ài }iiÀ> ÕÀ ÃÌÕ`Þ V>«ÌÕÀiÀiV>«ÌÕÀi >>ÞÃiÃ iÃÌ >Li vÀ «ÀÃ«iVÌÛi ÃÌÕìÃ q Ì Ài}ÃÌiÀ
VÌiÝÌ QxnR° " Ìi ÌiÀ >`] Ài}ÃÌiÀ `>Ì> >Ì} Ìi ÌÌ> ÕLiÀ v V>ÃiÃ ÜiÀi Ì `>Ì>L>Ãi` iÃÌ>ÌiÃ vÀ Ìi Ã>i Ã«Ì>Ã
>Ûi > `>Ì> ÃÌÀÕVÌÕÀi Ì>Ì >ÜÃ Ìi V >««V>Li] LiV>ÕÃi ÌiÀi Ü>Ã ÃÌÀÕVÌÕÀ> ì >` Ìi Ã>i Ìi «iÀ`Ã° Ì ÕÃÌ Li Ìi`
«iÌi LÃiÀÛ>Ì v «>ÌiÌ V>Ài ÃÌÀÞ] «iìVÞ LiÌÜii `>Ì> ÃÕÀViÃ Ìi VVÕÀ Ì>Ì iÛi ÌÃ ` v > ìvÌ >Ã Ãi
>` ÌiÀivÀi ÕÌ«iÀvÀÃ «ÀÃ«iVÌÛi `>Ì> ÀiVi v > «>ÌiÌ >Õ`Ì `>Ì> >Ã `V>Ìi` `À>ÜL>VÃ VìVi V>VÕ>ÌÃ° /i
ÌÃ ÃiÃi v Ãi ì>Þ `>Ì> «À`ÕV iÝ«iVÌi` VVÕÀÀiVi Ài}ÃÌiÀ `>Ì>® Qx] ÕLiÀ v «>ÌiÌÃ Ã Ã«Ì>Ã«iVvV] LÕÌ
Ì Ã ÌiÀ>Li° >} ÕÃi v ì ÈäR° /iÀivÀi Ìi V«iÌiiÃÃ Ü>Ã iÛ>Õ Ìi >Û>>Li ÀÃ««Õ>Ì `>Ì> ÀiviÀÃ Ì
}À>«V] ÃViVV >` i`V>Ì >Ìi` LÞ ìÌvÞ} > « vÀ>VÌÕÀi V>ÃiÃ }i}À>«V> >Ài>Ã >` i V> ÌiÀivÀi
ÃÌÀiÃ `>Ì> q ÜV >Ài] «ÀV«i] vÀ >Õ`Ì `>Ì> >` Ài}ÃÌiÀ `>Ì>] >` Ìi >À}Õi Ì>Ì «>ÌiÌÃ Û} Ìi Ã>i >Ài>
>Û>>Li ÌiÀ Ã Ài}ÃÌiÀÃ q VÕ` > V«>ÌLÌiÃ ÜiÀi ViVi` >}>ÃÌ LÕÌ «iÀ>Ìi` Ãi ÌiÀ Ã«Ì> ÃÕ`
>Ã «ÀÛi ViÀÌ> ìÌ>Ã v Ài}ÃÌiÀ Ìi vÀ>Ì i`V> ÀiVÀ`Ã° ÌÃ >Ã Li VÕÌi` >Ã V>ÃiÃ] >` «>ÌiÌÃ vÀ
`>Ì>] LÕÌ Ì ÀiµÕÀiÃ iÝÌiÃÛi ÜÀ Ì LÌ> ÃiÃi] Ìi `vviÀiViÃ V«iÌiiÃÃ Li ÌiÀ >Ài>Ã ÃÕ` Li iÝVÕì`°
>` «Ài«ÀViÃÃ ÃÕV q «ÃÃLÞ iÝ«iÃÛi q ÌÜii `>Ì> ÃÕÀViÃ `V>Ìi Ìi `vviÀiÌ "ÕÀ ìvÌ vÀ > }ì ÃÌ>`>À` v
`>Ì>° ìvÌÃ v « vÀ>VÌÕÀi À iÀÀÀÃ Ài} i>ÃÕÀiiÌ ÌÃ ÃÌÕ`Þ Ü>Ã iÀiiÕÌV°
ÌÃ ÃÌÕ`Þ] Ì Ü>Ã >ÃÃÕi` Ì>Ì > « ÃÌiÀ `>Ì>° «À>VÌVi] ÌÀii ÌÞ«iÃ v V«>ÀÃÃ
vÀ>VÌÕÀi «>ÌiÌÃ ÜiÀi ìÌv>Li vÀ Ìi /i V«iÌiiÃÃ v >Õ`Ì `>Ì> Ü>Ã iÝ LiÌÜii `>Ì> ÃÕÀViÃ ÜiÀi ÕÌâi` iÛi
Ài}ÃÌiÀ À >Õ`Ì `>Ì>° /Ã >ÃÃÕ«Ì Ã ViiÌ] >` >Ã Ìi V«iÌiiÃÃ v Ài} ÌÕ} Ìi LÕ`>ÀiÃ LiÌÜii ÌÞ«iÃ >Ài Ì
Û>` Ã v>À >Ã «>ÌiÌÃ ÜiÀi ÌÀi>Ìi` >Ì Ìi À ÃÌiÀ `>Ì> Ãiii` Ì Li ÛiÀÞ }`° v>VÌ] Vi>À VÕÌ®° /i vÀÃÌ ÌÞ«i Ü>Ã >««i` >V
Ì«i`V «>ÌiÌ Ü>À` >Ì Ìi ÕÕÃ>Ã ÌiÀi Ã > ÌiìVÞ Ì ÛiÀiÃÌ>Ì v VÕÀ>VÞ V«>ÀÃÃ\ Ìi Ã>i À>ÌiÀ i>ÃÞ
,i}> Ã«Ì>] LiV>ÕÃi Ì Ã ÛiÀÞ Õ >««À«À>Ìi `>Ì> >LÃÌÀ>VÌ ÀÕiÃ >Ài ÕÃi`° i>ÃÕÀ>Li Û>À>Li Ü>Ã >Û>>Li LÌ
iÞ Ì>Ì ÌÜ V«iÌi] LÕÌ ÕÌÕ>Þ `v À iÝ>«i] Ì Ã Ài Ài>Ã>Li Ì VÕÌ `>Ì> ÃÕÀViÃ] >` ìÌiVÌi` `vviÀiViÃ VÕ`
viÀiÌ Ài}ÃÌÀ>ÌÃ ÜÕ` ÃÃ > «>ÌiÌ° iÜ « vÀ>VÌÕÀi V>ÃiÃ Ì> iÛiÀÞ «>ÌiÌ Li vÕÀÌiÀ ViVi` vÀ Ìi i`V> ÀiVÀ`Ã
v>VÌ] iÛi Ìi «>ÌiÌÃ ÌÀi>Ìi` VÃiÀÛ>ÌÛiÞ ÜÌ > }} «vÀ>VÌÕÀiÀi>Ìi` V>Ài i°}° />Li Î®° /i ÃiV` ÌÞ«i Ü>Ã vÀ Ìi
À «iÀ>Ìi` >Ì >ÌiÀ Ã«Ì> >Ài ÀÕ i«Ãì QÈ£] ÈÓR° ÜiÛiÀ] ÕiÃÃ Ìi V V«>ÀÃ v ÌiÀiÌV>Þ Ã>À vÀ
ÌiÞ ÀiviÀÀi` Ì > ÀÌ«i`V «>ÌiÌ V> Õ`}iÌ v > « vÀ>VÌÕÀi V>Ãi Ã >VVÕ >Ì Ì>Ì >Ã ÃÕLÃÌ>Ì>Þ `vviÀiÌ «iÀ
Ü>À` >` ÜÕ` Li Ài}ÃÌiÀi`° ÌiÀÞ] À>ÌiÞ >` V«iÌiÞ ÀiVÀì`] Ài}ÃÌiÀ >Ì>â>ÌÃ LÌ `>Ì> ÃÕÀViÃ i°}°
Ãi «>ÌiÌÃ >`ÌÌi` Ì Ìi ÕÌ«>ÌiÌ L>Ãi` iÃÌ>ÌiÃ vÀ Ìi ÕLiÀÃ v « vÀ>V }Ã° £ >` Ó®° /i ÌÀ` ÌÞ«i v V«>ÀÃ
iiÀ}iVÞ ì«>ÀÌiÌ Ü ÜiÀi Ì ÀiÃ ÌÕÀiÃ >Ài «Ài Ì L>Ã° /i >««À«À>ÌiiÃÃ Ü>Ã > VL>Ì v Ìi vÀÃÌ ÌÜ ÌÞ«iÃ\
ìÌ Ìi «iÀ>Ì >Ài> v Ìi ÕÕÃ> v ÃÕV `>Ì> >LÃÌÀ>VÌ ÀÕiÃ V> Li «>À LÌ >««À>ViÃ V> Li >««i` i°}° >>
Ã ,i}> Ã«Ì> >` ÌÕÃ `ÀiVÌÞ Ì>Þ ÛiÀvi` LÞ V«>À} Ìi Ã«Ì> ÞÃiÃ v Ìi ìÌ>i` vÀ>VÌÕÀi V>ÃÃvV>Ì®°

xÈx
/i iÞ ì> Ü>Ã Ì >i ÛiÀÃ>Ìi V«>À Ài>Ìi` }ÀÕ« vÀ>Ì vÀ Ìi ÌÌ> «> È° -Ìiv>i ° Üi`}i >>}iiÌ Ì ÃÕ««ÀÌ
ÃÃ Ì>Ì >Ài iÌ`}V>Þ ÕÃÌvi` >` ÌiÌ ««Õ>Ì° /Ã >ÜÃ Ìi V>VÕ>Ì «iÀvÀ>ViL>Ãi` i`Vi° iÌ`Ã v i`
ÓääÓÆ {£ £®\ ÎÈ{Î°
«À>VÌV>Þ ÀiiÛ>Ì° v VÃÌÃ LÞ ÃÕ} Õ« Ìi ÃÕÌ>LÞ «ÀVi® Ç° iÃV ] ÕvÕÀ
] -Ì>VV *] ÕÛiÀiÌ ]
/i >VVÕÀ>VÞ v Ìi ÃÌ i>ÃÞ i>ÃÕÀ Üi}Ìi` vÀ>Ì Ìi ÕÃi v i>Ì Õ>``Õ "° i`V> ìVÃ ÃÕ««ÀÌ ÃÞÃÌiÃ\
>Li Û>À>LiÃ Ìi Ài}ÃÌiÀ Ã ÛiÀÞ }`° ÃiÀÛViÃ° /Ã Ã>i ì> vÀ V>VÕ>Ì} ` ì>Ã >` iÜ «>À>`}Ã¶ iÌ`Ã v
ÜiÛiÀ] Ìi `>Ì> ÃÌÀÕVÌÕÀi vÀ Ìi ÀiVÀ` VÃÌÃ V> Li vÕÀÌiÀ ÕÌâi`] ÌÕ} Ìi i` ÓääÎÆ {Ó Î®\ £än°
n° i>À` ° /Ì> µÕ>ÌÞ vÀ>Ì ÃÞÃÌiÃ
} v «iÀvÀi` «iÀ>ÌÃ Ã Ì «Ì> Üi}ÌÃ ÌÃ Ìi VÀÀiÃ«` Ì Ì «ÀViÃ >>}iiÌ\ ÃÃÕiÃ vÀ Ìi i>Ì V>Ài `ÕÃÌÀÞ°
Ìi Ài}ÃÌiÀ° À iÝ>«i] Þ i «iÀ LÕÌ Ì Ìi i> i>Ì À `Ã>LÌÞ ÃÌ>ÌÕÃ v iÌ`Ã v i` £nÆ ÎÇ Ó®\ £xÈÈä°
>Ì `>Þ Ã >Üi` vÀ i Ã«Ì>â>Ì «iÀÃÃ ÀiViÛ} Ã«iVvV vÀÃ v V>Ài] ° ÞBi *] >À>> ° -ÕVViÃÃ >` v>ÕÀi v>V
iÛi ÌÕ} Ìi «>ÌiÌ >Þ Li Ài«iÀ>Ìi` >` ÌÃ ÀiÃÕÌÃ i>Ì ÃÌ>ÌÕÃ i>ÃÕÀi ÌÀÃ Ìi Ài}> i>Ì vÀ>Ì ÃÞÃÌi ì
`ÕÀ} Ìi Ã>i Ã«Ì>â>Ì°/i V ÜV Ã ÕL>Ãi` Ìi Ã>i ÃiÃi >Ã Ã} «ÀViÃÃ q ÀiÃÕÌÃ vÀ > VÃÌÀÕVÌÛi iÛ>Õ
>Ì ÃÌÕ`Þ° iÌ`Ã v i` ÓääÈÆ {x £®\ nx°
ÌiÌ v ÌiÃi vi`Ã VÕ` >Ã Li «ÀÛi` VÃÌÃ V> Li ÕL>Ãi`® >Ì Ìi ««Õ>Ì £ä° -Ìiv>i ° Üi`}i >` «ÀViÃÃ >>}i
LÞ }Û} ÃÌ>`>À`âi` ÃÌÀÕVÌÃ Ì>Ì iÛi° Ý>>Ì v ÃÕV ì>Ã Ã ÕÌ v iÌ i>Ì V>Ài À}>â>ÌÃ° iÌ`Ã v
}Õì «À>VÌÌiÀÃ Ì >Ã ÀiVÀ` Ìi Ài Ìi ÃV«i v ÌÃ ÃÌÕ`Þ] LÕÌ >Þ vviÀ vÀÕÌvÕ i` Óää{Æ {Î x®\ xÓxÎx°
iÛ>Ì >``Ì> «iÀ>Ì VìÃ >Û>>Li `ÀiVÌÃ vÀ vÕÀÌiÀ ÀiÃi>ÀV° ££° >Àì -] 7vv
] ÕÌÃV> 1] 7iÌÌiÀ /] >Õ« *°
Ìi V>ÃÃvV>Ì] ÃÕV >Ã ÜV Ãì Ìi Ã > VVÕÃ] Ài}ÃÌiÀ >` >Õ`Ì `>Ì>
--
q
Vi«ÌÃ vÀ ÃÌ Ìi}À>Ì v
i>Ì V>Ài vÀ>Ì ÃÞÃÌi V«iÌÃ Ì
vÀ>VÌÕÀi Ã ° -ÕV >``Ì> iÌ>`>Ì> LÌ >Ûi ViÀÌ> Ì>ÌÃ >` «ÀLiÃ] iÃÌ>LÃi` «ÀViÃÃiÃ v «>ÌiÌ V>Ài° iÌ`Ã v
ÃÕ` >Ã Li vi>ÃLi Ì ÕÃi Ì «ÀÛi Ìi LÕÌ Ãii Ì Li ÃÕÌ>Li vÀ Ìi «iÀvÀ>Vi i` ÓääÈÆ {x £®\ £än°
µÕ>ÌÞ v Ài}ÃÌiÀ `>Ì> }iiÀ> V>ÃiÃ >ÃÃiÃÃiÌ v « vÀ>VÌÕÀi ÌÀi>ÌiÌ° 6Õ £Ó° "
° `Ãi>ÃiL>Ãi` V«>ÀÃ v i>Ì
ÜiÀi Ìi ÕÃi v ÃÌÀÕVÌÕÀi` `>Ì> iÌÀiÃ Ã Ì>ÀÞ «ÕÌ v >``Ì> « vÀ>VÌÕÀi iÛiÌ ÃÞÃÌiÃ° 7>Ì Ã LiÃÌ >` >Ì Ü>Ì VÃÌ¶ *>ÀÃ\ "À
}>Ã>Ì vÀ VV
«iÀ>Ì >` iÛi
Ü Ì Û>ÀÞ LiÌÜii VìÀÃ QÎ] ÈÎ] È{R° `>Ì> Ì Ìi Ài}ÃÌiÀ ÃÕ` Li >ì «ÃÃLi] «iÌÆ ÓääÎ°
À Ìi «ÕÀ«ÃiÃ v «iÀvÀ>Vi >ÃÃiÃÃ LiV>ÕÃi Ìi LiÃÌ V«ÀÃi vÀ «À>VÌV> £Î° À>`Þ ] -VÜ>ÀÌâ ] i`ÌÀÃ° i`V> ivviV
iÌ] vÜÕ« vÀ>Ì Ã «>ÀÌVÕ>ÀÞ «iÀvÀ>Vi >ÃÃiÃÃiÌ «ÕÀ«ÃiÃ ÜÕ` Li ÌÛiiÃÃ ÀiÃi>ÀV `>Ì> iÌ`Ã° ,VÛi\ }i
iÃÃiÌ>° ÀÌ>ÌÞ >` i}Ì v Ì> Ài>Vi` LÞ > «ÀÃ«iVÌÛi ÀiVÀ`} v « VÞ vÀ i>Ì
>Ài *VÞ >` ,iÃi>ÀVÆ £Ó°
Ã«Ì>â>Ì «iÀ` ÜiÀi ÀiVÀì` vÀ>VÌÕÀi iÛiÌ `>Ì> VLi` ÜÌ iÝÌÀ>V £{° Õ>À ] -Ì ] *Ã>i ] >}i ]
-Ìiv>i °
V> }ÕìiÃ >Ã «>Ã q >
iµÕ>Þ LÌ `>Ì> ÃiÌÃ] LÕÌ ÌiÀÜÃi Ì Ì v V>Ài ÃÌÀiÃ v ÌiÃi «>ÌiÌÃ vÀ Ì}V> ÌiÀÞ° iÌ`Ã v i` ÓääÈÆ {x Ó®\
Ü>Ã Ì i>ÃÞ Ì ìÀÛi vÀ>Ì Ì>Ì Ü>Ã Ìi Ài}ÃÌiÀÃ° Óä{£ä°
iÝ>VÌÞ V«>ÌLi° vÀ>Ì Ìi ÕÃi £x° > 7] > , ] Ã -] ,Ãi ] Ã
v «>ÌiÌ i>Ì ÃiÀÛViÃ > `>Þ L>ÃÃ VÜi`}iÌÃ ÜÌâ ° Û>Õ>Ì v V«V>Ì À>ÌiÃ >vÌiÀ
Ü>Ã >Û>>Li `ÀiVÌÞ vÀ Ìi Ài}ÃÌiÀ /Ã ÜÀ Ü>Ã v>V>Þ ÃÕ««ÀÌi` LÞ > ÀiÃi>ÀV VÀ>ÀÞ >ÀÌiÀÞ LÞ«>ÃÃ ÃÕÀ}iÀÞ ÕÃ} >`Ã
}À>Ì vÀ Ìi V>ìÞ v >`° /i vÕ`} ÌÀ>ÌÛi `>Ì>° iÌ`Ã v i` £nÆ ÎÇ Ó®\
Üi ìÀÛ} Ã>À vÀ>Ì vÀ Ìi >}ÀiiiÌ iÃÕÀi` Ìi >ÕÌÀÃ¿ ì«iìVi ì £ÓÓää°
>Õ`Ì `>Ì> ÀiµÕÀi` ViÀÌ> >ÃÃÕ«ÌÃ >` Ã}} Ìi ÃÌÕ`Þ] ÌiÀ«ÀiÌ} Ìi `>Ì>] >` ÜÀÌ} £È° ½Ài 7] -VÌÌi
] /µÕ
° ,Ã >`ÕÃÌiÌ
ÌiÀ«>Ì° /i ViVÌ v vÜÕ« >` «ÕLÃ} Ìi Ài«ÀÌ° ÕÌVi >ÃÃiÃÃiÌ\ Ìi
>ÀÃ VÀL`ÌÞ
vÀ>Ì vÀ Ìi >Õ`Ì `>Ì> ÀiµÕÀiÃ ìÝ° iÌ`Ã v i` £ÎÆ ÎÓ x®\ ÎnÓÇ°
£Ç° "½
i ,] ° 1ÌÌÞ v Ìi
>ÀÃ
iÝÌÀ> ÜÀ >` Ã «Ài Ì `vviÀiÌ L>ÃiÃ >Ã VÀL`ÌÞ ìÝ V«ÕÌi` vÀ ÀÕÌiÞ V
Ã Ìi ÀÕÌi ViVÌ v Ài}ÃÌiÀ `>Ì>°
Ûi ÌÕ} Ìi ÕÃi v i>Ì ÃiÀÛViÃ Ã
,iviÀiViÃ iVÌi` Ã«Ì> `ÃV>À}i `>}ÃÃ VìÃ°
iÌ`Ã v i` ÓäääÆ Î £®\ Ç££°
£° `>Þ ] i}iÞ
] >ÀÃ ,] ->ÌiÀ
°
ÀiVÀì` LiÌÌiÀ Ìi Ài}ÃÌiÀ] ÌiÀ Û>Õ>Ì} Ìi i>ÌV>Ài ÃÞÃÌi\ vviVÌÛiiÃÃ]
£n° iââ ° ,Ã `ÕÃÌiÌ vÀ i>ÃÕÀ} i>Ì
i>ÃÕÀiiÌÃ v ÕÃivÕ VVi«ÌÃ ÃÕV >Ã V>Ài ÕÌViÃ° /À` i`Ì°
}>}\ i>Ì
ivvViVÞ] >` iµÕÌÞ° Ó` i`Ì°
V>}\ `ÃÌÀ>Ì *ÀiÃÃÆ ÓääÎ°
ÕÃi v ÌiVV> >`] vÕVÌ> `Ã>LÌÞ i>Ì `ÃÌÀ>Ì *ÀiÃÃÆ £n° £° /Ài ] ÀiÀ * ° « vÀ>VÌÕÀi >Õ`Ì\ Ìi vÀ
>` «> vviÀ vÀ>Ì Ì>Ì Ã Ì ` Ó° "
° i>ÃÕÀ} Õ« q «ÀÛ} i>Ì ÃÞÃÌi ÃÌ>`>À`Ã Ì Li iÛìViL>Ãi`° Ì "ÀÌ« ÓäääÆ
ÀiVÌÞ >Û>>Li vÀ Ìi Ài}ÃÌiÀ° ÜiÛiÀ] «iÀvÀ>Vi "
VÕÌÀiÃ° *>ÀÃ\ "À}> Ó{ {®\ £n£Î°
Ã>Ì vÀ VV
«iÀ>Ì >` iÛi Óä° *>ÀiÀ ]
ÕÀÀi
/] ÕÌ> ] /À}Ài
LiV>ÕÃi Ìi Ài}ÃÌiÀ VÕìÃ > `Ã>LÌÞ «iÌÆ ÓääÓ°
i>ÃÕÀi] Ì >VÌÕ>Þ vviÀÃ V«iiÌ>ÀÞ ° -Ì>`>À`Ãi` >Õ`Ì v « vÀ>VÌÕÀi ÕÀ«i
Î° >Ã> ] ->vÀ>
] />i`> ° +Õ>ÌÞ v i>Ì - ®° « Ì £nÆ n £®\ £äx°
vÀ>Ì V>Ài\ vÀ>ÌVÃ vÕ`>ÌÃ° iÌ`Ã v i` Ó£° ii /] 7} ,] Bi ] Ãi ]
>Ì >`ÃÃ] `ÃV>À}i >` ViÃÕÃ «ÌÃ®° ÓääÎÆ {Ó x®\ xä£n° >ÃB *] >Û> ] iÌ >° « vÀ>VÌÕÀiÃ
{° iÜiÀÌ ] >ÕÝ ,] ÕÜÃ
] i]
/Ã vÀ>Ì ÜÕ` Li Ài Û>Õ>Li v À>ìÀ ,] À} ] iÌ >° i`V> vÀ>ÌVÃ
>` q > V«>ÀÃ v «>ÌiÌ V>À>VÌiÀÃÌVÃ
Ãi VÞ ÕÃi` vÕVÌ> `Ã>LÌÞ >` ÕÌViÃ ÃÝ Ã«Ì>Ã° -V>` -ÕÀ} Óää{Æ
>` Ìi µÕ>ÌÞ v i>Ì\ iÜ >««À>ViÃ Ì ÃÕ« Î Î®\ ÓÎ{{ä°
À µÕ>ÌÞ v vi i>ÃÕÀi >` Lii >Û>>Li «ÀÌ «>ÌiÌ V>Ài q v`}Ã vÀ Ìi 9i>À ÓÓ° ÕÀ ] >Ài ] ØÌi *] />i -°
ÃÌ
Ìi Ài}ÃÌiÀ ÃÌi>` v > Û>`>Ìi` L v i`V> vÀ>ÌVÃ ÓääÎ° iÌ`Ã v >>ÞÃÃ v « vÀ>VÌÕÀi ÌÀi>ÌiÌ >} Ìi
>` V i>ÃÕÀi QÈxR° i` ÓääÎÆ {Ó Ó®\ £nx° iìÀÞ vÀ Ìi «ÕLV i>Ì ÃiÀÛViÃ\ > £Þi>À
x° iÜiÀÌ ] iÀÃ ] VÃÌ>`ÌiÀ ,] >ÕÝ ,]
" Ìi ÌiÀ >`] Ìi Ài}ÃÌiÀ VÕìÃ * 1] ,iÃV ° -ÞÃÌiÃ >>ÞÃÃ i>Ì V>Ài\
«ÀÃ«iVÌÛi ÃÌÕ`Þ £äÈ VÃiVÕÌÛi «>ÌiÌÃ°
vÀ>Ì VViÀ} Ìi «ÀÛìÀÃ«i ÀV "ÀÌ«/À>Õ> -ÕÀ} ÓääÎÆ £ÓÎ £ä®\ xx£{°
vÀ>iÜÀ >` iÝ>«i° iÌ`Ã v i` ÓääÓÆ
VvV ÕÃi v i>Ì ÃiÀÛViÃ ÜÌ `>}ÃÌV {£ Ó®\ £Î{{ä°

xÈÈ
-Õ` iÌ >°
ÓÎ° ÕÀ ] >Ài ] ØÌi *] />i -° ÕV În° -ÀiÃi /] ->LÀi -] "Ãi ° vÀ>iÜÀ vÀ ÀiÃi>ÀV `>Ì>L>ÃiÃ° iÌ`Ã v i` Óää{Æ {Î
Ì> ÕÌVi >` ÃÕÀÛÛ> >vÌiÀ « vÀ>VÌÕÀi iÛ>Õ>Ì v ÃiV`>ÀÞ `>Ì> ÃÕÀViÃ vÀ i«ì Î®\ ÓnÇx°
iìÀÞ\ > «À«iVÌÛi ÃÌÕ`Þ v £äÈ VÃiVÕÌÛi «> }V> ÀiÃi>ÀV° Ì «ì £ÈÆ Óx Ó®\ xx° À>`Ì
]
i ] -v> ] iÀ *]
ÌiÌÃ° "ÀÌ«>i` /À>Õ>Ì Óää{Æ x £®\ Ç£{° {Îx{Ó° >`>À *] À>ÜiÞ -° ««À>ViÃ >`
Ó{° *Üi ] >ÛiÃ /] /Ã ,° 1Ã} ÀÕ Î° 7ÌiÀi ,° º7i v ÕÀ ÜÀ} ÕìÀ vÀ>ÌVÃ ÌÃ Ì >ÃÃÃÌ Ìi Ìi}À>Ì v Ã
Ìi V«>À>ÌÛi `>Ì> Ì >ÃÃiÃÃ Ìi µÕ>ÌÞ v ÃÌ>`}»\ VìÃ] V>ÃÃvV>ÌÃ >` Ìi «À >À VV> ÀiÃi>ÀV µÕiÃÌ>ÀiÃ° iÌ`Ã v
i>Ì V>Ài\ ÕìÀÃÌ>`} >` >Û`} V `ÕVÌ v >VVÕÀ>Ìi `>Ì>° iÌ`Ã v i` ÓääÎÆ i` Óää{Æ {Î Ó®\ £xÈÈÓ°
«Ìv>Ã° +Õ> ->v i>Ì
>Ài ÓääÎÆ £Ó Ó®\ £ÓÓn° {Ó {®\ {nÈ° xÈ° Û> ii °
ÃìÀ>ÌÃ vÀ Ìi Ài«Ài
Óx° ,ÃÃ>i *] -Õ` ,] À`L>V ] ,ÕÃ /] `B {ä° ÀLÀ
] ÕÀÌ>` 6] Ã , ° i>Ì ÃiÌ>Ì v iÌ>`>Ì> vÀ Ìi ÃÕ««ÀÌ v ÃÌÀÕV
«BBiB 1° ,i}ÃÌiÀ L>Ãi` ivviVÌÛiiÃÃ V>Ài i«ÃìÃ\ ìvÌ] i>ÃÕÀiiÌ >` ÕÃi° ÌÕÀi` `>Ì> iÌÀÞ° iÌ`Ã v i` ÓääÎÆ {Ó Î®\
i>ÃÕÀiiÌ Ìi V>Ãi v « vÀ>VÌÕÀi ÌÀi>ÌiÌ i`
>Ài ,iÛ £nxÆ {Ó Ó®\ £ÈÎÓ£n° ÓÓÈÎx°
Ã®° iÃ\ >Ì> ,iÃi>ÀV >` i {£°
] ÕÀ>>ÌÃÕ ° 7>Ì Ã Ìi µÕ>ÌÞ v xÇ° >ÕÝ ,° Üi`}iL>Ãi` ìVÃ ÃÕ««ÀÌ vÀ
Ûi«iÌ
iÌÀi vÀ 7iv>Ài >` i>ÌÆ ÓääÓ° µÕ>ÌÞ v i`V> V>Ài i>ÃÕÀiÃ¶ ,>Ãi `>}ÃÃ >` ÌiÀ>«Þ\ Ìi ÕÌ«i ÕÃ>LÌÞ v
ÓÈ° -Õ` ,° 1ÌÃ>Ì v >`ÃÌÀ>ÌÛi Ài}ÃÌiÀÃ Ài>ÌÛÃ >` Ài>ÜÀ` >««V>ÌÃ° *iÀÃ«iVÌ «>ÌiÌ `>Ì>° iÌ`Ã v i` £nÆ Ón Ó®\
ÕÃ} ÃViÌvV Üi`}i `ÃVÛiÀÞ° Ìi}iÌ i` ÓääÎÆ {È £®\ xÓäÆ `ÃVÕÃÃ Ó£Î° ÈÇÇ°
>Ì> >ÞÃÃ ÓääÎÆ Ç È®\ xä££° {Ó° Õ`}i ] >`>À ] i> ° Ü Ài>Li >Ài xn° iÀ} ] À> ° /i VÌiÝÌÕ> >ÌÕÀi v
ÓÇ° -Õ` ,] Ã ° +Õ>ÌÞ ivviVÌÃ v «iÀ>ÌÛi ì>Þ Ài>LÌÞ ÃÌÕìÃ v vÀ>VÌÕÀi V>ÃÃvV>ÌÃ¶ i`V> vÀ>Ì° Ì i` vÀ £Æ xÈ
ÀÌ>ÌÞ « vÀ>VÌÕÀi ÌÀi>ÌiÌ° +Õ> ->v ÃÞÃÌi>ÌV ÀiÛiÜ v ÌiÀ iÌ`}iÃ° VÌ> £Î®\ x£Èä°
i>Ì
>Ài ÓääxÆ £{ x®\ ÎÇ£Ç° "ÀÌ« -V>` Óää{Æ Çx Ó®\ £n{{° x° 7° ÌiÀ>Ì> 7À} ÀÕ« vÀ Ã
Ón° ÃÃiÀ ] >Õ> ° Ã i>Ì >` ÃV> {Î° <>} ° ,i«ÀiÃiÌ>ÌÃ v i>Ì VVi«ÌÃ\ > i>Ãi ÌÀ} >` ÀiV>ÃÌ}°
>«ÌÕÀiÀiV>«
Üiv>Ài Ài}ÃÌiÀÃ i«ì}V> ÀiÃi>ÀV° V}ÌÛi «iÀÃ«iVÌÛi° i` vÀ ÓääÓÆ Îx ÌÕÀi >` ÕÌ«iÀiVÀ` ÃÞÃÌiÃ iÃÌ>Ì \
ÀÃ «ì} Óää{Æ £{ £®\ ££ÎÓä° £®\ £ÇÓ{° ÃÌÀÞ >` ÌiÀiÌV> ìÛi«iÌ° «ì
Ó° ,Ã ] 7>`> ° ,iVÀ` >}i ÃÌÀ>Ìi}iÃ° *>ÀÌ {{° 7À>Þ *] ÃÌ
] ÕÞi`> ] }Ã £xÆ £{Ó £ä®\ £ä{Çxn°
\ ÃÌ>Ì} vÀ>Ì >` iÛ>Õ>Ì} >« ÜÀÌ
° 1Ã} >`ÃÌÀ>ÌÛi `>Ì>L>ÃiÃ Ì Èä° ÀiiÀ ° ««V>Ì v V>«ÌÕÀiÀiV>«ÌÕÀi
«À>ViÃ° iÌ`Ã v i` ££Æ Îä Ó®\ ££ÇÓÎ° iÛ>Õ>Ìi Ìi µÕ>ÌÞ v i`V> V>Ài\ > VVi«ÌÕ> iÌ`Ã vÀ `Ãi>Ãi ÌÀ}\ «ÌiÌ> ivviVÌÃ
Îä° 7>`> ] ,Ã ] >ÞivÃÞ ] -}iÌ ° vÀ>iÜÀ° -V -V i` £xÆ {ä £Ó®\ £ÇäÇ£x° v «iÀviVÌ ÀiVÀ` >}i° iÌ`Ã v i`
,iVÀ` >}i ÃÌÀ>Ìi}iÃ\ *>ÀÌ ° *ÀÌ>Li ÃvÌ {x° >i ,] i`ÌÀ° 1ìÀÃÌ>`} i>Ì V>Ài ÕÌ £{Æ ÎÎ x®\ xäÓÈ°
Ü>Ài >` ìÌiÀÃÌV >ÌV}° iÌ`Ã v ViÃ ÀiÃi>ÀV° >ÌiÀÃLÕÀ}\ Ã«i *ÕLÃiÀÃÆ È£° >}iÞ ] -Ìi«iÃ -]
ÀÞiÀ
] À> °
i` ££Æ Îä Î®\ Ó£ä{° £Ç° /À>«Ã vÀ Ìi ÕÜ>ÀÞ iÃÌ>Ì} «iÀÃ L>Ãi`
Î£° "LiÀ>}iÀ 7] -ÌÕ}iÀ 7° ,iVÀ` >}i {È° -ÀÞiÀ ] ` ] -iÌ ] >ÀÃ> ] ÕÀÞ VìVi ÕÃ} Ã«Ì> `ÃV>À}i `>Ì>°
Ìi
>ViÀ ,i}ÃÌÀÞ v /ÞÀ] ÕÃÌÀ>° iÌ`Ã ÀÛiÀ ] >iÀiÃÌiÀ ° ,i>ÌÃ«Ã *ÀiÛ ÓääÓÆ n {®\ ÎÎÓÇ°
v i` ÓääxÆ {{ x®\ ÈÓÈÎä° LiÌÜii «>ÌiÌÀi>Ìi` ÀÃ v>VÌÀÃ] «ÀViÃÃiÃ] ÈÓ° Ã«
] >}iÞ ° iÌiÀ} vÀÃÌ >`Ã
ÎÓ° iÃB ] À -° VVÕÀ>VÞ v `>Ì> `>} ÃÌÀÕVÌÕÀiÃ] >` ÕÌViÃ v V>À`>V ÃÕÀ}V> V>Ài° ÃÃ > Ã«Ì> `ÃV>À}i vi Û> ÀiVÀ`
ÃiÃ] «ÀVi`ÕÀiÃ >` >VVìÌÃ Ìi Ã
Vi«ÌÕ> ìÃ° i`
>Ài £xÆ ÎÎ £ä >}i° iÌ`Ã v i` £nÆ ÎÇ £®\ ÎÓÇ°
Ã«Ì> ÃV>À}i ,i}ÃÌiÀ° Ì i>Ì -V -Õ««®\ "-ÓÈÎ{° ÈÎ° Ã ,] ,Õi> ] Û> ii ] ì 7ì
££Æ Ó £®\ £xÓ£° {Ç° À«VÃ> ] iÌ> ° i>ÃÕÀ} >}ÀiiiÌ ] Û> ìÀ i ° Ài ÃÌÀÕVÌÕÀi` `>Ì> ÃÌÀÕVÌÕÀi`
ÎÎ° >Ì ,] *iÌÌB ° /i Û>`ÌÞ v ì>Ì ViÀÌv i`V> vÀ>ÌVÃ Ài>LÌÞ ÃÌÕìÃ° i` ìÌV>Þ¶ ÛiÃÌ}>Ì} Ìi ÕvÀÌÞ v «i`
V>ÌiÃ\ ÀÕÌi Û>`>Ì v ì>Ì ViÀÌvV>Ì >` vÀ ÓääÓÆ Îx Ó®\ ££ä° >ÌÀV «>ÌiÌ `>Ì> ÀiVÀì` ÕÃ} "«i- °
ÌÃ ivviVÌÃ ÀÌ>ÌÞ ÃÌ>ÌÃÌVÃ° ÀiÃV -V Ì {n° `ÀiÃ ] >Àâ *° iÌ>\ iÜ i>ÃÕÀi v iÌ`Ã v i` ÓääxÆ {{ x®\ ÈÎ£n°
Óää£Æ ££x £Ó®\ £xÎÓ° >}ÀiiiÌ LiÌÜii ÌÜ À>ÌiÀÃ° À >Ì -Ì>Ì *ÃÞ È{° 7>Ã ° i>ÃÕÀ} Ìi V«iÌiiÃÃ >`
Î{° i««BB ] 6ÀÌ> ] ii "*° 6>`>Ì V Óää{Æ xÇ £®\ ££° VÕÀÀiVÞ v V`vi` VV> vÀ>Ì° iÌ
v ÃÌÀi `>}ÃÃ Ìi >Ì> Ã«Ì> Ã {° `ÀiÃ ] >Àâ *°
>ViVÀÀiVÌi` i>Ã `Ã v i` ÓääÎÆ {Ó {®\ {nÓn°
V>À}i ,i}ÃÌiÀ >` Ìi ,i}ÃÌiÀ v
>ÕÃiÃ v ÕÀiÃ v Ài>LÌÞ >` Û>`ÌÞ y Ì>LiÃ° -Ì>Ì Èx° >Þ ] 6`À> 6° iiÀV ÛiÀÃÕÃ `Ãi>ÃiÃ«i
i>Ì >`° ÕÀ «ì £Æ £x Ó®\ iÌ`Ã i` ,iÃ ÓääxÆ £{ x®\ {ÇÎÓ° VvV ÃÌÀÕiÌÃ µÕ>ÌÞvvi >ÃÃiÃÃiÌ v
£xxÈä° xä° Õ`}i ] >`>À ] >Ã ] i> ° VÀV LÃÌÀÕVÌÛi «Õ>ÀÞ `Ãi>Ãi° iÌ`Ã
Îx° *>Õi *] ÕÕi ] iÌi ] iÀ> /] VVi«Ì vÀ Ìi Û>`>Ì v vÀ>VÌÕÀi V>ÃÃv v i` ÓääÈÆ {x Ó®\ Ó££x°
i,BB *] BÀBÃi>À *] iÌ >° V>ÌÃ° "ÀÌ« /À>Õ> ÓääxÆ £ È®\ {ä£È°
/i Û>`ÌÞ v Ìi Ã Ã«Ì> ÃV>À}i x£° À}Ài ] ÀiÃ ] Bi 1] ÀiÃ
,i}ÃÌiÀ >`
>ÕÃiÃ v i>Ì ,i}ÃÌiÀ `>Ì> °
>ÃiÝ >`ÕÃÌiÌ >` ivvViVÞ i>ÃÕÀi
VÀ>ÀÞ i>ÀÌ `Ãi>Ãi° ÕÀ
>À`Û>ÃV *ÀiÛ ,i iÌ° -V>` *ÕLV i>Ì Óää{Æ ÎÓ È®\ {È{Ç£°
>L ÓääxÆ £Ó Ó®\ £ÎÓÇ° xÓ° -V>Li ° ,i«ÀiÃiÌ} >` «ÀViÃÃ} i`
ÀÀiÃ«ìVi Ì\
ÎÈ° ØÌi *] ÕÀ ] >Ì>> ] iÛ>>À> ] ->Ì> V> Üi`}i ÕÃ} vÀ> VVi«Ì >>ÞÃÃ° ,i -Õ`
ÛÀÌ> -° VìVi v «iÛV vÀ>VÌÕÀiÃ >` iÌ`Ã v i` ÓääÓÆ {£ Ó®\ £ÈäÇ° >Ì> ,iÃi>ÀV >` iÛi«iÌ
iÌÀi
£nn° VÌ> "ÀÌ« -V>` £xÆ ÈÈ Î®\ Ó{xn° xÎ° i}ÕiÌ *] iÃV ]
>ÌiiÀ ° iVÃ ÃÕ« vÀ 7iv>Ài >` i>Ì -/ -®
ÎÇ° ÕÕÃ /] >À«« *] Û>i 6] >ÕÌ>i ] «ÀÌ ÃÞÃÌiÃ vÀ Ìi ÃÌ>`«Ì v Üi`}i "ÕÌViÃ >` iµÕÌÞ ÀiÃi>ÀV
-Õ>Û> ,° /i V>}} «VÌÕÀi v « vÀ>VÌÕÀiÃ\ Ài«ÀiÃiÌ>Ì° iÌ`Ã v i` £xÆ Î{ £Ó®\ *°"° Ý ÓÓä
`À>>ÌV V>}i >}i `ÃÌÀLÕÌ >` V>}i ÓäÓn° ääxÎ£ iÃ
>}i>`ÕÃÌi` VìVi ÜÌ £ä Þi>ÀÃ
i x{° À>`Ì
] -Õ ]
>À«iÌiÀ *] >`>À *° >`
ÌÀ> >`° i £Æ Ó{ Î®\ ÓxÇ° Ìi}À>Ì v 7iLL>Ãi` >` *
L>Ãi` VV> >\ Ài°ÃÕ`JÃÌ>iÃ°v

III
Epidemiologic Perspectives &
Innovations BioMed Central
Analytic Perspective Open Access

Utilization of routinely collected administrative data in monitoring
the incidence of aging dependent hip fracture
Reijo Sund*
Address: National Research and Development Centre for Welfare and Health (STAKES), Health Services and Policy Research, P.O. Box 220, FIN-
00531 Helsinki, Finland
Email: Reijo Sund* - reijo.sund@stakes.fi
* Corresponding author
Published: 7 June 2007 Received: 13 October 2005

Accepted: 7 June 2007
Epidemiologic Perspectives & Innovations 2007, 4:2 doi:10.1186/1742-5573-4-2
This article is available from: http://www.epi-perspectives.com/content/4/1/2
© 2007 Sund; licensee BioMed Central Ltd.
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0),
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Abstract
Societies are facing challenges as the public health burden increases in tandem with population
aging. Local information systems are needed that would allow a continuous monitoring of the
incidence and effectiveness of treatments. This study investigates the possibilities of routinely
collected administrative data as a data source for hip fracture incidence monitoring in Finland.
The study demonstrates that a straightforward use of register data results in biased estimates for
the numbers of hip fractures. An interpretation of hip fractures from the population aging point of
view offers an alternative perspective for hip fracture incidence calculation. This enables
development of a generalizable method for probabilistic detection of starting points of hip fracture
care episodes. Several risk factor and risk population extraction techniques required in register-
based data analyses are also demonstrated. Finally, it is shown that empirical evidence suggests that
hip fracture incidence is proportional to population level disability prevalence.
In conclusion, Finnish administrative data makes it possible to derive data for rather detailed
population level risk factor stratification. Certain limitations of register-based data can be partly
avoided by synthesizing data-sensitive methodological solutions during the analysis process.
Background In fact, there is a great deal literature on doing descriptive

Societies are facing challenges as the public health burden epidemiology with administrative data [1,2]. The main
increases in tandem with population aging. In order to problem with secondary administrative data is that the
fulfill the growing need for information for prevention straightforward application of standard epidemiological
and performance assessment purposes, local information practices may not be feasible, because the data collection
systems are needed that would allow a continuous moni- cannot be tailored to meet the needs of the exact research
toring of the incidence and effectiveness of treatments for problem, as with separate primary data collection. There-
important health problems. However, additional data fore, it is obvious that the validity of secondary data
production requires funding and resources, and it would depends on the research question in mind [3]. In other
be beneficial if the required information could be pro- words, an extra interpretation and preprocessing phase
duced using existing administrative data such as hospital that aims to find an adequate compromise between prob-
discharge registers. lem-driven and data-driven approaches becomes an
important part of the research process [4]. Even though
Page 1 of 11
(page number not for citation purposes)

Epidemiologic Perspectives & Innovations 2007, 4:2 http://www.epi-perspectives.com/content/4/1/2
this phase is commonly encountered while using second- is not a problem, because virtually all patients with hip
ary data in epidemiological studies, it is seldom reported fracture require hospital treatment, which results in a dis-
explicitly. In fact, a more coherent report of the results can charge record with a hip fracture diagnosis. Another
be achieved by stating that the required compromises potential pitfall is related to multiple hospitalizations of a
were known a priori even if those choices were actually single patient. In principle, record linkage allows detec-
"findings" of the preprocessing phase. The drawback in tion of multiple hospitalizations of the same patient, and
underreporting is that the data utilization may seem to be the real problem is to define which hospitalizations
easier than it really is. In the worst case, this may lead to should be considered as readmissions. There are also
uncritical and mechanical repetition of "good examples" other (more or less data-source specific) problems — such
in circumstances where the assumptions of those exam- as the use of diagnosis codes in the selection of cases —
ples are not valid. which are typically special cases of common challenges
for the use of secondary register data in (epidemiological)
This study examines the possibilities of routinely collected research, and require careful data preprocessing that
Finnish administrative data as a data source for an epide- incorporates tacit knowledge formalized in terms of
miologic surveillance system. The basic idea is to demon- appropriate data abstraction rules [2-4].
strate what kind of issues may require some "rethinking"
during a concrete empirical research process that utilizes Hospital discharge data in Finland
secondary data. Hip fracture incidence monitoring is used In Finland, hospital discharge data are available in the
as an example application. First, an example is given of Finnish Health Care Register which records data for all
how conventional use of data turns out to produce inpatient care discharges in institutions with 24-hour per-
improper estimates. Because data are not fully compatible sonnel and for outpatient surgical operations. Census
with the traditional conceptualization, some alternative data are also collected at the final day of each year in order
theoretical ideas originating from health services and to capture the ongoing care periods. The register is nation-
aging research are used as additional links between the wide, which in international terms is exceptional in that it
problem of hip fracture incidence monitoring and sec- has such extensive legislative coverage of all public and
ondary data. After that several pragmatic issues con- private service providers. Each record in the register
fronted during the empirical research process on hip includes data on patient and provider ID-numbers, age,
fracture incidence are examined from a methodological sex, area codes, and diagnosis and operation codes, as well
point of view. This study aims to offer practically useful as dates of admission, operation and discharge. Patient ID
methodological perspectives for the utilization of data by is a unique identification number given to all Finnish cit-
emphasizing the importance of continuous data-sensitive izens and permanent residents. A system of unique iden-
problem solving. tification numbers has been operating since 1968 and is
used in all Finnish registers. Importantly, this system
Analysis allows complete deterministic record linkage within and
Conventional approaches for determining the annual between Finnish registers [11]. Data quality is also shown
numbers of hip fractures by using hospital discharge data to be good [12-18]. In the case of hip fracture, the com-
Hip fractures represent a worldwide major public health pleteness is very good and the accuracy of easily measura-
burden whose impact is expanding as the population ble variables is at least 95% [18].
ages, with hip fracture incidence rates increasing exponen-
tially with age [5,6]. For an individual, hip fracture is a Register based numbers of hip fractures in Finland
serious and painful condition that requires and invariably Finnish register data have been previously used as a data
results in acute hospital treatment. Hip fracture is also rel- source for the calculation of the number of hip fractures
atively easy to diagnose (compared with many other [19,20]. For the purposes of this study, all discharge
health problems), and practically all recorded hip fracture abstracts with a primary or secondary hip fracture diagno-
diagnoses (regardless whether principal or secondary) in sis (10th revision of the International Classification of Dis-
hospital discharge abstracts reflect hip fracture treatment. eases diagnoses S72.0, S72.1 and S72.2) from 1998–2002
Therefore, it can be expected that hospital discharge data were identified in the Finnish Health Care Register. The
are a good source for the identification of hip fracture mean number of hospital discharge records (including
patients. census records) with hip fracture diagnosis was 14430 per
year during 1998–2002 in Finland (Table 1).
Pitfalls in the use of hospital discharge data
There are also several known potential pitfalls in using However, the annual numbers of hip fracture discharge
hospital discharge data for calculating the numbers of records do not tell the actual number of hip fracture
injures [7-10]. One general nuisance is that only hospital- events. A typical care chain for a hip fracture patient con-
ized injuries are observed. In the case of hip fracture that sists of acute hospital treatment on a surgical ward and
Page 2 of 11

Table 1: Number of hip fractures in Finland 1998–2002
1998 1999 2000 2001 2002 Mean

Records with hip fracture diagnosis in the Finnish Health Care Register 14089 13818 14192 14978 15071 14430
The number of patients with hip fracture diagnosis 7817 7661 7643 7924 8030 7815
The number of patients with hip fracture discharge 7575 7465 7425 7706 7854 7605
Admissions with hip fracture diagnosis 13219 12965 13381 14142 14145 13570
Of which:
- recorded hip fracture operation 5934 5972 6100 6221 6195 6084
- recorded hip fracture operation or at least two years gap to earlier record with hip fracture 6742 6631 6702 6924 6937 6787
diagnosis
- recorded hip fracture operation or at least two months gap to earlier record with hip 7246 7094 7136 7411 7405 7258
fracture diagnosis
Admissions with first record with hip fracture diagnosis in ten years 5990 5853 5932 6083 6141 6000
Of which:
- persons aged 50 or more 5551 5413 5543 5644 5667 5564
follow-up care on a general ward or in a specialist rehabil- and that the date of an acute admission for surgery can be
itation unit. In addition, a patient may be readmitted for used as an estimator for the actual occurrence date of the
the same fracture after initial discharge to home or to a hip fracture. The mean number of such operations was
long-term facility. In order to prevent multiple counting 6084 per year in Finland during 1998–2002 (Table 1).
of cases, an individual-based record linkage is recom- Unfortunately, this simple definition underestimates the
mended for the detection of multiple hospitalizations for true number of hip fractures. Even though a small number
the same fracture [10]. This is not a problem with Finnish of (incorrectly recorded) re-operations may be errone-
data, and the previous Finnish studies have identified all ously included, the definition excludes patients treated
hospital discharge records with hip fracture diagnosis and conservatively (non-operatively) as well as patients who
then used calendar year boundaries to exclude multiple die before the operation. It is also known that the record-
hospitalizations of each patient. The corresponding ing of hip fracture operations in The Finnish Health Care
annual number of patients with hip fracture discharge was Register has not been totally complete [18,22]. This
on average 7605 per year between 1998 and 2002 (Table means that even some hip fracture related admissions
1). without recorded operations should be considered indica-
tions of fresh hip fracture. The problem is to define which
Unfortunately, the use of calendar year boundaries makes admissions reflect fresh hip fractures, and an easy solution
such an exclusion approach artificial and has no epidemi- is to use a constant hip fracture free clearance period to
ological justification, because it is clear that calendar year sort out new admissions from readmissions. If the new
boundaries result in the fact that the related hip fracture admission is defined as a hip fracture related record with
free clearance periods (time from the beginning of the no other hip fracture admissions for that individual in the
year to a first fracture hospitalization of the year) vary per previous two years, the mean number of hip fractures is
patient. This has two serious consequences. First, it is 6787 per year (Table 1). As the use of long clearance peri-
likely that many patients having their fracture during the ods may exclude some true subsequent fractures, the
final months of each year are erroneously counted as sep- numbers of hip fractures were calculated also with a two-
arate cases for two years, because it takes at least four month hip fracture free clearance period. With this crite-
months until the maximum restoration in terms of resi- rion, the mean number of hip fractures was 7258 per year
dential status (hospitalizations) is reached at the popula- (Table 1).
tion level [21]. Second, one patient may have more than
one hip fracture per year resulting in some undercounting. In principle, further diagnosis-based exclusion rules could
These drawbacks are an indication of need for a more be utilized in order to exclude certain nonstandard cases,
appropriate estimate of the annual number of hip fracture such as pathological hip fractures, arthrosis related frac-
events. tures, and cases with multiple fractures or orthopedic
aftercare. However, variations in coding practices partially
It is reasonable to hypothesize that virtually all recorded invalidate the use of such data abstraction rules, which
hospitalizations with a hip fracture diagnosis combined assume recording of secondary diagnoses. In practice,
with hip fracture operation represent a new hip fracture, even the clinical criterion of hip fracture may have some
Page 3 of 11

regional variation which makes it impossible to obtain a ture should be rather easy to determine, because the actual
unique estimate for the number of hip fractures using reg- event of fracture (or hospitalization following that event)
ister data. is an obvious index point. However, in practice it is chal-
lenging to make a distinction between true subsequent
In any case, following the reasoning described above, it is hip fractures and hospitalizations due to ongoing treat-
likely that the mean number of hip fractures is somewhere ment episodes or reoperations in terms of data. In stand-
between 6800–7200 per year in Finland between 1998 ard practice, it is common to use the first health problem-
and 2002. The estimate for the number of hip fractures in related event available in the data as an index point.
1998 is 6742–7246. Even the upper limit is significantly Another widely used approach is to use certain clearance
smaller than in the earlier Finnish study [20], where the periods to determine the appropriate index points. A third
annual number of hip fractures in 1998 was reported to be technique is to use external complementary data in deter-
7698. The reasons for the small difference in relation to mining (the number of) incident cases [26]. Other
the number of hip fracture discharges in 1998 (Table 1) approaches seem to be rare [27].
are not known, but are most probably due to difficulties
in reconstruction of the exact operationalization of some Intuitively, it seems to be a good idea to determine an
inclusion and exclusion criteria used in the earlier study. individual's first hip fracture occurrence and consider it as
Also some overestimation is likely to occur because of the a starting point for a (chronic) care episode. In terms of
artificial definition of fresh hip fracture, as was described hospital discharge data this corresponds to the detection
above. In conclusion, the earlier study reports a number of an individual's first hospitalization with recorded hip
that seems to be about 10% too large (corresponding to fracture diagnosis. The problem is that the hip fracture
about 700 hip fractures). itself is not a chronic condition, but a remediable health
deficiency. Therefore, an alternative theoretical interpreta-
Alternative theoretical approaches tion for the first hip fracture is needed.
As was demonstrated above, it is difficult to derive ade-
quate estimates for the number of hip fractures even with Gompertzian interpretation of aging-related hip fractures
the careful use of register data. The main problem is that One particularly interesting fact is that hip fracture inci-
the secondary data are not completely compatible with dence increases exponentially with age [5]. This kind of
the theoretical objective of calculating the number of fresh functional dependency (Gompertz law) has very strong
hip fracture events. In other words, it would be beneficial interpretations from the point of view of population aging
if the theoretical objectives could be modified so that they [28-30]. In this sense, perspectives from aging research
are in agreement with the limits of the available register may strengthen the traditional epidemiological interpre-
data. Fortunately, certain methodological ideas originat- tations (and vice versa) [31,32]. Biologically, aging can be
ing from health services and aging research are also useful seen as a complex process occurring stochastically in
in epidemiological applications. organs and tissues after reaching adulthood (and mainly
after reproductive maturity), which results in irreversible
Care episode approach damage accumulation and vulnerability to the failures in
The definition for an incident case is a common problem maintaining the integrity of tissues and organs [33,34]. A
in epidemiologic research. At the conceptual level this particularly fruitful approach is to consider aging-related
means that all events related to the same underlying dis- events of interest as failures in components of a biological
order should be recognized. This is the fundamental idea system [35]. An exponential increase of incidence with age
in the care episode approach [23]. The care episode means that the intensity of failures is constant (as the
approach is widely applied in health economics and in probability of failure is constant in time, the cumulative
health services research. The key point is that the care epi- probability of failure is exponential). Therefore, it is not
sode approach offers a sound methodological framework surprising that most aging-related (Gompertzian) condi-
for dealing with multiple hospitalizations in terms of hos- tions are closely connected to (cumulative) alterations in
pital discharge data, and it also provides a sound basis for certain tissues or organs.
the measurement of incidence, because an appropriate
episode measure combining related hospitalizations is However, hip fracture is not just an aging-related failure in
less subject to over- or undercounting of cases [24]. some single tissue or organ, but is typically the result of an
accidental event (such as a fall). In other words, cumula-
Unfortunately, a sound methodological framework does tive damage in some tissues (such as bone and muscle)
not automatically solve empirical difficulties [25]. One of increases the probability of (accidental or pathological)
the biggest practical problems with linked register data is fracture, but it is also likely that some aging-related condi-
to determine the starting point of the care episode. In tions increase the risk of accident. In this sense, hip frac-
principle, the starting point of a care episode for hip fracture seems to capture effects of non-fatal failures in several
Page 4 of 11

components of the human body. Among the Gomperzian sequent hip fracture is significantly higher than the risk of
conditions, only death seems to have a similar multidi- first aging-related hip fracture. In other words, there may
mensional interpretation as it captures fatal failure(s) in be a cut-off point after which the probability of having a
any vital components of human body. In other words, an new hip fracture is reduced to the level it would be even
interesting (methodological) analogy can be seen without preceding hip fracture, i.e. preceding hip fracture
between death and hip fracture. is an "uninformative" predictor of the new hip fracture. By
examining the observed and expected probabilities of pre-
It is also obvious that serious but non-fatal failures in ceding hip fractures it becomes possible to give justifica-
components of the human body result in some kind of tion for a suitable clearance period to be used.
disabilities. Such disabilities — typically measured in Importantly, this probabilistic argument for the definition
terms of reduced functional capacity and coping with of the starting points of the care episodes also generalizes
physical activities of daily living — are important risk fac- easily to other disorders.
tors for hip fracture [36,37]. In other words, it is reasona-
ble to hypothesize that the prevalence of such risk factors Probabilistic determination of the starting point of care episode
in the underlying population is related to the incidence of In order to determine the cut-off point for a clearance
(aging dependent) hip fractures. Following this interpre- period, the first admissions (index points) with the (prin-
tation, it becomes clear that the occurrence of first aging- cipal or secondary) diagnosis of hip fracture in 1998–
dependent hip fracture gives an approximate upper limit 2002 were identified for each patient. To ensure sufficient
for the time of development of critical hip fracture risk fac- (backward) follow-up time for each patient, all available
tors, because the risk factors must have exceeded the criti- discharge records of the hip fracture population from
cal level before the event of (low-energetic aging 1987–2002 were obtained from the registers using the
dependent) hip fracture. From this it follows that the first unique personal identification numbers as linkage keys.
aging-related hip fracture can be seen as an indication of For each index point, backward time to the previous hip
a chronic (disability) condition, and any subsequent hip fracture admission of the same patient was calculated.
fractures represent just a nuisance for related population Time was measured in months. There were data from the
level interpretations. In conclusion, there are theoretical years 1987–2002, so the minimum (backward) follow-up
justifications for the determination of the first aging- time was 11 years.
related hip fracture occurrences, which are in concordance
with the restrictions of available hospital discharge data. The first task was to calculate the expected probabilities of
This interpretation is also of particular importance in a earlier hip fractures in the hip fracture population of inter-
theoretical sense, because it suggests that a commonly est. The problem here is that reasonable estimates for the
available time series of hip fracture incidence may also incidence of first hip fractures are needed for the calcula-
reflect the more general disability trends of the population. In this study, it was hypothesized that clearance peri-
tion, which is a testable hypothesis. ods between one and ten years may result in a reasonable
definition of first hip fracture, and the mean age- and sex-
Calculating the number of hip fractures group specific incidences between 1998 and 2002 were
So far, the methodological problem of the definition of calculated by using one year and ten year clearance peri-
the hip fracture episode has been reduced to a detection of ods. In order to calculate the actual expected probabilities,
first aging-related hip fractures in the population (starting the logarithms of the mean hip fracture incidence rates for
points of the first hip fracture care episodes). This is not different age-groups were used in estimating the exponen-
necessarily a straightforward task in practice, because the tial trend in age-group specific rates by using (log-)linear
available data allow only limited backward follow-up regression analysis for both sexes separately [38]. This
time and it remains unknown whether the first hip frac- functional relationship allows the interpolation of an
ture found in the data really is the first hip fracture of the incidence rate for persons who are over 40 years old [29].
person. Probabilistic methods can be used to correct the The individuals with an observed hip fracture were then
number of observed persons to the number of persons followed (backwards) in time, with age correspondingly
with the first appearance of a chronic condition [27]. corrected at each time point for all persons in the risk pop-
ulation. With a known age and sex distribution of the risk
However, if there is a need to identify the persons with a population, it was possible to predict the expected proba-
first hip fracture in terms of care episode (instead of just bility of hip fracture by using the estimated log-linear rela-
calculating the number of persons with a first hip frac- tionship between age and hip fracture incidence. For ages
ture), it is inconvenient that the actual backward follow- below 40 years, a constant incidence rate was used in the
up times (definitions of first hip fractures) vary between prediction. These probabilities were then summed across
persons. In addition, the subsequent hip fracture of a per- the risk population resulting in an expected number of
son is biasing interpretations only as far as the risk of sub- hip fractures that was finally divided by the size of the risk
Page 5 of 11

population, giving an expected (conditional) probability Estimating the number of new hip fractures with limited data
of hip fracture (for each time point). The expected proba- It is unfortunate that no shorter than a ten-year criterion
bilities are shown in Figure 1 using dotted lines. As can be seems to be suitable for hip fractures, because in most
seen, the difference between expected probabilities based countries there may not be the required data available for
on one- and ten-year clearance periods is quite small. such a long backward follow-up period. However, if an
estimate for the overall number of first-ever hip fractures
In order to calculate the observed probabilities, a hazard is enough, and there is no need to identify which patients
function giving the (conditional) probabilities of having a actually had their first hip fracture, data with limited back-
preceding hip fracture admission as a function of backward follow-up can also be used.
ward time (i.e. time was measured from index admission
to a preceding hip fracture admission) in months was esti- The numbers of patients without a preceding hip fracture
mated nonparametrically using a product-limit estimator, as a function of backward time in the logarithmic scale for
and smoothed using a polynomial moving average. As the 1998 are drawn in Figure 2. As can be seen, there is almost
estimated hazard function in Figure 1 shows there was an a linear relationship between logarithmic time and the
increased and non-stabilized (non-constant) risk for an number of patients without a preceding hip fracture. In
admission involving hip fracture occurring about 120 other words, even data with quite a short backward fol-
months after the previous admission related to hip frac- low-up time allow the estimation of such a linear trend
ture and the observed risk remains higher than expected resulting in a reasonable prediction for the number of hip
in the ten-year period. If a more formal way is needed to fracture patients at the ten year cut-off point (or at any
determine where the lines meet, it is straightforward to other desirable cut-off point) for a clearance period. For
calculate confidence intervals for the hazard function. In example, if only one year data for backward follow-up are
conclusion, it seems that a clearance period between available, the idea is to calculate time from index admis-
seven and ten years is needed until the risk is reduced to sion to previous fracture or to the beginning of (back-
the same level as that without a preceding hip fracture. In ward) follow-up. Then for each day (preceding the index
this study, a conservative ten-year clearance period was admission) the number of patients with a longer time dis-
selected. The ten-year criterion has also been used in other tance for the previous fracture is calculated. Finally, a log-
studies [39]. arithmic transformation is applied to the day variable,
and a linear regression model (where the number of frac-
tures is a response and constant and logarithmic day are
explanatory variables) is estimated. Predictions from this
linear model can be easily extrapolated to any cut-off
point of interest.
Risk factor extraction

Using Finnish register data it is feasible to identify hip
fracture patients who have their first aging-related hip
fracture in the sense described above. This makes it possi-
ble to determine the status of certain hip fracture risk fac-
tors available in the data — such as age, sex,
institutionalization, urbanity, season and year [40-45] —
in relation to these patients. Methodologically, three dif-
ferent risk factor extraction techniques can be separated:
internal, external, and empirical.
Internal extraction
Internal extraction corresponds to the use of data abstrac-
tion rules within the register data. For example, using the
2002 1 time
Figure
Probability
backward of having
in months
a preceding
from the
hipfirst
fracture
fracture
as ainfunction
1998– of
Finnish data it is not enough to determine the index hip
Probability of having a preceding hip fracture as a function of
backward time in months from the first fracture in 1998– fracture admission, but the actual day of hip fracture must
2002. The smaller picture is a tenfold magnification of the be inferred using appropriate algorithms. It is obvious
final months. Dotted curves represent the expected proba- that the care episode related to hip fracture starts from the
bilities of having a hip fracture (upper curve is calculated with first contact with the health care system after the actual
a one-year clearance period and the lower curve with a ten- event of fracture. However, the index admission could be
year clearance period, see text for more information). for a long stay in residential care, while the diagnosis
referred to an accident that happened near the discharge
Page 6 of 11

Empirical extraction
In empirical extraction, preliminary analyses of the avail-
able data are used to justify the definitions of variables
and data abstraction rules required in the internal extrac-
tion. For example, it is not obvious what kind of defini-
tion for seasons should be used. This problem can be
solved when using the Finnish register data, which allow
accurate calculation of daily numbers of hip fractures.
After smoothing out the random variation in absolute
numbers by a moving-average technique, the mean of
daily numbers of the new hip fractures was 15.2 for per-
sons aged 50 or older (Figure 3). There was small but clear
seasonal variation so that 53.5 per cent of fractures
occurred during the winter/spring season (from Novem-
ber to April) compared to 46.5% during the summer/fall
season (from May to October). Data also revealed that
there had been some very "hazardous" days during the
winter season, but seasonal variation was almost com-
pletely due to non-institutionalized persons (Figure 3).
Figure
Number
tion 2of patients
of backward time
without
in 1998
preceding
in Finlandhip fracture as a func- Using these results (of preliminary analyses) in the defini-
Number of patients without preceding hip fracture as a function of seasons is an example of empirical extraction. In
tion of backward time in 1998 in Finland. fact, the method developed above for the definition of
first aging-related hip fracture is another example of
empirical extraction.
day, or the index admission could be for a surgical period,
while the actual accident had happened before the admis- Risk population data
sion. After the detection of initial surgical admission, For incidence calculations data on risk population are also
more abstraction rules can be developed. For example, the needed. Typically, the official population figures are avail-
most accurate diagnoses can be extracted from the data able in administrative databases with stratification
corresponding to the surgical treatments following the according to age, sex and area of living. In this study, pop-
fractures, since those are the first ones based on x-ray and ulation figures (taken on the last day of the years 1997–
physical examinations. For the purposes of this study, a 2002) in 5-year groups were obtained for each municipal-
person was classified as an institutionalized long-term ity (local administrative unit in Finland) from the Social
care patient if he or she was admitted to the surgical ward and Health Service Statistical Database (SOTKA). Munici-
from some institution providing inpatient care, and if he pality works as an aggregate unit in external extraction,
or she also had a recorded administrative decision for and also allows easy determination of population figures
long-term care or had received inpatient care for more for any combination of these basic units.
than six months during the year preceding the fracture.
External extraction
In external extraction, variables that link individuals to
aggregate levels are first internally extracted and then
external data that describe aggregate units are linked to
each individual. For example, for the purposes of this
study, the area code (municipality) at the index admission
was used to classify each patient as rural or urban (includ-
ing semi-urban areas) using the official grouping defined
by Statistics Finland (rural municipalities are those
municipalities in which less than 60 per cent of the popu-
lation lives in urban settlements). In general, any other
patient characteristic attributable to area-specific phe-
nomena could also be used here instead of urbanity.
Figure
Daily 3
numbers of hip fractures in Finland 1998–2002
Daily numbers of hip fractures in Finland 1998–2002.
Page 7 of 11

Risk population for internally extracted risk factors population (no short-term fluctuations in migration or
It is more difficult to determine risk population for inter- mortality) is needed for interpolating census day popula-
nally extracted risk factors. In this study, long-term insti- tion figures to appropriate mean follow-up times.
tutional care was one risk factor of interest. Fortunately,
data from the Finnish Health Care Register can be used to In addition, as a ten-year clearance period was used in this
calculate the total numbers of clients in long-term institu- study, a person is not at risk of a new hip fracture for the
tional care on the last day of the years 1997–2002, since 10 years following a hip fracture and should be — in prin-
the register includes all inpatient hospital and nursing ciple — left out of the risk population. However, this cor-
home care in Finland. This individual level data can be rection was not done in this study, because adequate
easily aggregated to appropriate groups (such as stratifica- (group-specific) hip fracture prevalence data were not
tion by age, sex and municipality), and further subdivi- available. The bias resulting from keeping the prevalent
sion of population figures according to long-term pool in the risk population was considered to be insignif-
institutional care is possible. However, during the icant, as the number of hip fractures is very small in rela-
research process it turned out that the calculation tech- tion to the population in younger age-groups and
nique used in official statistics concerning institutional mortality following the hip fracture is very high for the
care was inappropriate for the purposes of epidemiologi- older age-groups. In fact, relative bias [1-(uncorrected
cal studies. Therefore, the significantly downwards biased incidence/corrected incidence)] of mean incidence
numbers were corrected using a technique reported else- seemed to be less than 2%. Moreover, even if some group-
where [46]. specific bias exists, the direction is towards conservative
estimates (underestimation rather than overestimation of
This procedure finally resulted in simultaneously incidences).
observed risk populations on the final day of the years
1997–2002 with stratification by sex, age, urbanity, and In conclusion, there is still room for improvement in
institutionalization. The derivation of such exceptionally deriving risk population data. However, the problems are
detailed nationwide population figures was possible insignificant in terms of the "iceberg phenomenon"
because of common aggregate units (municipalities) in involved in the denominator problem [47], and there is
the databases and registers. no reason to believe that the data used in this study would
not reflect the true epidemiology of hip fracture.
Trends for risk populations
Observed risk populations on the final day of the years of Hip fracture incidence
interest are not good approximations for the size of the After data preparation, it is simple to examine the univar-
risk population (or the follow-up time for risk popula- iate effects of population-level risk factors using standard
tion). A common method is to interpolate the mean pop- methodology. It is quite straightforward to extend such
ulation by using observed census data on two consecutive analyses for other stratifications of interest, such as area
years (approximating also the follow-up time of the risk specific estimates (possibly requiring empirical-Bayes esti-
population, if measured in person years). This method mation) revealing regional variation in hip fracture inci-
can be generalized by modeling the trends extractable dence. Results of such basic analyses are reported
from observed risk populations. In this study, regression elsewhere [48]. Only exemplary results on the simultane-
models with the constant, year and square(year) as regres- ous effects of risk factors with an interpretation offering a
sors were fitted for each group with a stratification by sex, new epidemiological perspective for hip fracture are given
age, institutionalization, and urbanity (with age groups here.
50 to 64, 65 to 74, 75 to 84 and 85+). These models were
then used in the approximation of the required risk pop- In standard practice, only a few risk factors are typically
ulations varying by year and–importantly–by (empirically considered simultaneously in population level incidence
extracted) seasons, too. Finally, there were the observed studies. Risk factors of interest are included into the
numbers of hip fracture events and the follow-up times model, and estimates are interpreted as independent
(in person years) for 320 groups. effects (effects adjusted for other risk factors in the
model). The problem is that there may be complicated
Limitations of risk population data interactions between the risk factors, which distort
It is well known that an accurate number of the popula- straightforward interpretations.
tion at risk is an essential requirement for incidence calcu-
lations [47]. In this study it was possible to derive In this study, the Poisson-regression model was used in
population-based denominators for each group of inter- the simultaneous examination of several risk factors.
est. These denominators are desirable in terms of accu- Other analyses have indicated that the (long-lasting)
racy, but not perfect, because an assumption of a stable increase in age-adjusted incidence had recently stabilized
Page 8 of 11

in Finland, with the age-adjusted incidence almost con- of (disability-related) hip fracture risk factors was not fal-
stant between 1999 and 2002 [48]. As the interest in this sified, and deserves further examination.
study was not in secular trends of incidence but in the
effects of risk factors, the final Poisson regression model Conclusion
was estimated using the combined data from 1999 to This study evaluated the feasibility of producing informa-
2002. All main effects (sex, age, institutionalization, tion for the purposes of monitoring hip fracture incidence
urbanity, and season) and statistically significant interac- based on Finnish administrative register data. This type of
tions up to the third degree were included in the model. information production is not a simple task, but requires
The goodness of fit was acceptable in terms of Pearson chi- creative use of rather complicated methodology. It was
square statistic. Because there were a lot of significant shown that straightforward use of register data results in
interactions between risk factors, it was rather difficult to biased estimates of the numbers of hip fractures, and that
report all relevant estimates in easily interpretable form. even the sophisticated use of data is dependent on the
Therefore, the group-specific rates adjusted for all other more or less ambiguous definitions of actual hip fractures.
factors were calculated by anti-logging the corresponding
linear predictors (without the offset) based on the esti- In this study, the definition of hip fracture was linked to
mated model. These adjusted rates represent the system- population aging. This reasoning resulted in the sugges-
atic effects extracted from the data by the model. tion that hip fracture-related disability has certain interest-
Overlapping confidence intervals of two groups indicate ing methodological similarities to the event of death.
that there is no statistically significant difference between Moreover, an example from the results empirically sup-
rates in these groups. ported this interpretation by demonstrating that hip frac-
ture incidence is associated with a certain type of aging-
Alternative epidemiological perspective for hip fracture incidence related disability. Actually it seems that hip fracture inci-
The results in Table 2 show that there is a higher hip frac- dence is directly proportional to the prevalence of popu-
ture incidence among older persons and women than lation level disability. In this sense, hip fracture incidence
among younger persons and men managing at home. trends also reveal the more general disability trends of the
However, there is no significant effect of sex on incidence population.
among institutionalized persons. The incidence is higher
during the winter time for all persons with a functional Methodologically, while using register-based data, the
status potentially good enough to allow for walking out- determination of first aging-related hip fractures is easier
doors (non-institutionalized under age 85). Urbanization than the determination of all hip fractures. Moreover, the
is not associated with significantly higher hip fracture detection of the first aging-related hip fracture event
incidence in Finland. As it is known that institutionaliza- closely resembles the determination of the starting point
tion is a sign of reduced coping with daily activities [49] of the hip fracture care episode. A method was developed
and that women tend to have more functional limitations based on an observed (and expected) hazard function,
and physical disability than men as age increases [50], the that allows probabilistic justification for the starting point
results are in concordance with the hypothesis suggesting of the care episode. This method is applicable in the gen-
that hip fracture incidence is proportional to the preva- eral care episode approach for a wide variety of health
lence of disability related hip fracture risk factors in the problems regardless of the intention to use care episodes
underlying population. in incidence calculations.
In order to test this hypothesis, a correlation was calcu- Furthermore, three different techniques for risk factor
lated between the nationwide data on the prevalence of (and corresponding risk population) extraction were
outdoor walking ability — being a strong disability- demonstrated. Finnish administrative data makes it possi-
related risk factor for hip fracture [51] — of persons aged ble to derive data for a rather detailed population level
65 to 84 and the hip fracture incidences for the corre- risk factor stratification, but such a feature has certain
sponding age and sex groups for each year between 1998– more or less data-specific limitations which can be partly
2002. Data on walking ability status for the population avoided by developing methodological solutions to the
was extracted from consecutive nationwide surveys on encountered problems. One area requiring further work
health behavior among older people conducted by the that is beyond the scope of this study is the more careful
National Public Health Institute [52]. Linear regression estimation of hip fracture prevalence in Finland.
analysis showed that the prevalence of individuals unable
to walk outdoors explained 97.5% of the variation in the Finally, this study tries to demonstrate that the traditional
hip fracture incidence. In conclusion, the hypothesized methodological paradigm with an assumption of theory-
association between hip fracture incidence and prevalence driven data collection and fixed methods has limited suit-
ability for data analyses that utilize secondary data. Two
Page 9 of 11

Table 2: Adjusted incidence rates per 100000 person years for hip fracture in Finland 1999–2002
Urban Rural
Winter Summer Winter Summer
Rate 95% CI Rate 95% CI Rate 95% CI Rate 95% CI
Non-institutionalized persons
Men 50–64 62 57, 67 49 45, 53 46 40, 53 36 31, 42
Women 50–64 45 41, 49 35 32, 39 46 39, 53 36 31, 42
Men 65–74 191 178, 205 158 147, 170 176 159, 195 145 131, 161
Women 65–74 223 211, 237 185 174, 197 234 214, 254 193 177, 211
Men 75–84 608 574, 644 503 474, 533 561 516, 609 464 427, 504
Women 75–84 966 932, 1002 799 769, 830 925 878, 974 765 725, 807
Men 85+ 1797 1668, 1936 1622 1504, 1748 1858 1680, 2054 1676 1515, 1854
Women 85+ 2805 2693, 2922 2531 2426, 2640 2567 2423, 2720 2316 2184, 2457
Institutionalized persons
Men 50–64 1380 1074, 1773 1277 993, 1643 916 689, 1218 848 638, 1128
Women 50–64 1214 908, 1622 1123 840, 1503 1096 795, 1511 1014 735, 1400
Men 65–74 1688 1437, 1983 1646 1400, 1934 1387 1149, 1674 1352 1120, 1633
Women 65–74 2018 1775, 2294 1967 1730, 2237 1884 1616, 2196 1836 1575, 2142
Men 75–84 2106 1883, 2355 2053 1836, 2296 1735 1511, 1992 1691 1473, 1942
Women 75–84 2495 2332, 2670 2432 2272, 2604 2134 1946, 2339 2080 1897, 2282
Men 85+ 2330 2066, 2628 2480 2200, 2795 2152 1861, 2489 2290 1981, 2647
Women 85+ 2529 2376, 2692 2691 2531, 2861 2068 1894, 2257 2200 2017, 2400
main ideas should be noted. First is the empirical justifi- References

cation of theory-driven operationalized definitions, and 1. Bright RA, Avorn J, Everitt DE: Medicaid data as a resource for
epidemiologic studies: strengths and limitations. J Clin Epide-
second is the theoretical model that is compatible with miol 1989, 42:937-945.
theory-driven as well as data-driven approaches. For 2. Baron JA, Weiderpass E: An introduction to epidemiological
example, in this study register data gave only uncertain research with medical databases. Ann Epidemiol 2000,
10:200-204.
estimates for the number of all hip fractures, so the defini- 3. Sorensen HT, Sabroe S, Olsen J: A framework for evaluation of
tion and theory underlying the hip fractures had to be secondary data sources for epidemiological research. Int J Epi-
revised. After theory revision, empirical justification for demiol 1996, 25:435-442.
4. Sund R: Utilisation of administrative registers using scientific
the extraction of risk factors was utilized, which finally knowledge discovery. Intelligent Data Analysis 2003, 7:501-519
gave results that supported the revised theory. In fact, it [http://iospress.metapress.com/openurl.asp?genre=arti
cle&issn=1088-467X&volume=7&issue=6&spage=501].
may be difficult to make a clear distinction between defi- 5. Melton LJ 3rd: Epidemiology of hip fractures: implications of
nitions, preliminary analyses justifying these definitions, the exponential increase with age. Bone 1996, 18:121S-125S.
and actual results of the study. In conclusion, it seems that 6. Cummings SR, Melton LJ: Epidemiology and outcomes of oste-
oporotic fractures. Lancet 2002, 359:1761-1767.
the more detailed (secondary) the available data is, the 7. Langley J, Stephenson S, Cryer C, Borman B: Traps for the unwary
more alternative perspectives are needed in the analyses of in estimating person based injury incidence using hospital
discharge data. Inj Prev 2002, 8:332-337.
such data and in the reporting of the results of the analy- 8. Boufous S, Williamson A: Reporting of the incidence of hospital-
ses. ised injuries: numerator issues. Inj Prev 2003, 9:370-375.
9. Smith GS, Langlois JA, Buechner JS: Methodological issues in using
hospital discharge data to determine the incidence of hospi-
Competing interests talized injuries. Am J Epidemiol 1991, 134:1146-1158.
The author(s) declare that they have no competing inter- 10. Brophy S, John G, Evans E, Lyons RA: Methodological issues in the
ests. identification of hip fractures using routine hospital data: a
database study. Osteoporos Int 2006, 17:405-409.
11. Gissler M, Haukka J: Finnish health and social welfare registers
Authors' contributions in epidemiological research. Norsk Epidemiologi 2004, 14:113-120
[http://www.ub.ntnu.no/journals/norepid/2004-1/2004(1)_21-
RS carried out the whole research process. Gissler.pdf].
12. Keskimäki I, Aro S: Accuracy of data on diagnoses, procedures
Acknowledgements and accidents in the Finnish Hospital Discharge Register. Inj
J Health Sci 1991, 2:15-21.
This work was financially supported by the research grant from the Acad- 13. Leppälä JM, Virtamo J, Heinonen OP: Validation of stroke diagno-
emy of Finland. The funding agreement ensured the author's independence sis in the National Hospital Discharge Register and the Reg-
in designing the study, interpreting the data, writing, and publishing the ister of Causes of Death in Finland. Eur J Epidemiol 1999,
report. 15:155-160.
Page 10 of 11

14. Lüthje P, Nurmi I, Kataja M, Heliövaara M, Santavirta S: Incidence of 39. Giversen IM: Time trends of age-adjusted incidence rates of
pelvic fractures in Finland in 1988. Acta Orthop Scand 1995, first hip fractures: a register-based study among older peo-
66:245-248. ple in Viborg County, Denmark, 1987-1997. Osteoporos Int
15. Makikyro T, Isohanni M, Moring J, Hakko H, Hovatta I, Lonnqvist J: 2006, 17:552-564.
Accuracy of register-based schizophrenia diagnoses in a 40. Brennan nee Saunders J, Johansen A, Butler J, Stone M, Richmond P,
genetic study. Eur Psychiatry 1998, 13:57-62. Jones S, Lyons RA: Place of residence and risk of fracture in
16. Pajunen P, Koukkunen H, Ketonen M, Jerkkola T, Immonen-Raiha P, older people: a population-based study of over 65-year-olds
Karja-Koskenkari P, Mahonen M, Niemela M, Kuulasmaa K, Palomaki in Cardiff. Osteoporos Int 2003, 14:515-519.
P, Mustonen J, Lehtonen A, Arstila M, Vuorenmaa T, Lehto S, Miet- 41. Sanders KM, Nicholson GC, Ugoni AM, Seeman E, Pasco JA, Kotow-
tinen H, Torppa J, Tuomilehto J, Kesaniemi YA, Pyorala K, Salomaa V: icz MA: Fracture rates lower in rural than urban communi-
The validity of the Finnish Hospital Discharge Register and ties: the Geelong Osteoporosis Study. J Epidemiol Community
Causes of Death Register data on coronary heart disease. Eur Health 2002, 56:466-470.
J Cardiovasc Prev Rehabil 2005, 12:132-137. 42. Sugarman JR, Connell FA, Hansen A, Helgerson SD, Jessup MC, Lee
17. Poikolainen K: Accuracy of hospital discharge data: five alco- H: Hip fracture incidence in nursing home residents and
hol-related diseases. Drug Alcohol Depend 1983, 12:315-322. community-dwelling older people, Washington State, 1993-
18. Sund R, Nurmi-Lüthje I, Lüthje P, Tanninen S, Narinen A, Keskimäki I: 1995. J Am Geriatr Soc 2002, 50:1638-1643.
Comparing properties of audit data and routinely collected 43. Chevalley T, Herrmann FR, Delmi M, Stern R, Hoffmeyer P, Rapin
register data in case of performance assessment of hip frac- CH, Rizzoli R: Evaluation of the age-adjusted incidence of hip
ture treatment in Finland. Methods Inf Med 2007, 46:(in press). fractures between urban and rural areas: the difference is
19. Kannus P, Niemi S, Parkkari J, Palvanen M, Vuori I, Järvinen M: Hip not related to the prevalence of institutions for the elderly.
fractures in Finland between 1970 and 1997 and predictions Osteoporos Int 2002, 13:113-118.
for the future. Lancet 1999, 353:802-805. 44. Crawford JR, Parker MJ: Seasonal variation of proximal femoral
20. Lüthje P, Nurmi I, Palvanen M, Kannus P: Epidemiology and pre- fractures in the United Kingdom. Injury 2003, 34:223-225.
vention possibilities of proximal femur fracture [in Finnish]. 45. Bulajic-Kopjar M: Seasonal variations in incidence of fractures
Suomen Lääkärilehti 2001, 56:1615-1619 [http://www.fimnet.fi/cl/ among elderly people. Inj Prev 2000, 6:16-19.
laakarilehti/pdf/2001/SLL142001-1615.pdf]. 46. Sund R, Kauppinen S: How to determine the number of older
21. Heikkinen T, Jalovaara P: Four or twelve months' follow-up in people in long-term care using register-based data? (in Finn-
the evaluation of functional outcome after hip fracture sur- ish). Sos Laaketiet Aikak 2005, 42:137-144.
gery? Scand J Surg 2005, 94:59-66. 47. Schlaud M, Brenner MH, Hoopmann M, Schwartz FW: Approaches
22. Sund R, Liski A: Quality effects of operative delay on mortality to the denominator in practice-based epidemiology: a criti-
in hip fracture treatment. Qual Saf Health Care 2005, 14:371-377. cal overview. J Epidemiol Community Health 1998, 52 Suppl
23. Solon JA, Feeney JJ, Jones SH, Rigg RD, Sheps CG: Delineating epi- 1:13S-19S.
sodes of medical care. Am J Public Health Nations Health 1967, 48. Sund R: Hip fracture incidence in Finland 1998-2002 (in Finn-
57:401-408. ish). Duodecim 2006, 122:1085-1091 [http://www.terveysportti.fi/
24. Hornbrook MC, Hurtado AV, Johnson RE: Health care episodes: xmedia/duo/duo95708.pdf].
definition, measurement and use. Med Care Rev 1985, 49. Laukkanen P, Karppi P, Heikkinen E, Kauppinen M: Coping with
42:163-218. activities of daily living in different care settings. Age Ageing
25. Rosen AK, Mayer-Oakes A: Episodes of care: theoretical frame- 2001, 30:489-494.
works versus current operational realities. Jt Comm J Qual 50. Murtagh KN, Hubert HB: Gender differences in physical disabil-
Improv 1999, 25:111-128. ity among an elderly cohort. Am J Public Health 2004,
26. Kegler SR: Reporting incidence from a surveillance system 94:1406-1411.
with an operational case definition of unknown predictive 51. Meriläinen S, Nevalainen T, Luukinen H, Jalovaara P: Risk factors for
value positive. Epidemiol Perspect Innov 2005, 2:7. cervical and trochanteric hip fracture during a fall on the hip.
27. Brameld KJ, Holman CD, Lawrence DM, Hobbs MS: Improved Scand J Prim Health Care 2002, 20:188-192.
methods for estimating incidence from linked hospital mor- 52. Sulander T, Helakorpi S, Nissinen A, Uutela A: Health behaviour
bidity data. Int J Epidemiol 2003, 32:617-624. among Finnish Elderly, Spring 2003, with trends 1993-2003.
28. Brody JA, Schneider EL: Diseases and disorders of aging: an Helsinki, National Public Health Institute; 2004.
hypothesis. J Chronic Dis 1986, 39:871-876.
29. Brody JA, Grant MD: Age-associated diseases and conditions:
implications for decreasing late life morbidity. Aging (Milano)
2001, 13:64-67.
30. Holliday R: The close relationship between biological aging
and age-associated pathologies in humans (with discussion).
J Gerontol A Biol Sci Med Sci 2004, 59:B543-53.
31. Khaw KT: Epidemiological aspects of ageing. Philos Trans R Soc
Lond B Biol Sci 1997, 352:1829-1835.
32. Brenner H, Arndt V: Epidemiology in aging research. Exp Geron-
tol 2004, 39:679-686.
33. Yin D, Chen K: The essential mechanisms of aging: Irreparable
damage accumulation of biochemical side-reactions. Exp Ger-
ontol 2005, 40:455-465. Publish with Bio Med Central and every
34. Holliday R: The multiple and irreversible causes of aging. J Ger-
ontol A Biol Sci Med Sci 2004, 59:B568-72. scientist can read your work free of charge
35. Gavrilov LA, Gavrilova NS: The reliability theory of aging and "BioMed Central will be the most significant development for
longevity. J Theor Biol 2001, 213:527-545. disseminating the results of biomedical researc h in our lifetime."
36. Walter LC, Lui LY, Eng C, Covinsky KE: Risk of hip fracture in dis-
abled community-living older adults. J Am Geriatr Soc 2003, Sir Paul Nurse, Cancer Research UK
51:50-55. Your research papers will be:
37. Stel VS, Pluijm SM, Deeg DJ, Smit JH, Bouter LM, Lips P: Functional
limitations and poor physical performance as independent available free of charge to the entire biomedical community
risk factors for self-reported fractures in older persons. Oste- peer reviewed and published immediately upon acceptance
oporos Int 2004, 15:742-750.
38. Kanis JA, Johnell O, De Laet C, Jonsson B, Oden A, Ogelsby AK: cited in PubMed and archived on PubMed Central
International variations in hip fracture probabilities: implica- yours — you keep the copyright
tions for risk assessment. J Bone Miner Res 2002, 17:1237-1244.
Submit your manuscript here: BioMedcentral
http://www.biomedcentral.com/info/publishing_adv.asp
Page 11 of 11

IV
-"- / / Ê1- /ÊÓääx\Ê{ÓÊ£ÎÇq£{{
-LKB@DLJK8@;FK
+UINKALASKEAIKËËNTYNEIDENPITKËAIKAISASIAKKAIDEN
MËËRIËREKISTERITIETOJENPERUSTEELLA
4ËSSËTUTKIMUKSESSAKUVATAANKAHDELLAERITAVALLALASKETUTIKËËNTYNEIDENPITKËAIKAISASIAKKAIDEN
LUKUMËËRËTIKË JASUKUPUOLIRYHMITTËINVUOSILLEn4UTKIMUSOSOITTAAETTËPELKËSTËËN
YHDENVUODENASIAKASLASKENTATIEDOISTALASKEMALLAPITKËAIKAISASIAKKAIDENLUKUMËËRËONSELVËSTI
PIENEMPIKUINSILLOINKUNOTETAANHUOMIOONMYÚSMYÚHEMPIENASIAKASLASKENTOJENJAPËËTTYNEIDEN
HOITOJAKSOJENTIEDOT+UMPIKAANTAPAEITËYSINVARMASTIOLETOISTAOIKEAMPIJARIIPPUUKIN
KËYTTÚTARKOITUKSESTAKUMPAATAPAAKANNATTAAKËYTTËË
REIJO SUND, SARI KAUPPINEN
*/($!.4/ 6À>ÃÊÌ>ÃÌÊÃ°Ê°ÊBBÌÞiìÊ
«ì}ÃÃÃ>ÊÌÕÌÕÃÃÃ>ÊÊÕÃiÊÌ>À ÃÃ>>Ê>ÊÌiÀÛiÞÃ«>ÛiÕÌÊÓääÓ®Ê«ÌB>>Ã>Ã>
«iiÃÌ>ÊÃÕÌiÕÌÌ>>ÊÌ>«>ÌÕiÊÕÕBBÀBÊ >>ÌÊ>ÃiÌ>>ÊÛ>ÊÞÃiiÃÃBÊiÛ>ÊÌ>ÃÌÛÕ
ÌBÊÛ>ÃÌ>>ÛÊÀÃÛBiÃÌ°Ê À>ÃÃÌ>ÊÌ>ÃÌÌi ìÊ>Ã>>Ã>ÃiÌ>ÌiÌiÊ>ÛÕ>°Ê/BÊ«ÌB
Ì>ÃÌ>ÊÞÌÞÞÊi«ÃÌÊ-ÕiÊÛBiÃÌÀ>i >>Ã>Ã>>ÃÊBBÀÌiBBÊiÊiÌ]Ê>Ê
iÌÌ>ÊÕÛ>>Û>ÊÌiÌ>ÊBÊ>ÊÃÕÕ«ÕÀÞÌ Ê>Ã>>Ã>ÃiÌ>ÌiÌiÊ«iÀÕÃÌii>ÊÛ>ÃÃ>Ê
ÌB°ÊBBÌÞiÌBÊÃiÛÃÃ>ÊÌ>À>ÃÌiÕÃÃ>ÊÊ iÛ>Ê«ÌB>>Ã`Ê«BBÌÃÊÌ>ÊÌ>Ê>Ã>>Ã
ÕÌiÊÕÃiÊÌ>ÀÌÕÃiÕ>ÃÌ>ÊiÀÌi>Ê >ÃiÌ>ÌÕÃiÊÌiÌiÊÕ>>ÊÛ>ÌÊÃiÊ
BÊ>ÊÃÕÕ«ÕiÊÃBÃÊ`ÊiÀÊ«ÀÌ>iÊ iiÃÃBÊiiÌÊÞÌB>ÃÃiÃÌÊ`ÃÃ>ÊÞÃiiÃÃBÊ
ÕÕÕÛ>Ê>Ã>>Ì>ÊÌÃÃÌ>>°Ê ÀÌÞÃiÊÌBÀiBÊ iÛ>ÃÃ>ÊÌ«>>ÃÃ>ÊÞÊäÊ«BÛBB°
ÀÞBÊÕ`ÃÌ>Û>ÌÊ«ÌB>>ÃiÃÃ>ÊÞ«BÀÛÕ ÕÊÃiÛÌiÌBBÊÀÃÛBiÃÌBÊi«ì}ÃÌ>Ê
À>ÕÌÃiÃÃ>Ê`ÃÃ>ÊiÛ>ÌÊBBÌÞiiÌ°Ê/BBÊ ÌÕÌÕÃÌ>ÊÛ>ÀÌi]ÊÊÌBÀiBBÊÃ>>`>ÊÌ>ÛÌiÌÌÕ>Ê
ÀÞBÊÕÛ>ÃÌ>>Ê«>ÛiÕBÀiÃÌiBÃÌBÊÀ««ÕÛ>ÃÌ>Ê ÞÃiiÃÃBÊ iÛ>Ê ÛBiÃÌÀÞBÊ >`Ã>Ê
»ÕÕÌÃÌ>»ÊÕ>]Ê>Ê>iÊB >ÌÌ>Û>ÃÌ°Ê BÊiÊÃ>>ÌÌ>>Ê>ÊÌ>À«iiÊBÞÌ
ÃiÌiÊiÞ«BÀÃÌÊÊÛ>ÀÃÊÃ>>>Ì>i°Ê ÌBBÊÞÛBÃÊÞÃÊÞi«iÊ>Ã>>Ã>ÃiÌ
*ÌB>>ÃÌ>Ê Þ«BÀÛÕÀ>ÕÌÃÌ>Ê Ì>Ê iÊ>Ê«BBÌÌÞiìÊÌ>ÃiÊÌiÌ>°
Ã>>Û>ÌÊBBÌÞiiÌÊ>Ã>>>ÌÊÛÛ>ÌÊ>ÊÛ>> /BBÊÌÕÌÕÃiÊÌ>ÀÌÕÃi>ÊÊÕÛ>Ì>Ê
`ÃÃ>]ÊÌiÀÛiÞìÕÊÛÕìÃ>ÃÌ`ÃÃ>Ê BBÌÞiìÊ«ÌB>>Ã>Ã>>ìÊÕÕBB
Ì>ÊÛ>>Ì>ÊÛ>ÃÌ>>Û>ÊÌiÃÌiÌÕÊ«>ÛiÕ ÀBÌÊBÊ>ÊÃÕÕ«ÕÀÞÌÌBÊÛÕÃiÊ£Çq
>ÃÕÃiÊ«ÀÃÃB°Ê BìÊ«>ÛiÕiÊ>Ã>>ÃÌ>Ê ÓääÓÊÃiBÊÃiÛÌÌBB]ÊÌiÊ>>ÃÌ>>ÊÞìÊÛÕ
Ã>>`>>ÊÌ>ÃÌ>ÃÌ>ÊÌiÌ>Ê-Ì>iÃÊ ìÊ>Ã>>Ã>ÃiÌ>Ìi`ÃÌ>Ê>ÃiÌÌÕÊ«ÌB>
ÌÌÕÃÀiÃÌiÀiÃÌBÊ-Ì>iÃÊÓääÎ®°ÊiÃÌÊ >Ã>Ã>>ìÊ ÕÕBBÀBÊ iÀ>>Ê ÃÌB]Ê >Ê
iÀBÌBBÊ ÛÕÃÌÌ>Ê >ìÊ ÌÞÞ««ÃBÊ ÌiÌ>\Ê Ã>>`>>]ÊÕÊÌiÌ>>ÊÕÊÞÃÊÞi
ÛÕìÊ«ÕÃÃ>ÊiÃiÊiìÊÌ>ÃiÊ «iÊ>Ã>>Ã>ÃiÌiÊ>Ê«BBÌÌÞiìÊÌ>
Ìi`ÌÊÊÃ>ÌÕÌÊ>Ã>>Ã>ÃiÌ>Ìi`Ì®ÊÃiBÊ ÃiÊÌi`Ì°Ê
«BBÌÌÞiìÊÌ>ÃiÊÌi`Ì°ÊÀÞ«Ì>ÌÕÊÞ
ÃÌÌÕÃÌiiÊ>ÛÕ>ÊÊ>`ÃÌ>ÊÞ`ÃÌBBÊ !).%)34/*!-%.%4%,-¯4
>ÊÃ>>Ê>Ã>>>ÊÌ>ÃÌÊÌÃÃ>°Ê /ÕÌÕÃÌ>ÊÛ>ÀÌiÊÃÃ>>Ê>ÊÌiÀÛiÞìÕÊ
iÃÌÊ>ÛÕ>ÊÛ`>>ÊÞÃÊ>Ãi>Ê>Ã>>Ã ÌÌÕÃÀiÃÌiÀÃÌBÊ«ÌÌÊ>ÊÞÊxä
BBÀBÌÊ«BÛBÌ>ÃiÃÌ° ÛÕÌ>ìÊ>Ã>>ìÊ«ÌB>>Ã`ÊÀÌii


ÀÌÊÌBÞÌÌBÛBÌÊÌ>ÃÌÊÃiBÊ>Ã>>Ã>ÃiÌ>Ìi ÛÕÃiÊ£ÇqÓääÓÊÛiÃBÊ«BÛB°Ê ÃÊÊ
`ÌÊiÌÌBÊ«BBÌÌÞiìÊÌ>ÃiÊÌi`Ì®ÊiÀ À>«ÀÌÌÕÊÃ>À>>>ÃÃ>ÊÌ>ÊÌiÀÛiÞÃiÃÕÃiÊÛÕì
ÃÃ>À>>`ÊÛÕìÃ>ÃÌ`ÃÃ>ÊÃ>À>> Ã>ÃÌ`ÃÃ>ÊiÛiÊÕÕBBÀBÌÊ«iiÊÞ
>®]Ê«iÀÕÃÌiÀÛiÞìÕÊÛÕìÃ>ÃÌ`ÃÃ>Ê ÃiiÃÃBÊiÛ>ÊÛÕìÊ>Ã>>Ã>ÃiÌ>ÌiÌiÊ«i
ÌiÀÛiÞÃiÃÕÃ®]ÊÛ>>`ÃÃ>ÊÌ>ÊÛ>ÃÌ>>Û>ÃÃ>Ê ÀÕÃÌii>Ê>ÊÃiÊBiiÊÊ>iÃÌÊ«iÀÕÃÌii>°Ê
ÌiÃÌiÌÕÃÃ>Ê«>ÛiÕ>ÃÕÃiÃÃ>ÊÛÕÃÌ>Ê£Çq >ÃiÌ>ÌiÌiÊÞÌiÞìÃÃBÊÊiÀÌÌÞÊBÞ
ÓääÓ°ÊÌ>Ã>Ê>Ê>Ã>>Ã>ÃiÌ>ÌiÌ>ÊÊ ÛÊÞÃÊÃÕÕÃÊÊ>iÃÌÊ«iÀÕÃÌii>Ê>Ãi
ÞÌiiÃBÊxÎÓÊ£È° ÌÕÃÌ>ÊÕÕBBÀÃÌB°Ê6>ÃÌ>>Û>ÌÊÌi`ÌÊÛ>ÌÊÞÃÊ
iÃÌÃÌ>Ê>ÃiÌÌÊÛÕÃiÊ£ÇqÓääÓÊÛ Û>>ìiÊÌ>ÊÛ>ÃÌ>>Ûi°Ê6iÃÃÃBÊÃ>À>
iÃÌiÊ«BÛiÊ«ÌB>>Ã>Ã>>ìÊÕÕBB iÃÃ>ÊÛ>ÌÊÊ««Õ>>ÌÊÕÕBBÀBÊ>Ê
ÀBÌÊBÞÌÌBiÊ«iBÊÌ>À>ÃÌiÕÛÕìÊ>Ã>>Ã ìÊ>Ã>>ìÊÃÕÕÃ]ÊÌ>Ê>iÃÌÊ«iÀÕÃÌii>Ê
>ÃiÌ>ÌiÌ>Ê"®Ê>ÊÛ>ÃÌ>>Û>ÌÊÕÕBBÀBÌÊÌBÞ Û>ÌÊÃ>>>Ê>>>ÊÃiBÊÛÕìÃ>ÃÌ>ÊiÌÌBÊ
ìÌBiÊÌ>À>ÃÌiÕÛÕìÊ>Ã>>Ã>ÃiÌ>ÌiÌ>Ê Û>>`ÃÃ>ÊÌ>ÊÌiÃÌiÌÕÃÃ>Ê«>ÛiÕ>ÃÕÃiÃÃ>°
«BBÌÌÞÃÌÕÃ>ÊÃiBÊÞiBÊ>Ã> />ÕÕìÊ£Ê>ÊÓÊÌi`ÃÌ>Ê>Û>Ì>>]ÊiÌÌBÊ
>Ã>ÃiÌ>Ìi`>Ê ®°Ê/>À>ÃÌiÕ«BÛBÊBÊ «iÊ>ÃiÌ>ÌiÌÊ«iÀÕÃÌÕÛ>ÌÊÕÕBBÀBÌÊ
«ÞÃÌÞÌÌÊ«BBÌÌiiBBÊÌ>À>ÃÌ]ÊÃ>ÊBÊ Û>ÌÊ£äqÎäÊ¯Ê«ii«BÊÕÊÛ>ÃÌ>>Û>ÌÊ>ÃiÌ
Ì>ÃÊ>ÕÃÃ>Ê>Ê«ÕÃÃ>ÊÃiBÊÌ>ÃÊ>ÕÊ ÌÕ>ÊÊ>iÃÌÃÌ>Ê>ÊiÌÌBÊÃÕÕÃÃÃ>ÊÊÃi
>Ê««Õ«BÛBÊÛ>ÌÊÌi`ÃÃ>°ÊÀÞ«Ì>ÌÌÕ>ÊÞÃ ÛBBÊÛ>ÌiÕ>ÊiÀÊÀÞiÊ>ÊÛÕÃiÊÛBB°Ê/>À
ÌÌÕÕÃÌ>ÊBÞÌÌBiÊÛ>ÀÃÌiÌÌ]ÊiÌÌBÊ> >ÃÌiÌ>iÃÃ>ÊÞÃÌÌBÃÌiÊÕÕÌÌÕiÊÛ>ÕÌÕÃ>Ê
iÊ>Ã>>ÃÊ>ÃiÌÌÊÕ>Ê«BÛBBÊÕ>>Ê ÃÕÕÃ]Ê>ÌÌ>ÛÕÕÃÊ«>À>iiÊÛ>>Ê
Û>ÊiÀÀ>° BÊ>ÃÛ>iÃÃ>°Ê>ÌÌ>ÛÕÕÃÊ«>À>iiÊÞÃÊÌ>À>ÃÌi
iÃÌÊ>iÌÌÊÃBÃÊÃÕÕ«Õi]ÊBÊ>Ê> ÕÛÕìÊ>ÃÛ>iÃÃ>°ÊiBÊ>ÌÌ>ÛÕÕÃÊÊÕ
ÌÃÌÞÞ«ÊÕ>>ÊÀÞ°Ê-ÕÕ«ÕÀÞBBÊ «Ê>ÊÛ>>ÌiÊÌ>ÊÛ>ÃÌ>>ÛiÊÃÕÕìÌÊ
ìÃÌÊ\BÊÊrÊ£]ÓÆÊ£ÊrÊiÃÊ>ÊÓÊrÊ>i®]Ê Û>ÌÊ«ii«B°ÊÃ>ÊÕÕÌÌÕiÊÛBBÊÊÕ
BÀÞBBÊ\BÊÊrÊ£]ÊÓ]ÊÎ]Ê{ÆÊ£ÊrÊxäqÈ{]ÊÓÊrÊÈxq ÌiÊÃiÛBÊÃBBÕ>ÃÕÕÃ>]Ê>ÌiÌ
Ç{]ÊÎÊrÊÇxqn{Ê>Ê{ÊrÊÞÊnxÛÕÌ>>Ì®Ê>Ê>ÌÃ ÌÊÃÕÕÃiÊBÞÌÌBÞÌÞÃÌBÊÞÃÊÕÕÌÌÕ
ÌÞÞ««BÊìÃBÊÊÊrÊ£]ÊÓÆÊ£ÊrÊÌiÀÛiÞÃiÃÕÃÊÌ>Ê >>Ê>ÛÕ>°
Ã>À>>>Ê>ÊÓÊrÊÛ>>ÌÊÌ>ÊÛ>ÃÌ>>Û>®°Ê BBÊ />ÕÕÃÃ>ÊÎÊÊÀ>«ÀÌÌÕÊ>iÃÌ>ÊÌ>À
iÀBÊÛÕìÊÝÊ>Ã>>Ã>ÃiÌ>ÌiÌÊ ÌÕÃiÕ>ÃÊÕÛ>>Û>Ê>ÊÌÕÃiÌ°Ê
«iÀÕÃÌÕÛ>ÊÕÕBBÀBÊÀÞBÃÃBÊÊÊ"Ý®Ê>Ê "ÃÕÕÌÌ>ÊÊÃÃÊÃiÌiÌÌÞÊBÀÞBB]Ê>ÌÃÌÞÞ«
Ê>iÃÌ>ÊÞ`ÞÌBiÊÃ>>ÌÕÊÛ>ÃÌ>>Û>ÊÕÕ «B]ÊÃÕÕ«ÕÌ>ÊÃiBÊBÊ>Ê>ÌÃÀÞÌÌBÃÌBÊ
BBÀBÊ Ý®°Ê ìiiÊ«Ý®ÊrÊ"Ý®É Ý®ÊÊ ÛÕÃiviÌBÊÕÛ>>ÛiÊÕÕÌÌÕiÊ>ÛÕ>°Ê>Ê
>ÃiÌ>ÌiÌiÊÃÕÕÃÊÊ>iÃÌÊ«iÀÕÃÌii>Ê ÞÌiiÃ«ÛÕÕÃÊ>iÃÌÊ>ÃÃ>ÊÌ>ÌÌÕ>ÊχÓ
Ã>>`ÕÃÌ>ÊÕÕBBÀBÃÌBÊÀÞBÃÃBÊÊÛÕ>ÊÝ°Ê ÌiÃÌBÊÊiÀ>iÊ«ÊrÊä]{Ç®]ÊiBÊBBÊ
/>ÛÌÌii>ÊÊÕ`ÃÌ>>Ê>]ÊÌ>ÊBÞÌ ÌiÀ>ÌÌiÀÊ ÃBBiÊ iBBÊ «>À>Ì>ÕÌÊ
ÌBiÊÃÊ>`ÃÌ>Ê>ÀÛ`>ÊÀÞiÊÊÛBÃBÊ >>ÊiÀÌÌBÛBÃÌ°Ê*iÀÕÃÀÞBBÊÊBÞÌiÌÌÞÊ
iÀ>ÊÃÕÕìÃÃ>Ê«Ê>ÊÌÃ>>Ì>ÊiÕÃÌ>>Ê ÊrÊ"É Û>>`ÃÃ>ÊÌ>ÊÛ>ÃÌ>>Û>ÃÃ>Ê>ÃÕÛ>ÊÞÊn{
« Ê >ÕÌÕiÊ ÛÕìiÊ Ý]Ê ÕÊ " Ê ÌÕiÌ>>°Ê ÛÕÌ>Ì>Ê>Ã>Ê>ÊÛÕÃÕÕÌÌÕ>Ê«iÀÕÃÛÕìÃÊ
"iÌÌ>>>]ÊiÌÌBÊ«ÊHÊ£É ®*ÃÃλ ®]ÊÃ ÊÛ>ÌÌÕÊÛÕÃÊÓäää°ÊBÀÞB]Ê>ÌÃÌÞÞ«Ê>Ê
Ã>ÊλÊÊÌiÀiiÌÌiÊÃÕÕÃÊÀÞBÃÃBÊÊ>Êλ Ê ÃÕÕ«ÕiÊ«BBÛ>ÕÌÕÃiÌÊÛ>ÌÊÌ>ÃÌÃiÃÌÊ
Ê"\Ê`ÌÕÃ>ÀÛ]Ê«BB`ÞÌÌÊÞiÃÌiÌÌÞÞÊ} iÀÌÃiÛB]ÊÛ>>ÊiÊÌÀ`>>ÊÌÃÌiÃ>Ê
i>>ÀÃiiÊ*ÃÃÀi}ÀiÃÃ>°ÊÃ>ÊÃi >ÊÀÞÌÌBÃiÊÛÕÃÕÕÌÌÕ>ÊÃÕÌii°Ê>ÌÌ>
ÌiÌÌBÛBBÊÛ>ÌÊÃÕÕìÌ]Ê«ÌÊ>ÃÃ>Ê«>ÌÌ>>Ê ÛÕÕÃÊÃÃÊ«>À>iiÊÃiÛBÃÌÊBÊÞÌBÊ«ÊÊä]äää£®°Ê
ÕÌ>Ê>Û>Ì>Ê«ÊÕÕBBÀBBÊ °Ê,ÞÌ "ÃÕÕìÌÊÛ>ÌÊÃ«>ÊÃ>À>>ìÊ>ÊÌiÀÛiÞÃiÃ
ÌiÞÕÕÌÌÕiÊÃBÃÊ>ÃÃ>ÊBÞÌiÌÌÊÃiÌÌB ÕÃÌiÊÛÕìÃ>ÃÌ>ÊÕÊÛ>>ìÃÃ>ÊÌ>Ê
ÛBBÊÕÕÌÌÕ>>ÊÞÃÊÛÕÌÌ>°Ê>>ÊiÃÌÌ> Û>ÃÌ>>ÛÃÃ>Ê«ÊÊä]äää£®Ê>Ê>ÌÌ>ÛÕÕÃÊÊ«>Ài«Ê
iÃÃ>ÊBÞÌiÌÌÊÛÕÃiÊ£ÇqÓää£Ê>Û>Ì>Ê>Ê >ÃÌiÊÃ>Ì>Ê«ÊrÊä]ääÓ®°Êi>>ÀiÊÛÕÃivi
«ÞÀÌÌÊ iÕÃÌ>>>Ê ÛÕìÊ ÓääÓÊ ÃÕÕìÌÊ ÌÊÊ>ÃÛ>Û>ÊÛ>>ÌiÊÌ>ÊÛ>ÃÌ>>ÛiÊÃ>Ì>Ê
«ÓääÓ®Ê>iÊÀÞiÊ° >ÃÃ>ÊBÀÞÃÃB°Ê->À>>ìÊ>ÊÌiÀÛiÞÃiÃ
ÕÃÌiÊÛÕìÃ>ÃÌ>ÊÊ>ÃiÛ>ÊÌÀi`ÊÕÀ
45,/+3%4 >ÃÃ>ÊBÀÞBÃÃB]ÊÕÌÌ>ÊÕÃÃ>ÊBÀÞÃÃBÊ
/>ÕÕÃÃ>Ê£Ê>ÊÓÊÊiÃÌiÌÌÞÊ«ÌB>>Ã>Ã> Ì>ÃÌÃiÃÌÊiÀÌÃiÛBBÊi>>ÀÃÌ>ÊÌÀi`BÊiÊ
>ìÊÕÕBBÀBÌÊBÊ>ÊÃÕÕ«ÕÀÞÌÌBÊ iÊ>Û>ÌÌ>ÛÃÃ>°


/>ÕÕÊ£°
BBÌÞiìÊ«ÌB>>Ã>Ã>>ìÊBBÀBÌÊ-ÕiÃÃ>Ê£ÇqÓääÓ]ÊiiÌ°
xäqÈ{ÊiiÌ
/iÀÛiÞÃiÃÕÃÊÌ>ÊÃ>À>>> 6>>ÌÊÌ>ÊÛ>ÃÌ>>Û> i>Ì
>ÃiÌ>£ "ÃÕÕÃÓ >Î >ÃiÌ>£ "ÃÕÕÃÓ >Î 9ÌiiÃB{ *BBiBÃBx
£Ç ÈÓÈ nä]{Ê¯ ÇÇ ÎnÎ Çx]£Ê¯ x£ä £Ón{ ä]{Ê¯
£n xÇÇ Ç]nÊ¯ ÇÓÎ Î{{ È]ÈÊ¯ {{ £Ó£{ ä]ÓÊ¯
£ È£ä n£]ÓÊ¯ Çx£ ÎÈ{ È]ÈÊ¯ xÓÎ £ÓÇÓ ä]ÓÊ¯
Óäää Èän Ç{]ÇÊ¯ n£{ În£ Çä]nÊ¯ xÎn £Î{Ç ä]{Ê¯
Óää£ xx Ç{]nÊ¯ Çx {Ó n£]Ê¯ Èä£ £Î{ ä]£Ê¯
ÓääÓ xÇÓ È
ÇÎ]{Ê¯È È
ÇÇÈ x£ È
ÇÇ]Ê¯È È
ÈÈÈÈ È
£{ÎnÈ
ÈxqÇ{ÊiiÌ
>ÃiÌ> "ÃÕÕÃ > >ÃiÌ> "ÃÕÕÃ > 9ÌiiÃB *BBiBÃB
£Ç ÇÓ nä]xÊ¯ £ÓäÇ £ÓÓ£ Çn]xÊ¯ £xxÈ ÓÇÎÇ ä]Ê¯
£n Èx nÎ]£Ê¯ ££È£ £Ó{Î nä]£Ê¯ £xx£ ÓÈÎ ä]ÇÊ¯
£ {Ç nÎ]ÎÊ¯ ££ÎÇ £ÓÇx n£]xÊ¯ £xÈ{ ÓÈnÈ ä]ÈÊ¯
Óäää n{ n{]ÎÊ¯ £äÈä £ÓÇ{ nä]Ê¯ £xÇx ÓÈÓx ä]{Ê¯
Óää£ Ó nÓ]ÇÊ¯ ££ÓÎ £ÓÈ{ nx]{Ê¯ £{nä Óxnx ä]ÇÊ¯
ÓääÓ n{{ nÎ]ÇÊ¯ £ään £Îxn nÎ]Ê¯ £È£ ÓÈ£{
Çxqn{ÊiiÌ
£Ç £În nÓ]{Ê¯ £ÈÇ Ó££Î nä]ÓÊ¯ ÓÈÎ{ {ÎäÈ ä]ÈÊ¯
£n £ÎÇ{ nÎ]äÊ¯ £Èxx Ó£Î£ Ç]£Ê¯ ÓÈ{ {ÎÓ£ ä]ÈÊ¯
£ £{Ó£ n{]ÈÊ¯ £ÈÇ ÓÎxÎ nÓ]ÇÊ¯ Ón{x {xäÓ ä]xÊ¯
Óäää £{{Ó nÎ]ÓÊ¯ £ÇÎ{ Ó{È{ nÓ]ÇÊ¯ ÓÇn {ÈnÎ ä]ÈÊ¯
Óää£ £{ÎÇ n{]{Ê¯ £ÇäÓ ÓÈxÈ nx]ÈÊ¯ Î£ä£ {ÇÇä ä]ÇÊ¯
ÓääÓ £{n£ nx]ÎÊ¯ £ÇÎÈ Ón£ä nx]ÎÊ¯ ÎÓx xääÈ
nxqÊiiÌ
£Ç {{ n{]ÇÊ¯ £££{ £nÎÎ n{]ÎÊ¯ Ó£Çx ÎÓÈ{ ä]nÊ¯
£n {£ nn]ÎÊ¯ £äÈÈ £Î n{]{Ê¯ ÓÓÇ ÎÎ{£ ä]ÇÊ¯
£ äÇ n{]ÈÊ¯ £äÇÓ £Çx nÎ]ÓÊ¯ ÓÎÇx Î{Ó ä]xÊ¯
Óäää n n{]xÊ¯ £äÈ{ ÓäÇn nÎ]nÊ¯ Ó{n£ Îxän £]£Ê¯
Óää£ nn nÇ]äÊ¯ £äÎÓ Ó£xÈ nÈ]nÊ¯ Ó{nÎ Î{Î ä]ÈÊ¯
ÓääÓ nÇ{ nn]ÎÊ¯ ää ÓÓxx nÈ]xÊ¯ ÓÈän Îxnä
£®ÊÞÃiiÃÃBÊiÛ>ÊÛÕìÊ>Ã>>Ã>ÃiÌ>ÌiÌiÊÕ>iÊ«ÌB>>Ã>Ã>>ìÊBBÀB
Ó®ÊÃ>>Ã>ÃiÌ>ÌiÌiÊÕ>iÊBBÀBÊÃÕÌiÕÌiÌÌÕ>ÊÕ>Ê«ÌB>>Ã`ÊÀÌiiÀÌÊÌBÞÌÌBÛBÊÌ
ÌÕÃ>ÊÌBÞìiÌÞÊ>iÃÌÊÕ>ÃiiÊBBÀBB
Î®Ê>iÊ«ÌB>>Ã`ÊÀÌiiÀÌÊÌBÞÌÌBÛiÊÌÌÕÃiÊ«iÀÕÃÌii>Ê>ÃiÌÌÕÊ«ÌB>>Ã>Ã>>ìÊ
BBÀB
{®Ê*ÌB>>Ã>Ã>>ìÊBBÀBÊÞÌiiÃBÊ«>ÛiÕÌÞÞ«ÊÕ>>ÊiÀÌÌii>ÌÌ>®
x®Ê ìÊ>Ã>>ìÊÃÕÕÃ]ÊÌ>Ê>iÃÌÊÕ>>ÊÛ>ÌÊÃ>>>Ê>>>ÊÃiBÊÌiÀÛiÞìÕÊÛÕìÃ>ÃÌ>Ê
iÌÌBÊÛ>>`ÃÃ>ÊÌ>ÊÛ>ÃÌ>>Û>ÃÃ>
È®Ê>Û`ÕÌÊÕÛÕÌÊÛ>ÌÊÛÕìÊÓääÓÊiÕÃÌiÌÌÕ>Ê>ÀÛ>Ê>ÌÃÊÃiÌÞÃÊÌiÃÌÃÌB®
/>ÕÕÃÃ>Ê£Ê>ÊÓÊÊÀ>«ÀÌÌÕÊÞÃÊÛÕ 0/($).4!
ìÊÓääÓÊiÕÃÌiÌÕÌÊÃÕÕìÌÊ>Ê>Ã>>ìÊÕ /BÃÃBÊÌÕÌÕÃiÃÃ>Ê>Û>ÌÌ]ÊiÌÌBÊ«iBÃÌBBÊ
ÕBBÀBÌÊ>Û`ÕÌÊÕÛÕÌ®°ÊÃ>ÊÃ>>>Ê> ÞìÊÛÕìÊ>Ã>>Ã>ÃiÌ>Ìi`ÃÌ>Ê>Ãi>
>>ÊÛÕìÃ>ÃÌ>Ê>ÊÛ>>`ÃÃ>ÊiÛÃÊ >Ê«ÌB>>Ã>Ã>>ìÊÕÕBBÀBÊÊÃiÛBÃÌÊ
iÀÌÌÞBÊÊiÃBBÀÊÊ«ÕÊ«ÀÃiÌÌ>]Ê «ii«ÊÕÊÃ]ÊÕÊÌiÌ>>ÊÕÊ
ÊiÕÃÌiÌÌÕ>Ê >ÃÕÕBBÀBBÊ À>ÌÌÕÊ ÞÃÊÞi«iÊ>Ã>>Ã>ÃiÌiÊ>Ê«BBÌÌÞ
Û>ÃÌ>>Û>ÃÌÊ>>Ã«BÊ>ÌÃÌÞÞ««Ì>ÃÌiÊi iìÊÌ>ÃiÊÌi`Ì°Ê/BBÊÊÌ>Ê
ÕÃÌiìÊÃÕ>>ÊBì° >`ÃÌ>ÊÃÞÞÌB°Ê


/>ÕÕÊÓ°Ê
BBÌÞiìÊ«ÌB>>Ã>Ã>>ìÊBBÀBÌÊ-ÕiÃÃ>Ê£ÇqÓääÓ]Ê>ÃiÌ°
xäqÈ{Ê >ÃiÌ
>ÃiÌ>£ "ÃÕÕÃÓ >Î >ÃiÌ>£ "ÃÕÕÃÓ >Î 9ÌiiÃB{ *BBiBÃBx
£Ç {xx Çn]ÇÊ¯ xÇn ÎÈ£ Ç{]ÎÊ¯ {nÈ £äx{ ä]Ê¯
£n {Î{ ÇÎ]Ê¯ xnÇ Î{{ ÇÎ]nÊ¯ {ÈÈ £ä{Ç ä]ÈÊ¯
£ {ÓÇ Çn]nÊ¯ x{Ó Îx ÇÎ]ÇÊ¯ {nÇ £äÓÎ ä]ÈÊ¯
Óäää {äÈ ÇÎ]ÇÊ¯ xx£ ÎÈÎ ÇÇ]ÓÊ¯ {Çä £ä£n ä]ÎÊ¯
Óää£ Î{ Ç£]{Ê¯ xxÓ {ÓÈ nÓ]ÈÊ¯ x£È £äÈ{ ä]{Ê¯
ÓääÓ {xÎ È
Ç{]ÈÊ¯È È
ÈäÇÈ {ÓÈ È
Ç]ÓÊ¯È È
xÎnÈ È
££ÎÈ
ÈxqÇ{Ê >ÃiÌ
£Ç £ÓnÈ n£]£Ê¯ £xnÈ £näÇ ÇÇ]Ê¯ ÓÎ£ ÎnÇÇ ä]ÇÊ¯
£n £ÓÓä nÓ]ÇÊ¯ £{ÇÈ £nä£ ÇÇ]ÈÊ¯ ÓÎÓÓ ÎÇÇÎ ä]ÇÊ¯
£ ££ÎÓ nÎ]Ê¯ £Î{ £nn{ nÎ]ÓÊ¯ ÓÓÈx ÎÈä£ ä]{Ê¯
Óäää ££ÎÓ nÓ]Ê¯ £ÎÈx £n££ n£]nÊ¯ ÓÓ£{ ÎxÈÎ ä]{Ê¯
Óää£ £ä{ä nÎ]ÇÊ¯ £Ó{Ó £än nÈ]ÎÊ¯ ÓÓ£Ó Î{Î ä]{Ê¯
ÓääÓ £ääÓ nx]£Ê¯ ££ÇÇ ÓäÎÈ nx]ÎÊ¯ ÓÎnÇ Îx{È
Çxqn{Ê >ÃiÌ
£Ç {äÓÇ nÎ]Ê¯ {ÇÇ ÈnÈ£ nä]xÊ¯ nxÓx £ÎÓÓÓ ä]nÊ¯
£n {äÇn nx]Ê¯ {Ç{x ÈÈÈ nä]nÊ¯ nÈÓÈ £ÎÓÈx ä]nÊ¯
£ ÎnäÎ nÈ]ÎÊ¯ {{äÇ Ç{£Ç nÎ]£Ê¯ nÓ£ £ÎÓxÓ ä]ÈÊ¯
Óäää ÎnäÓ nÈ]ÎÊ¯ {{äÈ ÇÈÈÎ n{]äÊ¯ £ÓÇ £Î{{x ä]ÇÊ¯
Óää£ ÎÇÇx nÈ]nÊ¯ {Î{n n££n nÇ]ÇÊ¯ ÓxÓ £ÎxÎÈ ä]xÊ¯
ÓääÓ ÎÈÎx nÈ]nÊ¯ {£ä nx£Ó nÈ]ÇÊ¯ n£x £ÎÎx
nxqÊ >ÃiÌ
£Ç {Ç{£ nÇ]ÎÊ¯ x{ÎÎ nn£Î nx]ÓÊ¯ £äÎ{È £xÈÇx ä]ÇÊ¯
£n {Ç{ n]ÈÊ¯ xÎxä £{È nx]äÊ¯ £äÇÈÈ £Èä£Ç ä]ÈÊ¯
£ {näÇ n]£Ê¯ xÎÈ xÎ{ nÈ]£Ê¯ ££äÇÓ £ÈÎnn ä]xÊ¯
Óäää {ÇÇÈ n]ÈÊ¯ xÎÓ Çäx n{]nÊ¯ ££{{£ £ÈÈ{Î ä]nÊ¯
Óää£ {n£{ ä]xÊ¯ xÎÓÓ £äÎxä n]nÊ¯ ££xÎÈ £ÈÇÈÇ ä]xÊ¯
ÓääÓ {{Çn n]nÊ¯ {nÇ £ä{Óä nÇ]Ê¯ £Ó{{È £ÇÎ{È
£®ÊÞÃiiÃÃBÊiÛ>ÊÛÕìÊ>Ã>>Ã>ÃiÌ>ÌiÌiÊÕ>iÊ«ÌB>>Ã>Ã>>ìÊBBÀB
Ó®ÊÃ>>Ã>ÃiÌ>ÌiÌiÊÕ>iÊBBÀBÊÃÕÌiÕÌiÌÌÕ>ÊÕ>Ê«ÌB>>Ã`ÊÀÌiiÀÌÊÌBÞÌÌBÛBÊÌ
ÌÕÃ>ÊÌBÞìiÌÞÊ>iÃÌÊÕ>ÃiiÊBBÀBB
Î®Ê>iÊ«ÌB>>Ã`ÊÀÌiiÀÌÊÌBÞÌÌBÛiÊÌÌÕÃiÊ«iÀÕÃÌii>Ê>ÃiÌÌÕÊ«ÌB>>Ã>Ã>>ìÊ
BBÀB
{®Ê*ÌB>>Ã>Ã>>ìÊBBÀBÊÞÌiiÃBÊ«>ÛiÕÌÞÞ«ÊÕ>>ÊiÀÌÌii>ÌÌ>®
x®Ê ìÊ>Ã>>ìÊÃÕÕÃ]ÊÌ>Ê>iÃÌÊÕ>>ÊÛ>ÌÊÃ>>>Ê>>>ÊÃiBÊÌiÀÛiÞìÕÊÛÕìÃ>ÃÌ>Ê
iÌÌBÊÛ>>`ÃÃ>ÊÌ>ÊÛ>ÃÌ>>Û>ÃÃ>
È®Ê>Û`ÕÌÊÕÛÕÌÊÛ>ÌÊÛÕìÊÓääÓÊiÕÃÌiÌÌÕ>Ê>ÀÛ>Ê>ÌÃÊÃiÌÞÃÊÌiÃÌÃÌB®
0)4+¯!)+!)3!3)!++55$%.+2)4%%2)4%)6¯46¯,44¯-¯44¯/,% ÌÕÕÊBÃÃBÊÌ>ÃÃÃ>ÊÌ>>ÕÌÕÛ>ÃÌÊÊ
6/)-!33!6)%,¯,!3+%.4!0¯)6¯.¯ Ì>Ã°Ê
*ÌB>>Ã>Ã>ÕÕÃÊiÊÛBÌÌBBÌÌBÊiÊÕÌÊÛ ÃÊÛ>ÀÃ>iÊ«BBÌÃÊ«ÌB>>Ã`ÃÌ>Ê
>ÃÃ>Ê>Ã>>Ã>ÃiÌ>«BÛBB]ÊÛ>>ÊÞi ÊÌiÌÞÊÛ>ÃÌ>Ê>ÃiÌ>«BÛBÊBii]Ê>Ã>>>Ê
«iÊÌÕÃÌiÊÕ>>ÊÌ>ÃÊÃÌÌ>ÕÌÕÕÊ iÊÕÕÕ>>Ê>ÊÕ>>Ê>ÃiÌ>«BÛBÊ«ÌB
ÞÊäÊÛÕÀ>ÕÌÌ>ÊiÃÌBÛBÃ°ÊÕÊ>iÃÌÃÌ>Ê >>Ã>Ã>>ìÊÕÕBBÀBÃÃB°ÊiÊ>Ã>
«Ì>>Ê«ÌB>>Ã`ÊÌÕÕÃiÀÌÊÌBÞÌ >ÃÀÞiÊ«ÌB>>Ã`Ê«BBÌÃÌBÊiÊiìÃÊ
ÌBÛBÌÊÌ>ÃÌ]Ê«ÌB>>Ã>Ã>ÕÕÃÊ`Ã Ìi`B]ÊÊ>Ã>>ÃÊÊ«ÌB>>iÊÛ>ÃÌ>]Ê


/>ÕÕÊÎ°
*ÌB>>Ã>Ã>>ìÊ>Ã>>Ã>ÃiÃÃ>ÊÀ>«ÀÌÌ>Û>>ÊÃÕÕÌiiÊÛ>ÕÌÌ>Û>ÊÌiÌBÊÕÛ>>Û>Ê
*ÃÃÀi}ÀiÃÃ>ÊÌÕÃiÌ£°
x¯ÊÕÌÌ>ÕÃÛB
*iÀÕÃÃÕÕÃ ä]nÈÓä ä]nxÈ ä]nÈ
"ÃÕÕÃ
ÀiviÀiÃÃÀÞBÊ
ÃÕÕìÃÌ> x¯ÊÕÌÌ>ÕÃÛB
BÀÞB
xäqÈ{ ä]nÇxx ä]nxÈ ä]nx
ÈxqÇ{ ä]{xn ä]ÎÎ ä]x
Çxqn{ ä]Èxx ä]xÈ ä]Çx
nx³ £]ääää ÀiviÀiÃÃÀÞB®
*>ÛiÕÌÞÞ««
6ÕìÃ>ÃÌ £]äÓnn £]ä£ £]äÎn
6>>Ì £]ääää ÀiviÀiÃÃÀÞB®
-ÕÕ«Õ
iÃ ä]nÎ{ ä]ÇÎ ä]{
>i £]äää ÀiviÀiÃÃÀÞB®
"ÃÕÕÃÊiìÃiÊÛÕìÊ
ÃÕÕìÃÌ> x¯ÊÕÌÌ>ÕÃÛB
6ÕÃiviÌÊ«iÀÕÃÛÕÃÊÓäää®
xäqÈ{ÊÛÕìÃ>ÃÌ ä]näÎ ä]ÈÓ ä]
xäqÈ{ÊÛ>>Ì £]äÓ{{ £]ääÓ £]ä{Ç
ÈxqÇ{ÊÛÕìÃ>ÃÌ £]ääÇÎ ä]{ £]äÓ£
ÈxqÇ{ÊÛ>>Ì £]äÓÓ £]ä£Ó £]äÎ{
Çxqn{ÊÛÕìÃ>ÃÌ £]ääÈÈ ä]n £]ä£x
Çxqn{ÊÛ>>Ì £]äÓän £]ä£{ £]äÓÇ
nx³ÊÛÕìÃ>ÃÌ £]ääÈÓ ä]n £]ä£x
nx³ÊÛ>>Ì £]ää £]ää{ £]ä£È
£®Ê />ÕÕÃÃ>ÊÀ>«ÀÌ`Õ>Ê>>ÊÛ`>>Ê>Ãi>ÊÌ>ÕÕìÊ£Ê>ÊÓÊiÕÃÌiÌÕÌÊÃÕÕìÌ°Ê ÃiÀÃÊÇxqn{

ÛÕÌ>>ÌÊÛÕìÃ>ÃÌ>ÊiÛ>ÌÊiiÌ\Êä]nÈÓäÊ«iÀÕÃÃÕÕÃ®Ê×Êä]ÈxxBÀÞBÊÇxqn{®Ê×Ê£]äÓnnÊÛÕìÃ>ÃÌ®ÊÊ
×Êä]nÎ{iÃ®Ê×Ê£]ääÈÈÛÕÃiviÌÊÓäääÊqÊÓää£®Ê×Ê£]ääÈÈÊÛÕÃiviÌÊÓää£ÊqÊÓääÓ®ÊrÊä]nxÎ
ÕÊÌ>ÃÊÊiÃÌBÞÌÊÞÊäÊÛÕÀ>ÕÌÌ>°Ê iÌiÊ«ÌB>>Ã>Ã>>ìÊ«ÕÕÌÌÕÃiÃÌ>Ê
ÃiÀÃÊÌiÃÌiÌÌÕÊ«>ÛiÕ>ÃÕiÊiÊiÊ >ÃiÌ>Ìi`ÃÃ>°Ê-iÊÃ>>ÊÛ>>ìÃÃ>Ê>Ê
>ÌÃÌ>]ÊÌiÊÃiÊ>Ã>>iÊiÊÌi`BÊ«ÌB Û>ÃÌ>>ÛÃÃ>ÊÊ>Ã>>Ì>]ÊìÊ«ÕÕÌÌÕiÊ
>>Ã`Ê«BBÌÃÌB°ÊÞÃBBÊÃ>À>>>ÌÊiÛBÌÊ >ÃiÌ>ÌiÌiÊ«ÌB>>Ã>Ã>>ÃÌ>ÊiÊÃiÌÞÊ
«BB>Ã>ÃÃ>ÊiÊÌ>ÀÌiÌÌÕÊ«ÌB>>ÃÌ>ÊÛ>À Ì>>ÕÌÕÛ>>ÊÀ>ÕÌÕÃi>°Ê BìÊ»ÃiÌÌBBÌ
Ìi]ÊÌiÊÃÃBÊÊ«iÀÕÃÌÕÕÊÛ>Ê>ÀÛÊ«ÌB Ìi»ÊÃÕÕÃÊÊÕÀÌ>ÊBÀÞBBÊxäqÈ{Û°®Ê
>>Ã`Ê«BBÌÃiiÊ>ÊÃÞÞÊ«ÌBBÊÌ> ÕÕÕÊÌÌ>>ÌÌ>ÊÕÕÌ>>Ê«ÀÃiÌÊÕ>>°Ê
ÃÊÊÛ>ÀÃ>ÃiÊÃ>À>>`Ê«ÌÌÌÞiÊ
iÃiÀÃÊ«>>ÌìÊÌ>>° (/)4/),-/)453!).%)34/.+!44!6553/.'%,-!4
ìBÊ>ÌÌÕiÊÃiiÊÛ>ÕÌÕÃÌiÊ>À 6>>ÊBBÀBÃiÃÌÊÃÕÕÀÊÃ>Ê>ÃiÌ>Ìi`ÃÌ>Ê
ÛÃiÃÊ>iÃÌÃÌ>Ê>ÃiÌÌÊÀiÌ>>ÊäÊ «ÕÕÌÌÕÛÃÌ>Ê«ÌB>>Ã>Ã>>ÃÌ>ÊBÞÌÌBBÊÃi
ÛÕÀ>ÕÌÌ>Ê>ÃiÌ>«BÛBBÊ`ÃÃ>ÊiìÊ ÌÌÞÛBÊÃB]ÊiÌÌBÊiÊiÛBÌÊÌìÃÕÕìÃÃ>ÊÛiBÊ
>Ã>>ìÊBBÀBÌÊiÀÊÀÞÃÃB°Ê6>>ÊBÊ iÊiiÌ>>Ê«ÌB>>Ã>Ã>>Ì>Ê>ÃiÌ>«B
Ã>>`ÕÌÊÕÛÕÌÊÃÃBÌBÛBÌÊÞÃÊiÊ>iÊäÊÛÕÀ ÛBB]ÊÃ>Ê>Û>ÌÕÃÌ>ÊiÀÃÌ>ÊÛÊÕÌiÊÌÕ>Ê
>ÕÌÌ>Ê`ÃÃ>ÊiiÌ]ÊiÊiBÊÊÌiÌÞÊ ÞÃÊ«ÕÕÌÌiÃÌ>ÊÌÌÕÃ>iÃÌiÊ>ÌÌ>
«ÌB>>Ã`Ê«BBÌÃ]ÊBÞÌÌBBÊÃÌB]ÊiÌÌBÊ«Ì ÛÕÕìÃÃ>°Ê/BBÊÛÃÊÃiÌÌBBÊiÀ>ÊÌiÀÛiÞì
B>>Ã>Ã>ÕÕìÊ»Ì>>ÕÌÕÛ>ÊÀ>ÕÌÕi»Ê ÕÊÌÌÕÃÌiÊÃ>Ì>]ÊÃBÊiÕÃ
ÞÃÊÌ>ÃÊ«ÌB>>Ã>Ã>ÕÕìÊÀÌii «iÀBÃiÃÌÊ>ÀÛ`>>]ÊiÌÌBÊÌiÀÛiÞìÕÃÃ>Ê
ÀÊÌBÞÌÌÞÃÌBÊiìÌBiiiÊ>>iÊÃiÌÌBBÊÛ>Ì> >Ã>>Ã>ÃiÌ>ÌÕÃiÌÊÃ>>ÌiÌ>>ÊBÌÌBBÊÕÃi
Ã>ÊÌiÀÛiÞÃiÃÕÃÌiÊ>ÊÃ>À>>ìÊ>Ã>>Ã>Ã >ÊÌÌ>>ÌÌ>ÊÕÊ«BBÌÌÞiiÌÊÌ>


ÃÌ]ÊÌ>ÊÌÕiÛ>ÌÊÕÃiÊÃÕÀ>>Ê>Ã>>ÃÌiÌBÀ iÌÌBÊ>Ã>>ÃÊÊi>ÃÌÊÀ>Ê>ìÃÃ>Ê«>>Ã
iÃÌiBÃÌB°Ê Ã>°Ê/BBÊiiiÊÃÕÕÀ>ÃÊÃ>ÃÊÃiÕÀ>ÕÃÌ>Ê
-Ã>>ÕÊÃ>Ì>ÊÌBBÊiÊÃiÊÃ>>ÊBÞ ÃÌB]ÊiÌÌBÊÞÞiiÊÃ>À>>>BÞÌiÊ>>ÃÊiÊ
ÌBÊÌìBÃiÌB°ÊÕÊÃÃ>>ÕÊÌ >Ê«>>ÃÃ>ÊÛBÌÌBBÌÌBÊÌi`BÊÕÃÀ>ÕÃÌ>Ê
ÌÕÃÌiÌ>ÊÛiÀÀ>Ì>>ÊÕÊÌ>ÃÌ]ÊBÞÌ Û>>`ÃÌ>°Ê *>ÛiÕ>ÃÕÃiÃÃ>Ê >Ã>>>Ê
ÌBÛBÌÊ>Ã>>Ã>ÃiÌ>Ìi`ÌÊÞÛÊ>ÌÌ>ÛÌ>°Ê-iÊ >ÃÕiÊ«iÀÕÃÌÕÕÊÛÕÀ>Ã«ÕÃii]ÊÌiÊ
Ã>>ÊÌ«BÛiÊBBÀBÊÊÃÃ>Ê«>ÛiÕÃ BÌBÊiÊÌiÌiBBÊÀÌiÌ>ÊÕÃÊiÃiÀÃÊ
Ã>ÊÌÌÕÃÃÃ>Ê>>Ãi«ÊÕÊÕÃÃ>ÊÌ Ã>À>>>BÞÊ>>Ã°ÊÌÌÕÃÌiÊÕ
>ÃÌÃÃ>]ÊBÊÛÌÌ>>ÊÃi]ÊiÌÌBÊ«BBÌÌÞiìÊ >>Ê>Ã>>ÃÌ>ÊÊ«ÕÊ«ÀÃiÌÌ>ÊÊ>Ãi
Ì>ÃiÊÌi`ÌÊiÛBÌÊiBÊÃ>>ÊÊ Ì>«BÛBBÊÀ>ÊÃiBÊÃ>À>>>ÊÌ>ÊÌiÀÛiÞÃiÃ
>ÌÌ>Û>°Ê/BBÊÌÕÕÊÌìBÃiÃÌÊÃÌB]ÊiÌÌBÊ ÕÃiÊÛÕìÃ>ÃÌ>ÊiÌÌBÊÛ>>`ÃÃ>ÊÌ>Ê
ÃÃ>>ÕÃÃ>Ê>ÃÊ>ÃÃ>>ÊÌ Û>ÃÌ>>Û>ÃÃ>°ÊÃ>Ê«BBiBÃÞÞÃÊÊBÊ«i
ÌÕÃÃÌ>ÊÌi`BBÊiÀBBÊ>Ã>>ÃÌiÌBÀiÃÌiBÃ ÌB]ÊiÊÃÌBÊÛÊ«ÌBBÊiÀÌÞÃiÊÛ>>Û>>Ê«ÕÕÌÌii>Ê
ÌBÊÊÌ>iÌ>iÊiÊ"*
i>>ÊÌ>Ê iÌiÊÕÊ>Ã>ÊÌi`ÃÌ>ÃiÊBiiÊÃiÊÊ
«>«iÀi°Ê/BBÊÌ>Û>>ÊÌiÌÞBÊ>Ã>>Ã>ÃiÌ>Ê i«ÊÀ>Ì>ÊBÞÌiÌÌBiÃÃBÊÌÌÕÃ>
ÊÌìBÃiÃÌÊi««Ê>iÃÌÊÌ>ÊÕÊ iÃÌ>ÊÌÕÌÕÃÌ>ÀÌÕÃ°
«BBÌÌÞiiÌÊÌ>ÃÌ°Ê
-Ã>>ÕÃÃ>ÊÌÌÕÃ>iÃÌ>ÊÊ 2!*/)445-).%.0)4+¯!)+!)30¯¯4¾+3%,,)3)).
ÕÌiÊiÀBÌÌÞÊÛ>ÃÌ>ÊÛ>>>ÊÞiiÊÛÕÌÌ>]Ê (/)4/),-/)4534)%4/)().
ÌiÊ Û>ÀÃÊ Ì>À>ÃÌiÕ>ÃÊ >Õ«Õi>Ê 6>>ÊiìBÊÊ«`ÌÌÕÊiÀBÌBÊÌÌÕÃÀi
>ÌÌ>ÛÕÕÃÊÊÕÌÊÞÞÃÌBÊi«>>°Ê-ÌiÊ ÃÌiÀÊÌÌÞÛBÊ«ÌiÌ>>Ã>Ê}i>]ÊÊÌBÀ
Ì>À>ÃÌiÕ>ÃÊ>ÕÛÕÃ>ÊBÞÛBÊÃÕÕÀi«Ê iBBÊ «ÌBBÊ iiÃÃB]Ê iÌÌBÊ ÌBÃÃBÊ ÌÕÌÕÃiÃÃ>Ê
iÀÊ>ÃiÌ>ÌiÌiÊ>Ê>iÊÞÃÊÞi BÞÌiÌÌÞÊ>iÃÌÊiÊÛ>ÀÃ>ÃiÃÌÊÌ`ÃÌ>ÊÌBBÊ
«i®ÊÌÕÃÌiÊ«iÀÕÃÌii>Ê>ÃiÌÕÃÃ>Ê«ÌB> ÌÌÕÃÀiÃÌiÀÊ>ÌÌ>ÛÕÕìÃÌ>]ÊÃBÊ«i
>Ã>Ã>>ìÊBBÀÃÃBÊÃiÌÌÞÞÊi>>Ê iÊ«ÌB>>ÃÕÕìÊÀÌiiÀÌÊÌBÞÌÌBÛiÊÌ
>ÌÌ>ÛÕÕìÊÛÕÃÌÌ>Ãi>Ê«>À>iÃi>°Ê >ÃiÊBÞÌÌÊ>iÕÌÌ>>ÊÊÃBÃBÊ}i>°Ê
ÃÊ«ÌB>>Ã>Ã>>>ÊÌ>ÃÌÊiÃiÀÃÊ
6)2(%%49+3)44¯)3)33¯(/)4/),-/)45+3)33! Ã>À>>>BÞÊÌ>>ÊiÃiÞÌÞÛBÌÊ>iÊäÊÛÕÀ
ÞÃÊÛÀiiÌÊÞÃÌÌBÃÃÃBÊÌÌÕÃÃÃ>ÊÛ >ÕìÃÃ>]Ê>Ã>>ÃÊ>ÊiÀÌBBÊ«ÌB>>Ã
Û>ÌÊ>ÊÃÞÞBÊÃ>>ÊÕÛÕÃÃ>ÊBÞÛÃÌBÊiÀÃÌ>°Ê `Ê«BBÌÃiÃÌBÊiÊÌÕiÊ>>>ÊÕ>>Ê«Ì
ÃÕÃÊ«ÌB>>Ã`Ê«BBÌÃÊÛÊ>ÊÌiÌÞ]Ê B>>Ã>Ã>>ìÊ«Ì>>°Ê/BBÊ}i>Ê
ÕÌÌ>ÊÃÌBÊiÊÕ>Ì>ÊÀ>Ì>ÊÌÌÕÃÌiÌ Êi>ÃÃ>ÊÀ««Õ>ÌÌ>ÊÃÌB]ÊBÞÌiÌBBÊ«Ì
°Ê ÀÌÞÃiÃÌÊÃÃ>>ÕÃÃ>]ÊÕÊÌi`ÌÊiÀB B>>Ã>Ã>>ìÊBBÀBÊ>ÃiÃiiÊ«iBÃ
ÌBBÊ«BB>Ã>ÃÃ>ÊBÃ]ÊÌB>ÃÌ>ÊÌiÌ>ÊiÊÛBÌÌB ÌBBÊÞìÊÛÕìÊ>Ã>>Ã>ÃiÌ>ÌiÌ>ÊÛ>Ê
BÌÌBÊBìÌBÊÌ>À>ÃÌ>>>]ÊÛ>>ÊÌ>ÊBÌiÌBBÊ ÞÃÊÞi«BÊÌÌÕÃ>°Ê*ÌB>>
ÌÞBÃ°Ê ÃÕÕìÊÀÌiiÀiBÊÀ>>ÌÕ>Ê>iÃÌ>ÊiÊ«ÞÃÌÞÌBÊ
ÃBÃÊÌÕÊ>ÊBÌ«BÛBÌÊÛÛ>ÌÊ>ÊÃÃ>Ê ÞÃBBÊÕÛ>>>>Ê>ÌÌ>Û>ÃÌÊ>Ã>>>ÊÊ
ÌÕÃÃÃ>ÊÛBBÀ°ÊÕÊÌÌÕÃ>iÃÌ>Ê ÌiÌÕ>]ÊÃBÊ>Ã>>>ÊÌ>ÃÊÃ>>ÌÌ>>Ê
-Ì>iÃÃÃ>ÊÌ>>]ÊÃ>>ÊÛÕìÊÌÕÃÃÌ>Ê ÌìÃÕÕìÃÃ>Ê>ÊÀiÃÌiÀÃÃB]ÊÕÌÌ>ÊiÊÌÕiÊ
Ì>À>ÃÌiÌ>>]ÊiÌÌBÊ«BBiBÃBÊÌ>Ã>ÊiÊ Õ>>Ê>iÃÌ]ÊÃÊÃiÊiÊÌBÞÌBÊ«ÌB>>ÃÕÕ
i°Ê ÀÊÛÕÃiÊÌÕÃ>ÊiÊÕÌi>>ÊÛiÀÀ> ìÊÀÌiiÀiB°Ê
Ì>]ÊÊÞiBÃÃBÊÌÕÃiÃÃ>ÊÛÊ>Ê BÞÌiÌÌBiÃÃBÊÃiBÊ«ÃÌÌÕÃÊiÌÌBÊ>Ãi
ÀÃÌÀÌ>iÊÌiÌÊÌÕ«BÛBÃÌB°Ê BÊiÊÞ Ì>ÌiÌ>ÊÃÊ>Ã>>ìÊÕÕBBÀBÌÊ«iÀ>>Ì
iBÊÌÕÃiÊÕ>>Ê>Ã>>>ÃÌ>ÊÌÕiiÊ ÌiiÃÃ>Ê>`ÃÌ>Ê«BBÌiBÊ>ÃiiÊ«BÛBiÊ
«ÌB>>Ã>Ã>>Ã]ÊÛ>>Ê>>Ãi«Ê>Ã>>Ã iÀÃii°Ê/BBÊÕÌiÊÌ>>Ê>>ÕÕÃ]Ê
>ÃiÌ>®ÌÕÃÊiÊÃÌBÊBÞÌB°Ê"Ê>`ÌÌ>Ê ÃBÊÀ>>ÌÌ>iÃÃ>Ê>Ã>>>ÌÊÀiÃÌiÀÃÌBÊ«ÌB>>Ã
Ã>>]ÊÛ>ÌÊ>>ÃiÊÛ>ÊÞiÊÌÕ «BBÌÃiÊ«iÀÕÃÌii>]Ê«BÛÌÌBÃÌiÊ««Õ>>Ì
iiÊÌÕÃiÊÌÕ«BÛBÌi`ÌÊi°Ê ìÊÕÛ>>>ÃÌ>ÊÌÕiiÊ»Ã>>>i»ÊÃÌi]ÊiÌÌBÊ
ÕÊÛÕìÊÛiÃiÊ«BÛBÊ>ÊÃiÕÀ>>Û>Ê
0¯¯,,%++¯)3%4(/)4/*!+3/4 ÛÕìÊiÃBÃiÊ«BÛBÊ««Õ>>ÌÌÊiÀ>Û>ÌÊ
9ÃÊÌÌÕÃÀiÃÌiÀÌiÌiÊ}iÃÌ>Ê ÌÃÃÌ>>Ê«>ÊÞiiÊ«ÀÃiÌÌ>°Ê BÊÃÕÕÀÊ
Û>ÌÊ«BBiBÃiÌÊÌ>ÃÌ°ÊÞÃiÊiÊ«BB>Ã>Ã «BÛBÕÕÌÃÊiÊiÊÌìÃÕÕìÃÃ>ÊÀi>ÃÌi°Ê
Ã>ÊiÊ>iÃÌÊ«ÕÕÌÌiiÃÕÕìÃÌ>]ÊÛ>>ÊÃÌB]Ê ÊÌÕÕÊÃÌB]ÊiÌÌBÊ>iÊ>Ã>>iÊiÊ


Þ`ÞÊ>Ãi>ÊBiiÊ`Ê«BBÌÌÞÃ iÌ>Ìi`ÌÊÃÛ>ÌÊÕ>>ÊÌ>À>ÃÌiÕÃÃ>ÊÀ«
ÌÕÃÌ>ÊÌ>ÊÃiÕÀ>>Û>ÊÛÕìÊ>ÃiÌ>ÌiÌ>]Ê>Ê «Õ>ÌÌ>ÊÃÌBÊÌBÞÌÌBÛBÌÊiÊ«ÌB>>Ã`Ê
ÌBÞÌÌBÃÊ«ÌB>>Ã`ÊÀÌiiÀÌ°Ê/ÃÊÃ> ÀÌiiÀÌÊÛ>ÊiÛBÌ°Ê/B>ÃiÊ>iÃÌÊ>ÛÕ>ÊÃ>>
iÊÃÃÊÃ>iÊ>Ã>>ÃÌ>ÊiÊÞ`ÞÊ>ÊÛÕìiÊ Ì>ÃÊÃBÌiÌ>ÊÞÃÊÃÌB]ÊÌiÌ>>ÊBÌÊ
«ÌB>>Ã`ÊÀÌiiÀÌÊÌBÞÌÌBÛBBÊ>ÃiÌ>Ìi >ÊÌÕ«BÛBÌÊi>ÃÕÕÌ>ÃiÃÌ°Ê ìiiÊÌB>
Ì>Ê>ÊÌÃ>>Ì>ÊÃ>iÊiÊ>ÃiÌ>Ìi`ÊBiiÊ ÃiÊ>Ã>iÃÌÊ>ÛÕ>ÊÃÊ>`ÃÌ>Ê
Þ`ÞÊÀÌiiÀÌÊÌBÞÌÌBÛBBÊ`Ê«BBÌÌÞÃ >Ãi>Ê>Ã>>>ÊÊ>ÌÃ`ÊiÃÌ]Ê>Ê
ÌÕÃÌ>°Ê->>>iÊÊÊ>Û>ÌÌÕÊ>>Ãi ÊÌÌÕÃÀiÃÌiÀÊ»}>»ÊrÊÌi`ÌÊ
ÊÞÃÊÃÌÃvÀi>`>}Ã>ÊÀ>>ÌÕiÊ>Ã> Û>ÌÊ>ÃÌ>Ã>]ÊÌiÊiÃiÀÃÊÃ>À>>>
>ÃÕiÊ-Õ`ÊÓäää]ÊxÓqxÇ®°Ê BÞÌÊ>Ì>ÃiiÊÛ>>ÌÌ>Ã®ÊÌ>>Ê
ÕÕÌiÊ}i>ÃÌ>°Ê
0)4+¯!)+!)3!3)!++55$%.-¯¯2)4%,-¯33¯+%()4%44¯6¯¯
,iÃÌiÀ«Ì>ÊÛ>ÀÌiÊBBÀÌiBÊÊ>Ê */(4/0¯¯4¾+3%4
Ì>«>ÕÃiÃÃ>ÊÌ>Û>ÊÞÃÃiÌÌii]ÊÊÃÌBÊ 6>>ÊÞìÊÛÕìÊ>Ã>>Ã>ÃiÌ>Ìi`ÃÌ>Ê
ÌÕiiÊÕÌiiÌ>>Êi«ÃÌÊÞÛÊÌii°Ê Þ >ÃiÌÌÕÊ«ÌB>>Ã>Ã>>ìÊÕÕBBÀBÊÊ
ÞiÊ «Ì>Ì>«>Ê Ì>>Ê Ãi]Ê iÌÌBÊ iiÊ ÃiÛBÃÌÊ«ii«ÊÕÊÃ]ÊÕÊÌiÌ>>ÊÕ
>Ã>>iÊiÊÌiÌÞÊ«ÌB>>Ã`Ê«BBÌÃÌB]Ê ÊÞÃÊÞi«iÊ>Ã>>Ã>ÃiÌiÊ>Ê
iÊÌÕiÛ>ÌÊÕ>>Ê«ÌB>>Ã>Ã>>ìÊBB «BBÌÌÞiìÊÌ>ÃiÊÌi`Ì]ÊiÊÃiÊÌ`ÃÌ>]Ê
ÀBBÊÛ>]ÊÃÊiÊÛ>ÌÊiiÌÊÞìÃÃBÊ«>>ÃÃ>ÊÞÊ iÌÌBÊ>Ã>>Ã>ÃiÃÌ>Ê«ÌÌÕÊ«ÌB>>Ã
iÊÕÕ>ÕÌÌ>°ÊBÞÌBÃÃBÊ>Ã>>ìÊ«> >Ã>>ìÊBBÀBÊÃÊÃiÛBÃÌÊÌìÃÌ>Ê«ÌB
ÛiÕiÌÕÌÊÕÌiÊiÀ>Û>Ì\ÊÌÕÌÊ>Ã>>>ÌÊ >>Ã>Ã>>ìÊBBÀBBÊ«ii«°ÊÞi
Û>ÌÊÞìÃÃBÊ«>>ÃÃ>ÊÛÕÃ>Ê>ÊÌÕÌÊÌ>>ÃÊÛ> «iÊÌÕÃÌiÊÌi`ÌÊiÛBÌÊ>ÕÌ>>ÌÌÃiÃÌÊiÊ
Ì>Û>ÌÊÌ«>>>ÊÕÃi>°Ê iiÃi«>Ê >Ê Ã>Ê «ÌB>>Ã>Ã>ÕÕÃÊ
ÃBÃÊ«ÌB>>Ã«BBÌÃÊÛ`>>ÊÌi`BÊiÀÊ BÞÌÌBBÊÞi«BÊÌÕÃ>ÊÞ`ÞiÌÌBiÃÃBÊ
ÕÃÃ>]ÊiÀÊ>ÌÃÌÞÞ«iÃÃBÊ>ÊiÀ>ÃiÊ>Ã> ÌÕiÛ>ÊÛ>>Ê»>Ê>>Ã»]ÊÛÊÞi
>iÊiÀÊÌ>Û>>°Ê ÃiÀÃÊÌBÃÃBÊ>iÃÌÃÃ>Ê «iÊÌÕÃÌiÊ>ÛÕ>Ê>ÃiÌÌÕÊ«ÌB>>Ã>Ã>
>Û>ÌÌ]ÊiÌÌBÊÛ>iÃÃ>ÊBÀÞÃÃBÊ>Ãi >ìÊBBÀBÊ>Ê«>Ê>ÊÃ°Ê6ÕÃÌ>ÃÌ>Ê
Ì>>iÃÌiÊ>ÌÌ>ÛÕÕÃÊ«ÌB>>Ã>Ã>>ìÊ >ÃiÌÌ>iÃÃ>ÊÊÕÕÌiÊ«>Ài«ÊBÞÌÌBBÊ>Ê
BBÀBÊÕÛ>>>>ÊÊ«>Ài«ÊÕÊÕÀiÃ ÛÕÃÊÃ>>>Ê«Ì>Ì>«>>Ê>ÊÃ>>>Ê>iÃÌ>°Ê
Ã>]ÊÃ>Ê>Ã>Ê>ÌÌ>ÛÕÕÃÊÊ«>Ài«ÊÕÊ ÕÌiÊiÃiÀÃÊ>ÃiÌÌ>iÃÃ>Ê>
iB°Ê6>iiÊ>Ã>>iÊ>Ê>ÃiÊ«BB ÕÀÌÕ>Ê>>ÌÕÛÕÕÌÌ>Ê«ÌB>>Ã`ÃÃ>Ê
ÌÃÊÊÌi`BBÊ«i>ÊÌ>ÊÃiÊÊ«>ÀiÊ iÛiÊiìÊÃ>Ì>ÊqÊÌ>Ê>ÀÛÌ>iÃÃ>Ê«Ì
Ìi`ÃÃ>]ÊÊÃiÊÌÕiiÊÛ>ÀiÊÕ>>Ê B>>Ã>Ã>>ìÊÌ>ÀÛÌÃiiÊÛÕì«>
ÞÃÊ>ÃiÌ>ÌiÌ°ÊÃ>Ê«ÌB>>ÃiiÊ> iÊBBÀBBÊqÊÊ>iÃÌÊÕÛ>ÌÌ>Û>Ê>`
ÌÃÌÊÃÀÀÞÌBBÊÊnÓÛÕÌ>>>]ÊÛ>ÌÊ Ã>Ê>ÌÌ>Û>ÃÌÊ«ÌB>>Ã>Ã>>ìÊ««Õ
iÃiÀÃÊnxÊÛÕÌÌ>ÊÌBÞÌÌBiiÌÊiiÌÊiÃBB >>ÌÌ>°Ê/B>ÃÌ>ÊBÞÌÌÌ>ÀÌÕÃÌ>ÊÛ>ÀÌiÊ>
ÀÊ>ÌÃ`ÃÃ>ÊÊÕÕÌ>>ÊÛÕì°Ê*BBÌÃÊ >ÌÌ>>ÊÃÃÊÞ`ÞÌBBÊ>Ãi>ÃÃ>ÊÞÃÊÞ
ÊÌBÊÌiÌÞÊÊ>Õ>ÊÃÌÌiÊ>ÊÞÃÊ>ÃiÌ> i«iÊ ÛÕÃiÊ ÌÌÕÃÌiÌ>°Ê Ã>Ê
ÌiÌÊÌiÌÊiÀÌBBÊÌìBÃiÊÕÊ ÌÕÀi>iÊÛÕìiÊiÊÕÌi>>ÊiÊBÞÌÃÃBÊ
ÕÕìiÊ>Ã>>i°Ê-ÕÕ«ÕÌiÊÛBÃÌBÊiÀ>Ê Þi«BÊÌÕÃ>]ÊÕ`ÃÌiÌÌÊÌBÃÃBÊÌÕÌ
ÊÛ>i>«ÊÃiÌÌBB]ÊÕÌÌ>ÊÃBÊÛÊ>ÊÞÃiÊ ÕÃiÃÃ>ÊÞÃÊÌ>ÃÌiÊ>]Ê>Ê>ÛÕ>Ê
iÃÌiÊ>Ê>ÃÌiÊiÀ>ÃiÃÌ>Ê«>ÛiÕiÊBÞÌÃÌBÊ Ê >`ÃÌ>Ê iÕÃÌ>>Ê ÞÃÊ ÌÕÀi>iÊ
BBÌÞiÃÌBÊ>ÃÃÌ>ÊÊ>ÌÃ`ÃÃ>ÊÃ«Ê ÛÕìiÊÕìÊÛÕÃiÊ>ÃÃ>ÊÛiÀÌ>Õi«
Ã>ÊÕÊiÃÌB®°Ê iÊ«ÌB>>Ã>Ã>>ìÊÕÕBBÀB°
>Û>ÌÌÕiÊ«ÌiÌ>>ÃÌiÊ«ÕÕÌÌiìÊÃÞìÊ ««ÕiÊ«ÕÃÊÀiÃÌiÀÌiÌ>ÊBÞÌiÌÌBiÃÃBÊ
Ì>Ài«ÊÃiÛÌÌBiÊÃÊÌÃ>ÊÃBÌ>À ÊÃÃÊ>>ÊÌiÌÌ>Û>ÊÕ]ÊÌiÊ>Ê>
>ÃÌiÕiÊì]ÊÃBÊÃiÊiÀÌÃÊÌ>ÀiÊÃiBÊ ÃÃÌ>ÊBÌiÃÌBÊiÊÊÌÌÕ°ÊiÃÌiÊBÞÌÌBBÊ
ÌÌÕÃ>iÃÌÊÕÌiÌÌ>ÛÕÕìÃÌ>ÊiÌÌBÊ«Ì ÊBÀiÛiÊÌÕÌiÊ>`ÃÌ>ÃiÃÊ>Ê
B>>Ã`ÊÀÌiiÀÊÌ>ÀÌÕÃiÕ>ÃÕÕ Ì>«>ÕÃiÃÃ>ÊÃiÛÌiÌÌBÛBÊÌÃiii]ÊiÌÌBÊ>iÃÌÊ
ìÃÌ>°Ê/BBÊÕÌiÊÛ>>ÌÃÊ>iÃÌ>]ÊÃÃ>Ê ÕÛ>>ÊÛ>À>ÃÌÊÃÌBÊ>Ã>>]ÊÌ>ÊBÊiÌÌ>>ÊÃiÊ
«ÌB>>Ã>Ã>>ìÊÀÌÊ>ÊÞÃ ÕÛ>>Û>Ê-Õ`ÊÓääÎ]Ê-Õ`Ê>Ê Þ>ìÀÊ>Ê*>
ÌÌÕÕÃi>ÊìÌwÌÕÛ>ÌÊÌ>ÃÌÊ>Ê>Ã iÊÓää{®°Ê


3UND2+AUPPINEN3(OWTODETERMINETHENUMBEROFOLDERPEOPLEINLONG TERMCAREUSINGREGISTER BASEDDATA
3OSIAALILËËKETIETEELLINENAIKAKAUSLEHTIn*OURNALOF3OCIAL-EDICINEn
/ÃÊÃÌÕ`ÞÊÀi«ÀÌÃÊÌiÊ>}iÊ>`ÊÃiÝÊÃ«iVwVÊ «i>ÀÊÌÊLiÊviÜiÀÊÌ>ÊÜiÊÕÃ}Ê>``Ì>Ê
VÕÌÃÊvÊìÀÊ«i«iÊÊ}ÌiÀÊV>ÀiÊÊ vÀ>ÌÊvÀÊÌiÊ`ÃV>À}iÊ>LÃÌÀ>VÌÃ°Ê
>`ÊvÀÊ£ÇqÓääÓ]ÊL>Ãi`ÊÊÌÜÊ`v iÌiÀÊvÊÌiÃiÊÌiVµÕiÃÊÃÊLiÌÌiÀÊÌ>Ê
viÀiÌÊV>VÕ>ÌÊÌiVµÕiÃ°ÊÌÊÌÕÀÃÊÕÌÊ ÌiÊÌiÀ]Ê>`ÊÌiÊÌiì`ÊÕÃiÊvÊ`>Ì>Êì
Ì>ÌÊÜiÊÕÃ}ÊÞÊÌiÊÃ}iÊÞi>ÀÊViÃÕÃÊ ÌiÀiÃÊÜVÊÃÊÀiÊÃÕÌ>Li°
`>Ì>]ÊÌiÊÕLiÀÊvÊ}ÌiÀÊViÌÃÊ>«
+)2*!,,)3553
BBÌÞiìÊÃÃ>>Ê>ÊÌiÀÛiÞÃ«>ÛiÕÌÊÓääÓ°Ê , "Ê-1
-ÕiÊ6À>iÊÌ>ÃÌ]Ê-Ã>>ÌÕÀÛ>ÊÓääÎ\£°Ê 6/]ÊÌ>ÃÌÌiÌiB
-Ì>iÃ]ÊiÃ°
6>ÕÌÌ>ÛÕÕìÊ>ÊiÕìÕ>ÃÕÕìÊ
-Õ`Ê,°Ê/>ÃÌÃ>ÊiiÌiBÊ`Þ>>ÃÌiÊ
«Ì>Ã««Õ>>ÌìÊ>Ì>Ãii°Ê ÌÕÌÕÃÀÞB
/>«>ÌÕ>ÃÌÀ>>>ÞÞÃ>Ê -Ì>iÃ
ÌÌÕÃÀiÃÌiÀÊÃÌÃvÀiii°ÊiÌ>Ê
ÓÈÉÓäää°Ê-Ì>iÃ]ÊiÃ°
-,Ê1**
-Õ`Ê,°Ê1ÌÃ>ÌÊvÊ>`ÃÌÀ>ÌÛiÊÀi}ÃÌiÀÃÊÕÃ}Ê ]ÊiÀÃÃÕÕÌÌi>
ÃViÌwVÊÜi`}iÊ`ÃVÛiÀÞ°ÊÌi}iÌÊ>Ì>Ê
>ÞÃÃÊÓääÎ\Ç\È\xä£qx£° />ÃÌÌÀÞB
-Õ`Ê,]Ê Þ>ìÀÊ"]Ê*>iÊ/°Ê,>>½>ÃÌ>Ê -Ì>iÃ
ÀiÃÌiÀ>iÃÌÃÌ>ÊÌiÀÛiÞÃ«ÌÌÃiÃÌÊÀiiÛ>ÌÌÊ
vÀ>>Ì°Ê9ÌiÃÕÌ>«Ì>Ê
Óää{\È\{\ÎÇÓqÎÇ°
-Ì>iÃ°Ê-Ã>>ÕÊ>ÊÌiÀÛiÞìÕÊ
ÌÌÕÃÊÓää{ÊqÊBBÀÌÌiÞÌÊ>ÊiÃÌÕÃ°Ê
-Ì>iÃ]ÊiÌ>Ê>ÊÕÌÕÃ>ÊÓääÎ\Î°Ê-Ì>iÃ]Ê
iÃ°


V
Alkuperäistutkimus
Reijo Sund
Lonkkamurtumien ilmaantuvuus
Suomessa 1998–2002
Väestön ikääntyessä lonkkamurtumat ovat merkittävä terveydenhuollon haaste Suo-
messa. Lonkkamurtumien määrä on pitkän aikaa kasvanut suhteessa nopeammin kuin
vastaava riskiväestö, mutta viime vuosina vakioidun ilmaantuvuuden kasvun on rapor-
toitu tasaantuneen monissa maissa. Ikääntymiseen liittyvien lonkkamurtumien ilmaantu-
vuus ei kasvanut Suomessa vuosina 1998–2002. Talvisin murtumia tapahtui enemmän
kuin muina aikoina, poikkeuksena pitkäaikaisessa laitoshoidossa olevat. Pitkäaikaisessa
laitoshoidossa tapahtuneiden lonkkamurtumien ilmaantuvuudessa ei ollut sukupuolten
välisiä eroja. Lonkkamurtumia ilmeni poikkeuksellisen paljon Helsingissä, mutta muuten
sairaanhoitopiirien väliset ilmaantuvuuserot olivat varsin pieniä. Koska lonkkamurtuman
ilmaantuvuus heijastaa väestön toimintakykyä, on vakioidun ilmaantuvuuden tasaantu-
minen selitettävissä väestön odotetun eliniän ja toimintakyvyn muutosten yhteisvaikutuk-
set huomioon ottavalla yleisellä ikääntymisteorialla.
V
äestön ikääntyessä reisiluun yläosan mur‑ tutkimusten tulokset eivät ole yhdenmukaisia
tumat eli lonkkamurtumat ovat merkittävä Keski‑Suomen tilannetta kuvaavan tutkimuk‑
terveydenhuollon haaste Suomessa, koska sen kanssa; viimeksi mainitussa tutkimuksessa
lonkkamurtuman ilmaantuvuus kasvaa ekspo‑ ei havaittu ikävakioidun ilmaantuvuuden kas‑
nentiaalisesti suhteessa ikään (Melton 1996, vua ajanjaksojen 1982–83 ja 1992–93 välillä
Brody ja Grant 2001). Lonkkamurtumien il‑ (Huusko ym. 1999). Suomessa ei ole tutkittu
maantuvuudessa on suuria eroja eri maiden lonkkamurtumien ilmaantuvuutta pitkäaikai‑
välillä (Kanis ym. 2002), ja lonkkamurtumien sessa laitoshoidossa eikä myöskään murtumiin
määrän on havaittu kasvavan nopeammin kuin liittyvää kausivaihtelua. Tässä tutkimuksessa
riskiväestön (Kannus ym. 1999). Ikävakioidun tarkastellaan ikääntymiseen liittyvien lonkka‑
ilmaantuvuuden kasvun on kuitenkin viime murtumien ilmaantuvuutta Suomessa vuosina
vuosina raportoitu tasaantuneen monissa mais‑ 1998–2002 edellä mainittujen tiedon tarpeiden
sa, ja syiden on arvailtu liittyvän murtumien tyydyttämiseksi.
onnistuneeseen ehkäisyyn tai ikääntyneiden
kohentuneeseen terveydentilaan (Rogmark ym.
Aineisto ja menetelmät
1999, Boufous ym. 2004).
Lonkkamurtumien ilmaantuvuudesta Suomes‑ Stakesin ylläpitämästä valtakunnallisesta hoitoilmoitus‑
sa ei ole julkaistu koko maan kattavia alueellisia rekisteristä poimittiin vuosien 1998–2002 lonkkamurtu‑
matapaukset käyttäen rajauskriteerinä ICD‑10:n pää‑ tai
tietoja ja koko maan tiedotkin ovat usean vuo‑ sivudiagnooseja S72.0, S72.1 ja S72.2. Rekisteristä poi‑
den takaa (Kannus ym. 1999, Lüthje ym. 2001). mittiin henkilötunnusta käyttäen tiedot kaikista näiden
Lisäksi näiden rekisteritietoihin perustuvien potilaiden hoitojaksoista vuosilta 1987–2002. Hoitojakso
Duodecim 2006;122:1085–91 1085

luokiteltiin lonkkamurtumaan liittyväksi, jos sen pää‑ tai murtumien sairaanhoitopiireittäiset määrät olivat varsin
sivudiagnoosina oli lonkkamurtuma (ICD‑9: 820, ICD‑10: pieniä, käytettiin laskennassa empiiristä Bayes‑estimointia
S72.0–2), ja vuosien 1998–2002 hoitojaksojen osalta edel‑ (Greenland 2000).
leen leikkaushoitojaksoksi, jos rekisteristä löytyi merkintä
lonkkaleikkauksesta (Nomeskon toimenpideluokituksen
suomalainen versio: NFB10‑99, NFJ42‑64). Saman hen‑
kilön lonkkamurtumaan liittyvien hoitojaksojen väliset
Tulokset
erot laskettiin päivän tarkkuudella rekisteriin kirjattujen
tulo‑ ja poistopäivien perusteella. Vuosien 1998–2002 aikana lonkkamurtuma‑
Koska tässä tutkimuksessa ensisijaisena kiinnostuksen diagnoosillisia hoitojaksoja alkoi vuodessa noin
kohteena olivat ikääntymisen seurauksena ilmenevät lonk‑
13 500 (taulukko 1). Alkaneita hoitojaksoja oli
kamurtumat, rajoituttiin tarkastelemaan 50 vuotta täyttä‑
neitä henkilöitä. Murtuman luokiteltiin tapahtuneen pit‑ keskimäärin noin 7 200 potilaalla vuodessa. Re‑
käaikaisessa laitoshoidossa, jos murtuma rekisteritietojen kisteriin kirjattuja lonkkamurtumaleikkauksia
perusteella tapahtui laitoshoidossa (henkilö tuli sairaalaan tehtiin vuosien 1998–2002 aikana keskimäärin
pitkäaikaista hoitoa antavasta laitoksesta) ja henkilö täytti
runsaat 6 000 vuodessa. Määrä on luultavas‑
murtuman tapahtumahetkellä väljennetyt pitkäaikaishoi‑
don kriteerit (yli kolme kuukautta yhtäjaksoista laitoshoi‑ ti murtumien todellista lukumäärää pienempi.
toa tai yli kuusi kuukautta laitoshoitoa murtumaa edeltä‑ Murtumaleikkauksen lisäksi uudeksi murtumak‑
neen vuoden kuluessa). Vuosien 1998–2002 aikana tapah‑ si voidaan tulkita lonkkamurtumadiagnoosilli‑
tunutta lonkkamurtumaa pidettiin henkilön ensimmäisenä
nen hoitojakso, jos edeltäneestä lonkkamurtu‑
lonkkamurtumana, jos hänellä ei edellisen kymmenen vuo‑
den aikana ollut lonkkamurtumaan liittyviä hoitojaksoja. man hoitojaksosta on kulunut riittävän kauan.
Näin määritelty lonkkamurtuman tapahtumahetki ilmaisee Käyttäen kriteerinä kahden vuoden eroa edelli‑
samalla myös sen, mihin mennessä ikääntymiseen liittyvät seen lonkkamurtumahoitojaksoon murtumien
lonkkamurtuman riskitekijät ovat kullakin potilaalla vii‑
lukumääräksi saatiin keskimäärin noin 6 800
meistään ylittäneet kriittisen rajan, jos kyse on vähäener‑
giaisesta murtumasta. vuodessa (taulukko 1). Höllennettäessä aika‑
Riskiväestön laskemiseen käytettiin virallisia ikä‑ ja su‑ kriteeriä kahteen kuukauteen murtumien kes‑
kupuoliryhmittäisiä väestötietoja. Pitkäaikaispotilaiden kimääräiseksi lukumääräksi saatiin noin 7 200
riskiväestötiedot muodostettiin rekisteritietojen avulla (ks.
vuodessa. Vaikka kliiniset kriteerit täyttävien
Sund ja Kauppinen 2005). Ilmaantuvuusluvut saatiin jaka‑
malla lonkkamurtumien lukumäärä vastaavalla riskiväes‑ murtumien lukumäärää ei rekisteritietoja käyt‑
tön henkilövuosina mitatulla seuranta‑ajalla. Vertailuihin täen ole mahdollista yksiselitteisesti laskea, vuo‑
sopivat vakioidut ilmaantuvuusluvut laskettiin suoraa va‑ sien 1998–2002 aikana näyttäisi tapahtuneen
kiointia käyttäen. Ne kertovat, mikä on kokonaisilmaan‑
keskimäärin noin 7 000 lonkkamurtumaa vuo‑
tuvuus vakioväestössä, jos ryhmittäiset ilmaantuvuudet
oletetaan samoiksi kuin tarkastelun kohteena olevassa vä‑ dessa.
estössä. Vakioväestönä käytettiin Suomen keskimääräistä Ensimmäisten lonkkamurtumien määrä oli
väestöä vuosina 1998–2002. Vakioidut ilmaantuvuusluvut vuosien 1998–2002 aikana keskimäärin noin
laskettiin myös sairaanhoitopiireittäin (Helsingin ja Uuden‑
6 000 vuodessa (taulukko 1). Näistä runsaat
maan sairaanhoitopiiri jaettiin edelleen sairaanhoitoaluei‑
den mukaan neljään osaan). Puutteellisten tietojen takia 5 500 tapahtui vuosittain 50 vuotta täyttäneil‑
Ahvenanmaa jätettiin pois tarkasteluista. Koska lonkka‑ le. Kyseisistä lonkkamurtumapotilaista valta‑
TAulukko 1. Lonkkamurtumien määrät Suomessa vuosina 1998–2002.
1998 1999 2000 2001 2002
Alkaneet hoitojaksot, joissa diagnoosina lonkkamurtuma 13 219 12 965 13 381 14 142 14 145
merkintä lonkkaleikkauksesta 5 934 5 972 6 100 6 221 6 195
merkintä lonkkaleikkauksesta tai vähintään 2 vuotta edellisestä
hoitojaksosta, jossa diagnoosina lonkkamurtuma 6 742 6 631 6 702 6 924 6 937
merkintä lonkkaleikkauksesta tai vähintään 2 kuukautta edellisestä
hoitojaksosta, jossa diagnoosina lonkkamurtuma 7 246 7 094 7 136 7 411 7 405
vähintään 10 vuotta edellisestä hoitojaksosta, jossa diagnoosina
lonkkamurtuma 5 990 5 853 5 932 6 083 6 141
kuten edellä, mutta vain 50 vuotta täyttäneet henkilöt 5 551 5 413 5 543 5 644 5 667
1086 R. Sund

KuvA 1. Lonkkamurtumien päivittäiset määrät Suomessa 1998–2002.
osa (72 %) oli naisia. Noin 29 % murtumista keästi poikkeava alue on Helsinki (taulukko 2).
tapahtui laitoshoidossa ja 19 % pitkäaikaisessa Muuten alueittaiset ilmaantuvuuserot ovat var‑
laitoshoidossa. Liukuvalla keskiarvolla tasoi‑ sin pienet. Keskimääräinen ilmaantuvuus vuosi‑
tetut lonkkamurtumien päivittäiset määrät on en 1998–2002 aikana oli suurinta Helsingissä ja
esitetty kuvassa 1. Lonkkamurtumia tapahtui Keski‑Suomessa ja pienintä Etelä‑Karjalassa, Ete‑
keskimäärin 15,2 päivässä, ja havaittavissa oli lä‑Pohjanmaalla ja Kainuussa (kuva 2). Etelä‑ ja
pienehköä mutta selkeää kausivaihtelua: noin Pohjois‑Savossa vakioitu ilmaantuvuus näyttäisi
54 % murtumista tapahtui marras–huhtikuus‑ kasvaneen vuosien 1998–2002 aikana lineaarises‑
sa ja noin 46 % touko–lokakuussa. Aineiston ti ja Kanta‑Hämeessä sekä Helsingissä taas pie‑
mukaan talvikauden aikana murtumia tapahtui nentyneen varsin lineaarisesti (taulukko 2). Mui‑
muutamina päivinä poikkeuksellisen paljon. On den alueiden osalta ilmaantuvuuden trendi vuosi‑
myös huomion arvoista, että pitkäaikaispotilai‑ na 1998–2002 oli käyräviivainen tai vakio.
den murtumissa ei ollut havaittavissa kausivaih‑
telua (kuva 1).
Pohdinta
Ikä‑ ja sukupuolivakioitu ilmaantuvuus oli
vuonna 1998 suurempi kuin muina vuosina, Suomessa käytännössä kaikki lonkkamurtumat
joiden aikana ilmaantuvuus pysyi varsin vakaa‑ hoidetaan sairaalassa, joten on odotettavaa, että
na (taulukko 2). Tämä on merkki lonkkamur‑ lonkkamurtumaan liittyvät hoitojaksot kirjau‑
tumien ilmaantuvuuden tasaantumisesta myös tuvat valtakunnalliseen hoitoilmoitusrekisteriin.
Suomessa. Vakioitu ilmaantuvuus oli naisilla Lonkkamurtumapotilaiden kattava identifiointi
suurempi kuin miehillä ja erityisen suuri pitkä‑ rekisteristä on siis periaatteessa mahdollista sillä
aikaisessa laitoshoidossa olleilla (taulukko 2). oletuksella, että tiedot on kirjattu kattavasti ja
Pitkäaikaisessa laitoshoidossa olleiden miesten luotettavasti rekisteriin. Tuoreen validointitutki‑
ja naisten ilmaantuvuusluvuissa ei kuitenkaan muksen mukaan kattavuus on lonkkamurtuman
ollut merkitsevää eroa. osalta erittäin hyvä ja vain noin 2 %:iin kirur‑
Sairaanhoitopiireittäisissä ikä‑ ja sukupuoliva‑ gian erikoisalan hoitojaksoista on liitetty vir‑
kioiduissa ilmaantuvuuksissa ainoa muista sel‑ heellisesti muu kuin lonkkamurtumadiagnoosi
Lonkkamurtumien ilmaantuvuus Suomessa 1998–2002 1087

ovat merkittävästi pienempiä kuin aikaisemmis‑
sa samaan rekisteriin perustuvissa tutkimuksissa
(Kannus ym. 1999, Lüthje ym. 2001). Pääasial‑
linen syy raportoitujen lonkkamurtumien mää‑
rien eroon on lonkkamurtuman määritelmien
erilaisuus. Aikaisemmista tutkimuksista poike‑
21 ten tässä tutkimuksessa tarkasteltiin kaikkien
lonkkamurtumien sijaan ensimmäisiä lonkka‑
Vakioitu ilmaantuvuus murtumia, koska ne pystytään poimimaan re‑
> 320 (2)
kisteristä huomattavasti yksiselitteisemmin kuin
320–325 (5) 20
310–320 (5) kaikki lonkkamurtumat. Ensimmäisiin lonkka‑
305–310 (6) murtumiin perustuvissa ilmaantuvuuslaskelmis‑
< 305 (3) sa myöskään saman henkilön toistuvien lonkka‑
18 murtumien välinen ilmeinen riippuvuus ei pääse
19 vääristämään tuloksia.
Vertailun vuoksi tässä tutkimuksessa arvioi‑
17 tiin lisäksi kaikkien lonkkamurtumien vuosittai‑
16 set lukumäärät. Tiedetään, että vuoden aikana
13
15 12 alkaneiden lonkkamurtumadiagnoosillisten hoi‑
14 tojaksojen määrä ei vastaa lonkkamurtumien
todellista määrää, koska lonkkamurtumapoti‑
11
4
6 10 laan hoitoepisodi on voinut alkaa edellisen vuo‑
7 den puolella, sillä se koostuu tyypillisesti useas‑
9
5 ta hoitojaksosta, ja samalla potilaalla voi olla
8
3 useita murtumia saman vuoden kuluessa. Käy‑
2 1
tännössä eri hoitojaksoille kirjautuneita lonk‑
kamurtumaleikkauksia on mahdollista pitää
KuvA 2. Lonkkamurtumien sairaanhoitopiireittäinen ikä‑ ja
merkkinä uudesta lonkkamurtumasta, vaikka
sukupuolivakioitu keskimääräinen ilmaantuvuus Suomessa
1998–2002 sataatuhatta henkilövuotta kohden (50 vuotta täyt‑ pieni määrä uusintaleikkauksia voidaan tällöin
täneiden ensimmäiset murtumat). virheellisesti tulkita uusiksi murtumiksi ja toi‑
saalta esimerkiksi konservatiivisesti hoidetut
kiilautuneet murtumat ja sellaiset murtumat,
(Sund ym., julkaisematon käsikirjoitus). joiden jälkeen potilas menehtyy ennen leikkaus‑
Suurempia hankaluuksia kuin tavanomaiset ta, jäävät havaitsematta. Lisäksi puutteet leikka‑
kattavuus‑ ja luotettavuusongelmat aiheuttaa usten kirjaamisessa rekisteriin voivat aiheuttaa
rekisteritietoihin perustuvissa ilmaantuvuus‑ harhaa. Esimerkiksi vuonna 1998 lonkkamur‑
laskelmissa lonkkamurtuman hoitoepisodin tumaleikkauksia ei muutamissa sairaaloissa kir‑
koostuminen useasta toisiinsa liittyvästä hoito‑ jattu lainkaan rekisteriin. Tässä tutkimuksessa
jaksosta, joista vain ensimmäinen liittyy uuteen saatu kaikkien lonkkamurtumien lukumäärä
lonkkamurtumaan (Langley ym. 2002). Myös vastaa sellaisten tarkasteluvuonna alkaneiden
pitkäaikaispotilaiden tunnistamiseen liittyy vas‑ lonkkamurtumadiagnoosillisten hoitojaksojen
taavankaltaisia ongelmia (Sund ja Kauppinen lukumäärää, joiden osalta oli merkintä lonkka‑
2005). Tässä tutkimuksessa käytetyt määritel‑ leikkauksesta tai joita edelsi kahden kuukauden
mät perustuvat teoreettisten lähtökohtien lisäk‑ tai kahden vuoden lonkkamurtumadiagnoosi‑
si lukuisiin alustaviin analyyseihin, ja herkkyys‑ ton aika. Kliinisesti uutena pidettävien lonkka‑
analyysien perusteella tulokset ovat luotettavia. murtumien lukumäärä on hyvin todennäköisesti
On kuitenkin huomattavaa, että tässä tutki‑ edellisillä kriteereillä saatujen arvioiden osoitta‑
muksessa raportoidut lonkkamurtumien määrät malla välillä.
1088 R. Sund

TAulukko 2. Lonkkamurtuman ilmaantuvuus Suomessa vuosina 1998–2002 sataatuhatta henkilövuotta kohden (50 vuotta täyttä‑
neiden ensimmäiset lonkkamurtumat).
1998 1999 2000 2001 2002
50 vuotta täyttäneiden ensimmäiset lonkkamurtumat

Ikä‑ ja sukupuolivakioitu ilmaantuvuus 330 315 316 315 311
Miehet
Ikävakioitu ilmaantuvuus 216 193 193 194 190
Ilmaantuvuus 50–64‑vuotiailla 57 59 55 52 54
Ilmaantuvuus 85 vuotta täyttäneillä 2 114 1 796 1 957 1 910 1 707
Naiset
Ikävakioitu ilmaantuvuus 421 412 413 413 408
Ilmaantuvuus 75–84‑vuotiailla 1 014 995 1 008 988 959
Ilmaantuvuus 85 vuotta täyttäneillä 2 482 2 598 2 572 2 561 2 552
Pitkäaikaisessa laitoshoidossa olleet
Ikä‑ ja sukupuolivakioitu ilmaantuvuus 1 650 1 582 1 485 1 371 1 747
Muut kuin pitkäaikaisessa laitoshoidossa olleet
Ikä‑ ja sukupuolivakioitu ilmaantuvuus 307 294 293 293 286
Sairaanhoitopiireittäinen ikä‑ ja sukupuolivaikoitu ilmaantuvuus
3 – Varsinais‑Suomi 320 328 320 318 329
4 – Satakunta 336 329 304 310 327
5 – Kanta‑Häme 332 312 302 313 301
6 – Pirkanmaa 341 302 313 321 316
7 – Päijät‑Häme 337 308 311 327 306
8 – Kymenlaakso 343 326 317 313 312
9 – Etelä‑Karjala 313 304 295 294 311
10 – Etelä‑Savo 284 297 311 321 312
11 – Itä‑Savo 320 303 316 298 301
12 – Pohjois‑Karjala 342 299 314 309 313
13 – Pohjois‑Savo 297 295 309 317 311
14 – Keski‑Suomi 340 325 320 330 332
15 – Etelä‑Pohjanmaa 333 294 282 285 295
16 – Vaasa 324 313 315 302 324
17 – Keski‑Pohjanmaa 303 308 312 307 298
18 – Pohjois‑Pohjanmaa 305 294 320 315 316
19 – Kainuu 309 294 289 306 310
20 – Länsi‑Pohja 331 311 317 306 339
21 – Lappi 301 314 304 310 302
1a – Helsinki 396 361 346 350 338
1b – Jorvi‑Peijas 337 325 305 305 328
1c – Hyvinkää‑Porvoo 336 325 314 317 320
1d – Lohja‑Länsi‑Uusimaa 326 322 318 307 321
Näennäisesti samanlaisista määritelmistä kiksi vuonna 1998 lonkkamurtumien todellinen

huolimatta tämän tutkimuksen arviot lonkka‑ määrä oli välillä 6 800–7 200, eli aiemmat arviot
murtumien lukumääristä ovat selvästi pienempiä ovat noin 6–14 % liian suuria. Aikaisemmissa
kuin aikaisemmissa tutkimuksissa raportoidut tutkimuksissa jokaisen henkilön ensimmäinen
(Kannus ym. 1999, Lüthje ym. 2001). Esimer‑ rekisteristä löytyvä lonkkamurtumadiagnoosilli‑

nen hoitojakso kunakin vuonna on tulkittu liit- Myös vuodenaikojen välisen kausivaihtelun
tyvän uuteen lonkkamurtumaan. Näin laskettu ilmeneminen vain muilla kuin pitkäaikaisessa
lukumäärä kertoo kyllä, kuinka useaa henkilöä laitoshoidossa olevilla on selitettävissä toimin-
kunakin vuonna on lonkkamurtuman takia hoi- takyvyn perusteella. Kirjallisuudessa kausivaih-
dettu (kuinka monella henkilöllä hoitojakso on telulle on esitetty syyksi sään aiheuttaman liuk-
alkanut tai päättynyt kyseisenä vuonna). Itse kauden lisäksi esimerkiksi kylmyyttä, hitaampia
murtumien määrän todentamiseen laskutapa reaktioaikoja ja pimeyttä (Douglas ym. 2000),
kuitenkin sopii huonosti, sillä uuden murtuman jotka varmasti myötävaikuttavat kaatumisris-
kriteeri vaihtelee henkilöittäin sen mukaan, mi- kiin ulkona liikkumaan kykenevien osalta.
hin aikaan vuotta ensimmäinen lonkkamurtuma-
diagnoosillinen hoitojakso sijoittuu. Jos se sijoit-
Lopuksi
tuu vuoden loppuun, on lonkkamurtumadiag-
noositon aika noin vuoden pituinen, kun taas Ulotettaessa toimintakykytulkinta vakioidussa
vuoden alussa alkaneen hoitojakson osalta diag- ilmaantuvuudessa havaittuun trendin katkea-
noositonta aikaa ei ole välttämättä päivääkään. miseen ei ole lainkaan yllättävää, että samoihin
Tästä seuraa, että vuoden loppupuolella lonk- aikoihin myös toimintakyvyn heikkenemisen
kansa murtaneet henkilöt lasketaan murtumata- on raportoitu tasaantuneen Suomessa (Sulan-
pauksiksi sekä todellisena murtumavuonna että der ym. 2003). Tasaantuminen on sopusoin-
seuraavana vuonna, vaikka kyse olisikin kiistat- nussa myös odotetun eliniän ja toimintakyvyn
ta samasta hoitoepisodista. muutosten yhteisvaikutuksia selittävän, väestön
Tässä tutkimuksessa todettiin, että naisilla oli ikääntymistä kuvaavan yleisen teorian kanssa,
suurempi lonkkamurtuman riski kuin miehillä joka kuitenkin ennustaa tasaantumisen olevan
paitsi pitkäaikaisessa laitoshoidossa, joka oli vain väliaikainen ilmiö (Robine ja Michel 2004).
sukupuoltakin voimakkaammin yhtey-
dessä lonkkamurtuman riskiin. Pitkä-
aikaisessa laitoshoidossa olevien osalta
ei ollut havaittavissa vuodenaikoihin Y D I N A S I AT
liittyvää kausivaihtelua lonkkamurtu-
missa, vaikka muiden osalta talvikausi ➤ Suomessa tapahtuu nykyään noin 7 000 lonkka-
oli yhteydessä lisääntyneeseen murtu- murtumaa vuodessa.
mariskiin. Löydöksiä yhdistävä tekijä ➤ Lonkkamurtumien vakioidun ilmaantuvuuden
näyttäisi olevan toimintakyky. pitkään kestänyt kasvu on tasaantunut samoihin
Toimintakyky heikkenee iän myötä aikoihin kuin väestön toimintakyvyn heikentymi-
ja naisilla nopeammin kuin miehillä nen on pysähtynyt.
(Murtagh ja Hubert 2004). Laitoshoi-
toon joutuminen heijastelee kyvyttö- ➤ Pitkäaikaisessa laitoshoidossa olevien osalta lonk-
myyttä selviytyä kotona ja kasvanutta kamurtumien ilmaantuvuudessa ei ole eroa suku-
palvelujen tarvetta, joten yhteys pitkä- puolten eikä vuodenaikojen välillä.
aikaisen laitoshoidon ja heikentyneen ➤ Lonkkamurtumien ilmaantuvuudessa ei ole mer mer-
toimintakyvyn välillä on odotettava kitseviä sairaanhoitopiirien välisiä eroja.
(Laukkanen ym. 2001). Lonkkamurtu-
➤ Rekisteritietojen huolellinen käyttö tarjoaa
mien ilmaantuvuuksien samanlaisuus
erinomaisen mahdollisuuden lonkkamurtumien
miehillä ja naisilla pitkäaikaisessa lai-
(alueellisen) ilmaantuvuuden jatkuvaan seuran-
toshoidossa on myös ymmärrettävää,
taan ja alueellisten ehkäisyprojektien vaikutta-
sillä jos laitoshoitoon valikoidutaan
vuuden arviointiin.
hoidon tarpeen perusteella, ei laitok-
sessa olevien toimintakyvyssä pitäi-
sikään olla eroa sukupuolten välillä.
1090 R. Sund

Jos toimintakykytulkinta pitää paikkansa, ker‑ liian varhaista tehdä päätelmiä lonkkamurtumi‑
tovat lonkkamurtumien alueelliset ilmaantuvuu‑ en ehkäisyn vaikuttavuudesta. Varsinaisten pää‑
det samalla myös väestön toimintakyvyn alueel‑ telmien tekoon tarvittaisiin pitempiä alueellisia
lisista eroista. Tässä tutkimuksessa raportoidut aikasarjoja lonkkamurtumien ilmaantuvuudesta
vuosien 1998–2002 alueelliset trendit ovat var‑ ja toisaalta tietoa alueellisista murtumienehkäi‑
sin monimuotoisia ja kuvastanevat toimintaky‑ syprojekteista, joissa on systemaattisesti pyritty
vyn murrosta hiukan kehityksen eri vaiheissa soveltamaan käytäntöön tehokkaaksi todennet‑
olevien väestöjen osalta. Tässä mielessä on siis tuja ehkäisykeinoja (Kannus ym. 2005).
***
Tämä tutkimus on osa Suomen Akatemian rahoittamaa projektia.
Kirjallisuutta
Boufous S, Finch CF, Lord SR. Incidence of hip fracture in New South Laukkanen P, Karppi P, Heikkinen E, Kauppinen M. Coping with activities
Wales: are our efforts having an effect? Med J Aust 2004;180:623– of daily living in different care settings. Age Ageing 2001;30:489–
6. 94.
Brody JA, Grant MD. Age‑associated diseases and conditions: implica‑ Lüthje P, Nurmi I, Palvanen M, Kannus P. Reisiluun yläosan murtu‑
tions for decreasing late life morbidity. Aging Clin Exp Res mien epidemiologia ja ehkäisymahdollisuudet. Suom Lääkäril
2001;13:64–7. 2001;56:1615–9.
Douglas S, Bunyan A, Chiu KH, Twaddle B, Maffulli N. Seasonal variation Melton LJ 3rd. Epidemiology of hip fractures: implications of the expo‑
of hip fracture at three latitudes. Injury 2000;31:11–9. nential increase with age. Bone 1996;18(3 Suppl):121S–125S.
Greenland S. Principles of multilevel modelling. Int J Epidemiol 2000; Murtagh KN, Hubert HB. Gender differences in physical disability among
29:158–67. an elderly cohort. Am J Public Health 2004;94:1406–11.
Huusko TM, Karppi P, Avikainen V, Kautiainen H, Sulkava R. The changing Robine JM, Michel JP. Looking forward to a general theory on population
picture of hip fractures: dramatic change in age distribution and aging (with discussion). J Gerontol A Biol Sci Med Sci 2004;59:
no change in age‑adjusted incidence within 10 years in Central M590–620.
Finland. Bone 1999;24:257–9. Rogmark C, Sernbo I, Johnell O, Nilsson JA. Incidence of hip fractures in
Kanis JA, Johnell O, De Laet C, Jonsson B, Oden A, Ogelsby AK. Interna‑ Malmo, Sweden, 1992–1995. A trend‑break. Acta Orthop Scand
tional variations in hip fracture probabilities: implications for risk 1999;70:19–22.
assessment. J Bone Miner Res 2002;17:1237–44. Sulander TT, Rahkonen OJ, Uutela AK. Functional ability in the elderly
Kannus P, Niemi S, Parkkari J, Palvanen M, Vuori I, Järvinen M. Hip frac‑ Finnish population: time period differences and associations,
tures in Finland between 1970 and 1997 and predictions for the 1985–99. Scand J Public Health 2003;31:100–6.
future. Lancet 1999;353:802–5. Sund R, Kauppinen S. Kuinka laskea ikääntyneiden pitkäaikaisasiak‑
Kannus P, Sievänen H, Palvanen M, Järvinen T, Parkkari J. Preven‑ kaiden määriä rekisteritietojen perusteella? Sosiaalilääk Aikak
tion of falls and consequent injuries in elderly people. Lancet 2005;42:137–44.
2005;366:1885–93.
Langley J, Stephenson S, Cryer C, Borman B. Traps for the unwary in
estimating person based injury incidence using hospital discharge Reijo SunD, VTM, tilastotieteilijä
data. Inj Prev 2002;8:332–7.
reijo.sund@stakes.fi
Stakes, vaikuttavuuden ja oikeudenmukaisuuden tutkimusryhmä
PL 220, 00531 Helsinki

VI
371
ORIGINAL ARTICLE
Quality effects of operative delay on mortality in hip fracture

treatment
R Sund, A Liski
...............................................................................................................................
Qual Saf Health Care 2005;14:371–377. doi: 10.1136/qshc.2004.012831
Background: Most hip fracture patients undergo surgery, but there is conflicting evidence on the relation
between the timing of surgery and the outcome of treatment. There is considerable variation in the length
of surgical delays between hospitals, possibly reflecting the quality of care.
Aim: To examine the associations between in-hospital surgical delay and the mortality of hip fracture
patients from a practical quality assessment perspective.
Methods: The effects of operative delay on mortality were estimated using various statistical methods
applied to observational data from 16 881 first time hip fracture patients aged 65 or older from 47
hospitals (providers) in Finland in 1998–2001.
Results: A prolonged in-hospital operative delay was associated with a higher mortality of hip fracture
See end of article for patients in individual level analyses, but the instrumental variable approach indicated that the individual
authors’ affiliations
....................... level effect was not caused by the operative delay but by inappropriate methodological assumptions. There
was extensive variation between providers in the proportion of late surgery patients. Provider level
Correspondence to: analyses showed that the effects of the provider of operative delay on mortality are quite small, but there is
R Sund, STAKES, P O Box
220, FIN-00531 Helsinki, a clear association between the proportion of late surgery patients and non-optimal treatment.
Finland; reijo.sund@ Conclusions: If provider level heterogeneity is not explicitly taken into account, studies of the effects of
stakes.fi surgical delay on outcomes are prone to serious bias. The proportion of patients with prolonged waiting
Accepted for publication time for surgery at the provider level seems to work as an effective evidence-based quality indicator.
7 August 2005 Providers should reduce unnecessary delays to surgery and identify more carefully patients not suitable for
....................... early surgery.
H
ip fractures are common injuries among older people effects of operative delay on outcomes has also not received
and are associated with substantial morbidity and sufficient attention.
mortality.1–3 Ageing of the population has resulted in an This study was undertaken to examine the associations
increase in the mean age of hip fracture patients which is between the in-hospital operative delay and the mortality of
likely to cause additional problems in the treatment and hip fracture patients in Finland in 1998–2001. Since previous
rehabilitation of patients in the future.4 5 A surgical operation studies concerning the effect of operative delay on mortality
is performed for most patients during the acute management had reported conflicting results, special attention was paid to
of hip fracture. A typical hip fracture patient is confined to methodological issues. More specific goals for empirical
hospital bed rest before surgery. A fairly short operative delay analyses were (1) to identify a suitable definition for early
is suggested in the clinical guidelines,6–10 but it is known that and late surgery groups, (2) to describe the postoperative
there is considerable variation in the operative delays excess mortality associated with prolonged operative delay,
between providers.11–17 In fact, operative delay is a commonly (3) to illustrate in which respect the early and late surgery
used process measure in health system performance assess- groups differ from each other and whether adjustment of
ment.18 Minimisation of unnecessary preoperative inpatient these observed factors changes the association between
care can be used as an economic justification for shorter operative delay and mortality, (4) to compare the risk
operative delays. For quality measurement purposes, the adjusted proportion of late surgery patients between provi-
operative delay should also have confirmed effects on ders, and (5) to investigate provider level heterogeneity in
outcomes otherwise improvements may possibly be targeted terms of 1 year mortality and the proportion of patients with
to issues which do not improve health.19 The evidence late surgery.
concerning the optimal timing of surgery in relation to the
overall outcomes of hip fracture treatment is quite mixed.20–29
In short, the main reasons for variations in the operative METHODS
delay are provider level system factors (availability of Data set
required resources such as operating rooms, support services, The total population of hip fracture patients in 1998–2001
or personnel with expertise) and clinical decisions based on was identified in the Finnish Health Care Register. The
patient specific factors (co-morbidity, medical stabilisation). medical histories (1987–2002) and deaths (1998–2002) of the
In some studies conflicting results may be due to methodo- hip fracture population were extracted from the Finnish
logical shortcomings (small numbers of observations, no Hospital Discharge Register, Finnish Health Care Register,
adjustment for confounding factors), but more sophisticated and from the National Causes of Death Register using the
evidence as to whether operative delay has an independent unique personal identification numbers of the patient
effect on primary outcomes is also conflicting.30–38 Another population. Each record in these registers includes data such
open question is the definition of the time period for delayed as patient and provider ID numbers, age, sex, area codes, and
surgery.18 The impact of provider level heterogeneity on the diagnosis and operation codes, as well as dates of admission,
www.qshc.com

372 Sund, Liski
operation, and discharge (or death). The completeness and 0.10

accuracy of the registers is known to be good.39–41 0.09
0.08
Data preprocessing
Excess mortality
0.07
Data were preprocessed so that information concerning 0.06
patients aged 65 or older with a first hip fracture could be 0.05
accurately identified.42 For the purposes of this study, the 0.04
data were restricted to patients operated on using an internal 0.03
fixation, a prosthesis, or a total hip replacement. Patients
0.02
with incomplete data were excluded from the analyses. A few
0.01
providers (hospitals) had not reported operation dates for the 0
years 1998–1999, but otherwise no systematic bias was found 10 30 60 90 120 180 240 300 365
in the sensitivity analyses. The final study population Days
consisted of 16 881 patients, which represents 83.3% of all
first time hip fracture patients aged 65 or older. Forty six Figure 2 Unadjusted excess risk of death (with 95% confidence
providers had at least 50 cases, while the remaining 154 cases interval) for late hip fracture surgery.
from other providers were combined into a single figure. The
existence of possible co-morbidities was extracted for each Provider level heterogeneity was examined in terms of the
patient from his or her medical history using the diagnosis proportion of late surgery patients and 1 year mortality. The
codes recorded in the data.34 43 The extraction method was provider specific overall mortality rates were stratified
adapted and updated from the original Charlson co-morbid- according to early and late surgery, recognising the fact that
ity categories44 and applied to the current data set. the weighted average of these mortalities gives the overall
mortality. The provider level associations between these three
Statistical analyses mortality rates and the proportions of late surgery patients
Cumulative probabilities of unadjusted mortality were were estimated using Stein estimation with an assumption of
calculated using the product limit estimators. Excess risk of linear trends for overall and early surgery mortality rates.50
death was defined as difference in cumulative probabilities of The associations were also analysed separately according to
mortality between two comparable groups. Hazard ratios severity. The five severity groups were established by
were obtained using the proportional hazards model allowing classifying the patients into five equal sized groups in terms
the adjustment of observed confounders. Since only a limited of their predicted 1 year mortality based on the logistic
number of covariates were observed, the biasing influence of regression model.51
unobserved heterogeneity was statistically controlled by
incorporating the gamma frailty term into the model.45 RESULTS
The significance of the differences between the unadjusted Effect of length of operative delay on mortality
patient characteristics was determined using the z test The cumulative probabilities of unadjusted mortality rates
(binary variables) or x2 test (categorical variables). The attributable to different lengths of operative delay are shown
logistic regression model was used to calculate odds ratios in fig 1. Waiting times of 0–2 nights had the same effect on
and significances of differences between adjusted patient mortality, but there was a significant increase in mortality
characteristics. The same logistic model was also used for with waiting times of 3–4 nights (p,0.001), and waiting
predicting the late surgery probability for each patient. These times of 5 or more nights resulted in an even higher mortality
probabilities were then aggregated to the provider level, rate. Early surgery was defined as a waiting time of 0–2
resulting in the expected number of late surgery patients for nights (n = 14 426, 85.5%) and late surgery as a waiting time
each provider.46 The provider specific ratios between observed of at least 3 nights (n = 2455, 14.5%). The unadjusted hazard
and expected numbers of late surgery patients were further ratio between late and early surgery was 1.24 (95%
modelled using the hierarchical gamma-Poisson model to confidence interval (CI) 1.15 to 1.34, p,0.0001).
obtain the comparable risk adjusted proportions of late
surgery patients for providers.47 The weekday adjusted Excess risk of death
proportion was calculated using the predictive margin The unadjusted excess risk of death for late surgery compared
technique.48 In the instrumental variables analysis49 the with early surgery quickly rose to 3%, then slowly increased
1 year adjusted mortality was calculated for each admission during follow up to about 5% at 1 year (fig 2).
day of the week and the null hypothesis of equal mortalities Complementary analyses revealed that the excess risk was
was tested using the x2 goodness of fit test.30 quite stable and near to its maximum at 1–3 years and then
began to decrease slowly. As the lower confidence limits in
fig 2 indicate, the unadjusted excess risk of death became
0.4
5+ nights statistically significant 2 weeks after the operation.
Proportion of death patients
3, 4 nights
0.3
0, 1, Characteristics of early and late surgery groups
2 nights Characteristics of the patients with hip fracture classified into
0.2 early and late surgery groups are shown in table 1, together
with the corresponding odds ratios for late surgery adjusted
0.1 for the other characteristics in the table. Early surgery was
associated with being 85 years or older, of female sex, having
a pertrochanteric fracture, and receiving long term inpatient
0
10 30 60 90 120 180 240 300 365 care during the year preceding the admission. Admission on a
Days Wednesday and a medical history of dementia or cataract
were also associated with early surgery. The risks for late
Figure 1 Cumulative probabilities of unadjusted mortality following hip surgery were likely to be increased by admission from a
fracture for operative delays of various lengths. nursing home, health centre or readmission shortly after
www.qshc.com

Effects of operative delay in hip surgery on mortality 373
Table 1 Characteristics of patients with hip fracture

Early surgery Late surgery Unadjusted Adjusted
Characteristic (n = 14 426, 85.5%) (n = 2455, 14.5%) p value Odds ratio 95% CI p value
Age (years) ,0.001 0.13

65–74 2721 (18.9) 504 (20.5) 1
75–79 2995 (20.8) 491 (20.0) 0.885 0.770 to 1.017 0.086
80–84 3483 (24.1) 616 (25.1) 0.954 0.833 to 1.092 0.494
85–89 3379 (23.4) 556 (22.6) 0.867 0.753 to 0.998 ,0.05
90+ 1848 (12.8) 288 (11.7) 0.836 0.706 to 0.989 ,0.05
Sex ,0.001 ,0.01

Male 3443 (23.9) 696 (28.4) 1
Female 10983 (76.1) 1759 (71.6) 0.844 0.762 to 0.935 ,0.01
Fracture type ,0.001 ,0.05

Pertrochanteric 4529 (31.4) 694 (28.3) 1
Subtrochanteric 835 (5.8) 165 (6.7) 1.241 1.027 to 1.500 ,0.05
Femoral neck 9065 (62.8) 1596 (65.0) 1.135 1.029 to 1.253 ,0.05
Medical history
Cancer 1676 (11.6) 299 (12.2) 0.223 1.006 0.878 to 1.151 0.935
Diabetes 1470 (10.2) 290 (11.8) ,0.01 1.006 0.873 to 1.160 0.931
Dementia 2761 (19.1) 411 (16.7) ,0.01 0.863 0.762 to 0.976 ,0.05
Hypertension 2188 (15.2) 444 (18.1) ,0.001 1.104 0.979 to 1.244 0.106
Cardiovascular disease 5110 (35.4) 1108 (45.1) ,0.001 1.462 1.329 to 1.609 ,0.001
Cerebrovascular disease 2807 (19.5) 562 (22.9) ,0.001 1.110 0.995 to 1.238 0.062
Peripheral vascular disease 632 (4.4) 174 (7.1) ,0.001 1.447 1.207 to 1.736 ,0.001
Chronic pulmonary disease 1205 (8.4) 265 (10.8) ,0.001 1.148 0.991 to 1.330 0.066
Anaemia 1161 (8.0) 236 (9.6) ,0.01 1.135 0.972 to 1.319 0.109
Diseases of the nervous system* 852 (5.9) 139 (5.7) 0.333 0.895 0.740 to 1.084 0.257
Eye diseases� 4182 (29.0) 673 (27.4) 0.058 0.903 0.816 to 1.000 ,0.05
Diseases of the digestive system` 2316 (16.1) 396 (16.1) 0.474 0.940 0.833 to 1.060 0.310
Other diseases1 2429 (16.8) 432 (17.6) 0.185 0.986 0.877 to 1.108 0.814
,0.001 ,0.01
Pre-admission residence
Home (.1 week) 9444 (65.5) 1559 (63.5) 1
Home (max 1 week) 685 (4.7) 153 (6.2) 1.300 1.076 to 1.570 ,0.01
Nursing home 2508 (17.4) 416 (16.9) 1.215 1.016 to 1.451 ,0.05
Health centre� 1249 (8.7) 143 (9.9) 1.177 1.000 to 1.384 ,0.05
Hospital 540 (3.7) 84 (3.4) 0.888 0.694 to 1.135 0.341
.180 days inpatient care during year 2569 (17.8) 402 (16.4) ,0.05 0.823 0.691 to 0.979 ,0.05
preceding fracture
Day of admission ,0.001 ,0.001

Monday 2479 (17.2) 307 (12.5) 1
Tuesday 2289 (15.9) 288 (11.7) 1.026 0.864 to 1.218 0.770
Wednesday 2248 (15.6) 220 (9.0) 0.803 0.668 to 0.964 ,0.05
Thursday 2091 (14.5) 401 (16.3) 1.567 1.334 to 1.840 ,0.001
Friday 1898 (13.2) 678 (27.6) 2.970 2.559 to 3.447 ,0.001
Saturday 1744 (12.1) 358 (14.6) 1.686 1.429 to 1.990 ,0.001
Sunday 1677 (11.6) 203 (8.3) 0.989 0.818 to 1.195 0.907
Year** ,0.001 ,0.001

1998 3275 (83.1) 667 (16.9) 1
1999 3325 (84.8) 597 (15.2) 0.877 0.775 to 0.992 ,0.05
2000 3869 (86.1) 625 (13.9) 0.801 0.709 to 0.904 ,0.001
2001 3957 (87.5) 566 (12.5) 0.700 0.619 to 0.793 ,0.001
*Includes Parkinson’s disease, hemiplegia or paraplegia, and epilepsy.

�Includes glaucoma and cataract.
Ìncludes peptic ulcer disease, abdominal hernia, liver disease, disorders of gallbladder and biliary tract, and irritable colon.
1Includes hypothyroidism, gout, renal disease, arthrosis, and rheumatic disease.
�Health centre refers to the inpatient ward of the local primary health care unit
**Percentages in parentheses are the proportions of the categories in the current year.
discharge, admission between Thursday and Saturday, and unobserved covariates, the hazard ratio remained significant,
co-morbidities including cardiovascular disease and periph- indicating an increased mortality for late surgery. However,
eral vascular disease (and, prior to adjustment, diabetes, after the pseudo randomised assignment of patients into
cerebrovascular disease, chronic pulmonary disease, hyper- early and late surgery groups using admission day of the
tension, and anaemia). The proportion of patients having late week as an instrumental variable, the difference in the
surgery substantially decreased during the period of the adjusted 1 year mortality rate was found not to be
study. attributable to operative delay (p = 0.069).
Adjusted effect of operative delay on mortality Proportions of late surgery patients

The hazard ratio for the late surgery reduced to 1.18 (95% CI There was extensive variation in the proportions of late
1.09 to 1.28, p,0.0001) after adjusting for the provider and surgery patients between providers (fig 3). The conservative
the characteristics in table 1. Even after controlling for the 95% confidence interval crosses the potentially achievable
www.qshc.com

374 Sund, Liski
0.7
0.6
One year mortality

0.5
0.4
0.3
0.2
0.1
0
0 0.1 0.2 0.3 0.4 0.5
Share of patients with late surgery
Figure 4 Scatter plot of adjusted mortality for surgically treated hip

fracture patients and the share of patients undergoing late surgery for
different providers (The triangles represent the overall mortality of
operated hip fracture patients for each provider and the crosses the
mortality of late surgery patients for each provider. The line with a
positive slope is the trend of overall mortality and the dotted curved line
indicates the association between the share of late surgery patients and
late surgery mortality across providers).
0 10 20 30 40 50 60
% The unadjusted excess mortality for late surgery patients
increased during the year following the operation. It is
Figure 3 Provider-specific risk adjusted percentages of hip fracture
interesting that the short term mortality differed only slightly
patients undergoing late surgery. The thin parts of the lines correspond to
95% confidence intervals and the thick parts correspond to 50% between early and late surgery groups, but the difference
confidence intervals. The dashed line refers to the potentially achievable became much clearer after the perioperative period. This may
percentage (7.7%) of patients undergoing late surgery. be partly due to the fact that almost half the perioperative
deaths are unavoidable in an unselected population.53 If
patients are going to die shortly after the fracture—regardless
(20th centile)52 92.3% share of patients undergoing early of the quality of the hip fracture treatment provided—it is
surgery for less than 50% of the providers. obvious that the treatment effect turns out to be small. The
treatment effect becomes observable for long term mortality
Trend analysis at provider level but is potentially biased because of heterogeneity between
The simultaneous examination of the provider level propor- the early and late surgery groups in patient characteristics
tions of late surgery patients and 1 year mortality rates affecting the outcome.20 30 33 36
showed that there was a statistically significant positive trend Several patient characteristics differed between the early
for an association between a larger share of late surgery and late surgery groups, but the prolonged operative delay
patients and 1 year overall mortality (p,0.05), and that a seemed to increase the (long term) mortality significantly
smaller proportion of late surgery patients was (non-linearly) even after adjustment for these observed patient character-
associated with a higher mortality rate for these patients istics. There were no detailed clinical data available which
(fig 4). obviously makes the adjustments only partial. The bias
attributable to the unobserved covariates was controlled by
allowing extra variation in the model, but the treatment
Trend analyses for severity groups
effect remained significant. However, this kind of model is
The associations were further analysed according to severity
still prone to bias if the relationships between the risk factors
(measured as predicted 1 year mortality based on the same
and mortality are not correctly specified in the model.54 The
observed patient characteristics as in table 1) in five severity
pseudo randomizing instrumental variable approach was
groups of patients (table 2). The proportion of patients
therefore applied and indicated that the increased mortality
undergoing late surgery was only slightly higher in the more
was not due to operative delay. Since the interpretations
severe patient groups but the 1 year mortality rate increased
based on the instrumental variable approach also require
significantly with severity, as expected. The slope of the trend
strong and questionable assumptions,30 33 the sizeable differ-
of overall mortality increased with the severity but was
ence in the estimates of the effect of operative delay on
statistically significant only for the most severe patients. The
mortality between methods means that there is a need for
positive slope of early surgery mortality was statistically
hypotheses to explain such differences.
significant in severity groups 3 and 4, but for the most severe
For instance, one could try to separate the acceptable
patients this association disappeared. A smaller proportion of
delays from the unacceptable ones27 36 or one could record the
late surgery patients was associated with a higher late
reason for the delay in the data.17 31 32 In the strictest sense,
surgery mortality rate in all groups except the most severe
this kind of approach needs the assumption that the
patients.
acceptable and unacceptable delays can be defined and
measured uniquely.55–57 In practice, an acceptable delay
DISCUSSION corresponds to a clinical decision to postpone the operation
In this study the definition of late surgery turned out to be 3 and the assumption is approximately fulfilled if the clinical
or more nights in hospital after admission, since shorter practice remains constant. Because the data used in this
waiting times were not associated with higher unadjusted study did not include the reason for the delay, the hypothesis
mortality. A similar ‘‘long’’ delay is commonly used as the of constant clinical practice was examined indirectly using
definition of late surgery in other studies.17 18 20 34 provider level analyses.
www.qshc.com

Table 2 Effect of proportion of late surgery patients on mortality in different severity groups of hip fracture patients
One year mortality
10% of patients 15% of patients 20% of patients 30% of patients Slope*

n (%) with late surgery with late surgery with late surgery with late surgery (p value)
All patients (adjusted mortality)�

Overall 16881 29.0 29.3 29.5 30.0 0.5 (,0.05)
Early surgery 14426 (85) 28.4 28.8 29.3 30.2 0.9 (,0.01)
Late surgery` 2455 (15) 35.0` 31.9 30.5 29.7
Patient group 1 (least severe)
Overall 3376 (20) 9.4 9.3 9.2 9.0 20.2 (0.232)
Early surgery 2931 (87) 8.8 8.6 8.4 8.1 20.3 (0.168)
Late surgery 445 (13) 15.4 13.4 12.4 11.1
Patient group 2
Overall 3376 (20) 18.0 17.9 17.8 17.6 20.2 (0.283)
Early surgery 2932 (87) 17.4 17.5 17.7 17.9 0.2 (0.310)
Late surgery 444 (13) 23.8 20.1 18.4 16.7
Patient group 3
Overall 3376 (20) 26.9 27.2 27.6 28.2 0.6 (0.101)
Early surgery 2875 (85) 26.6 27.5 28.5 30.4 1.9 (,0.01)
Late surgery 501 (15) 29.7 25.4 23.8 23.1
Patient group 4
Overall 3377 (20) 37.9 38.3 38.7 39.5 0.8 (0.126)
Early surgery 2869 (85) 37.4 38.3 39.2 41.1 1.8 (,0.05)
Late surgery 508 (15) 43.0 38.5 36.7 35.8
Patient group 5 (most severe)
Overall 3376 (20) 51.4 51.9 52.5 53.5 1.1 (,0.05)
Early surgery 2819 (84) 50.7 50.6 50.6 50.5 20.1 (0.455)
Late surgery 557 (16) 57.9 59.3 59.9 60.6
*Change in percentage of deaths when the share of patients with late surgey increased 10%.
�Corresponds to fig 4. Example: Expected mortality for late surgery patients in provider with 10% of patients with late surgery is 35%. These percentages can be
interpreted as coordinates (0.10, 0.35) which hit the dotted curve in fig 4.
`Calculated from the equation: ml(X) = (m(X) 2 (12X) * me(X))/X, where X is the proportion of late surgery patients, m(X) refers to overall mortality at X, me(X) to
early surgery mortality at X, and ml(X) to late surgery mortality at X. Example: ml(0.1) = (0.2903 2 0.9 * 0.2837)/0.1 = 0.3497.
The provider specific proportion of late surgery patients, the medical problems which require late surgery and the
adjusted for observed patient characteristics, showed that prolonged surgical delay does not itself increase the
there was extensive variation between providers, which is a mortality.28 30 31 36 For patients in the most severe group
common finding.11–17 A simple performance assessment mortality was also higher for the late surgery patients for all
interpretation is that the percentage of late surgery patients providers. However, in this group the mortality rate in the
can be reduced to a potentially achievable level52 that can also late surgery patients was lower in the providers with a small
be interpreted as the upper limit for the proportion of proportion of late surgery patients and there were no
acceptable delayed patients. Correspondingly, the expected differences in mortality rates among early surgery patients
proportion of unacceptable delayed patients is the proportion between providers. This means that, in this group, it is
of late surgery patients exceeding this upper limit. This leads essential to perform early surgery in all patients who can
to the hypothesis that the overall mortality of hip fracture withstand it because the significantly prolonged surgical
patients should increase with an increasing proportion of late delay makes the patient’s condition worse and increases
surgery patients given that the longer operative delay would mortality.22 23 29 32 33 For groups 2–4 the interpretation is more
have an adverse effect on mortality. In fact, a statistically difficult than for patients in the extreme groups. Since the
significant trend between the provider specific proportions of mortality rate in early surgery patients increases significantly
late surgery patients and overall 1 year mortality gives and is higher even than the mortality rate of late surgery
empirical evidence for the hypothesis, but the actual volume patients for providers with a large proportion of late surgery
of the effect was small. Another provider level hypothesis is patients, it seems that in these groups an operation
that the long term mortality of late surgery patients is higher performed too early may cause more harm than the
if only the patients unfit for surgery are delayed, since the prolonged waiting time.21 22 26
unfit condition for surgery is also a risk factor for 1 year Other differences between providers in the proportion of
mortality. This hypothesis was also supported by empirical patients with late surgery result mainly from lack of
evidence which revealed a non-linear association between resources such as temporary unavailability of the operating
the provider specific proportion of late surgery patients and rooms or surgeons and problems with obtaining medical
the 1 year mortality rate for these patients. clearance from other specialities in off hours.57 58 The
A more careful examination of the groups of patients with importance of these factors can be illustrated using the effect
different predicted mortality (severity) gave even more of the admission day on the in-hospital operation delay; the
insight into the provider level association between the mean proportion of patients undergoing late surgery would
proportion of late surgery patients and mortality. Intuitively be reduced from 14.5% to 9.0% (p,0.0001) if all patients had
speaking, the least severe patients are young, come from been admitted on ‘‘the best day’’.
home, and are without severe medical conditions. In spite of the large nationwide database and the advanced
Correspondingly, the most severe patients are older with methods, the observational study design and administrative
much co-morbidity and come from residential care.21 31 53 For data require a special orientation to the analyses and caution
patients in the least severe group, mortality was higher in the should be exercised when interpreting the results. Data were
late surgery patients for all providers. This probably indicates available for 1998–2001 and the practices may have
that, in this group, hip fracture related mortality is caused by disproportionately changed for the different providers during
www.qshc.com

376 Sund, Liski
that time. The results must therefore be interpreted as the 12 Freeman C, Todd C, Camilleri-Ferrante C, et al. Quality improvement for
patients with hip fracture: experience from a multi-site audit. Qual Saf Health
mean situation between the years 1998–2001 for each Care 2002;11:239–45.
provider. Non-operated patients were excluded from the 13 New Zealand Health Information Service (NZHIS). Fracture of neck of femur
analyses since the operative delay was not defined uniquely services in New Zealand hospitals 1999/2000. Wellington: New Zealand
Health Information Service (NZHIS), 2002.
for these patients. Inclusion of these patients would certainly 14 Bottle A, Jarman B, Aylin P, et al. Some way to go for consistent
increase the mortality rate since the typical non-operated implementation of guidance on hip fracture. BMJ 2004;328:1097.
patient is in a worse condition and dies before any operation 15 Scottish Hip Fracture Audit Steering Group (SHFASG). Scottish hip fracture
audit report 2002. Edinburgh: NHS Scotland, 2002.
can even take place. The trend analyses are intended to
16 Heikkinen T, Willig R, Hanninen A, et al. Hip fractures in Finland: a
summarise the systematic parts of the variation between the comparison of patient characteristics and outcomes in six hospitals.
providers. However, there seems to be some unexplained Scand J Surg 2004;93:234–40.
variation, especially in the proportions of late surgery 17 Todd CJ, Freeman CJ, Camilleri-Ferrante C, et al. Differences in mortality after
fracture of hip: the East Anglian audit. BMJ 1995;310:904–8.
patients, because of the small number of patients for some 18 Fairbank J, Goldacre MJ, Mason A, et al. Health outcome indicators:
providers, so the results of the trend analyses need to be fractured proximal femur, Report of a Working Group to the Department of
confirmed and validated in further studies. Health. Oxford: National Centre for Health Outcomes Development, 1999.
19 Brook RH, McGlynn EA, Cleary PD. Quality of health care. Part 2: measuring
Study designs which explicitly take into account the quality of care. N Engl J Med 1996;335:966–70.
heterogeneity at the provider level are a prerequisite for 20 Zuckerman JD, Skovron ML, Koval KJ, et al. Postoperative complications and
unbiased results since the selection bias may vary between mortality associated with operative delay in older patients who have a fracture
of the hip. J Bone Joint Surg Am 1995;77:1551–6.
providers. One could also include the provider level char- 21 Mullen JO, Mullen NL. Hip fracture mortality. A prospective, multifactorial
acteristics in the analyses,35 37 but easily observable char- study to predict and minimize death risk. Clin Orthop Relat Res
acteristics such as hospital type or number of surgeons 1992;280:214–22.
typically allow only indirect interpretations. A more direct 22 Kenzora JE, McCarthy RE, Lowell JD, et al. Hip fracture mortality. Relation to
age, treatment, preoperative illness, time of surgery, and complications. Clin
approach would be to observe the performance of providers Orthop Relat Res 1984;186:45–56.
continuously and to ask the providers themselves to clarify 23 Casaletto JA, Gatt R. Post-operative mortality related to waiting time for hip
the reasons for their performance.12 58 59 fracture surgery. Injury 2004;35:114–20.
24 McGoye PF, Evans J. Fractures of the hip: Immediate versus delayed
In Finland the mean proportion of late surgery patients has treatment. Can Med Assoc J 1960;83:260–3.
decreased during the years of the study, but extensive 25 Davis FM, Woolner DF, Frampton C, et al. Prospective, multi-centre trial of
variation between providers exists. In addition to possible mortality following general or spinal anaesthesia for hip fracture surgery in
the elderly. Br J Anaesth 1987;59:1080–8.
system problems, there also seems to be variation in the 26 Kern IB. Fractures of the neck of the femur: a review of 308 cases treated at
clinical practices for judging which patients are suitable for the Prince of Wales Hospital from January 1960 to June 1961. Med J Aust
early surgery. Identification of good practices and the release 1962;49:554–7.
27 Bredahl C, Nyholm B, Hindsholm KB, et al. Mortality after hip fracture: results
of a national clinical guideline for hip fracture treatment will of operation within 12 h of admission. Injury 1992;23:83–6.
probably improve the situation. In this sense, the proportion 28 Holt EM, Evans RA, Hindley CJ, et al. 1000 femoral neck fractures: the effect
of patients with a prolonged waiting time for a hip fracture of pre-injury mobility and surgical experience on outcome. Injury
operation seems to be an effective evidence based quality 1994;25:91–5.
29 Hoenig H, Rubenstein LV, Sloane R, et al. What is the role of timing in the
indicator, since it clearly indicates an area for improvement. surgical and rehabilitative care of community-dwelling older persons with
The effects of operative delay on mortality at the patient level acute hip fracture? Arch Intern Med 1997;157:513–20.
are quite small, but at the provider level the association 30 Ho V, Hamilton BH, Roos LL. Multiple approaches to assessing the effects of
delays for hip fracture patients in the United States and Canada. Health Serv
between the proportion of late surgery patients and non- Res 2000;34:1499–518.
optimal treatment is clear. 31 Grimes JP, Gregory PM, Noveck H, et al. The effects of time-to-surgery on
mortality and morbidity in patients following hip fracture. Am J Med
..................... 2002;112:702–9.
32 Moran CG, Wenn RT, Sikand M, et al. Early mortality after hip fracture: is
Authors’ affiliations delay before surgery important? J Bone Joint Surg Am 2005;87:483–9.
R Sund, A Liski, National Research and Development Centre for Welfare 33 McGuire KJ, Bernstein J, Polsky D, et al. The 2004 Marshall Urist award:
and Health (STAKES), Outcomes and Equity Research, Helsinki, Finland delays until surgery after hip fracture increases mortality. Clin Orthop Relat
Res 2004;428:294–301.
This research was supported by a research grant from the Academy of 34 Roos LL, Walld RK, Romano PS, et al. Short-term mortality after repair of hip
Finland. fracture. Do Manitoba elderly do worse? Med Care 1996;34:310–26.
35 Weller I, Wai EK, Jaglal S, et al. The effect of hospital type and surgical delay
on mortality after surgery for hip fracture. J Bone Joint Surg Br
2005;87:361–6.
REFERENCES 36 Orosz GM, Magaziner J, Hannan EL, et al. Association of timing of surgery
1 Zuckerman JD. Hip fracture. N Engl J Med 1996;334:1519–25. for hip fracture and patient outcomes. JAMA 2004;291:1738–43.
2 Roberts SE, Goldacre MJ. Time trends and demography of mortality after 37 Hamilton BH, Hamilton VH, Mayo NE. What are the costs of queuing for hip
fractured neck of femur in an English population, 1968–98: database study. fracture surgery in Canada? J Health Econ 1996;15:161–85.
BMJ 2003;327:771–5. 38 Hamilton BH, Bramley-Harker RE. The impact of the NHS reforms on queues
3 Liporace FA, Egol KA, Tejwani N, et al. What’s new in hip fractures? Current and surgical outcomes in England: Evidence from hip fracture patients. Econ J
concepts. Am J Orthop 2005;34:66–74. 1999;109:437–62.
4 Marks R, Allegrante JP, Ronald MacKenzie C, et al. Hip fractures among the 39 Keskimäki I, Aro S. Accuracy of data on diagnoses, procedures and accidents
elderly: causes, consequences and control. Ageing Res Rev 2003;2:57–93. in the Finnish Hospital Discharge Register. Int J Health Sci 1991;2:15–21.
5 Thorngren KG, Hommel A, Norrman PO, et al. Epidemiology of femoral neck 40 Lahti RA, Penttila A. The validity of death certificates: routine validation of
fractures. Injury 2002;33:C1–7. death certification and its effects on mortality statistics. Forensic Sci Int
6 Royal College of Physicians. Fractured neck of femur. Prevention and 2001;115:15–32.
management. Summary and recommendations of a report of the Royal 41 Luthje P, Nurmi I, Kataja M, et al. Incidence of pelvic fractures in Finland in
College of Physicians. J R Coll Physicians Lond 1989;23:8–12. 1988. Acta Orthop Scand 1995;66:245–8.
7 Chilov MN, Cameron ID, March LM. Evidence-based guidelines for fixing 42 Sund R. Utilisation of administrative registers using scientific knowledge
broken hips: an update. Med J Aust 2003;179:489–93. discovery. Intelligent Data Analysis 2003;7:501–19.
8 Scottish Intercollegiate Guidelines Network. Prevention and management of 43 Romano PS, Roos LL, Jollis JG. Adapting a clinical comorbidity index for use
hip fracture in older people. Edinburgh: Scottish Intercollegiate Guidelines with ICD-9-CM administrative data: differing perspectives (with discussion).
Network, 2002. J Clin Epidemiol 1993;46:1075–90.
9 New Zealand Guidelines Group. Acute management and immediate 44 Charlson ME, Pompei P, Ales KL, et al. A new method of classifying prognostic
rehabilitation after hip fracture amongst people aged 65 years and over. comorbidity in longitudinal studies: development and validation. J Chronic Dis
Wellington, NZ: New Zealand Guidelines Group, 2003. 1987;40:373–83.
10 Thorngren KG, Dolk T, Jarnlo GB, et al. National guidelines for care and 45 Hougaard P. Frailty models for survival data. Lifetime Data Anal
treatment of hip fractures (in Swedish). Stockholm: The National Board of 1995;1:255–73.
Health and Welfare, 2003. 46 Goldstein H, Spiegelhalter DJ. League tables and their limitations: statistical
11 Audit Commission. United they stand, Co-ordinating care for elderly patients issues in comparisons of institutional performance (with discussion). J R Stat
with hip fracture. London: Audit Commission, 2000. Soc Ser A Stat Soc 1996;159:385–443.
www.qshc.com

47 Howley PP, Gibberd R. Using hierarchical models to analyse clinical 54 Shwartz M, Ash AS. Estimating the effect of an intervention from
indicators: a comparison of the gamma-Poisson and beta-binomial models. observational data. In: Iezzoni LI, eds. Risk adjustment for measuring health
Int J Qual Health Care 2003;15:319–29. care outcomes. 3rd edn. Chicago: Health Administrative Press,
48 Graubard BI, Korn EL. Predictive margins with survey data. Biometrics 2003:275–95.
1999;55:652–9. 55 Schultz RJ, Whitfield GF, LaMura JJ, et al. The role of physiologic monitoring
49 Newhouse JP, McClellan M. Econometrics in outcomes research: the use of in patients with fractures of the hip. J Trauma 1985;25:309–16.
instrumental variables. Annu Rev Public Health 1998;19:17–34. 56 Charalambous CP, Yarwood S, Paschalides C, et al. Factors delaying surgical
50 Greenland S. Principles of multilevel modelling. Int J Epidemiol treatment of hip fractures in elderly patients. Ann R Coll Surg Engl
2000;29:158–67. 2003;85:117–9.
51 Daley J, Iezzoni LI, Shwartz M. Conceptual and practical issues in developing 57 Orosz GM, Hannan EL, Magaziner J, et al. Hip fracture in the older patient:
risk-adjustment methods. In: Iezzoni LI, eds. Risk adjustment for measuring reasons for delay in hospitalization and timing of surgical repair. J Am Geriatr
health care outcomes. 3rd edn. Chicago: Health Administration Press, Soc 2002;50:1336–40.
2003:179–205. 58 Beupre LA, Johnston DB, Jones CA, et al. Continuum of care and
52 Gibberd R, Hancock S, Howley P, et al. Using indicators to quantify the integration of services for treatment of elderly patients with hip
potential to improve the quality of health care. Int J Qual Health Care fracture. Alberta: Alberta Heritage Foundation for Medical Research
2004;16:i37–43. (AHFMR), 2003.
53 Foss NB, Kehlet H. Mortality analysis in hip fracture patients: implications for 59 Parker MJ, Currie CT, Mountain JA, et al. Standardised audit of hip fracture in
design of future outcome trials. Br J Anaesth 2005;94:24–9. Europe (SAHFE). Hip Int 1998;8:10–5.
www.qshc.com


Methodol

Transféré par

Informations du document

Description originale:

Titre original

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

Methodol

Transféré par

Droits d'auteur :

Formats disponibles

Reijo Sund

Methodological Perspectives for

National Research and Development Centre

Research Report 174

© Author and STAKES

Translation of the Swedish abstract: Done Information Oy

Cover design: Harri Heikkilä

ISBN 978-951-33-2114-7 (nid.)

Research Report 174 Methodological Perspectives for Register-Based 

Reijo Sund. Methodological Perspectives for Register-Based Health System

Keywords: administrative registers, methodology, secondary data, hip fractures,

Research Report 174 Methodological Perspectives for Register-Based 

Reijo Sund. Methodological Perspectives for Register-Based Health System Perfor-

Terveydenhuoltojärjestelmän resurssit ovat rajallisia. Jotta näitä resursseja pystyt-

Avainsanat: hallinnolliset rekisterit, metodologia, toisen käden aineistot, lonkka-

 Methodological Perspectives for Register-Based Research Report 174

Reijo Sund. Methodological Perspectives for Register-Based Health System Perfor-

Nyckelord: administrativa register, metodologi, andrahandsmaterial, höftfrakturer,

Research Report 174 Methodological Perspectives for Register-Based 

2 Register-based data analysis............................................................................. 19

3 Register-based health system performance assessment............................ 28

4 Register-based data on hip fractures – sources and validity.................. 33

5 Hip fracture incidence....................................................................................... 43

Research Report 174 Methodological Perspectives for Register-Based 11

7 Operative delay as a performance indicator................................................ 66

12 Methodological Perspectives for Register-Based Research Report 174

[1] Sund R: Utilisation of administrative registers using scientific knowledge

[2] Sund R, Nurmi-Lüthje I, Lüthje P, Tanninen S, Narinen A, Keskimäki I:

[3] Sund R: Utilization of routinely collected administrative data in monitoring

[4] Sund R, Kauppinen S: Kuinka laskea ikääntyneiden pitkäaikaisasiakkaiden

[5] Sund R: Lonkkamurtumien ilmaantuvuus Suomessa 1998–2002. Duodecim

Research Report 174 Methodological Perspectives for Register-Based 13

Research Report 174 Methodological Perspectives for Register-Based 15

1.1 Aim of the study

16 Methodological Perspectives for Register-Based Research Report 174

1.2 Structure of the study

This summary synthesizes the methodological ideas from these contributions.

Research Report 174 Methodological Perspectives for Register-Based 17

18 Methodological Perspectives for Register-Based Research Report 174

2 Register-based data analysis

A register is an information system which continuously records subject-based data

2.1 From technical data processing to a

Research Report 174 Methodological Perspectives for Register-Based 19

Figure 1. Research process schema

information are, for instance, evidence, non-redundancy, simplicity, novelty, and

2.2 Secondary data as a methodological

20 Methodological Perspectives for Register-Based Research Report 174

Phenomenon Coding Register Decoding Data

Figure 2. Schematic diagram of information communication via administrative registers

I have proposed that the methodological challenges related to the utilization

2.3 Prerequisites in the register-based data

Research Report 174 Methodological Perspectives for Register-Based 21

2.3.1 Principles of measurement

22 Methodological Perspectives for Register-Based Research Report 174

2.3.2 Information science

Research Report 174 Methodological Perspectives for Register-Based 23

Phenomenon Knowledge component

Theoretical measurement properties

Figure 3. An overview of the extended pyramid framework

More generally, data modeling can be used to construct (computer-based)

24 Methodological Perspectives for Register-Based Research Report 174

describe what is actually observed. The (extended) pyramid framework provides a

Research Report 174 Methodological Perspectives for Register-Based

Research Report 174 Methodological Perspectives for Register-Based

Methodological Perspectives for Register-Based Research Report 174

Research Report 174 Methodological Perspectives for Register-Based