Vous êtes sur la page 1sur 9

The American Statistician

ISSN: 0003-1305 (Print) 1537-2731 (Online) Journal homepage: http://www.tandfonline.com/loi/utas20

The Second Course in Statistics: Design and


Analysis of Experiments?

Natalie J. Blades, G. Bruce Schaalje & William F. Christensen

To cite this article: Natalie J. Blades, G. Bruce Schaalje & William F. Christensen (2015) The
Second Course in Statistics: Design and Analysis of Experiments?, The American Statistician,
69:4, 326-333, DOI: 10.1080/00031305.2015.1086437

To link to this article: http://dx.doi.org/10.1080/00031305.2015.1086437

Accepted author version posted online: 11


Sep 2015.
Published online: 29 Dec 2015.

Submit your article to this journal

Article views: 1471

View related articles

View Crossmark data

Citing articles: 1 View citing articles

Full Terms & Conditions of access and use can be found at


http://www.tandfonline.com/action/journalInformation?journalCode=utas20

Download by: [CAPES] Date: 15 October 2016, At: 20:07


The Second Course in Statistics: Design and Analysis of Experiments?

Natalie J. BLADES, G. Bruce SCHAALJE, and William F. CHRISTENSEN

in 2010. In this same time period, enrollments in first semester


Statistics departments are facing rapid growth in enrollments calculus dropped from 349,300 to 334,000 students; introduc-
and increases in demand for courses. This article discusses the tory statistics enrollments have increased from 34% of calculus
use of design and analysis of experiments (DAE) as a nontermi- enrollments in 1990 to 93% of calculus enrollments in 2010
nal second course in statistics for undergraduate statistics ma- (Albers et al. 1992; Blair, Kirkman, and Maxwell 2013a, b).
jors, minors, and other students seeking exposure to the practice This growth is also observed in the nations high schools: AP
of statistics beyond the introductory course. DAE is a gateway Statistics lags behind only AP Calculus and AP Biology among
to approaching statistical thinking as data-based problem solv- the STEM AP exams offeredand it is growing twice as fast as
ing by exposing students to statistical, computational, data, and the AP Calculus exam (The College Board 2011, 2013).
communication skills in the second course. Given the somewhat This growth in introductory statistics enrollments, together
antiquated view of design and deemphasis of classical design of with a developing understanding of pedagogy, has led to changes
experiments topics in the new ASA curriculum guidelines, DAE in the introductory statistics curriculum. The introductory statis-
may seem an odd choice for the second course; however, it ex- tics curriculum is delivered in many ways; however, the gen-
poses students to the breadth of the statistical problem-solving eral consensus converged upon over the last half century holds
process, explores foundational issues of the discipline, and is that the first course in statistics should roughly cover descrip-
accessible to students who have not yet finished their advanced tive statistics, elementary sampling and design of experiments,
mathematical training. These skills remain essential in the data elementary probability, the sampling distribution of the sam-
science era as students must be equipped to understand the po- ple mean, confidence intervals, hypothesis tests, one- and two-
tential and peril of found data using the principles of design. sample problems for means and proportions, and inference for
While DAE may not be the appropriate second course for all simple linear regression. While this consensus is documented
statistics programs, it provides a strong foundation for causal with little variation in the contents of most standard textbooks
inference and experimental design for students pursuing a B.S. (Freedman, Pisani, and Purves 2007; Agresti and Franklin 2012;
in Statistics in a program housed in a department of statistics. DeVeaux, Velleman, and Bock 2012; Moore and McCabe 2012),
recent efforts have emphasized: use of active learning techniques
KEY WORDS: ANOVA; Design of experiments; Statistical ed- to focus on broadly defined conceptual understanding of statis-
ucation; Undergraduate curriculum tics; statistical literacy and statistical thinking using real data
rather than rote cookbook knowledge of statistical procedures
(ASA 2005); module-based approaches that use authentic prob-
lems to drive and organize the required statistical skills (Lock
and Meng 2010); and exposure to multivariate statistics (De
1. INTRODUCTION
Veaux 2015).
The emergence of big data from every sector of the global The standard introductory statistics topics have served as an
economy demands that future scientists, educators, managers, introduction to an often unexpectedly useful subject for students
and policy makers are equipped to manage, understand, and act in a broad variety of majors and with a wide range of mathe-
upon that data deluge. This increase in demand for data sci- matical competence. These introductory topics have spurred
entists, the new catchphrase for statisticians and data-literate interest in further study of statistics for those with quantita-
analysts, preceded a huge growth in enrollments in statistics tive aptitudeand the media is promising rewards for people
courses and a displacement of calculus as the lingua franca of trained in increasingly sophisticated statistical methods (Lohr
science and commerce (Meng 2009; Utts 2015; Wasserstein 2009; Davidian and Louis 2012). This trend is already manifest
2015). Introductory statistics enrollments in the United States here at Brigham Young University (BYU) where record num-
more than doubled from 118,700 students in 1990 to 312,000 bers of incoming first-year students are declaring statistics as a
major (46 and counting this year) and annual enrollment in our
second course has tripled from 60 to 180 since 2010; nation-
Natalie J. Blades (E-mail: blades@stat.byu.edu) is Associate Professor, G. wide, statistics has been the fastest growing STEM major since
Bruce Schaalje (E-mail: schaalje@stat.byu.edu) is Professor, and William F. 2010 (ASA 2015).
Christensen (E-mail: william@stat.byu.edu) is Professor, Department of Statis-
A consensus has not, however, developed about what to do
tics, Brigham Young University, Provo UT 84602. The authors thank the editor,
the associate editor, and the reviewers for many useful suggestions. The authors next with these students. What should be the content of a sec-
also thank H. Dennis Tolley and Del T. Scott, current and former chairs of their ond course? Even if first-year students have had AP Statis-
department, for facilitating development of this second course. tics and AP Calculus, many programs still require additional

326 2015 American Statistical Association DOI: 10.1080/00031305.2015.1086437 The American Statistician, November 2015, Vol. 69, No. 4
mathematical preparation for courses like mathematical statis- Lovedays (1961) Statistics a Second Course and Mosteller and
tics or regression analysis. If students are excited by statistics, it Tukeys eclectic and inimitable Data Analysis and Regression:
seems a pity to abandon them to a year or two of pure mathemat- A Second Course in Statistics (1977) through Regression: A Sec-
ics before education in statistics resumes. Even if their interest ond Course in Statistics (Wonnacott and Wonnacott 1986) and
in statistics does not wane during this period, students are liable Regression with Graphics: A Second Course in Applied Statis-
to develop the attitude that statistics is just a subset of mathe- tics (Hamilton 1991) to Dielmans Applied Regression Analysis:
matics. The foundational fact that statistics forces a brave but A Second Course in Business and Economic Statistics (2004),
risky confrontation of mathematics with the real world can be Lomaxs Statistical Concepts (2007), and the most recent edition
swamped or lost. of Mendenhall and Sincichs A Second Course in Statistics: Re-
The purpose of this article is to propose design and analy- gression Analysis (2011), the titles and subtitles of the last three
sis of experiments (DAE) as an especially appropriate second decades point to the pervasiveness of this opinion; however,
course for undergraduate statistics majors and to relate our ex- these texts presuppose a potentially outdated view of the math-
periences with DAE as the second course over the last 4 years. ematical preparation of students taking a second coursethat
This proposal may appear to be either radical by its departure these students already have the mathematical maturity encapsu-
from most programs or archaic in its foundational emphasis on lated in integral and differential calculus, multivariate calculus,
design. Undoubtedly some statistics programs use DAE as the and linear algebra.
second course (though none of the top ten programs as defined The last few years have seen the publication of several
by number of statistics majors do) but the topic deserves fresh new textbooks updating the regression curriculum for second
eyes and fresh discussion. courses with progressive active-learning approaches: Kuiper and
Sklars Practicing Statistics: Guided Investigations for the Sec-
2. CURRENT SECOND COURSES IN STATISTICS ond Course (2012) aims to provide a case-based introduction to
statistical modeling and Cannon et al.s STAT2: Building Mod-
While the introductory statistics course prepares students as els for a World of Data (2012) provides a modular approach
critical consumers of data and analyses, it is insufficient to pre- to linear regression, analysis of variance (ANOVA) and exper-
pare students to apply statistical methods in their own work imental design, and logistic regression. By focusing on broad
(ASA 2005). Students may understand basic statistical concepts conceptual understanding these texts introduce simple and mul-
after a first course but they have not mastered the application tiple regression, logistic regression, ANOVA, and experimental
of those ideas. The second course must begin to equip students design without including the mathematical details: both texts
with tools to produce rather than merely consume statistical present the material assuming students have only an exposure
analyses. The expected learning outcomes of the second course to algebra.
are not as standardized as those for the first course in statistics; In addition to these regression approaches there are many
these learning outcomes will depend on whether the second other possibilities that could be considered for a second course:
course is the terminal statistics course of the students under- introduction to probability, statistical computing, data science,
graduate training or the next step in an undergraduate statistics data scraping, or algorithmic predictive modeling. Some pro-
core for a statistics or analytics degree and they must reflect grams are now allowing for flexibility in the entry into the
whether advanced training in mathematics or another discipline major: the undergraduate statistics major at Harvard can choose
is presupposed. to take linear models or data science or finance after an intro-
Second courses in statistics have been developed for many ductory statistics courseor he or she may take probability as
audiences: second courses for students in statistics majors, mi- the first course; at Berkeley a student may choose computing
nors, and concentrations, second courses for math majors, sec- with data, concepts of probability, or concepts of statistics as a
ond courses for economics and psychology and sociology ma- second course. In our own program, while most students take
jors, and second courses for graduate students in a wide variety the DAE course described here as the second course, students
of disciplines. The new ASA curriculum guidelines (2014) ac- could instead take discrete probability (an option popular among
knowledge, There is not a single definition of what is appro- students who would like to prepare for the actuarys probability
priate as a second course in statistics, and a number of different exam) or an introduction to statistical computing (in either R or
options can be found at many institutions. This is a big space SAS) as his or her second course; Figure 1 briefly displays these
that is dominated by mathematical statistics: More than half entry points into the statistics major at BYU. A second course
of nonintroductory statistics courses are taught in math depart- that focused on algorithmic modeling would address Breimans
ments and 72% of the nonintroductory statistics courses taught (2001) concern regarding the heavy emphasis on data models
in math departments are mathematical statistics or probability in our profession; Draper (1987, 1995) also called for more
(Blair, Kirkman, and Maxwell 2013a, b). predictive modeling to identify how the past and future are
A terminal second course in statistics should be focused on connected by comparing predictions to observable reality and
breadth, not depth; it should give students tools for building Cobb (2015) further addressed this conflict between algorithmic
models with quantitative or categorical responses and quantita- modeling and generative data models.
tive or categorical predictors. The relatively small (but growing) When the second course is not terminal, but rather part of
number of textbooks with the title or subtitle Second Course the statistics core for a B.S. in Statistics, the material could be
in Statistics apparently without exception imply that the ap- presented at greater depth. Rather than a survey of advanced
propriate topic for a second course is regression analysis. From methods taught without accompanying mathematical detail, the

The American Statistician, November 2015, Vol. 69, No. 4 327


Figure 1. Entry points into the B.S. in Statistics at BYU. After the introductory course, students may enter the undergraduate core through
a methods sequence, a theory sequence, a computing sequence, or a mathematics sequence. Arrows represent prerequisites; the first classes
in methods, theory, and computing do not require prerequisite calculus. All tracks must be completed but the order can be flexibly adapted to
the interests, preferences, and experience of the student. After completion of the 27-credit core students take roughly 24 upper-division credits
that could include undergraduate courses in Bayesian statistics, survey sampling, experimental design, nonparametric statistics, quality control,
reliability, and survival analysis as well as classes from computer science, mathematics, and areas of application.

course could facilitate a deeper understanding in what will be calculus and linear algebra as part of the course, require concur-
the second of many applied and theoretical statistics courses. rent enrollment, and so forth.
Such an approach would sacrifice breadth in the second course; In 2010, a second course in DAE was developed to provide
however, that breadth will be acquired by the end of the program. a platform for students with only a single introductory statistics
Additionally, the appropriate second course must work within course (possibly AP Statistics) to start understanding fundamen-
the structure of the university. At BYU we are concerned with the tal concepts of statistics. This DAE-based second course covers
curriculum for the second course for statistics majors housed in the scientific method, statistical thinking, sources of variation,
a department of statistics with 19 full-time tenure-track faculty. randomized factorial designs, power, and sample size calcula-
Many of our majors have not finished calculus at the time they tions. Students are encouraged to enroll concurrently in differen-
declare a major in statistics. We require roughly 50 credits for tial calculus; students who have already completed differential
a B.S. Statistics degree (out of the 120 credits required for calculus are encouraged to enroll in integral or multivariate cal-
graduation). This allows for more extensive study beyond the culus, as appropriate. This second-course DAE covers the ma-
introductory course: a required core of 6 credits of methods, 6 terial in sufficient depth that an undergraduate statistics major
credits of theory, 6 credits of computing, 814 credits of math, would have appropriate familiarity with the design of experi-
and 24 upper-division credits. In this context, we suggest DAE ments without requiring additional courses; students who wish
has proven to be an appropriate second course. to explore design of experiments further are well prepared for
an elective course (using, e.g., Lawson 2014).
The expected learning outcomes for second-course DAE
include
3. DESIGN OF EXPERIMENTS AS THE
SECOND COURSE

From the founding of BYUs Department of Statistics (with 1. Defining the experimental unit, response variable, factor(s),
concomitant creation of a B.S. in Statistics) in 1960 until 2010, and level(s) of a basic experiment;
the second course in statistics at BYU was regression analysis. 2. Understanding the role of randomization and replication in
The specific content of the second course drifted as textbooks inferring causation;
and software changed, but not in a major way. The question of 3. Performing a completely randomized design and construct-
how to balance the introduction of interesting statistical topics ing the ANOVA table in SAS and R;
with the need for mathematical rigor was initially not a problem 4. Computing the minimum number of replicates in a com-
because most statistics majors were sophomore- or junior-level pletely randomized design to achieve a given level of power;
transfer students from engineering or mathematics who had al- 5. Computing pairwise tests of differences in means in SAS
ready completed multivariate calculus and linear algebra. As and R to understand a significant overall F-test;
the major began to attract first-year students, this balance prob- 6. Performing a randomized complete block design and con-
lem was approached in several ways (none of which were fully structing the ANOVA table in SAS and R;
satisfactory): enforce prerequisites that delayed students from 7. Performing a factorial design and constructing the ANOVA
taking the second course until multivariate calculus and lin- table in SAS and R;
ear algebra had been completed, cover requisite concepts from 8. Explaining a statistically significant interaction;

328 Statistics and the Undergraduate Curriculum


9. Describing the differences between a split plot and a two- experience conducting and analyzing experiments, these
way ANOVA; and projects help students begin to develop statistical practice skills:
10. Working in teams to write a technical report and make a the group work with other statistics majors generates a nice ca-
technical presentation of a designed experiment. maraderie and the project reports are refined in a subsequent
one-credit writing course that focuses on clear and ethical pre-
Class begins with an intense two-week review of introduc- sentation of results.
tory statistics; the focus is on the two-sample t-test and con- Students are remarkably adept at finding interesting term
fidence intervalbut developing them in greater detail and projects that are sufficiently simple and innocuous to obviate the
using them as a springboard for deeper discussion of power need for ethical review beyond the instructors required preap-
and sample size. In weeks 3 and 4, the students cover prin- proval. Example projects include experiments that: (1) evaluate
ciples of experimentation and experimental design including the ability of subjects to reproduce a list of words after being as-
planning a three-factor experiment; this provides a platform for signed a specific time for studying and a specific font color for
discussing sources of variability and how to choose the response, the study materials; (2) compare respondents opinions about
treatments, and experimental units. Week 5 is spent setting up quotes when they are reported to have originated from either a
ANOVA tables, partitioning sums of squares, and performing political or religious group for which they may or may not have
F tests. Weeks 6 and 7 focus on randomization and balanced affinity; and (3) evaluate the heating potential of a microwave
two-factor designs. Weeks 8 and 9 cover multi-way ANOVA oven based on the presence/absence of a rotating table and the
including interpreting interactions and main effects. Weeks 10 centering of the heated object. Given the complexity of the ex-
and 11 cover blockingincluding randomized blocks, split-plot perimental material (and in an effort to show their statistical
designs, nesting, crossing, and latin squares. Week 12 is spent chops), most students opt to use blocking designs (blocking on
discussing multiple comparisons and contrasts. subject, on microwave oven, and so forth), and many students
While the ability to perform analyses in SAS and R is in- use relatively sophisticated designs such as a split-plot design
cluded as a learning outcome, second-course DAE is not meant with two between-block factors and one within-block factor.
to teach statistical software: software programs are used to offer Second-course DAE differs from the traditional design of ex-
students exposure to SAS and R and to serve as tools for teach- periments (DOE) course in the sense that the course objective for
ing the structure of statistical problems. Students are given SAS second-course DAE is not necessarily to develop students com-
and R code snippets to use in their homework that require very petence in the field of experimental design. There are many top-
minor modification. This indirect, immersive introduction to ics in DOE that are not covered in this second-course DAE: A/B
SAS and R takes students from zero to at-least-familiar in only testing, triangle taste tests, random effects, multilevel models,
two months. At the end of second-course DAE, students can and adaptive designs. Our students gain facility with concepts
analyze a continuous response variable from one-way, blocked, and tools associated with DAE, but developing such expertise
crossed, and even balanced split-plot designs in both SAS and R; is not the motivation for using DAE per se.
this includes inputting and transforming the data and performing
model diagnostics. Statistical computing is explicitly covered in 4. ADVANTAGES AND DISADVANTAGES OF DAE AS
a 6-credit sequence that students may begin while enrolled in THE SECOND COURSE
DAE, and it is expanded and reinforced in subsequent applied
courses. DAE is an appealing second course for both pedagogical
One of the culminating experiences of DAE is a project where and logistical reasons: DAE provides a framework for statistical
the students synthesize the course material and gain experience thinking and exposes students in their second course to a unified
in the practice of statistics by designing, executing, analyzing, approach subsuming the full spectrum of statistical problem
and presenting the results of their own experiments. Students solving. But it is also mathematically accessible to first-year
see the practical issues of turning a research question into a university students.
clearly defined experiment and grapple with amorphous issues DAE emphasizes to students the primacy of both design
related to choosing factor levels, sample size, blocking, and so and analysisteaching students that statistics is not just data
forth. After refining these questions in several meetings with the analysis but rather a framework for problem solving that in-
instructor, execution of the experiment is fairly straightforward. cludes formulating questions, collecting and organizing data,
Students report that they feel much more confident participating analyzing data, and interpreting results (Franklin et al. 2005).
in later research having had to wade through practical problems Studying DAE before regression places more emphasis on fun-
on their own in DAE. We also employ preliminary structured damentals of statistical thinking: the simple ANOVA model
projects where the students use catapults, windmills, helicopters, yij = + i + ij introduces modeling while emphasizing no-
or activities from Stat2Labs (Kuiper 2015) to conduct their tions of components of variability, experimental units, and
own experiments. Sometimes we use one of these structured statistical significance. These concepts are often lost in the
experiments in class to solidify rudimentary concepts of ran- regression-first approach as the students try to explain y with
domization, data collection, and moving data into software. On X. Once students have studied the linear model they often
occasion we have used the catapults multiple times throughout believe y = X +  is the answer to everything statistical;
the second course to demonstrate different designsas the stu- this leaves them less open to thinking about the structure of
dents gain familiarity with the devices they can focus on just the datasubsampling versus sampling and split plots and intelli-
newest layer of complexity. In addition to providing tangible gent blocking. Students who go on to study multilevel modeling

The American Statistician, November 2015, Vol. 69, No. 4 329


or hierarchical modeling build on this DAE foundation to further pragmatism and a heightened sensitivity to the need for ethical
develop a complex appreciation of the data model. communication.
DAE provides a convenient vehicle and an overarching narra- Second-course DAE reviews and builds on what students
tive for understanding many fundamental questions of statistical learned in an introductory statistics course without requiring
thinking: How is the statistical model a tool for understanding calculus. As statistics has grown from a graduate discipline
factors influencing an observed variable? How does a random to include undergraduate majors, many undergraduate courses
error component differ from a factor of direct interest? How can (particularly regression courses) have retained the math prereq-
we compare components of variability? What is an experimental uisites of their graduate counterparts. This is not required: the
unit? Second-course DAE offers students a unique opportunity Ramsey and Schafer (2012) text successfully uses this approach
to internalize these foundational concepts. For example, char- to teach multiple regression and ANOVA without requiring cal-
acterization of variability in air quality is a motivating example culus or linear algebra; graduate service courses often take this
on the first day of the course. The students begin by trying to approach. Gradually undergraduate programs have been ques-
characterize the variability of 12 measurements of local ambient tioning the importance of these prerequisites. The new ASA
particulate matter. But then it is revealed that the 12 measure- guidelines (2014) assert both that [f]ew undergraduate statistics
ments were obtained by taking four measurements on one day, students need the mathematics used to derive classical statistical
four measurements on a second day, and four measurements formulas, but also that undergraduate students should be able
on a third day. So the class attempts to incorporate that in- to explain the interplay between mathematical derivations and
formation in their characterization of variability of particulate statistical applications. Many programs, including ours, still re-
matter. Finally, it is revealed that the first two measurements quire calculus as a prerequisite for regression. But DAE can be
are two different measurements of the same filter using X-ray taught at an appropriate mathematical level for first-year univer-
fluorescenceand the discussion of characterization of vari- sity students without obscuring mathematical detail or treating
ability continues. This forces students to think about the nature output like black-box magic; consequently, students can take
of numbers, observations, and the genesis of variability. This DAE while they are concurrently developing their mathemati-
discussion of decomposition of variability is revisited at the cal skills. Decoupling the mathematical prerequisites indicates
end of the course after studying the split plot design. Students to students that statistical training is different from mathematical
then understand that the definition of the experimental unit de- training: It is not an extension of math. It is not a subdiscipline.
pends on the chunk of material being randomly assigned to Statistics is an applied science that uses math. Concurrent sta-
different levels of a factor. Students do not seem to really un- tistical and mathematical training keeps students engaged with
derstand experimental units until this pointand this is a valu- statistics while they acquire the requisite mathematical tools.
able concept for all studentseven those in nonexperimental Second-course DAE also provides many opportunities to con-
fields. tinue emphasizing ethical elements of statistical practice that
Another major advantage of second course DAE within a are introduced in introductory courses (Verhoeven and Baker
broad program of statistical training (i.e., a curriculum consist- 2008; Hubert and Wainer 2012). Ethics are particularly ger-
ing of many more than two applied statistics courses) is that mane when discussing experiments with human subjects: we
students begin their journey of statistical thinking with excep- discuss whether it is ethical to use a placebo to treat sore throats
tionally high standards for data quality and exceptionally strin- and whether it is ethical to randomize babies (or anyone) to
gent grounds for asserting causality. The gold standard of DAE smoke. We talk about the (sometimes amusing) ethical aspects
proves to be an invaluable tool when contextualizing analyses of previous student project proposals that were submitted but
of data arising from natural (but nonrandomized) experiments not approved: for example, a study of methods for asking some-
or even purely observational studies. While we discuss the use one on a date when the students had no intention of following
of ANOVA for the analysis of data arising from observational through with the dates. We discuss ethical considerations in test-
studies in second-course DAE, together with the limitations of ing a priori hypotheses versus post hoc tests and data snooping.
the conclusions we can make in those settings, second-course We talk about reproducibility and documentation as guidelines
DAE is followed in the applied arm of our curriculum with and for statistical practice: in second-course DAE we explain that
prerequisite to a course focusing on models for observational if your description of your design has insufficient detail for
data, confounding, and causal inference. The experience of stu- the instructor to replicate the experiment, it is not sufficiently
dents in DAE prepares these students to understand the limits of well documented; subsequent courses in our curriculum address
observational data. This subsequent regression course begins by additional aspects of creating reproducible code.
considering a randomized experiment we would like to perform While second-course DAE has many advantages for statis-
but cannot. The students consider why we are unable to perform tics majors, it may not be the most obvious choice for all
some experiment and how a given observational study might studentsparticularly for students in nonexperimental fields.
emulate a randomized experiment. Then they discuss what con- Second-course DAE was designed for students who will take a
ditions are required to identify causal effects in observational two-semester methods core: first, design and analysis of experi-
studies (Hernan and Robins 2015). The students learn that if ments; then, regression. Minors who do not elect to take regres-
observational data cannot reasonably emulate an interesting ex- sion as well as students who discontinue the major after taking
periment, the data may still be useful for prediction a la Breiman DAE will have seen a fairly narrow view of the discipline. While
(2001); however, their foundational training in DAE provides this may not disserve manufacturing and engineering students
grounds for considering results from such studies with nuanced who want a little more statistics, it may be too circumscribed for

330 Statistics and the Undergraduate Curriculum


students in nonexperimental fields: many economics, psychol- Student performance on expected learning outcomes is con-
ogy, and epidemiology students take courses in our department tinually assessed by measuring the proportion of possible points
beyond introductory statistics. For a student planning on only earned by each student on exam questions corresponding to
one other statistics course after introductory statistics, we en- each expected learning outcome. With the exception of block-
courage enrollment in a different service course with a broader ing and factor crossing, all expected learning outcomes had me-
scope (that begins with regression and incorporates point-and- dian achievement above 0.80. For all learning outcomes, student
click software); however, for students who have indicated a de- performance is skewed with large variation among students in
sire to become statisticians and who are enrolled in a program the lower tail. In our previous math-first-then-regression second
offering a large suite of courses for a statistics major, we believe course, with the exception of the time series/logistic regres-
that the foundational underpinnings of second-course DAE best sion outcome, the median for each expected learning outcome
prepare a student for his or her educational career. was also over 0.80. We also track performance in second-course
A minor disadvantage of teaching DAE as a second course DAE as it relates to subsequent courses: overall grade in second-
is the small set of suitable textbooks. We have used Cobbs course DAE is positively correlated with the three expected
Introduction to Design and Analysis of Experiments (2008); learning outcomes in the subsequent regression course, which
however, it is self-admittedly unorthodox in its conversational are most related to topics covered in second-course DAE (ex-
style. While we like the book, many students find it difficult to plaining the difference between ANOVA and regression and
use as a reference for technical details. An alternative would their computational methods, applying appropriate transforma-
be Oehlerts (2010) freely available DAE digital textbook that tions to the response variable to improve agreement with re-
has been developed for students who have completed only an gression assumptions, and creating sets of indicator variables
introductory statistics course (though it is a bit heavy on formu- for categorical explanatory variables).
lae and information and lighter on ideas). Montgomery (2012) While student learning is critically important, as we try to
and Hicks and Turner (1999) could also be used for a second nurture a students interest in statistics through the second course
course; however, both seem a bit inaccessible for a first-year stu- we believe student enjoyment of the learning environment is also
dent armed only with AP Statistics or a sophomore with only an important if we are to develop effective and engaged learners.
introductory statistics course. Due to the sparse textbook options In our previous second course, between 10% and 25% of the
(and lack of computing in the text we selected), the instructors students either failed or withdrew after the add/drop deadline;
of this course have done more course preparation than may be in second-course DAE only 6% to 8% of the students fail or
typical for a 200-level course (a course schedule and course withdraw. Additionally, students in second-course DAE report
notes are available at the website, tofu.byu.edu/stat230/ ). greater satisfaction with the course: average course satisfaction
Another disadvantage is that DAE does not necessarily de- is a full point higher (on an 8-point scale, comparison based on
velop the mathematical skills of students who arrive with well- sections of our old second course with new second-course DAE
developed math skills. We do not capitalize on the beauty of taught by the same instructor); student perception of intellectual
the general linear model; instead we use the pictorial approach skills developed is unchanged.
of Cobb (1998). It is unclear that this is unequivocally a dis-
advantage: students seem to gain a deeper understanding using 5. CONCLUSIONS AND RECOMMENDATIONS
the pictorial approach; in contrast, when students are first in-
troduced to the matrix formulation of the linear model they The McKinsey Global Institute report (Manyika et al. 2011)
approach it more mathematically/algorithmically. Nonetheless, estimates that by 2018 the United States needs 140,000 to
for better or worse, reduction of the mathematical barriers to 190,000 more people with strong analytical skills and 1.5 mil-
the statistics major has been accompanied by a decline in the lion more data-savvy managers and analysts to respond to the
average math competence of our B.S. in Statistics graduates. tide of big data into every sector of the global economy. The
Ultimately, we have found that DAE is a better choice for importance of meeting this demand is underscored by the NSF
the second course in a university where many students arrive announcement in March 2012 to dedicate $200 million to a
without having finished calculus. We are not avoiding topics federal initiative to improve our ability to deal with big data
that need mathbut our first-year students often need to take (Mervis 2012). To address this need we must develop a flexible
math, and DAE as the second course gives students the flexibility approach to providing statistical educationbeyond the intro-
to learn statistical ideas while acquiring math skills. ductory courseto increasingly diverse groups of students.
Given the competing options for a second course (as discussed
in Section 2), DAE may seem like the least progressive or mod-
ernizing approach. With the explosion of data that has never
4.1 Evidence Supporting Second-Course DAE
heard of good experimental design (Efron 2001), focusing on
To support claims of the benefits of second-course DAE, we good experimental design in the second course may seem naive
present evidence that student learning in second-course DAE is or even archaic: as Horton (2015) explained, Students are gen-
acceptable and that students are more satisfied than students who erally taught that if data arise from well conducted randomized
took our previous second applied course (a six-credit regression trials, they can make causal conclusions using a two-sample
course that required calculus as a prerequisite and introduced t-test. All too often, this is considered the pinnacle of statis-
enough matrix algebra and computing to cover the topic at the tics.... But most data that students see are not derived from a
level of Kutner et al. (2004). randomized trial with no dropout, full adherence, and sufficient

The American Statistician, November 2015, Vol. 69, No. 4 331


blinding. Focusing on DAE for the entire second semester of jors the ability to answer authentic questions with real data early
study would seem to extend this limited view. And if the second in their programand this process is repeated as they continue
course were simultaneously the last course, second course DAE through the core and elective courses of the major.
would be regressive. Second-course DAE reflects the guiding principles of the new
In fact, second-course DAE does not modernize the curricu- ASA guidelines that emphasize the scientific method and its re-
lum by virtue of its subject matter: DAE is stodgy; it only lation to the statistical problem-solving cycle; real applications;
obliquely equips students to face-off with new and interesting the need to compute with data in the context of answering a sta-
data types. But a modern curriculum is not just defined by the tistical question; flexibility; and a relationship with mathematics
coverage of modern data types within its coursework; rather, a that acknowledges undergraduate training in statistics has dif-
modern curriculum must also respond to changing student de- ferent demands than Ph.D. preparatory training. Ultimately, the
mographics. The increasing demand for statisticians and data variety of second courses that are taught reflects the demands
scientists has led to a change in the profile of entering statistics of the students at each university and the interests and expertise
majors: these students no longer have the mathematical sophis- of the faculty; however, DAE should be considered a strong
tication of previous generations of statistics majors. Second- contender among the set of possible second courses.
course DAE trumpets the idea that a fundamental understand-
ing of the data-generating model provides a good foundation for [Received December 2014. Revised July 2015.]
students of statisticsand is accessible to students with limited
mathematical training. So DAE is paradoxically a modernizing
influence because it addresses the new profile of statistics ma-
REFERENCES
jors: not transfer students from mathematics but new first-year
students who have been educated in primary and secondary Albers, D. J., Loftsgaarden, D. O., Rung, D. C., and Watkins, A. E. (1992), Sta-
schools that recognize and promote the importance of statistics. tistical Abstract of Undergraduate Programs in the Mathematical Sciences
We could respond to these majors in a traditional way: finish and Computer Science in the United States, Washington, DC: American
Mathematical Society. [326]
your math, then come back. Or we can respond to these majors
in a way that nurtures their interest in statistics while concur- American Statistical Association (ASA) (2005), Guidelines for Assessment
and Instruction in Statistics Education: College Report, available at
rently developing their mathematical skills. So second-course
http://www.amstat.org/education/gaise/. [326,327]
DAE is progressive because it explores how to take a traditional
(2015), A Peek into the Largest, Fastest-Growing Undergraduate Statis-
topic that seems less cutting-edge than data science or analytics
tics Departments, Amstat News, 452, pp. 1822. [326]
and finds a place for ita place that recognizes some of the
American Statistical Association Undergraduate Guidelines Workgroup (2014),
foundational truths within its bounds.
2014 Curriculum Guidelines for Undergraduate Programs in Statistical
We are not advocating design of experiments as the most Science, Alexandria, VA: American Statistical Association, available at
important statistical skillsuch a position encapsulates an an- http://www.amstat.org/education/curriculumguidelines.cfm. [327,330,332]
tiquated view of the discipline. In fact, the new ASA guidelines Blair, R., Kirkman, E. E., and Maxwell, J. W. (2013a), Con-
(2014) do not even include classical design of experiments as ference Board of Mathematical Sciences (CBMS) Survey Reports
requisite for undergraduate training in statistics. The guidelines 2010, Providence, RI: American Mathematical Society, available at
only explicitly mention a few topics from design of studies (ran- http://www.ams.org/profession/data/cbms-survey/cbms2010 [326,327]
dom assignment, random selection, data collection, efficiency, (2013b), Statistical Abstract of Undergraduate Programs in the Mathe-
bias, causality, confounding, and coincidence, page 11). Rather, matical Sciences in the United States: Fall 2010 CBMS Survey, Providence,
RI: American Mathematical Society. [326,327]
we have found that DAE is a stable platform for introducing
general statistical methodology in the context of a simple prob- Breiman, L. (2001), Statistical Modeling: The Two Cultures, Statistical Sci-
ence, 16, 199231. [327,330]
lem that straddles the entire statistical analysis process from
question formulation and data acquisition through analysis and Cannon, A. R., Cobb, G. W., Hartlaub, B. A., Legler, J. M., Lock, R. H., Moore,
T. L., Rossman, A. J., and Witmer, J. A. (2012), Stat 2: Building Models for
communication of results while gently introducing students to
a World of Data, New York: Freeman. [327]
statistical computation along the way. In our program the DAE
Cobb, G. W. (2008), Introduction to Design and Analysis of Experiments, Hobo-
is the first in a series of real data-based experiences culminat-
ken, NJ: Wiley. [331]
ing in elective courses that contain mini-capstones designed to
(2015), Mere Renovation is Too Little Too Late: Its Time to Rebuild the
collectively prepare students for the workforceand it is acces-
Undergraduate Curriculum from the Ground Up, The American Statistician,
sible to majors and minors and students from other disciplines 69, this issue. [327]
who want to take just one more statistics course.
Davidian, M., and Louis, T. A. (2012), Why Statistics? Science, 336, 12. [326]
The ASA guidelines (2014) assert that, No matter how inno-
De Veaux, D. (2015), Intro Stats in the 21st Century, Presented at the
vative the approach, we believe that it is not possible to develop
U.S. Conference on Teaching Statistics, State College, Pennsylvania, 28
a comprehensive understanding of the range of key statistical May. Available at https://www.causeweb.org/uscots/uscots15/presentations/
concepts after only two courses. Statistical training for under- DickOpening.pptx [326]
graduate statistics majors must go beyond this second course; Dielman, T. (2004), Applied Regression Analysis: A Second Course in Business
however, by the end of DAE our students have an encapsulated and Economic Statistics, Belmont, CA: Brooks/Cole. [327]
view of statistical problem solving and the challenges of col- Draper, D. (1987), Comment: On Exchangeability Judgments in Predictive
lecting, analyzing, and interpreting data. Positioning DAE as the Modeling and the Role of Data in Statistical Research, Statistical Science,
second course in our curriculum allows us to give statistics ma- 2, 454461. [327]

332 Statistics and the Undergraduate Curriculum


(1995), Inference and Hierarchical Modeling in the Social Sciences, Lomax, R. G. (2007), Statistical Concepts: A Second Course, New York: Rout-
Journal of Educational and Behavioral Statistics, 20, 115147. [327] ledge. [327]
Efron, B. (2001), Comment on Breimans Statistical Modeling: The Two Cul- Loveday, R. (1961), A Second Course in Statistics, Cambridge, UK: Cambridge
tures, Statistical Science, 16, 218219. [331] University Press. [327]
Franklin, C., Kader, G., Mewborn, D., Moreno, J., Peck, R., Manyika, J., Chui, M., Brown, B., Bughin, J., Dobbs, R., Roxburgh,
Perry, M., and Scheaffer, R. (2005), Guidelines for Assess- C., and Byers, A. H. (2011), Big Data: The Next Frontier for Innova-
ment and Instruction in Statistics Education (GAISE) Report, tion, Competition, and Productivity, McKinsey Global Institute. Available at
Alexandria, VA: American Statistical Association, available at http://www.mckinsey.com/insights/business_technology/big_data_the_next_
http://www.amstat.org/education/gaise/GAISEPreK-12_Full.pdf [329] frontier_for_innovation [331]
Hamilton, L. C. (1991), Regression with Graphics: A Second Course in Applied Mendenhall, W., and Sincich, T. (2011), A Second Course in Statistics: Regres-
Statistics, Belmont, CA: Duxbury. [327] sion Analysis, Upper Saddle River, NJ: Prentice Hall. [327]
Hernan, M. A., and Robins, J. M. (2015), Causal Inference, available Meng, X.-L. (2009), Desired and FearedWhat Do We Do Now
at http://cdn1.sph.harvard.edu/wp-content/uploads/sites/1268/2015/02/ and Over the Next 50 Years, The American Statistician, 63, 202
hernanrobins_v1.10.27.pdf [330] 210. [326]
Hicks, C. R., and Turner, K. V. (1999), Fundamental Concepts in the Design of Mervis, J. (2012), Agencies Rally to Tackle Big Data, Science, 336, 22. [331]
Experiments, New York: Oxford University Press. [331] Montgomery, D. C. (2012), Design and Analysis of Experiments, Hoboken, NJ:
Horton, N. J. (2015), Challenges and Opportunities for Statistics and Statistics Wiley. [331]
Education: Looking Back, Looking Forward, The American Statistician, Mosteller, F., and Tukey, J. W. (1977), Data Analysis and Regression: A Second
69, 138145. [331] Course in Statistics, Reading, MA: Addison Wesley. [327]
Hubert, L., and Wainer, H. (2012), A Statistical Guide for the Ethically Per- Oehlert, G. W. (2010), A First Course in Design and Analysis of Experiments,
plexed, Boca Raton, FL: Chapman and Hall. [330] available at http://users.stat.umn.edu/gary/book/fcdae.pdf [331]
Kuiper, S. (2015), Stat2Labs, available at http://web.grinnell.edu/individuals/ Ramsey, F., and Schafer, D. (2012), The Statistical Sleuth: A Course in Methods
kuipers/stat2labs/ [329] of Data Analysis, Pacific Grove, CA: Duxbury. [330]
Kuiper, S., and Sklar, J. (2012), Practicing Statistics: Guided Investigations for The College Board (2011), AP Examination Volume Changes (20012011),
the Second Course, Boston, MA: Pearson. [327] available at http://media.collegeboard.com/digitalServices/pdf/research/
Kutner, M., Nachtsheim, C., Neter, J., and Li, W. (2004), Applied Linear Sta- AP-Exam-Volume-Change-2011.pdf [326]
tistical Models, Boston, MA: McGraw-Hill/Irwin. [331] (2013), The 10th Annual AP Report to the Nation: Subject Supple-
Lawson, J. S. (2014), Design and Analysis of Experiments With R, Boca Raton, ment, available at http://apreport.collegeboard.org [326]
FL: Chapman Hall. [328] Utts, J. (2015), The Many Facets of Statistics Education: 175 Years of Common
Lock, K., and Meng, X.-L. (2010), Real-Life Module Statistics: A Themes, The American Statistician, 69, 100107. [326]
Happy Harvard Experiment, in Data and Context in Statistics Ed- Verhoeven, P., and Baker, H. M. (2008), An Approach for Incorporating the
ucation: Towards an Evidence-Based Society (Proceedings of the Topic Ethics in Research into an Introductory Statistics Course, Decision
Eighth International Conference on Teaching Statistics (ICOTS8, July, Sciences Journal of Innovative Education, 6, 489493. [330]
2010), Ljubljana, Slovenia), ed. C. Reading, Voorburg, The Nether-
lands: International Statistical Institute. Available at http://iase-web.org/ Wasserstein, R. (2015), Communicating the Power and Impact of Our Pro-
documents/papers/icots8/ICOTS8_4D1_LOCK.pdf [326] fession: A Heads Up for the Next Executive Directors of the ASA, The
American Statistician, 69, 9699. [326]
Lohr, S. (2009 August 5), For Todays Graduate, Just
One Word: Statistics, The New York Times. Available at Wonnacott, T. H., and Wonnacott, R. J. (1986), Regression: A Second Course
http://www.nytimes.com/2009/08/06/technology/06stats.html [326] in Statistics, Malabar, FL: Krieger. [327]

The American Statistician, November 2015, Vol. 69, No. 4 333

Vous aimerez peut-être aussi