Vous êtes sur la page 1sur 5

Experimental Designs by W. G. Cochran; G. M.

Cox
Review by: D. F.
The Incorporated Statistician, Vol. 8, No. 3 (Apr., 1958), pp. 140-143
Published by: Wiley for the Royal Statistical Society
Stable URL: http://www.jstor.org/stable/2986414 .
Accessed: 28/06/2014 10:44

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

Wiley and Royal Statistical Society are collaborating with JSTOR to digitize, preserve and extend access to The
Incorporated Statistician.

http://www.jstor.org

This content downloaded from 91.213.220.109 on Sat, 28 Jun 2014 10:44:57 AM


All use subject to JSTOR Terms and Conditions
leads up to the questionnaire,which includes projective tests where
necessary,and is again piloted before a full-scalequantitativesample is
taken. This last phase contains the most importantdifferencerelativeto
other forms of researchactivityin this field, especiallyabroadwherethe
psychologistperhapspredominatesaridwheretoo oftena subjectiveappeal
is madeto "psychologicaluniversals"regardlessof the (usually)unknown
weightswhichneed to be attachedto themin solvingmarketingproblems.
The second half of the book goes on to say where the method (for it
is a way of thinkingand askingquestionsas well as a way of measuring)
may be applicable to production policy, packaging, marketing and
advertisingmatters. A finalchapterroundsoff with a discussionof points
relevant to the present controversyround the new ideas, and which
servesto stressthe author'sstandpoint.
The book can be read with interestby experts,with profit by students
or laymenand, probablywith a grainof salt (and a bit moreresearch),by
those expert in the marketingof product groups referredto in the case
histories. One point inevitablynot stressed,is that even to-day not nearly
enoughmarketresearchof otherfact-findingtypesis carriedout before,or
at the sametimeas, usingthe relativelyexpensivenewform. Oneor two of
the solutions developedmight perhaps seem equally self-evidentif the
market,etc., had been clearlyoutlined in other ways; or again, without
access to details, it might be felt occasionallythat alternativelines of
actionought not to be ruledout. The firsttableson page 12 and 13 do not
clearlyenough distinguishbetweenthe numberof purchasersin different
classes and the sales they make, while there seems to be some inversion
in the argumenton pages 42 and 43, but these are minor criticisms.
The book is commendablywrittenin a clear,non-technicalstyle, is very
readable,and seemsgood valuefor moneycomparedwith the unsubstan-
tiated piffle which popular expositors are producing,without practical
experience,in a field whichcould so easily becomea charlatan'sparadise,
unless the now establishedethics of the professionare rigorouslyupheld.
The serious reader will go on thinking round the questions asked; in
itself a tributeto, and a justificationof, the author'sefforts.
J. I. M.
"ExperimentalDesigns" by W. G. COCHRAN and G. M. Cox. Wiley,
New York (Chapman & Hall, London). Second Edition, 1957.
Pp. xiv + 617. $10.25 (82s.).
The first edition of ExperimentalDesignsby Cochranand Cox has for
some time now been an authoritative,even if not wholly exhaustive,
source of informationon experimentalstrategiesavailableto the experi-
menter. The book had its roots in the frequentquestioningof the authors
by researchworkerswantingeitherguidanceon the appropriatestatistical
techniqueto be used in a specificexperiment,or an experimentalplan
for carryingout a techniquealready chosen; the authors felt that the
requestsindicateda need for a handbookof experimentaldesign setting
out, and evaluating,those designsthat experiencehad shown to be most
useful. As a handbookof experimentalarrangementsit does not try to
present a unified exposition of the subject, proceedingfrom the basic
140

This content downloaded from 91.213.220.109 on Sat, 28 Jun 2014 10:44:57 AM


All use subject to JSTOR Terms and Conditions
principlesand assumptionsthroughto a generaltheory of experimenting
and then particularizing:throughout,the book assumesa knowledgeof
the principles of analysis of variance and also of the computational
methodsto be used in the analysis. The second edition of this book has
taken advantageof the lapse of six years to include some more recent
workon the explorationof responsesurfaces,includingthe workfollowing
on the paper by Box and Wilson in 1951. It has also been able to draw
on a largerfund of experimentalmaterialand this led to a reassessment
of the usefulnessof fractionallyreplicatedfactorial designs and to the
inclusionof these designsin a new chapter.
The general structureof the book remains as before but additional
materialhas been includedon many topics. This new materialis easily
identifiedby a systematicuse of the suffixesa or A to section, plan and
chapter numbers. Thus, Chapter 6A is a new chapter devoted to the
method of fractionalreplicationin factorialexperiments.
Chapters 1 and 2 act as an introductionto the general principles
underlyingsound experimentation.The first chapterdiscussesthe use of
experimentsto provide tests of hypothesesand to provide intervalesti-
matesof a parameterof interest,from the viewpointof what is practised;
the readeris referredto Kendall's"AdvancedTheory"for the viewpoint
of that which is preached. The role of randomizationas an interference
with any unknown biasses acting in the experimentis introducedbut
overemphasized to the detrimentof its morefundamentalrole as envisaged
by Sir R. A. Fisher, where it is used to validate the usual statistical
procedures.Emphasizingthe bias-destroyinguse of randomizationwould
lead to rejectinga randomsample that occurredin a naturalsequence,
such as steadilyincreasingfactorlevels, as being not sufficientlyrandom.
This kind of random samplingwas in the mind of a candidatefor an
examinationsome time ago who describedtakingrandomsamplesin this
way: "Writenumberson pieces of card and put the cards into a box.
Shufflethese cardsand then select the requirednumberof cards; if these
cards aren't randomreturnthem to the box and take anothersample."
The use of restrictedrandomizationis discussedand so is the effect of
restrictedrandomizationupon the statisticalanalysis.
A section on the initialsteps in the planningof experimentsis included,
dealing with that so often neglected part of experimenting-a definite
purpose. When setting up an experimentthe suggestionis that the first
step in carryingout the programmeis to set out clearlythe objectivesof
the experimentand then follow with a descriptionof the experimental
materialand the treatmentsto be appliedto this material. The reviewer's
experienceof experimenters'behaviouris that the ideas given in this
section are not well enough dispersedamongstexperimentersand much
effort is lost in recoveringinformationfrom experimentsthat change
directionduringthe course of the experiment.
The second chapterexplainsthe methodsthat can be used to increase
the precision of experimenting,whetherby increasedreplication,more
efficient designs, selection of treatment levels in regression analysis,
selectionof homogeneousexperimentalmaterial,use of groupsinternally
homogeneousthough with externalvariabilityand covarianceanalysis.
141

This content downloaded from 91.213.220.109 on Sat, 28 Jun 2014 10:44:57 AM


All use subject to JSTOR Terms and Conditions
Chapter3 derivesfor a simple experimentthe least squaresestimate
of the parametersinvolvedin a specifiedmodelwiththe objectof providing
a guide to the inquisitivereaderwho wishes to know how the computa-
tional proceduresused in later designs are derived. Also dealt with in
this chapterare covarianceanalysis.and the treatmentof missingdata.
The rest of the book is given over to detailingexperimentaldesigns
together with the computational procedures required to analyse the
experimentalresults. Chapter4 covers completelyrandomizeddesigns,
randomizedblock design, the Latin Squareand cross-overdesigns. An
addition to the first edition is the treatmentof residual effects when
treatmentsare appliedsuccessivelyto the experimentalunits.
Chapters5 and 6 describethe factorialmethodof experimentationand
interpretsthe results of two experimentsto show how to extractusable
informationfrom the analysis of variancetable. The principlesof con-
founding are given for 2", 3" and mixed factorials. Plans are given for
confounded designs up to 26, 34, 42 and also some mixed designs.
6A is a new chapterdevotedto fractionallyreplicatedfactorialdesigns
and providesgeneral informationon both the uses and the hazards of
fractionalreplication. The designs,which allow for a certainamount of
confounding between blocks within a fractional replicate, include the
2" series:
24 replicate
25 i, j replicate
26 i, i, j replicate
27 ,i, i, i replicate
28 *, i, i, i replicate.
3n series:
34 i replicate
36 j replicate.
Mixed series:
3 x 26 i replicate
4 x 24 i replicate.
Chapters7 and 8 continue the use of factorialexperimentationwith the
use of split plot designs,quasi-latinsquaresand the plaid designs.
The study of responsesurfacesfollows in Chapter8A. This chapteris
completelynew as far as this book is concerned; the chapterprogresses
from the use of polynomialsto approximatesurfaces,throughthe tech-
niques of multiple linear regressions,to the use of designs to find the
coefficientsof the polynomial. Composite and rotatable designs are
introducedand the analysisof a particulardesignis given in detail. After
this expositionof methodsused to obtain polynomialapproximationsto
the response surface there follows the problem of determiningoptimal
combinationsof factorlevels. Among the methodson whichinformation
is given is the single factor method in which the optimumlevel of each
factor is found in turn, the whole processbeing iterativeuntil the change
in responseis sufficientlysmall. Also discussedis the methodof steepest-
approach proposed by Box and Wilson in 1951, and since greatly
142

This content downloaded from 91.213.220.109 on Sat, 28 Jun 2014 10:44:57 AM


All use subject to JSTOR Terms and Conditions
developed. An interestingmethod is given which relies upon a random
samplingof the responsesurfaceand using the sampleoptimalcombina-
tion as the workingvalues.
Chapters9, 10, 11, 12 and 13 deal with variousbalancedand partially
balancedincompleteblock designswith the chapterheadingsbeing:-
Chapter 9 IncompleteBlock Designs.
10 LatticeDesigns.
11 Balanced and Partially Balanced Incomplete Block
Designs.
12 LatticeSquares.
13 IncompleteLatin Squares.
Chapter 14 deals with the problemof combininga series of experiments
to provide over-all estimates and the difficultiesthat may arise when
handlingsuch data. Chapter15 describesthe use of randompermutations
of 9 items and 16 items.
The great advantage of this book is its extensive use of practical
examples to illustratethe working of particularexperimentalschemes.
Theoreticalexperimentalarrangementsnevercarrythe convictionof one
that has been tried and proved. At the end of each chapter there are
useful bibliographiesproviding both experimentalresults and experi-
mentaldesignsso that the text can be amplifiedfrom the originalsources.
D. F.

"Women's Measurementsand Sizes." A study sponsored by Joint


ClothingCouncil, Ltd. Board of Trade. H.M.S.O., 1957. 30s.
Giventhreemainpracticalmeasurements,how shouldthe manufacturer
of clothing decide on specificationsfor the remainingelementsin the cut
of a garmentto ensurethat the maximumnumberof potentialcustomers
with that "size" will be efficientlyand satisfactorilyfitted when they
purchase "off the peg"? Which three measures really give the best
foundation (sic) on which to build a range of "patterns"or "shapes,"
subjectto the practicalrequirementsof staff and customersin the shop?
Will these "shapes"be merelysimilarfiguresin varyingscalesor will they
be too diverse to serve as a base? Is any measureof standardization
possible(so that ideallya customerknowingher "size"might even order
garmentsby post with a high degreeof confidence)?
Providingsatisfactoryanswersto these questions is obviously a rich
field for the ingenuityof the practicaland theoreticalstatisticiansaided
as they are to-day by electronic machinery. These problems have, of
course,worriedthe progressivesin the clothingtradesfor well over thirty
years. Some companieswere lucky enough to have reasonablyadequate
internal records so that they could see what they had sold and plan
accordingly. But none knew what they could or should have been able
to if they could have developedmore scientifically-based cuts. The field
is essentiallyone for art and craft, and of course the various tastes and
styles likely to be selected by differentwomen, and the varying shop
distributionenjoyed accordinglyby the individualmanufacturer,makes
questions such as the actual frequencyof the various "shapes," in the
143

This content downloaded from 91.213.220.109 on Sat, 28 Jun 2014 10:44:57 AM


All use subject to JSTOR Terms and Conditions

Vous aimerez peut-être aussi