Visit https://ebookfinal.

com to download the full version and

explore more ebooks

Encyclopaedic Companion to Medical Statistics

2nd Edition Brian S. Everitt

_____ Click the link below to download _____

https://ebookfinal.com/download/encyclopaedic-
companion-to-medical-statistics-2nd-edition-brian-s-
everitt/

Explore and download more ebooks at ebookfinal.com

Here are some suggested products you might be interested in.
Click the link to download

Medical Statistics from A to Z A Guide for Clinicians and

Medical Students 2nd Edition B. S. Everitt

https://ebookfinal.com/download/medical-statistics-from-a-to-z-a-
guide-for-clinicians-and-medical-students-2nd-edition-b-s-everitt/

Cluster analysis 5ed. Edition Brian S. Everitt

https://ebookfinal.com/download/cluster-analysis-5ed-edition-brian-s-
everitt/

Applied Multivariate Data Analysis Second Edition Brian S.

Everitt

https://ebookfinal.com/download/applied-multivariate-data-analysis-
second-edition-brian-s-everitt/

A Whistle Stop Tour of Statistics Everitt

https://ebookfinal.com/download/a-whistle-stop-tour-of-statistics-
everitt/
A handbook of statistical analyses using Stata 3rd ed
Edition Brian S. Everitt

https://ebookfinal.com/download/a-handbook-of-statistical-analyses-
using-stata-3rd-ed-edition-brian-s-everitt/

An Introduction to Medical Statistics 4th Edition Martin

Bland

https://ebookfinal.com/download/an-introduction-to-medical-
statistics-4th-edition-martin-bland/

A Companion to Medical Anthropology 1st Edition Merrill

Singer

https://ebookfinal.com/download/a-companion-to-medical-
anthropology-1st-edition-merrill-singer/

The Complete Idiot s Guide to Football 2nd Edition Brian

Tarcy

https://ebookfinal.com/download/the-complete-idiot-s-guide-to-
football-2nd-edition-brian-tarcy/

Handbook of Statistics Epidemiology and Medical Statistics

1st Edition C.R. Rao

https://ebookfinal.com/download/handbook-of-statistics-epidemiology-
and-medical-statistics-1st-edition-c-r-rao/
Encyclopaedic Companion to Medical Statistics 2nd
Edition Brian S. Everitt Digital Instant Download
Author(s): Brian S. Everitt, Christopher Palmer (editors)
ISBN(s): 9782010018145, 1119957400
Edition: 2nd
File Details: PDF, 34.77 MB
Year: 2011
Language: english
Encyclopaedic Companion to Medical Statistics

Second Edition

Edited by

BrlanS.Ev.nt
Professor Emedtus, KIIig's CoIIege,. London, UK
and

ChrIstopher R. Palmer

DItecIot Of the CenIte for App1Ied MetJIt:!II StatIstics,

UnIvetsIty of CsmbIIdgs, UK

With a Foreword by Richard Horton

~WILEY
AJohn Wiley and Sons, Ltd, Publicalian
This editiaa 6rst palllisbcd 2011
l' 2011 . . W'1Iey a. Saas. Ltd

I14lJ1tm1l1jJ1R .
JaIaa Wiley a: Saas IJd., The Alrium,.Soulbma ~ CIUc:hestcr. West Suaex, P.OI9 ISQ. ",lted KiQIIIom
Far demOs 01_ PaW editOrial a8Iccs, ·far CIIIIaaIa' ~I aacI far iDformIIdaD . . bow
copyriPt IIIIIaW iD dIis book please see oar wbsik: at www.wiley.ODID.
to....,. far permissioa to mase tha

I,..
The riPI of abe IIIlbar: to be ideati8ed as the IIIIhar or this ?laik has beat utCdcII ill ~ with 1hc CClpyricM, Dcsipu aad
PIIInIs Act

All riPIs raerwd. No part of Ibis publiadi_ may .., n:pmduI:cd. stmaI ia a Idrieval SJStaII, GI' tnInsIaittaI, in lIlY farm GI' by
10)' . . . . . .~ 1DCICI.aica1, pbolocapyiD,l. m:CIIdiag Dr adawise., czccpt as pa1IIiUai by Ibc UK .CapyrqIII~ Dcsipi
IUd PU:ats·Ad 1988, wiIhaut abc prior ~ of abe pubUsher.·· . .

Wiley also publishes its boab fa a ~ of elOCInIaic fOlllllll. Same C'OIIIeDt dial appears in priDt may ~ .., ·a\llilable iD
ckcIraaa"e books.
DcIipaIians used by compIIIics 10 cIistinpi'lb 1hcir pnIducIs In oftca clabncclll tradcmub.. AlIIaad DIIIIICS IDCI pmdud _
Used in dIis book an: InIIIc: Dillie&, scrvD -ks, Indc:IiiIub ar.staaI bldelDlrb of Ibeir 1apCCtiw: __II.. The pJbIidM:r Is DOl
iDfonnaIiDn ill IqIId to *'
IlSOCialed wida aa)' pmduca or veaclar IDeIItiaDad in dill book. 'dais public:adaa iI.Biped to pnwide accurIIe aad authmi~ .
subject maIIa ~ Il is sold _ die .......ini thai the- publisher il DDt cappel iD tadi.iq
...,reaiaaallClvic:cs. If JIR)fasiaaaI acIvice or ada aped usisIaace is .aiui1ed..1hc ~ of a ~ profcaional ~
be BDqbt.. .

Librory"fl/CIIIIBWU OIlIllo'''''''·l'IIbIicaliDn Dtlltl

The CDC)'CIopaadic campaaiaD to ..... staIiiIiI:s I cdiIaI by Brim S. Emtt
aad CJn1apIaer R. PahDc:r; with • fcRwanl by RicbanlIIadaIL - 2IId ell.
p.;cm.
lacludes IIiIIIiDpapIIic n:f'cmxcl.
~ ..... Elicyclopaalic Campaaian .10 MtcUc:.I Stlllisties, ....i · rcadibIe
~ or almost 400 sbdistical topicl ceatial to cam:III medical n:sean:b.
Each catry _1Icca wriIIat by aa iDdiYidilal chasca far bath 1hcir expc:rIiIcia
1hc field aad tbdr abilitY to.CDIIUIIIIIIicale sbdilitical aIIICCpIS succ:cssrully 10
madical mtai"'lI.. Real a .....cs tiUm 1hc bIomcdicaI ~ .........
ililstndioas feam mID1ft)' entries, aad ~ cnu n:fmnciDI sipposts
Ihc iader to ....,. CIIlrics"'-PRJvidai by puIiIisbu. . .
ISBN 978-0470-61410-1 (!lib)
I. Medical ~Iopedjal I. Ewritt,. __. D. Palmer.
aIrisIopbi:r .Rat,h. .
(DNlM: l. S1aIiItics as ~"'s-EagIisIL 2. ~
'J'bcaIaic~~ WA 13 E562 2010)
RA409.ES2120IO
6Io.120~·
2010018141

A c....... 1tICCIId for this .... 1s IlftiIIbIc tiaai Ibc British LiInIy.
PrinlISBN: 978-0470-61419-1
cPDF ISBN: 9'fI.O..470-ddD74-7
 To Mary-Elizabeth
Brian S. Everitt
To Cailry-Joan, Lalll'a, Carolyn and David
Christopher R. Palmer
 Contents

Foreword ........ I .......... I ......... I ......... I .......... I .......... I .......... I .......... I .......... I .......... I • • ix·
Preface 10 the Second Edition •••••••••••••••••••••••••••••••••••••••••••••••••••••• I •• xi
Preface • . . • • • . • . . • . • . . . . • • . . . . . • • . . . . . • . . . . . . • . . . . . . • . . . . . . • . . . . . . • . . . . . . • . . . . . • xiii
Biographical Infonnalion on the Edilors ••••.••••••.•.••••.•.•••••••••••.••••••.•••••••••• xv
LiSl or Conbibtltors . • .. .. . . • . • • .. . .. • . • • .. . .. • • • .. .. . .. • . • .. .. . .. • . • .. .. . • • • • .. .. . .. • • • .. .. . .. • • • .. .. . .. • • I vii
Abbreviations arad Acronyms ......... I • • • .. .. .. • • • .. .. .. .. • • Xli
• .. .. .. .. • • • .. .. .. • • • • .. .. .. .. • • • .. .. .. .. • • • .. .. .. .. • •

Eocyclopaedle ColDpaaioD to Medical Statlsllcs A-Z . . • • • • • . . • • • • . . • • • • • • . • • • • • • . • • • . . . 1-491

vii
 Foreword

This cac:yclopaediac:ontains nocnlly ror 'Pe«~iew' .In my of infonnation. quality assuranc:c of ~vicwcrs and editors.
small comer of the medical statistical uni\lCnlC, Ibis seems aulhorshipand conlributanhip. conflicts or interest. scientific
Iikcagrasssinoromission.lnstead.lhcpn:x:cssorcw)ualing misconduct. peer ~vicw of grant proposals. economic as-
MSCIII'Cb papers is discussed under 'Critical appraisal'. Arc pects of peer ~icw. and the fUlU~ of scicntific publication.
these two proc:edun:s synonymous? And. im:speclive or In other words. peer ~view is a ~mendously elastic
whether they 1ft or 8M not. should anybody cam? concept. allowing editon to stretch it to mean whatever
I believe Ibat peer review and critical appraisal do dilTer. inle~ts them at a given (whimsicaJ) moment in lime and
that these dille.n:na:s mattcr a,real deal when considering place. Indeed. its claslicity is seen by many of us as its ,~t
the ways in which n:aclcrs should intClpn:l the medical ~th. ThcCODClCpI pows in richncss and undenlanding as
Iitcral~ and lhat an undcnIandin, or these difrc~nces our OWD appreciation of its complexity and nuance soars. 11le
hclps to place medicaJ Slatistics in its proper (lCJftlext when impenClrable nature of pccrnmcw. and the obscu~ and hard-
surveying Ihc wide horizon of c:IinicaJ and public health to-learn cxpertise iI demands. rceds our brilllc c,os. The
n:scardI. notion of critical appraisal. by contrast. is far dUnner in
Thc editors of this quite wonderfully rewarding uatise on meanin,. wilh much less room for cdilorial manipulation
slatisticaltenns havedcftncdcritical appraisal as 'the process and BlP'Bndiscment.
of evaluatinJ; n:search n:poItS and assessing their contribu- EVcn if peer review and crilical appraisal do diller. should
tion to scientific: knowledJ;c·. 'Ibis statement follows natu- anyone actually can:? Yes. they should. and for a very simple
rally from Ihc mean in, of the words 'criticism' (the an of rason: the idea of peer ~icw is now bankrupt. Its ~tention
judgiq)and 'appraisal' (lheestimationofquality). Tbatisto as an operation within the biomedical sciences ~Rects Ihe
say. critical appraisal is an cstimation of wonh followed by inte~sts or those who wish to preserve their own pow« and
some kind of judJ;ment- ajudt;ment that leans IIKR towards position. Peer ~icw is fundamentally anti-democratic. II
an art lhan a science. As a non-statistician, I rather wann to elcwtes the mediocre. It asphyxiates originality and it kills
the precise imprecision of Ibis definition. careers. How so?
Now consider the more commonly embeddc:d term 'peer Peer ~view is DOl about intelli,cnt cnJ;agcment with a
n:vicw' and look how infcrior it is! Who is this anonymous piece of rac~h. It is about defining the margins of what is
idealised peer? Ocncrally. one would consider a peer 10 be an acceptable and unaa::eptable to the ~vicwcr. 'I1Ie mythical
equal. somebody whocamc:s frum a groupcamparable to Ihat 'peer' is bein, asked to view again. after Ihe cditor. the work
from which Ihc penon under scrutiny has cmcrgc:d. 1bis in queslion and to oller a eomment about thc gcogaphical
intellectual cgalitarian is subsequently setlhclask ofviewing location of Ihat work on the map or existing knowledp:. If
apin (to take '~vicw' at its mosl literal meaning) the work ~ is space on Ihis map, and providc:d the work does not
underconsiclcration. But 10 view with what purpose? None is disrupt (too much) the terrain cslablishcd by others. its
specified. location can be secured and marked by sanctioniq publi-
Despite these practical shortcomings. editors of biomccl- cation.lfthe disruption is toog~at. the wed's wish to seck a
ical joumals ~maiD wedded 10 "peer review'. Wc feci place of ~st must be "CIocd. Peer review is about the DleDeY
uncomfonablc with the notion of critical appraisaJ. The of power to praervc csbIblished onhodoxy. It has nolhin, 10
cmbodiment of peer ~view as a distinct scientific discipline do with science.•1 has cverylhinJ; to do with ideology - and
is the series of inlcnlalional cOnJ;~sses devoted to peer the maintenance of a quiel Iifc of privilcp: and mysliquc.
n:vicw in biomedical pubUealion. or,aniscdjointly by JAMA Instead. critical appraisal is aboul incrementally wedin,
and thc BMJ. These conpcsscs have spawned hundrals of one's way towards lrulh'. It can new:r be about lruth itselr.
abstracts. dozens orrescarch papers. and fourthemc issues or The essencc of biomedical ~sean:h is cstimation. Our world
JAMA. They are enti~lycommendable in every WDy. For the ~sists CCltainly. Crilical appraisal is aboutlJ'aDsparcnt. mea-
cdilon of JAMA and the 8MJ, peer revicw encompasses a surablc analysis Ihat cuts a path towards grater precision.
broad nmsc of activities: mechanisms of editorial decision
making. toscthcr with their quality. validilY, and pncticality. I. Honan.. R. 2002: Poi1pubItcaliCIII criticism _ the shapi", of
online peer review and publication. pm-publication posting dintcal ~. JAMA 217.2143-7.

Ix
RDREWORD _______________________________________________________________________

Critical appraisal refuSC5 to \'eil itself in the gaudy adorn- editors of this elK')'dopaedia is therefore a triumph of liberty
ments that editors pin to peer re'liew in order to embellish against the forces of confonnity.
their own imporlance in the carlography of scientific inquiry. Yet still today. too much of medicine takc:s medical
A far more robust instrument critical appraisal is for that statistics for granted. Time and a~ain. we sec research that
refusal. has dearly not bc:cn within a hundred miles of a statistical
What do these differelK"CS tell us about the propcr place of brain. Physicians usually make poor scientists. and physi-
medical statistics in biomedicine: today'! In my \'iew, as a cians and scientists to~ether too onen play the: pari of amatc:ur
lapsed doctor and a now wrinkled editor. medical statistics is statistician - with appalling consequences. The future of a
the most important aspect of our critical appraisal of any piece successful biomedical research enterprise: depends on the
of new resc.an:h. The e\'aluations by .so-called peers in the f1ourishin~ oflhe discipline we call me:dical statistics. It is not
clinical specialties that concern a particular l'CSCW'Ch paper at all clear to me that those who so depend on medical
prm'ide valuable: insight into how that work will be reeeh'cd statistics appreciate either that dependence: or the: fragility
by a community of practitioners or scholars. Howevc:r. as an of its foundation.
editor I am less intere:slc:d in n:ccpCion than I am in meaning:. If this magnificent elK'yclopac:dia can be deployed in the
I want a tough intcno~ation of new work before its ongoing ar~ument about the: future of twenty-first century
publication. according to commonly a~reed standards of academic medicine. then not only the research enterprise: but
questioning - standards that I can sec and e',-aluate for myself. also the public's health and well-being will be far stronger
To return to my personal definition of critical appraisal. I tomorrow than it is today.
want an estimation of quality combined with ajudgment.1 do
not want a view from the dub culture of one particular Richard Horton
academic discipline. The rc:jection of peer re\'iew by the Editor. Lanc'el

2. Horton. R. 2000: Common sense and fig~: the mc:toric of

\'ulldity in medicine. Solisl Med 19, 3149-64.
 t
111 "C
'.Jlil i lIlt, I:i rl'l~ It
'Ii~i l"~tJ f
rIs=tJllrl'IJKJl~,1 ·wI II

tlr~ .,rJllifil 11ft

, It
till Iltlll!l .J 111;n:
r i'r i It ..,, i· ; I ~ I f. II ·1.
I-S' :IJttI5f~rltll.
II I· ,(.'

f'J t I 'II- ,r I. r,ll" ·1:1 if • r

""alrilifil'" 'I;,.lr
a '&.1 1 - f I
i'I: I t I tL·a! II II J
' 0 D
l f,!~ B-
rII IIiiI'Ii i'. 'JI,nli 0
II; •
lt •
'S,. i a
a i '=t ' IJ~ I
i ' .. eI'" l:U· ~lllil 1
ff~f~Jalli(- ~r -
11~llfr!
l t I i . '~t' f e.·1 I;'J !.
"II ",tlll·
~~ ~~frlltl l~ll:l' I'

;r •
II f. •. 1;.-1 B
:~1 r I!Ii! S: I IIs,r ii':1.: II r
lilt ~1!!sf~!I~~
iJJi 'li 1 ;11:~f!§1 •
IS. ~t II ~1~li~=11 r
 Preface

Sialistical science plays .. ilDporlanl role in medical pcnpc:clive it has lDade publicalion in leading journals
JaelRh. Indeed a major pall of lhc key 10 lhc pmgras in more challcnginglhan ever befOM.
medicine tian lhc 17th ceatlll)' 10 the pn:scnl day has been So slalistics is (and an:))RYalcnl in the medical world now
the collcetion and valid inlCqRtalion of evidence.. panic... and is sello ranain so ror the future. Clearly. clinicians and
larly quantitative cvidence. pmvidccl by Ihe appIicalion of medical n:scan:hcn need to know somcthiq about Ihe
stalislical methock to medical inYeSlipiioas. Cum:nl med- subject. cven ifonJy 10 maIcc their discussion with a friendly
ical journals ~ full of statislical lDIIICriai. both rdalively Slalislician more rruitful. The article on consulting a stalilr
simple (for eumple.. t-lcsls. p-wlues.lincar lqIaSion) and. tician quolcs one of the fGRfatbcn or mocIenI slalillics, R.A.
increasingly, mo~ complCJC. (for cxample. genenalised c:sti- FISher who- back in 1938. obsc:ncd wryly: "To conn,lt lhe
maliDl equalians. cluSla" analy. Ba)eSiaD IDClhods). The $Itll&lIcilBr tlJln till e."cpnimerrl isjin&lred u often merely 10
laller material reRects the vilnnl 5IaIc or stalistical research ale him ID conduct II po.,-mor'em e.'ttlmintlliDR. He Cllll
with many new mdhods having praclical implicati_s f~ per/uJps my ...·btlI tire e.'fpe,imerrl dim of.' Thus. one or our
medicine heiDI cIc~oped iD thc IasllWO cIcc&des ar so. Bul hopes for the usefulness and helpfulness of the EncycloPllff-
why is slatistics impallaDl in medicine? Some possible dic COmptllliDIf 10 MediCtlI Sltlli,sl;cs is lhat it may SCrYC 10
answers are: cncaurqe both productive and timely intaaclioas bcawcca
medical racan:hc:rs and statiSlicians. Anolhcrsincc~ hope is
(I) Medical praclice and medical rescan:h gcneralC IaJp that il ftlls a pp between, on the one hand. lextbooks thai
amounts of clara. Such data an: generally full of uncer- delve into possibly 100 much theory and. on the olhcr haad..
tainly and \·arialion. and e&tIacling the "sipal· from lhc shader cliclianarics thai may not necessarily focus on Ihc
"Daisc' is usually nollrivial. needs of medical racan:hc:n.. or else have c:alrics Ihat an:
(2) Medicine involves asking quclli_s lhaI have slraDg IanlalisiDlly succinct. 1b meet these ends. lhc presenl ~f­
slalislical overtones. How common is lhc discasc? Who cn:nce wark conlains concise. informative.. n:lali~ly non-
is especially likely 10 conInIcl a particular condition? technical. and hence, we IrUsl. readable acc:ounlS of over ]So
Whal an: lhc chances thai a patienl diasnoscd with breast topics ccatrallo modem medical statistics.
cancer will survive man: tha fi\IC yean? Topics are coVCRd either brieRy or mo~ extensively. iD
(3) The evaluation of campcliq In:allDCnts or pn:vc:alative general. in accordance with the subject matter's pcn:eivcd
IDCIISIRS relics heavily on slatillical cOllCc:pl' in both lhc ilDportance. although we acknowlcdlc then: will be dis-
clcsip and analysis phase. agn:cmcnt. incvilabJy. about our ChaiClC of DIIicle lengths.
Many entries benefit from containing real-life, clinical
Recognilion orlhc: importance ofstalislics in medicine has exalDples. Each has been wriltcn by an individual chosen
increased considerably in n:ecnl years. The lasl decade. in not only for subjecl-maller expertise in the ficld bUI.just as
padicular. has seen the emergcnce of cvidence-based mccI- ilDportantly. also by ability 10 communicale statistical
icinc.., ..d with it the need for clinicians to keep one step concepls to olhcn.
ahead or lheir paticnts. lDay or whom nowadays ha\IC 11ac extensive cl'OSlrrefcn:nc:ing supplied usiq SMAU.
access 10 virtually unlimiled inronnation (lDuch ofil being CAPII'ALS 10 indiclllC tcnns thai appear as separate entries
virtual, yct SOnIC of it being limiled in its reliabilily). should help the radcr to find his or her way amuncI and also
Comparcel with ~vious gencrations of medical students,. serves to point out associated topics thai might be of intcral
today's prc-cliniclil undergraduales arc heiDI taughl elsewhere within lhc EncycloptJedic Comptllf;tIII. All but the
more about statistical principlcs than lhcir predecessors. shadcst caines contain rerCI'CIICCS to fUJthcr resources when:
Furthermore. today's clinical researchers 1ft faced (hap- the interested n:aclcr can learn in paler cIcpIh about Ihc
pily. in our vicw) with growing numbcn or biomedical palticular topic.
journals utiBsing slatislical rcfCl'CCS as part of their peer Thus. while hoping this work is found to be mostly
Mview proccsses (sec amCAL APPRAISAL and STATImCAL comprehcasiblc we do nol claim illo be rullyCGIDprchcnsi~.
REFEllEEDKJ). This enhances the qualily oflhe papersjoumal As c:o-editan we lake joint n:spansibilily far ..y enurs ('sins
editors select. although from the clinical n:scardJcr's or commission') and would positively weIc:omc sugcslions
xiii
PREFACE _____________________________________________________________________

ror possible new topics to consider for futw-c: inclusion to and encouragemenl throughout the preparation of this book.
R:ctiry perceivc:d missing entries ("sins or omission'). F"mally. our family memben deserve especial thanks ror
Our thanks ~duc to numerous people - first. to all onhe ha\ing been exba toleranl of our lime spenl on developing
many conlributors for providing such excellenl material. and executing this extensive projecl from beginning 10 encLll
mostly on time (mosaly!) with panicular gratitude: extended is our hope thai Ihe Encyclopaedic- Companion pnn'cs all
to those who contributed multiple &nicles or who handled these efforts and sucriHces to be well worthwhile. becoming a
requests for additional &nides so gracefully. NeXI. we up- userul. rqularly-thumbc:d reference added to the bookshelf
prcciatc:d Ihc lrcmendous and indispensable effom of staff at or many or those involvc:d in contemplating. conducting or
Arnold. cspccially Liz Ooostc:r and Liz Wilson, and not least contributing 10 mc:dical rcscan:h.
ror their remaining calm during an editor's moments or
anxiety and neurosis about the enlire projecL In addition we Brlaa S. Everitt aDd Clartstopber R. P......r
would like to thank Harriet Meteyard for her eonslanl support /tlnlllllY 2005
 Biographical Information on
the Editors

BrI_ S. Everitt - Professor Bmerllal, KIDR·s eo... Cbrlstapher R. Palmer, roundins Din:ctor of Cambridge
Londo.. After 3S years al the: InSlitulC of Psychiatry. Uni- University's Centn: for Applied Medical Statistics. rqularly
versity ofLondoD. Brian Ewritt n:tired in May 2004. Author teaches and collabanlc:s with cum:al and fuaun: cIoctan. His
of approximately 100 journal articles and O\'Cl'SO books on first clcln:e was from Oxford. while graduate and post-
statistics. and also co-c:ditor of SllIlisllml Melhods in Met/· doctoral studies \VCR in the USA (at UNC-Cbapel Hill and
im/ Reselll'cir. Writing continues apace in n:lin:ment but now Harvani). He has shined frvm mathcmaticaltowards applied
puncluatc:cl by lenn~ walks ia the counlry.luitar playing and Slatistics.. with particular inten:st in lheethics ofclinical trials
visits to the Dna. rather lban by committees, committees and and Ihe use or ftexible designs whenever appropriale. Fun-
mon: committees. damentally. he likes to promote soUDd statistical thinking in
all areas of medical raean:h and hopes this volume might
help towards dlat end. Chris SClrved as Deputy or Acting
Editor far Slatistics In Medicine, 1996-2000. and is a long-
standinlslatistical n:viewer for 71re Lancel. He and his wife
have three childn:n they ClClllsider to be IIICR than statistically
sipificant.

xv
 List 01 Contributors

K. R. Ab...... (DA). Cenln: for Biaslalislics aad Ocnelic. I.ac)r M. CarpeD.... (LMel, DeparlmeDt of Public Heallh.
BpiclemioJOI)', J)cpaIbDent.ofHeallh Scic:nccs. UniwDit)' of Univenil)' of oxronl,. and Nullield CoIIcp,. oxranl ox I
Leicesla',....,icestc:r.
LBI 7RH. UK INR UK .

Calla ....... (CB), Cli.aI nial Servic:e JJnil and iuaa CIdaD (SC), Respiratory Epidc:miok., aad Public
Health. Imperial College. Emmaaucl Ka~. Building.
EpiclcmioJogicai Studies Unit (CTSU). Richanl Doll
Manraa Road. LaacIon SW3 6LR,. UK
Buildilll. Olel Raad Campas. Raosevelt Drive.
oxranl OX] 7LP,. UK '11m CelIe (TJC), MRC C'enIrc of EpidemioiOD rar aaJcI
Health. UCL Instiblte or Child Health. 30 Ouilrord SIn:et.,.
Alua ....... (AB), QuantilatiWl Scienc:cs. OIuoSmith- LandaD WC1N lEH,. UK
Kline. Medicines Rcscarch Ce~ Gumaels Woad Raacl.
Stevcaace. Hertrordshill' SOl lNY. UK Chris CGftOI8D (CCo), J)cpaIbDent of Madlematies and
Los-. UT 84322-3900.
Statistics. Ulall Slate Univenity,
TlJI De Ble (TOB), ISIS Rc:sean:. 0nIup. Buildiq 1. USA
UniWlnity of Sauduunplon. SouIb....pIaI S017 IBJ. UK or
NeIIo CrIIIIaaIaI (NC), UC Davis Departmc:Dt Slalisties..
KII.... Bjork CD). Primelric:s. Inc:. AmIcIa, Colanclo, 360 ICar Hall. One Shields A~ Davis.. CA 9S616. USA
USA (btheOprimebics.net) Sanb CnaIer (SRC), MRC urec:ounc EpidelllioiOlY
Unit. University or Southampton,. Southampton
J. MuIID ....... (JMB), Pmfessor or Health SlDIistics. General Hospital. Southampton $016 6YD, UK
DepadJnent of Health Sciences. UniWlmity of yadt,
HesliDgtaa. yadt YOlO SDD,. UK CaraIe O.mml" (CLC). 'I'be University of BinniaPam.
Departmeat ofPublic Heal.... BpidemiolOl)' ancIBiotali*s,
Ma.......... (MtB), DivisiOll of Bioltatics. AnIoId go VlIIL'ICnt Drive. Ecl&ballOll. Binni~ B 15 lTH
School of Public Heal.... Univenity of South Carolina.
800 Sumter Sbeet. Columbia. SC 29208. USA. and also Geo.... DllYII,-8mUb (ODS), Schaol or Social
Unit of BiosIaIisIics. InstitulC of BnvinHlll1ClllaI and Communit)' Medicine.. Uniwnit)' of' Bristol,.
Medicine. Kmulinslca Institutet Nobels ric 13. Stackholm. Oakfield Hause. Oakfield Onnre. Bristol BSS 2BN, UK
Sweden
Sbaaa Da, (SO), Raclae Products LilnilCd.. Welwyn 0anIen
MkIIIUe Bndley (MMB), Health Ial'onnatiaa and Quality
Cit)'. HertfonIshiM. AL7 ITW. UK.
AUlharit)', Oc:orp's Court. Ocoqe's lane. Dublin 7.1n:1anc1
........... (SB), Departmenl or Sacial Medicine. DaaIeIa·De A"'IIII (DBA), Statillics.. Modellin& and
UniWlnity or Bristol. C_ynge HaD. Whiteladies Road. Economics Departmc:llt. Healdl Pmta:tion Apncy, Cenln:
Clifton, Brislol Bsa 2PR. UK far Infections. Landon and MRC BiCl5lalistics Unit. Jnstilute
of Public: Health. Uniwnity Pantie SilC. Robinson Way.
Marc aa,.. (Ma), Intenational Dna, DeWllapmenl
Cambridp: ca2 OSR, UK
"stitute (IDDI). 30 avenUc provineialc..
1340 Louvaia-la-Neuve. BelPum J......... Dee.. (JD), Public Health. EpiclcmioJogy and
M..J. CampIIeII (MJc)' School of Heahh aiad Related BiastatiSlics. Univenit)' or Binningham. EcIgbasIon.
ReseaKh. University of Sheflielcl. Resent Court. 30 Relent Birmingham BI5 21T. UK
Sbeet. SlIellleld SI 4DA,. UK
0 ....... Daaa (OD), Health Scienc:cs Resc:an:b Gnlup.
, ..... R. C......ater (JRC). Mcdic:aI SlaIistic:s Unit. Schaal or Canmaunily Based Medidae, UniWlnily of
London School or Hygiene and 'l"rapkal Medicine:. Keppel Manchester,. lean McFarlane Buildin& Oxford Road,.
Sbeet,. Lonclan welE 71fT. UK MaacbesIer MI3 9PL. UK
xvii
USfOFCONTRlmnoRS _________________________________________________________

Daua E8Itoa (DE), Department of Publi~ Health and THY Johnso. (TJ), MRC Biostatistics Unia.lnslitute
Primary OR, University of Cambridge. SlI1Ingeways of Public Health. University Fonie Site. Robinson Way,
Resean:h I..abanlory. Worts Causeway. Cambridp Cambridge CB2 OsR. UK a MRC Clinical Trials Unit. 222
CBIIRN. UK Euston Road. London. NW 1 2DA

Jo....... Bmbenoa (JE), Cinical Trial Servitle Unit Karea KaIMIar (KKa), Department of Mathemalics.
and Epidemiological Studies Unit (CJSU). Ricbard Doll University of Colorado at Denver. PO Box 173364. Campus
Building, Old Road Campus. Roosevelt Drive. Oxfard OXl Box 170. Denver. CO 80217-3364, USA
7~UK
lC.,.........•• lC.bn (KlC.), Department of Biostatistics and
IUcharcI EauIey (HE), Health Sciences Resean:h Oraup. Medical Informatics, University of WISConsin Medical
School of Community Based Medicine.. University of School. 60D Highhiad Ave., Madison. WI 53792-4675, USA
MancheslCr. JeaD McFarlane Building. Oxford Raad.
Manchester M 13 9Pt.. UK bib Kina (RlC.)t School of Mlllhematics and Statistics.
Mathematit:al Institute. UnivcrsityofSt Andrews. Fife KYI6
Bri. . S. Everlll (851£), Biastalistics Dcpadment. Institute 9SS. UK "
of Psycbially. Denmadt Hill. London SES lAP. UK
WoJlek IC.JozaDDwsti (WK), School of Engineering,
David FaraaII (OF), Depadment of Statistics. Univenity Mathematics and Physical Science. University of Exeter.
of Haif... Haifa 31905. Jsrael Harrison Building. North Park Road. Exeter EX4 4QF. UK

W. Harper GUDIOur (WHG), Section far Public RaDjlt LaD (RL), Warwick Emergency Care and
Health and Health Policy. Division of CommuniI)' RebabililaliCID. Division of Health in the Community.
Based Sciences. University of GIIISIOW. GlasJow Warwick Medical School. Univenity of Warwick.. The
G128RZ, UK PannhoWie. Gibbet Hill Campus, Coventry CV4 7AL. UK

81. Gaet.......ur (EO), Department of Applied Sabine Landau (SL), BioslDlislics Department.
Mathematics and Statistics. OIIent Universaly. Krijgslaan Institute of Psychiatry.. King's College. Denmark Hill.
281-89,9000 Ohcna. Belgium London SBS lAP. UK

AlKlnw Orl..e (AO), Division of Health a Social Andrew 8. La. . . (AL), Division of Bioslalistics and
Can: Research. Depanment of Primary ~ and Public Epidemiology, College of Mectacine. Medical University
Health Scieaces. School of Mc:dicine.. King's College. of South Carolina, Charleston. SC 29415. USA
Floor 7. Capital House. 42 Weston sa.Loadon
SEI 3QD. UK Marwn Leese (ML), Health sCrvitle and Population
Resean:h Department. Institute of Psychiaby, King's
Julilua P. T. ......_ (JPTH), MRC Biostatistics Unit. College. Denmark Hill. Loncloa SBS lAP. UK
Institute of PUblic Health, University Forvie Sileo Robinson
Way. Cambridge CB2 OSR. UK AmI1 Lyada(AGL), DeparlmentofOncology. University of
Cambridge. Li Ka Shing Cenam., Robinson Way. Cambridge..
11Ieodore .. HoIfanl (TRH), Division ofBiosiatistics. Yale CB20RE. UK
School of Public Health~ Yale.. New Haven. CT 06520. USA
Cyrns M..... (CM), President. Cytcl Software
SaB)'HaUls (SH), AstraZc:aeca, Parklands. Aldedey Part. Corporation. 675 Massachusetts Avenue. Cambridge.
Macclesftclcl. ~eshire SKIO 4TF. UK MA 02139, USA

Tonte. Hotbana (TH). Institut fUr Slalistik. Ludwig- RI..... Morris (RM), Department of Primary Can: and
Muimilians-Univenitit MUnchen. LudwiplraS&e 33. Population Health. UCL Medical SdJooI, Royal Free
DE-80S39 Miinchen. G:nnany Campus, London NW3 2PF. UK

Hazel 1_ldp (HI), MRC Lifecounc Epidemiology Unit. Paul M......... (PM), Department" of Statistics, The
Uni~ty of Southampton, SouthamplOD Cieneral Hospital. University of Auckland. Private B8I92019. Auckland.
~~nSOI66YD.UK NewZcaland
_______________________________________________________ USTOF~~

CIIrIstopIIer R. ......r (CRP), J)epadlDent of Aaden SknIadIII (AS), DiYisiem of EpidcmioloJy.

Public Health and Primuy Care, Institute: of Public Health. Norwepua InsIituteorPublic Health. PO Box 4404 Nydalen.
Uni~nity Fantie Site. Rabiason Way. C8mbridp CB2OSR. N~. Oslo. Nanvay
UK
NIaeI Smeetaa (NCB). Kina's CoIlqe London.
Max ......... (MP), MaC CUnicai Trials Unit. 222 Busloft Departmc:Dt of Prillllll)' Care and Public Health
Road. London NW 1 mA. UK Scieaces.. DiYisiem of Heal... and Social Cam Rc&eaIdI.
7th Floor Capital Haase. 42 Weston SlRIe~ Landon sal
Nlla ..... tNP), Cytel Softwan: Carpanlioa. 615 3QD.UK
Massachusetts A\lellue. C'ambridp. MA 02139-3309, USA
NIaeIStdanl (NS), Wanvick Medical Schaol, UniYersilyof
J_a Pow-UP), Dcpanmc:at of Public Heallb and Primary Warwick. COW:nII')' CV4 7AL. UK
ellie,Iaslitalc of PubUc Health.. Uni~ly Fcnie Site.
Robinson Way. C8IIIbriqc Ca2 OSR. UK
Joaat. . . . . . . (JS), School of Social and Comaumil)'
Medicine. Uni~nily or BristoL Canynge HaIL
39 Whatley Road. Bristol BSS lPS, UK
P. PracaIt (pP), Faculty of Malhematical Studies,
UniWl5ity or SauahamplOD. SouIhaaaplOft SOl7 IBI. UK
.......... SWaIIDa (88), Swedish Business Schaoll
Slalislics. oR::bro Univenily. 0n:In. Swedc:n
SopIda ............. (SRH). Graduate School of
Educatioa and Gndu_ 0nJup in Bi~. M ...... SJdes(MS), MRCClinicai Trials Unit. 222 EUSloII
UniWl5ity or California. Berb1ey. 3659 Tolman H•• Road. I..aDcIoa NWI 2DA. UK
Califomia 94720. USA and InstilUte of Educalion.
UniWl5ily or London J......, Taylor (NOT), Depanmc:at or BiaslalilliCs.
Univenity of Michipn. 1420 WIIIhiDIton Heilhts.
8m ...... (BR), Department or SIalistics.. Uni~ or Ann Arbor. MI4BI09-2D29, USA
Haif.. Haifa 31905. Israel
Kale TIIIIaa (KT), School or Social and Comm_ily
ShaaII . . . . . (sas), MRC BiasIalislic:s Unit. Institute Medicine. Uni~ty of Bristol, OmYlile Han, 39 Whalley
of Pablic Health, UnivCnily Fonic Site. Robinaaa Way. ~ Bristol BSS 2PS. UK
Cambridge CB2" OSR, UK
81t1ua TGID (BT), MRC Biaslalistics Unil.lnstilUle of Public
M ........ (MRS), Division of aiaslalistics.. Uniwnity Health. University Fanie Site, Robinson Way. CambridJe
of Califamia, 185 8c:n)' SIreel. Suite 5100. San FraDcisco. CB2OSR, UK .
CA 94107, USA
ReIM!ca 1'araer (RT), MRC Biostatistics Unit. Institute
PraI..,. .........art (PIe), CyteI Software Carpanlion.
or PublicHealth, Uniwnily Fame Site. Robinson Way.
Cambrid&e CB2 OSR, UK
675 MassachuseUs Avcnue.. Cambriclce. MA 02139-3309.
USA Aady V. (AV), Bio.dptista 0nJup. Uniwmdly or
Maachcster, Oxfanl Road. Mancheller M13 9PL. UK
............. (SS),DepartmentofStlllistics. The Uni~ity
of Olascow. Olasgow 012 8QQ. UK SII,Ia V...........t(SV),Gheal Uni~lSity, DcpLof AppIic:cI
Mathematics and Computer Science. Krijplaan 281, S9,
~ SIuIID (PS), Deputmcllt of Psychialr)'. 11Ie Uni~ily 8-9000 Ghent, BelPum
of Hoa& Kan" Queen Mill)' Haspital, 102 Pakfulam ReI.
HOIIIICaIl& Sanb L Vowier (SL\I), BioinfCll1lllltics CoR, Cancer
Racan:h UK. Cambridge Rescan:h lnIIitute, Robinson Way.
c .... Sbarp (CS), Computiq Deputment. Institute Cambridp CB2 ORE., UK
of Psychially. Daunut Hil~ LoncIaa se 8AP. UK
Stepbla J. W...... (SJw)' Medical SIaIistics Graup.
Anid SjalaDder (AlS,. Depadment of Medical School or Health and Related Raarch, University of
Epiclemiolo" and Biasllllillics.. Karvlinsb InstilUlet. Sheffteld. Regent Court. 30 Repnl Stn:et. Shemeld
Nabels Vic 12A. 171 17 Stockholm. Sweden SI4DA.UK
xix
USTOFCONTRIBUTORS _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ __

J.G. Wheeler (JG\V), Quanticate Ud. Bevan House. 9-11 J .... \VII. . (JW), 1710 Rhode Island Ave. NW. Suite 200.
Bancroft Court. Hitchin. Herts.sas
lUi. UK Washington DC 20036. USA

Bnmdon WbItcbI!r (aW). Clinicallmaginl Centre. Mark Woad....... (MW" The George iMlibiac for
GluoSmilhKliae, Hammersmith Hospital. Inlemalional Health. PO Box M20I. Misscaclen Road.
Du Cane Road. Lonclaa WI2 OHS. UK Sydney NSW 2OSO, Australia

laD WhIte (IW), MRC Biostatistics Unil.lnslitulc or Public Ra-F_ Yeb (BY), University or California
Health.
Univenit)' Forvie SiRe. RObinson Way. Cambridge San Fl'IIIICisco. Campus Box Number 0560. Soo PanlB!isus"
CB2OSR. UK 420 MU-W. San Francisco, CA 94143-0560. USA
 Abbreviations and Acro.nyms

ACES Actiw conbvI cquivalcnc:c &Iud, DZ Diz)'lGlic

ACET Ac:Iiw COIIIIVi cqui~ I&:It 88M EvicIencc-bued aalicinc
AD Adapliw des.ip BOA BxplanlOl)' data analysis
AI Altildal inlellipnc:e EM Expeclalion-aaaimilalian
AlC Alaike's information Crilaion £MEA Ell...... Medicines EwllIIIIioa Apacy
ANCOVA AnaI,sil or COWIiaDt!e FDA Food and DIu& Adminillnlliaft
ANOVA Analysil or variance GAM OcneraIised additive madel
AR AuloJqn:sSive Gi!B GencnlilCd csliaIaliDl eqlllltioDs
ARMA Aulolqlaslve lIIDVinI'avcrap GFR CJeaeraI t'CItility nile
AUC Ala " ' r c:urYe GIS ~ infOl1ll&llion sysICID
BlC Ba)'Clian infonnaliaft crilCrian GUM Gcnaalisallinear. illlenK:tiw .modcJIilll
BUGS BaycSiaD inraa.ce Usia& Gibbs SIImpIinc (SDftWIR)
(softw8re) GUMM CleIIcraIisc:d Unear mW:cI .....
CACe. CompUer awnp causal ell'ect GLM OcneraIised linear model
CARr Classiftcal~n and RpasiDn ftc GLMM Gcnaalisallinear mIUcI ......
.CAT CoaapuICr..-laplive tcsIiJII ORR Gmss ..".aduclion mte
CBA Cosl-bc:neftl ",,.1 GWAS C1enorno-wicla associlllion ·slUdics
CEA Cosl-ctrec:li--. analysis HALE HcalIh-a4iUSlcd lire cxpecllUlcy
CI Caiaftdencc inIavaI HMM Hidclen Markov model
CONSORT CGasolidalioa of s&ancIanIs or ftpIItinJ HPDl Hilliest (IOIIaior·dcnsity ~
IriaIs HREC H..... IaClRh elbics conunitlec
COREe ~:OI1ice roi.RaeaIdI HRQo~ Hcallbcllllcd qualily or lire
Ethics ComniiIIecs IBD. fclenlily-by-clesc:cal
CPMP CaaunillCC fal' PmpricIary ICC IDlradus (or inIIaclustcr) com:laIioIa
Medicinal Praducll cadIlcicat
CPO Coiaditiaaal )RdicIiYC- onIi...e ICER ~1I1CIIIaI CDII-clra:liveaess ratio
Cd C1dble inlcmII ICH InlcmalioDal ConrCRnCC GD
CRM Coaliaua11aS1C1S111e11t mcIhod Hannaaizali_
CSM c.uaillac GD Safety or Medicines IQ laslilulianal .mew board
.CUE Cosl-ulility analysis rrr InlcnlioD-~
CV C'ocOicient or wrilllian IV lllslnanallal variabIc
CWT Conli.ous wavelel .......... KDD ICnowlqe cIisccway ill clatabuel
DAG Din:c:lCd acyclic paph KM ICaplan-Meil;r
DALY Disability alij8led life-)'Car IcNN k-ncan:1t aeilhbaur
DAR Dlapaal at nnIom LDF u.. discrinainanI ftIIIcliDn
DeAR DIapauI mmplcrely aI mndaIn LR ~11IIio
DDD
DE
DaIa-clepcadcnt
Dclipefred
"p LREC
LS·
Local n:aean:b c:lhici c:ommillee
Lcasl·.._
OaF Dc:pa:s of rn:cclam LST l.aIp simple IriaI
DIC DeviaDcc informalion crilcrion MA MoviI1l awrap
DM Dalaminill& MANOVA MukiWriate ...., . or variance
DMC Dada.monitorial cammilfeo MAR Missialld mndom
DIMC Data and saf'c:Iy monilariD& cammiuce MeA Mc:cIiciael ConIIvI At,crrl:y
I)WT DiI!Clde wavclcllnlnlf'orm MCAR Misslnl. mmpIctcly aI. raacIam

xxi
~~AND~ ______________________________________________

M.m. cliaiD MoRa.: Carlo ~Y· ~ty. ad)i~ ure:-~.

M~·_·~.~ PnJducis· QoL Quality or lire
.RepIalDay Apacy ~ Quiaai~1IIlIiIe
. M~m_ JikemioacI CIIIinuiIe _ .~
Quanlitali~ .lloci
(- C8IimalicIa) RCT· bndomIsed conIIOIIaIlriai
MREC
MsB
MTQ
MZ
Mulliccalre ·lUCarch ethiCs caaunillee
Mc.a.1IqUIR e~
MliXim
.. . .... Iolerali:d
Manozyplic:
. cloSe
••
REB·

RBML,
·ROC·
Rc~.I.~ .
llcscaIda.eIhics cammilleC
b ..... max_UIII. UbiihOOd
~eaciwropifttiq ~
NI ~ (arnoninfanaalhe) ROt ~_ofintaat
~. ·:Ncl mcmeIaIy ~ftt. RPW ~isacl play~. . .
NNH
NNT
.NulDJM:r .aated·IO.hann
Niam....... lOlMat
RR
.SD
RcIaIiWJ .Ii*
SIancIanI deviation
NPV Ncpii~ pmlicii~ value !S. SI8nCIanl~
NUs
Mdt
.OLS
Naiiaaal
. ' . '

·Net . . . .1CtiDn . .
.0Idin.y: inst sipuua
. . .SerVice
Racan:h ElJiic:s . SSM
SMR
_ur81
slIuIdIud ciJcr .Of'the mCan;:

slanclaidi.d
~ DDIcJ
l'IIOIIIiIily. raIio .
·S8.t
QR
PeA
PDP
·Odds,.aio
anal,.
-PriDciplil companc:al.
......,.litY "'Y runciIa.
.SPRT
.SS
S.aultical parameIric nap
Seq..,."
.
SUmofsquans
psababilily
.ratio . tcit
.
PEsr ...... ad B.arian of ,S8 Sum of.squaR:S ~ 10 error
~~(IOft~D) SVM s~ -reCtor madu_
fGM ,..,.
. ~. nicuuni ·TDT
.1M TransmiisioD dlsiciltiDn lest·
ftl PnIpOrtiDaaI . . . . 1bbiJ. fertility rata . -
PKIPD PIiarwnac:oJci_:-.....lo_ _- - - ' - _ : -
• • • 1~ . . . ~H~t7. . . . . . . . . . Tail ·~rncihod
POP -~miit pn,bability n· TiiquIar rat
pp
p'-p
.,. .,...,col·
~.JlCl'CCntile
. . VAS
wLSi!
Y""" aDaJOpe...Je .
W~ .... squan:a estimate
ppp. ~n-·;miIl
, .r-" ~hiJit
l"" ........ Y c. (or lIati~)
PPV Positive paalic:live ·alae
 A

accelerated factor See SURVIVAL ANALYSIS maceuticals: that they are of sufficienl quality. arc safe and
emcacious. Efficacy isdcmonstrated iflhc treatment is better
accelerated failure time models See Slm1\'AL than placebo. even if it is not as good as some other treat-
ANALYSIS. 'JRANSRlStAnON ments. The comparison of a new drug 10 an active In:alment
may be dictated by ethics but the object of the trial may
active control equivalence studies The classic simply be an indin:ct pRlOf that the treatment is beuu
ranclomiscd aJNICAI.TRIAL seeks 10 prove superiorily of a new than placebo through ClOIDparison to an agent whose emcacy
lmItmenl to an existing one and a successful conclusion is is accepted.
one in which such proof is clcmansll'alcd. The famous MRC Rctlently the issue of the indirect comparison to platlCbo
trial of slIqJIOmycin is a case in poinl (Mc:dical Research has been taken more seriously. Consider Ihe cue when: we
Council Streptomycin in 'lUben:ulosis liials Committee. ha\IC a single effectiye treatment on the market, say A. whose
1948). The trial concluded with a signiftcanl difference in emcacy has been demonstrated in a series ofbials comparing
outcome in favour oflhc group glven streptomycin compared ilto placebo. We now run some new trials comparing a furtha
to the group that was noI. In n:ceal yc:an., however. there has treatment. B. to A. 1bking all these trials together, they then
bcc:nan iDCMaSing inletat in IriaIs whose objeclive is to show have lhc structure of an incomplete blocks design. The effect
thai some new therapy is DO worse as repnls some outcome of B compared 10 placebo can then be estimated USing Ihe
than an existing treatment. Such trials ha\'C particular fea- double contrast of 8 comparai 10 A and A compared to
twes and difftc:ulties that wen: described in an important platlCbo. This approach has been examined in detail by
paper by Makuch and Johnson (1989) in which lhcy used the Hasselblad and Kong (2001). A CXJIIscquence of taking this
term 'active control equivalence studies' (ACES). particular view of malleI'S is that the precision with which
Actually. lhc term is nul ideally chosen since. unlike the effect of A was established compared to platlCbo cannot
bioequivalence studies. when: Ihe object is to show that the be excec:dc:d by the indirect comparison of B to placebo. since
bioavailability of a new formulalion is not only at least 204Jt the variance of this indirect conlnSt is the sum of Ihe
less than that of an existing formulation. but also at most2S 4Jt variances of the two din:c:l CXJIIlraslS.
more, ad he:nce w~ f!qUilYllence 10 some dcgn:c: is This is. however, nol the only difficulty with such studies.
genUinely the aim, in ACES it is almast always Ihe case that The folloWing are some of those that apply.
only noninferiority is the goal. II may be questioned as to why
the: rather modest goal of noninferiorily should be of any
inlclal in drug regulation. There arc several reasons. 11Ic flnt Est.liming Q cliniCtlIl,. irrelevant difference. Ifthe route of a
is lhat the new drug may have advDDtagcs in lenns of fannal anaIysisClOlDpamd 10 placebo via an indirect contrut is
tolerabililY. Second. the new drug. while showing no net taken. this particular difficulty may be finessed. Tbe new
advantage to lhc existing one. may increase patient choice RalJnent is shown to be 'sipiflcantly' better than placebo.
and this can be useful. For example. many people have an albeit using an indiru1 argument. and the extent of its
aspirin allergy. Henee. it is desinble to ha\'C altcmalive inferiority 10 the campanlOr is only of relevance 10 the
analgesics, even if no better on average than aspirin. Third, extent that it impinges on the proof of el1icacy compared to
it may became necessary to withdraw tn:abnents from the placebo. If this proof is provided.. then the comparison to the
market and one can never pR:dicl whe:n lbis may happen. actiYe compamtor is 'walei' under the bridge' . If this particular
'I1Iere IR now several stalins on the market. 1be facl that lbis approach is nol taken. howcvu. lhen any proof of eflicacy of
is so means that withdrawal of "riYaSlatin does not make it the new tn:atmc:nt rests on a dcmonslrlllion that it is not
impossible for physicians 10 continue to treat their patients 'substanliaUy inferior' to the CompandOr. which camp;uarar
with this class of drug. Pourth. introduction of further equiv- is accepted as being cfftcacious. 1bis nises the issue as to what
alent therapies befon: patent expily of an innovator in the it means for a ckug to be not substantially inferior to anolhcr
class may pennit price compelition to the advantage of ODe. This appears to rcqui~ that some naargin .d • .:I> O. be
reimbunors (although such competition is probably not adopted such that if" is theelllent by which the new trcaIment
particulariyeffccti\IC; Sean ad Rosati. 2(03). However. the is inferior to the slandard (where r < 0 indicates inferiority)
nOh nmon is probably the most important Dnq; regulation is then it is judged subslanliDlly inferior if 1" :5 -.:I and not
designed to satisfy some minimum requirements for phar- subslantially inferior or "equiYalent' if T > -.d.

&qdOfNllldjt CtNHpIIIfioIr It) Mftd"1IYI1 Slalislia; S«rMd EdiliufJ Edited by Briaa S. Everitt and ChrisIGph« R. P'dmeI'
C 2011 JohD Wiley & ~ ....

1
ACTIVEOONI'ROL EOUIVALENCE STUDIES _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ __

Tecl",;clI/ slolislim/ IDpeds. In a Neym~Pc:anoD infcn:ntial framework one decides 10 usc. An analogy may be
hmewarlt (sec Salsbury. 1998) the lesl of noninfcriorily uscrul hc:rc. In a pme of hUnI the thimble. a found thimble
raauira one to use a shifted NULL HYPOrHESIS. One miPt. renders the quality or the Slrately used for finding it
thcn:fon:. adopt Ho.~~-A. The silaalion is DOl lIS irrelevant. II is no marc 'found' ir a goad stnIcgy wen:
canlnn'asial as thai for InIe biacquiyalence. where the used than if a bad ODe were. However. a failun: 10 find a
fiact thai two hypotheses have 10 be rejected. that of thimble docs not automatically justify the conclusion that
inferiority and that of superiority. means that an inlaiU", the room doc:s nat contain one and the quality of the SCBIdI
approach of seeing thai the confidence limits for the employccl is a CI1ICiaI consideration in any judgemenl that
ditreraace lie within the limils of c:quivalc:ace is not it doc:s noL
'optimal' (Beller and Hsu. 1996~ althoUlh the 'optimal'
lest may in pnldia: be: worse (Perlman and Wu. 1999: Senn.
2(01). In praeticc. in the case of ACES if the lower TIre ejfed of DROI'OU1S, NONCOJIPUANCE tmd lire role of
CXlDYenlionai I - a ,wo-sided CONFlDENtE INIBlVAL for ~
INTEN1ION-ro-TIlEAT tnltllysu. It is plausible that in many
excc:cds -A. Ihc hypothesis or substantial inferiority may
cin:UJI1Itancc:s in conventional superiority trials if
be: n:jcclcd .. the level a ad noninferiority asserted. n might IICIIICOmpiiana: 01' clmpauts an: a problem an intcnlion-l~
be: thoughl that a o~sided conficleace inlCrVal would be IRat analysis will give a more modest estimate of the
suftlcient for this purpose. Howc:Yc:l', the general rqulalory
IRllllllcnt effecl than will a PER PIIOMCOI. analysis. In
CXlDYenlion is Ibal all tests clc:signed to show superiority
ACES. it is at least plausible thai this may nat be the case.
an: lwo-siclc:cl (despite appan:nt purpose) and. since such
tests an: a special case of a noninfcriorily lell with.d =O. usc
or onHdcd tests for noninferiority would lead to
COIfjlid ojrequiremenls ofodt/ilil'/Iy IIIfdclinical re/eFtlllt:e.
inconsistcncies (Sean, 1997; Committee for Prapric:bUy
It may be that the clinically irrelevanl diffen:nce is most
Mcclicinal Pnxiucas. 2000). In a Bayesian framework (sec
meaningfully established on a scale that is nat additive. For
BAYESIAN .IEIHODS) one might RqUin: that the posterior
example. in a trial of an anti-inrective. it could be most
probabilily of noninferiority wen: less Ihan same specific:cl
appropriate tocstablish thai theditrerenc:e in cun: rate an the
amounl. Abanalivcly. use of a loss functian would permit a
PIObability scale was DOl gJaIIuthan lOme specified amounL
clcc:ision analytic meIhod. such as has been proposed for
Contrariwise. the log-odds scale mighl lend itself marc
bioequiva1ence (Undley. 1991). to be used.
n:adily to staUstical modelling. This can lead 10
consich:rablc climcullies (Holmgn:n. 1C)g9). in panicular
Po...~r of ,rio/£. Nole that the rcasan one daes not employ a because a trial doc:s not Rendt a nmdom sample fmm the
value of A = 0 in practice is that unless it is expected that lalgct papulation. It may be that funhcr modelling usinI
the new ImIImeat Rally is better than the slandanl. the additianal data may be necessary (Sean, 2000).
power of the resulting test could nevc:l' exceed SO fJ,. A CGIIUIIOft ein:umstana: likely to make n:gulalOly author-
However, the clinically im:levant difl'cn:nce is likely to be ities ask questions is that a trial thai was designed with
less than the clinically n:levanl difl'en:nce used in optimism to show superiorily to an active co. . . .t... fails
CXlDYentionalliiais. Hena:. if the new tn:aImcDt is ac""'ly to do so. but then is used 10 aIIc:mpIto dcmonstndc noninfai-
no bc:_1han the stanclanl IRalment. then. for a given sample orily. 11Iis parlic:ular set of cin:umstances has bctvmc Ihc
size. the IIOIICICntndity panamcler. cl == AISE(J) is likely to subject ofoncofthe European Medic~EvaluatiaD Ageacy's
be: smaller f... ACES than for lliaJs designed to show 'poinls to consider' (Scan. 1997; CommitlCc f... Proprietary
superiority. COIIICqucntly. ACES either have lower power Meclcinal Pmducts. 2000). This stresses the desirability of
... higher sample sizes than conyentional trials. elilablishing the trial's purpose prc-paf'armanoc and also
warns apinsl establishing the clinically irrelevanl diffen:ncc.
A, after the trial is ClDIDplcte. II ~ putlinl a trial thai was
A.uay :lensili"ily. A pmbIem with ACES is thai if the trial designed to show superiority to the puI'JJCl5C ofncminfcriority
appears to show noninferiority or the new IIQtmcnt. then as an unac:a:ptableusc butKCqJtstheconvcnc.111e guideline
then: are thn:e plausible explanations. 11ac lint. thai of recognises thai then: an: no issues of multiple Ic:sting involw:d
cluance. is one: that statistical analysis is designed to with such switches (Bauer and Kic:sc:r. 1996) but that estab-
addn:ss. 111e second. that the new IIQtmcnl is indeed lishing values of A mraspc:clively may be biasing. 111us. it is
noninfcrior. is what was dcsimlto prow. However. a thinl preferable ror invcstigalOlS to specify in adwncc (e.g. by
possibility. that the experiment was not sensitive 10 find a mcansof fannal chanp to the QlNlCAL TRL\LS JIIOIOCOL) Ihcir
dift'cmacc. is difficulllO exclude. "111 is issue bas been n:rcm:d inlenclcd switch or purpcI5C and to nx the yalue of A prior 10
to as anc of ·competenc:e' (Senn. 1993) and atrects whalc\'CI' data unblinding. This. however. raises the ilSUC as 10 whether
______________________________________________________________ ADAPnVEDESIGNS

the value of.d is not somclhing the regulator shoulddeclarc for were to usc a traditional fully nmdomiscd approach to
given indications rathcI' than relying on the sponsor to do so. runnirq; the trial. which is not adapth-c. we would probably
Otherwise,. a regulator could be raced with the following not look at the data until the end or the trial. thereby risking
position. Drug B is rqislcnXl on the basis of comparison to exposing subjects to toxic doses and also possibly failing to
a standarclln:allnent A bc:c:ause the lower confidence interval produce an optimal estimate of the RXluired dose. Another
ror the lrealment effect. TIJ_A' excc:eds same pre-spccifted such example of an adaptive design isgival in Rosenberger
,·a1ue..d. HoweVCl', a further drug. C. which has also been and Lachin (1993). whereby there arc two trcalments in the
compared to A, is notgranled a lia:ace because a superiority study, A and B. and as inrormation emerges from the triaI.he
trial was planned. Although superiority to A was not proven. treabncnt assignment probabilities arc adapted in an aucmpt
the lower confidence inlerval for the In:aImeni etrc:ct 1'C_A to assign IDOI'e patients to the treatment pc:IftXllling beucr
excludes a smaller possible ditT~1It'e between C and A than is thus far. 11Jereforc. when a patient enters the study. ir
excluded for the difference between B and A by the lriallhat treallnent A appears to be better than treatment B, a patient
has led to rc:;istralion of B. SS has a greater than 50 Cit chance of being allocated In:alment
A - and vi" vena.
B...... P. aad Kieser, M. 1996: A urufying appRl8l:h f«confidence
inlm'als and testing of equivalence ud difremICC. Biometrika 83. 4. Because adapti"e designs modify the allocation of treat-
~7. . . .r.1t. L aad Bsa. J. C. 1996: Biocquivalcnce mals. ment on an ongoing basis, and thus protect patients from
intcrscClion-union ICSts and equivalence confidence sets. Slatiftical inefTeclive or toxic doses, they can be said to be more
Sdenre II. 4. 283-302. CeaualUee for ProprIeWJ l\IedldaI ethical than traditional designs. Rosenberger and Palmer
Prodad.s 2000: Points to oonsider on Switching between superiority (1999) consider the ethical dilemma between collective
and non-inrcriority. HUllibIad, V. aad 1CoDa. D. F. 200 I: Stati.stkal and individual ethics (see mucs .~ND CUNlCAL TRIALS) and
methods ror compcuiscn to placebo in acliveoCOiD)1 SlUdies. Drug argue thai in a clinical trial setling indh'idual ethics should
In/ormation Jorunttl35. 435-19. RoIIngreD, E. B. 1999: Establish- be uppennosl; i.e. consiclendion should be towards doing
ing equivalence by shwing that a speciftcd perce8IagC of thc cffect what is best ror patients in the current trial as opposed to
or the active control over placebo is maintained. Joumal of BiD-
doing what is best for future patients who stand to beneftt
plrarmat't'llticalStalisti(,..J9.4. 65 1-9.IJndI..,.. D. V. 1998: Decision
analysis and biocquj\'aJcncc lrials. Slatisim Sri~lfCe 13.2. 136-4J. from the l'Csults of cUlTCnt trial. The Declaration or
Makudl, It. aad Jabasaa, M. 1919: Issues in planning and Helsinki of October 2000 outlines the tension between
interpming adiYe control equivalence studies. Joumal of Clinical these two types of ethics by stating: ·Considerations
Epitlenriology42. 6. 503-1 I. MedkaI Researdl CoaDdI sa.... related to the well-being of the human subject should take
m)'dn III Tabercu1a511 Tdais CGlllllllttee 1948: ~ptomycin prcc.x:dence over the interests of science and society: It is
batment for pulmonary tubcmalosis. Brilish Medical JoufIJQl ii. adaptive designs thal address the indiYidual ethics, as
769-82. Pedmaa,M. D. aad Wu, L. 1999: 11ac empeRll"s nc\\' tests. opposed to fully randomised designs. which address those
Slalistical Sciencr 14.4.355-69. Sal....,., Do 1998: Hypothesis collective ethics.
testing. In Armitage. P. and Colton. T. (eels). EnC)'tiopedia of We will be dealing primarily with n:sponse adaptive de-
biDslatisti('s. Chichester: John Wile)' & SOlIS. Lad. Sea&. S. J.
signs here. such as those just outlined. and will not be
1993: InbaaIt difficultics with active control equivalence studies.
StalUtia in MeJitUre 12.24,2367-75. SeDII,S.J. 1997: Slatiftical describing those designs that atlempt dynamically to balance
ismr.f in drug tlrvelopnrent. CUchestcr: John Wiley & Sons. Lad. the randomisation forcovarialc information. such as oUllined
s.m. S. J. 2000: Consensus and con~ny in phaimaceutical by Pbc:ock and Simon (1975) (see D.~TA-DEPENDEHT DESIDNS.
statistics (\\ith discussion). 1M Stalistician 49. 135-76. Sean. S. J. MJmMISA11ON).
2001: Slalislic:al issues in biocquivalencc. StDtiftia in Medidne 20. The randomised play winner (RPW) design attempts to
17-18. 2785-99. Seaa. S. J. ad ItaIad. N. 2003: Editorial: Phar- a1loeale trcaIments to patients sequentially based on a simple
mac:eutic:als. paIcnIs and competition - some stalislical issues. probability model. Rosenbeq;er (1999) emphasises that the
Joumoi 0/ lire Royoi Stalislical Sot'iel}' sma A - Statisti('s in RPW design speciftcally applies to the situation whel'C the
Sotiety 166. 271-7. outcome from a trial is binary, i.e. either 'success' or 'failure'
and where there arc only two lrealments. e.g. chug A and drug
adaptive designs a.JNJC\L 1RLWii that arc adapli\'C arc B. At the start of the trial there is an assumed urn of a baDs of
modified in some way by the data dud hayc aln:ady been type A (which rehde to drug A) and fJ balls or type B (which
collected within that trial. The most common way the designs relale to drug B). When a subject is recruited. a ball is drawn
adapt is in the allocation of treatment. as a function or the from the urn and then l'Cplac:ed. If the ball is type A then the
n:sponse. For example. we may be interested in a dose that subject is allocated to drug A. if type B then the subject is
givcs a 20., chDlKlC of toxicity. whel'CCX"SSCS to this level of a1loeBled to drug B. When the subject's outcome is available
toxicity would be harmful. Thcrcrol'C, we may want to design (and we assume that the outcome is available befol'C the next
the trial in such a way that. as more infonnation is gath~d. subject is randomized), the WD is updaacd. If the response is a
doses are a1loealed to optimise lhc estimate of Ihat dose.lfwc success on drug A.then a ball ortype A is put into the urn, and
3
ADAPnVEDESIGNS ______________________________________________________________

similarty fora success on drug B.lfthe ouk:omc is a failure on each or the series of doses. The first patient is then ~aIcd
drug A. then a ball of t)'pC B is put into the urn. and spin with the dose that is aJllsidcred to be the closest 10 the TOlD.
similarly for a failure on dl1ll B.ln this way. the balls build up Once the OIIk:ome is obsc:rved the FROB.o\BlUTYof Ioxic:ity at
such thai a new subject has a better chance or being allocated each of the doses is recalculaled using the Bayesian method
to a better lreatment. of statistics. The proccd&R continues in this way until it
Rosen~er (1999) concludes with a table of eonditions wtles on a single dose. Whitehead el QI. (2OOIa) point out
UDder which the RPW rule is reasonable and provides a that the CRM does home in 011 the ro20 quickly and
rQlistic allenlalivc to the standard cliniul trial design. These efficiently, but then: has been concern that early on in the
an: given in the table. lrialsubjc:cts could be allocaled 10 too high a dose. leading 10
palentia) toxicity problems. This has led to a number of
modifications. such as starting at the lowest dose and never
adaptive designs CondItions under which the RPW Is skipping a dose during the escalation.
reasonable (Rosenberger, 1999) Whitehead el aL (200lb) suggelil practical exlensions to
the CRM for pharmacokinclic data. employing the use of
• The therapies ha\"C been evaluated previously for toxicity Bayesian decision theory 10 allocate ~alments optimally
• The raponse is binary 10 subjects. They argue that conventional dose-escalatiOll
• Delay in n:sponse is mociemte, allowing adapting to take studies carried out in healthy volunteers do not normally
plaee employ statistical methodology or fonnal guidelines for dose
• Sample siza an: moderate (at least SO subjects) escalation. As such the studies can take a long time 10
• Duration of the lrial is limited and recruitment can take complc:lc with little opportunity to skip doses. The methods
plaee during the entire trial proposed allocate doses in anlcr to maximise the information
• The trial is carefully planned with extensive computations about the ~response curve. gi\'Cn a pre-speCified safely
done under dilTerent models aDd initial urn compositions constrainL They use two simple utility or gain functions. one
• The experimental therapy is expec:led to have signiflcanl that allocates the highest allowable dose under the safely
beneftts to public health if it proves effective constrainl and the other that allocates doses in order to
optimise the shape of the dosc-n:sponse curve.
Krams el al. (2003) also use a Bayesian decision theory
Traditional dose-n:sponse studies. where patients are approach with sc:quential cIase anocation 10 a Phase II study
allocated to a limited number of doses along an assumed in acute SIroIcc therapy by inhibition of neutrophils (ASTIN).
dose-rcsponse cun'C. are limiled and. some would say. which employs up 10 15 dose levels. They usc a response-
wrong. For example. if the assumc:d dose-n:sponse model adapti\"C procedun: in order to find a dose that gives an
is inCOlTCCI then palients may be allocated to ineffc:cli\"C or improvement over that of placebo in the primary ENDPOINT.
unsafe doses. One answer could be to increase the number of allocllling the next subject eilhu to the optimal dose or
doses. However. this would resull in many patients allocated FLo\CEBO. Slopping nales were employed by which if the
to wasted doses. It would be much belter 10 increase the pD5lerior probability of an effectivc drug or ineffective drug
number of doses and allocate doses to a subjC:CI based on were greater than 0.9 then the dc:cision would be made eithu
cum:nl knowledge of the dose-response curve. which best 10 go on 10 a confinnatory lrial (effeclive dnag) or to stop
optimises some IR-spc:cified criteria. This is precisc:ly what development (inelTc:ctive drug). In this way. they were able to
Bayesian response adaptive designs attempllo do. by em- stop dc\'Clopment or a compound more quickly than would
ploying Bayesian DECISION THEORY to a utility function. Thus. have been possible under the traditional panuligm.
the dose thai most optimally addresses the utility is allocated In 2006. the Pharmaceutical Research and Manufacturers
to the next available subject or cohort of subjects. of Amc:rica (PhRMA) Adaptivc Design Working Group
One of the first BA~ r.tE1lIDDS described was the pUblished a series of papers in an issue of the Drug Infor-
continual mlSsessmc:nt method (CRM), inll'Oduced by mation Association (DIA) joumal detailing various aspects
O'Quigley. Pep.: and Fisher (1990), and originally devised of these lrials. Topics included terminology and classifica-
for dose-escalalion studies in oncology. Whitehead el QL tion: implementation; conftdentiality and trial integrily:
(200la) suggest that the method ClOUld also be used for adapti\"C dose response: seamless Phase IIIIlI: and sample
applications in other serious diseases. The CRM c:nvisages size n:estimation (see Drug In./OTnJlllion JOIImal40. 425-84.
a study whereby human voIunlc:ers are lrealed sequentially. in 20(6). In addition. and re8ecling the growing inleresl in
order 10 detc:ct a dose with a probability of loxicity of 20 CJt. adaptive designs. there have been numerous special editions
i.e. TD20. The ~sponse is a binary response, 'Ioxicity' or 'no of other jounaals devoted to these trials. includi~ JoumQI
toxicity'. Before the study staIts, investiptors are asked to 0/ Slalislical Pltllllling ad In/erence. issue 136(2). 2006:
proVide what their best guess is of a probability of toxicily at JOUTnQI of BioplwrIPlQt:eUliml Slalislics. issues 16(5).
_ _ _ _ _ _ _ _ _ _ _ _ ADJUSTMENT FOR NONCOMPlIANCE IN RANDOMISED CONrROLLED TRIALS

2006 and 17(6). 2007; and Stalislics in Medicine. issue In:abnent policies as actually implemented in the llial -
27(10). 2008. AS e.g. "drug A plus changes' venus ·dlUg B plus changes'.
Unlilcc efl'ectivc:aess. f!//k1lC)' ..,11lleS 10 the eITects or the
KnIll., Mot Lees, JC., IIac:b, W., Grine. A. p.. Oraoaau,J..l\oL
aDd Font. O. A. 2003: Acule IIIakc therapy by iabibilian of In:alments themselves. and is not estimatc:d by an intc:Dtion-
nculnlphils (ASTIN). An adaptive cIose-Iapoasc study or UK- to-trat analysis.. Rc:scan:ben may also be inlclated in the:
279.276 in Kute iscbanic stnIke. SlroJcr 34, 2543-8. Paeoek, S. eO'ectiveness oran intervention in othc:rcilaunslancc:s. e.g. if
&ad SImaa, R. 1975: ScqueDliaillatmcnt assipmeDl with ~ public suspicion of the intervention had bec:a ~duced by the
iltl or prolDDSlic factCIIS in CGIdIOllcd clinicallrials. Biome"k~ l I. positive aawls of a clinicallrial. In these circumstances. the:
IOJ-IS. O'QaIIIe". J., PIpe, Me aDd FIIbIr, L 1990: Coalinual ratc:s orcompliance may be improved and actiustment for this
reassessment mcIhod: • pl'ldieal design far fhasc 1 clinical trials change may be aIlc:Inpted.
in cancer. BionretTia ~ 33-11. _ ...... W. F. 1999: R..- It is imponant to define the aim or adjustment for non-
cIomiJJCd piay-lhc-winDc:r clinicallrials: rmcw and m:onuneada- compliance. For example. in a trial of immediate venus
tians. COIIIroil. Clilriml Trab 20, 321-12................ W. F.
c1crc:nat zicIovudine in asymptomatic HIV infc:ction. the:
&ad tar..... J. M. 1993: n.e usc or rapanse-adaptive designs in
clinicallrials. Co""oIledClinimlTrillh 14. 471~............., initial plan was to derer zidowdine until the onset of symp-
W. F.... hIBler, C. R. 1999: Ethics 8IId practice: alternative tomatic disease. However. rollowing a pnJtocol amendment.
designs far Phase 10 ranclamised clinicallrials. COIflrollfti Clinical some individuals startc:cI zidowcline beron: the onset of
Triola 20, 172'-. WUeMad, J., Y......... z., P......... So, symptomatic disease (White et Ill.• 1997). Then: was interat
We""", D.... Fraadr, S. 2mla: Easy-lO-implemcat Bayesian in estimating the eITect Ihat would have been absc:m:d undc:r
methods far dasc-esc:alalian studies in healthy volunteers. Siolla· the original protocol. Zidovudinc: beron: the onset of symp-
lirl;cr2.47-61. WMtebead,J.. ZIIDa, y ..staIIard,N..Tadd,S. ... tomatic disease was thc:n:rore regarded as ·noncompliance'.
WbIIeIIad A. 2OO1b: Lamial fram JRYious responses in Phase 1 Other individuals stopped zidovudinc Imdmc:IIt because of
cbc-esc:aI.oa 1lUdics. BTitUir JDIInIIlI of ClininrJ Pbsrnlll(,o/ogy advc:rse events. Additional adjustment ror stopping tn:almc:nt
S2. 1-7.
would not answc:r a clinically relevant question. so the
analysis did DOl aim 10 estimate efficacy.
adaptive rancloml.. aon Sec ADa\P11YE DESlCINS. Adjustment rar noncompliance: is useful in a variety of
JlANI)CaUSATlON situalions. Patients may be most intc:n:sIcd in In:alment
efticacy. DiO'e~nc:cs in compliance may help to explain
adJU8Imeni for noncompliance In randomlaad variation or a lreatment effm with time. between subgroups
controlled btala In clinical medicine. ·noncompli- in a Irial or betwc:c:n trials in a .mAo-ANALYSIS. Reconciling
ance' occurs when a patient does not rully rollow a prascribed llial daIa with obscr'Vational data may laIuire adjustment rar
course: of lIaImenL The alternative terms "adhcmlcc' and noncompliancc: in the: trial. Policy analysis may lellui~
"concanlance' attempt toaw.icltheautharilarian cwertonc:s or projc:ctions for situadons with improYed eampUancc:.
"compliance'. In randomisecl QINIC'AL TRL\U. we an: can- Most llllempls 10 allow ror noncompliance use on-IRal-
cemc:d with any ~ rrom a randomisc:d lRalmcnt. ment analysis or PER PIIDI"OCOL analysis.. This only proviclc:s a
whether due to noncompliance CJI' a In:alment change q~ed valid comparison of the In:alments themselYc:s (efflcacy) if
with mc:cIical staff. In a trial to eamlNR two types or complic:rs and noncomplien do not cliffeI' systematically in
medication (drug A and dnag B. say) for the: tn:aImeat or their disease stale or prognosis. In practice this is unlikely to
hc:art disease. far eumple~ patients may ~ftlse or rorgc:t to be the case. so selection bias occurs. Heart clisc:ase palic:nts
take any of their medicalion or £orgc:a to Iakc: it some or the who comply with their pn:scribc:cl medication. rar example.
1ft also those who an: likely to impro~ thc:ir diet or lab
time (partial compliance). PatienlS aliocalcd 10 ra:cive dl1lg
A might switch to drug B~ mad vice versa. Some of the patients man: exc:n:isc and thesecllanges. in lOm.1Ire likely to lead toa
mayevc:a take anaIhcr mc:dication altogether (drug C. say) or. better outcome. SELECTION BIAS may often be n:duced by
particularly ir the Ihcrapy appears to be failing. RXlCive a adjustment ror baseline m\'ariates~ but thc:n: is still no
much more radical intc:r'VCntion such as surgery. A rwther guanntcc of an unbiased analysis. For elUUllple. in the:
complication for the estimation of In:aImenl effects arises Coronal)' Drug Pmjc:cl. 5-year manaUty of poor eampUc:n
when patients who fail to comply with thc:ir prescribed was 28.2 CJt compared with 15.1 CJt in good mmplic:~ and
tratment an: also Ihose who an: IIIIR likely to be last to adjustment far 40 baseline ractors only raIu4"ed the diITer-
follow-up. ence to 25.8 CJt vc:nus 16.4 CJt (The Coronary Drug Project
Rc:scan:h Group. 1980).
Newer ·randomisation-basecl' methods am estimate em-
Ratiollllie. Conventionally. trials with c1cpanun:s ftvm cacy while aw.iclinl selection bias by din:dly comparing the
mndomisc:d tn:atment lire analysed by IN1ENI'JON-~TREAT. groups as randomiscd as in an ination-IO-tn:at analysis
This clim:lIy compares the e;8'rclirmess of the diO'e..,nt (While., 20(5). This is made possible: by considc:ring the
5
AGE-PERIODCOHORTANALYSIS _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ __

subpuup or 'compliers' who would ha\'e n:ceivcd Ihcir abaat comparability or noncompliers and compliers. In athc:r
mndomised batmen.. whichever group they \\'ere nndo- situations. some gain in power is theoretically possible. but
miscd to. For example. a trial in Indonesian childn:n this is unlikely to be appn:ciable in plKticc (Becque and
campan:cI vitamin A supplemc:alalion with no inlerVention. White. 20(8). Signiftcance testing should therefon: rely on
the outcome beil1l 12-month manalily. Vitamin A supple. intention-to-lmll analysis even when other melhacls are used
mentation was actually received by only 8Oe.t or the to estimate emCat'y. IWIGD
intervention ann and by DOlle or the caatnJl ann. Sammer
and 2qcr (1991) cansideml the subgroup who did not AnpIst,J. Do, IID....O. w..... RaItIa,D. B. 1996: ldentificlllioa
n:ceive vitamin A in die intervention ann and a corn:spond- of causal effects usiq illSlnllDeDtal variables (with discussiOll).
illlsubJrouporlhe conbol ann who ~·ould rrol/raJ,-e ,eaiJ,-eti JoumaloJlbe AmtriamStaiinicalAs.ftJdalitHI91.~72. B....e.
l'ilanrin A if Ihey bad hem allocaled 10 receire il. These T."'" WIllIe. L R. 2008.ltepininc pG'A'Cr lost by non-campIiance
'noncomplier' subgroups were assumed to be unalJ"ecled via full proIIability modcllinc. Slatistiu in Medicine 27. 5640-63.
by allocalion to vitamin A. It is then straiplfolWani to DulIn. G............. R. 2007: ModdlinglJatmcnt-cffc:ct hcaem-
~ity in rancIamizcd CGIdIOllcd trials of complc:x intcnentiaas
estimate abe number or nancomplic:n in the conlrol arm
(psydlolCJCical tn:aImcnlS). StaiulicJ in Meditillt 26. 47(4)....15.
and their mean outcome. and hence the risk difTen:ace. risk .....-....." K. .... ~.... B. 1999: PmcticaJ pnJpCrties
maio or odds ratio in complien. This is oRen called of same structural mean analyses of the effect of compliance in
the 'complier .~ causal effect' (CACE) estimate randomized trials. COIrtroiletl CliRitll' Triu& 20. S31-46.1..1tt1e. R-
(Utile and Rubin. 2000). This approach is • special case of .... RuIIIn. 0. B. 2000: Causal effects in cliDical and epidemiolog-
PRINCIPAL STRAnFlCAtICN. ical studies via potential oukClmC5: CCJIICCIIlS aad analytical a~
A man: gc:aeral approach requires a model n:laling praachcs. A""""I Bern· oj PIIbIir Health 21. 121~S. 'Ibe C.....
polential oUIcomes for each individual under different ..,. Draa PmJtd R..ardI Graup 1980: InIIueaoe of adhaaace
to lRallllent and rapoIIle 10 cholcsaeml on mortality in Ihc CORJRII)'
caunterractual In:almc:ats. A simple model mipt sa)' that
Drug PIojed. Nn' Eng/ad./tJumQ1 tJj Akdkine 303, 103s-J1.
each individual would have blood pressure b nunHg lower if
WIllie, L R 2CJ05: Uses aad limitations or randamizaIioa·1Jascd
they lOok abe cIru& with perfect compliance than if they did
eflicaty cstilDldol5. Sialislital MellrDt/s in Medial' Relml"ch 14,
not take the cInas. with pmpodional blood pn:ssun: mluclions 327-47. \\'Idte. L It, Walbr,5..BaIIIker, A. G...... .,.,.,.........
for pallial compliance. Such • maclcl may be lilted by J. R. 1997: Impad of tn:abncnt chanFs OD the inlcrpmaaion orthe
abservil1lthat untn:ated blood pn:ssure must have the same Conconle trial. AIDS 11. 999-1006.
dislribulion in each rancIomisc:cl group (Fischer-Lapp and
0adJhebc:ur. 1999). An important ad\'8Rlage of these .ge-perlocI cohort ...lysls To understand die
methods is that no assumption is rcquin:d about the relalion- effect of lime on a particular oulcome ror an individual it
ship bc:Iwc:c:a compliance and polc:IIliai outcomes. 'I1Iey are is c:ssenlialto n:alisc the n:levanl temporal praspc:ctive. Ale
closely related to the use: or JNSTIt1}),IE1Il VAlUABlES mdhods affects mIlD)' aspects of life. including Ihc risk of dise:ase. sa
(Dunn and 8entall. 2007). this is an essential componc:at or any analysis of time trends.
The approaches just described are genc:rally only able to Period denotes the elate of Ihe outcome and if Ihc outcome
estimate one tn:almenl effect in a Iwo-ann trial. They ad to varies with period it is likely to be due to some undedyilll
be: hopelessly impra:ise: in situations such as EQUIVALEICE fat'tor that affects the outcome and varies in Ihe same way for
STUDIES where patic:ats may Slopallbatment during the trial. Ihc entire population under study. Cohort. aJIItrariwise.
so that the analysis rcquin:s estimation of the elrect or both refers to lencrational effects caused b)' factors that
lIaImenlS.ln this case it is possible to adjust abe nndomised only affect particular &Ie groups when their level changes
comparison usil1l observational estimation of one or more wi'" time.
treatment elTects - i.e. assuming the~ are no unmc:asun:d An eumple ora period effm would be a potential elTecl of
caafounden for tn:aIment. Methods such as mtlrginlJ/.rlnlc- an air contaminant thai affected all qe poups in the same:
Iliral modellirrg can work e'VCII when at'tuaillalmeni is both way. If Ihc leyel of exposure to thai fat'tor incrcascdlde.-
a consequence of symptomalic deterioration and a cause of creased with time. ellClting a change in the outcome in all
slowa' disease progression (se:e Little and Rubin. 2000. for age groups. then ~ would expc:ct • relalcd paltel'll aclDSS all
~ferences 10 this literalUre). age poups in Ihe stud)'. In studies that take place over IOIIJ
A trial wilh noncompliance has less POWER than one with periods or lime. the technology for measuring Ihc outcame
perfect compliance. as a n:sult of Ihc mlUClCd effect size as may change. giving rise: to an arlifactual effect thai was nol
estimalc:d in an intcDtion-lO-Rai analysis. and it is natural to due to change in exposure to a causative ageaL For example.
want to n:cover the last power. Howe~r. many or the DeW intensive scn:ening for disease can identirydisc:asc: cases that
pnx:ed1RS preserve the intcntion-lo-Cn:aI SKOOF1CANQ LEVEL would not previOusly haw: been identified. thus artificially
and thererore do not affect power. In some cases. il is increasilll the disease rate in a population "'aI
has had no
impossible to relain power without makil1l some assumption change in exposure over time.
___________________________________________________ AG&P~ODCOHORT~~S

Cohort (also called birth cGbart) eJrcc:ts may be due to aenlS or ownIllinear tn:Dd and curvallR ar cIc:paItIft fram
fllClorS n:1ated to exposun:s assacialed willi Ihc dale oI'birlh. I.ar_ In:nd. For c:uaaplc. • can be gi\len by
suc:b.dIe inlRMluctianofaparticulardrul_pnldiceduriDl ai = i P.. + ci ;. when: i=i - 0.5(1 + I). tI" is die overall
pn:p8DC)' dial was bmuchl in at a parlicular point in tilDe. slape and ;;; the cunalule. The ownIl model can be:
FarexampIc. aprepancy pmcticc assaciated with iDcn:asc:cI expn:ssccI as:
risk and adapk:d by the popuIaIian or mothers cluriDg a
panicular tilDe pcriodcould aired the risk during Ihc lire. . .
of Ihc c:nIiR: ICncnllioa bam cluriq thai period. While it is
conunaD ton:l'crto thc:sc: efl'eclSas bc:inlassodatc:d with Jar
ofbinh..lhcy could also be: &be ",sak or chanps in c:.x~ 2010
thai occ:unalaRcrbirth.ln .....y inclividuals..lifell),le fadexs
thai may affect disease risk ewer a lifetime an: Iixc:cI .Ihcy
approach aduldlaacl.. Aqullllliftcalion ofthesc cft'c:ctson such
a pnc:ndian would give rise 10 a COlDpuisan oflhc:se cohort
or ICncndioDaI efra:ls.
An inhc:n:nt n:dunclancy'amGIIglhcse thrc:c Icmponl fae.
Ian arises rram Ihc facllhat knDwin& any lwo fadon implies
the _lie of &be third. Farc:.xaaaple. if we: know an individuals
• Cd) at a given date or period, (P). Ihc:n dac: cabart is the
difrCIaICC (c =p - II). This linearclepcndc:ncc liYC:Srise to an
idc:nliftalH1ity pmblc:m in a fonn" ...-c:ssion model that
atlc:lllpis to obtain quanlilali~ c:stimates of rqrasion .,....
mc:Icn lIIIOC:iated with each IalDporai c:1emc:at:

E[Y) = flo +II/l. + PIl, + r/J"

Using Ihc linear n:latiansbip between Ihc ~ f~exs 30 40
lives rise ID: Age(yearB)
ElY] = Jlo + 4. + p/l, + (P-")Ilt __period cahort . . . . , . lsIs diagtam showing
= /10 + d(IJ,,-!Jt) +p(JJ, + /le) the RJItJtionship beIween age, period and oohott. The
dagonaIline tnIc8s III/8'11IHiOd lfelime toran intIvkIuIII
'which has only two identifiable piaramcIcn bc:sidc:s the bam In 1947
inlcR:c:pI instead of die c:xpectccl tIRe.. Another way of
Yisualising allis phc:aolllCllDll is Ihat all eambinatians or SIC.
period and cohort II1II)' be displayc:cl in &be LExIs DIACIWI
... .... .. -
e["gt] =/l + (;!J" +a i) + (jll" +~i)
(_&be ftpn:)~ which is obviously a n:pn:sentaliDn ofa two-
+ (lcll" + y.:)
-iUJ.. +/l,,) + -i(JJ" +/l.,) +a;-
dimc:nsionaI plaac inslCad of die: tine dimensions cxpc:c1Ccl
for tine separate factors.
=JI +
In pnc:nl. dlCse analyses an: IIDllimited to linear cfl'c:cIs
appIiccllo a eaalinuous mc:asun: oflimc:~ but iDslead dle:y 1ft
applied to b:alporal intervals.. such • disc:aso IIIICS observed
-
+.7ri+1".
.....

for 5- or l~ycar intcmds or.., and period. When Ihc widths because Ic = } - I. Th..... caeh of the curvalun:s can be:
of these iatcrvals an: equal. the model may be c:.xpn:liICd as: lDIiquely delalnined. but dae 0\'e11l11 sq,cs 1ft hopelcssly
c:ntaqlc:d so that only CCIIain CICIIIIbinatians can be uniquely
£(Yp) =/1 + a; +.7rj + Yk
cllimated (HolfonlI913).
when: /l is the inll:RcpL a, the efl'eet or . . f_ die ilia '1111: implication or the: identifiability pnJblcm is thallhc:
(/= I •.... I) inlen'aI. lrJ Ibe effc:cl of period far &be jlh ewerall dim:tion of the c:O'c:cI far any or the ..... 1CmparaI
U= I ..... J) inlCmll and r.die eJTect orIhc kth callOd companellts cannaI be cIc:b:nniaecI fium a ~lR'ssion anal),-
(k=i-j+l= t ..... K=I+J-I). 'lheusual c:onsIIaints in sis. Thus.. we: cannoI evc:a detcnnine whether the tRmds an:
this madel imply dIa. Ea,= E~J= EYA =0. '1111: _ntiaa- inm:asiq ar declasiq with cobad. far instanc:lc. '!'he
bility plUblc:. manircSlS illelfdnugh a single: unidentifiable sc:cond ftgun: on PIIIc I displays snc:nI combilUllians 01'
palBlllCtc:r (Faenbc:rg and Mason. (979). which can be ~ age. period and cohort parameters,c:Kb SCI ofwhich JllDvicIcs'
easily sc:cn if we partitioa each tcmparaI efl'eet into compo- an iclc:nlical Ic:I of filled rates. Nalicc that as Ihc: period
7
AG&RELATEDREFERSNCERANGES ________________________________________________

panuaeIeIS an: IObdc:cI clockwise:. the age and cobalt panI- 2004: TCIIIpCII'II flldals in public heal... survcillance: SCIItiDg aut age.
mdc:n an: eomparably IQlaIcd in abe counterclockwise period and CCIhart dfeds.1D Bsookmc)'CI', R. and Sbaup. D. F. (cds).
dim:tion. Each of these parameters can be IObdc:cI a fuU aVDllilDrinl tbe bealtb 11/ populstitNu. Odani: Oxford Uai\'CISiI)'
181)0. but it is importDDl also to n:alise that they cannol be PIal. lIP- 99-126.
IOlIIII:d one at a time.. only all together. n..s. even thOUlh the
speaRe tIaIds cannaI be uniquely cstimalcd. certain combi- age-ralatedreference ranges 1besc an: ranges of
natio.. of the overallln:nd can be uniquely delennincclsuch values or a mcasu~mc:at that identify the upper and Iowc:r
as!J" +{J". which is callecl Ibc nel drijJ (Clayton and Schil1lcn.. limit of nannality in lhe population. w~ the range varies
1987a. 1987b). Alk:maIive drift a.timata covering shaner according to Ihc subjecl'S age. Rcfcn:nce l'BIIIes are an
timcspmas can also be dclcnniDcd and these have pnldical important put of medical diapasis. w~ a conliDUDUS
significance in ta.abey describe the experience orfoUowing a mcasU~lI1CIIt (e.1- blood pn:ssun:) needs converting 10 a

palticular age paup in lime, because bath period and cohort binary wriable fardccision-making purposes. If the patient's
will advantIC lOgeIhcr. Curvatura. by way of aJIIlnIsl. are value lies outside Ihc measurement's reference range it is
campldcly cldcnnincd. including paI)'DDIDiai panunden for tn:atccI as abnormal and the palicnt is invcstiptc:d further.
the sq~ and higher powers. changes in slopes and second The canSlJUclion or refcrcacc I1IIIgcs involves estimatinl
difrc:n:nc:es.. The signiftcancc test ror anyone of the Ic:nlporaI lhe range or values that CO\lCJ'i a spc:ciftc:d pm:enlqc or Ihc
~ren:ace papulation. often 9S .... Usually this is the ccntnl
eft'ccts iD the pracnce of the 0Ibcr two wilileacrally be aacst
of abe conaponding curvatun: and naI the slope. Holfonl part of lhe distribution with equal rail an:a probabilities.
piovidc:s rudhcrdcrail an how software can beset up for fitting allhaugh in somccascs the ~ren:nce range is baundccI at zero
or infinity. For nonnally distributed cIaIa the range can be
these models (Holford. 20(4). ' TRH

aa,toa,D.............. E. 1987a: Models for temporal wriation

cIeri_ fnmI the population ~ and STANDARD DEVL\11O.'f
(SD). die 9541, nmgc. for example. being the meaD plus or
in caliccr I1IIC5 I: Ap-periocland ~ahort madcls.. SlatUli('s in minus 2 SOs. For nannormal data the simplest approach is 10
MftlidM6, 449-67. a..., D...... SchIfIInI, E. 1987b: Madcls use quanliles. i.Co rank and caunt the data. then the 2.5CJ. and
for temporal ~ in cancer riles D: ~ ClDhon nladelL
'¥IoSCJ, points are Ihc lower and upper limits or Ihc 95 CJ,
St,,'isliaill aVecIi('ille6,469-8J.I1eJIbIq,5.E..........., w. M.
1979: lcIeatilicalion and cstimaliaa or ~period-cobod models in referaICC I1IIIgCo However. this is indlicicnt and n:quircs
the aaaI)'sis of discmc. .bMIl data. Sociolori('tll'MelhoiolDgy a large sample. If the data an: sIcew they can be iransronncd..
1971.1-61. Halford, T. R. 1983: Tbc cSlimaliall of age. pcrioII and e.g. to lopritlumi. and then the refcrcacc range can be
cohorI drects rar vital rates. BiDlrwtri('s 39. lll-24.1IoIfonI. T. R. calculllled f..... the mean and SD on the transformed scale

2.------------------------------------------------------------

1 ~ _
........- ........--........- .........--........--........- .......... ......... _ ......._.
0.5 .......- .......- ........--........- ....-..- ........--.....~.~.-.........- .........- .......-.

-1 ..M._......._.M...._.M...._ ........_ ........_ .........M_....... M ....... ......

PerIod slope
-1.5 . . M._........._ ........_ ......._ ......._ ......M_........_ ........ -0.00
-0.05 --
-0.10
-0.15 I--
-0.20
~.5~------------------------------------------------------~
age period cohort 1InIIIys. AiJB. pIHiodand cohof1 efIeds torpm-menopausal bnHlsI cancer incidence for SEER
19~1~ ,
_________________________________________________________ AG&S~F~RATES

1m~------------------------~
180

f
-140
12OJ------
-
I! 100 L-------~
1m.··············
I. .................................................. ..
1 80
40 .......................................... . ........................
20
o~----~------,-----~------~
1.. 19 24
Age(J8ars)
.....,..... ....iance . . . . AgtHe/ated 95" reIetence ranges for blood Pf8SSUte In boys: systolic (solid lines}
and diIIstoIIc (dolled Hnss}

andlnnsfGllllCd back 10 Ihc: original scale. A JIIOIe ftexible UIS MEIIIDD isa papuI.-way todo~is, or altemmvelythcEN
variant is 10 use a 8oJt-Cox power tnmsfomudion (of' which method of Royllon and Wnpa (1998). For IIIIR extn:me
the laprithm is a special case). which adjusts for slcewaess IlDllnormai elata, a nOnpanuneIric appmach based on QUA.N1D.E
more prmscly (sa: 'IIlANSRJIW.VIO). PElWfSSIlN i5 ncecIecl. a fonn of least absaIUb: enors n:p:s-
A&e«laIcd refen:nc:e nnps an n:feIaIcc l1IIIIes that sioD. when: smoaIh curves are cOlllbuc:led for the age-
dc:pend on DIe. They arise mast commoal)' in paediatrics. related upper and IGWa' limits of' the n:feraace ranp. 'I1Ie
notabl), for apHelalccllllCaStRS of'bod)' size like hei&ht and figure livcs ase-n:JaIcd refelalce raqes for systolic and
weight. which can be displayed as CIIOW11I awns. The diastolic bloacl preIIUfC in boys.apd 4-24, estimated by the
priac:iples or reference IBIIIC eslilDalion ale CSICIIliaU)' the LMS methacI.
SIIIIIe wilen they an: age re'aIaI. exC:Cpllhat Ihe JDD&CS Cor ~ an: two advanIaps of'refeninte mnps based on an
adjacent BlC JIUUPS aeeclto be consisteat. To awid clilCOllli- undedyiq fmauenc)'distributioa. uapposcdto~cIeri_
nuiIic:s at the 8IC lmap boundaries .requires the sUIIUIUIIY . . . quantile rcp:ssion. TIle lint is eflic:iency - the standard
slatistics toclefine the refereDCIC nmce (e.I.the mean ancISO) c:mmI ofb refen:noe I1IIIJC limits are smaller. 1b: SCICDIId is
andtocbanpsmaothl)' with lIIe butimpasiq this constraint anaI)'licai conVCIIiencc-daIa for indivicluals can beconvenccl
compIicarcs die fitling piQCess. For normaIl)' distributed loz-SCORES.. indicating how many SDs they an: above or below
homosccdastic data, when: the SO is coastanl ac:nJSS BIC. the median of the c:IisIribulion. wllic:h is a convenienl way of
the lIIe-n:lated mean can be eslimllb:d by LINEAR IWIRESSION adjullilll for. prior 10 fUrther analysis. TIC
and Ihc: Jd'en:nc:c map cxmsllUcted 8IOUIId the regrasiOD (Sec also 0I0W1II awnsJ
cane using the n:sidual SO. 11M: rqrcssiDn curve is estimated
AlIda, M. 1987: Madcllilll Y8riIIxe hc~ in IIGIIDIIIqIa-
asinl a smoaIbiq repalion fUnction. e.1- a pol)'llDlllial. sicIa asilll GUM. App/iIIJSlflIlJIlcl36. 332-9. Mala, Do O. 1993:
Iiac:IionaI polynomial or lencrarllCd addilhc (cubic spline) CalSlructiaaafap«laIaIicLaxeccntilcs......... aidulJs.
cunc.lfthe SDchanps witb . . u is often the case. a c~ StlllirlnillM«Iit.W 12. 917-24. Cole, T.J.... Ona, •• J.I992:
of Ihc: ap-related SO also needs to be estimated by the 5'nIDaIhiac n:fCJCIII!C ccnlilc CIII\'CS: ~ LMS IiIdhoIIIIIII peaaIizcd
n:p:ssiaa methods or Aitkin (1981) or Allman (1993) and IibIihDocl. Sltllistin iIr IIft1lrine II, 1~19. KoII*tr, R. W....
the Ble-reIaIccI me.. obtained usiq wciJhtcd linear relR:S- D'O"'t V.I987: ~"",,""'.ApJIIWSttIt&lirl
sian willa weights com:spollding to the imene squan: or the 36, 383-93• . . . . . , P..... WrlPt, JL M. 1998: A IiICIhod far
Ble-n:lated SD. 1'bc age-relatccl Jd'en:nc:c mqc is apia cstimd.lIMPCCific Id'CftIICe illlCnlls ('nonDaI raps') based an
COIISII'IICIaI araund the repssioD CIII\'C using abe SD curve. fraI:tiaaaI paIynamiIk and apanealiallIIIIsfarmIIIia JourtIIIIof"
When the cia.. are skew it may be possible toacljUSl for the
IlDJYII SIIIIiMiaII SDci6y sma A, un. 79-101.
skewness ..... a single. col. logarithmic. 1ransf0000000on at
all asCI. However. often the cIepec or skewness is iiself' age .......,aeHIc rates These an: rates calculated within
~ althouP this needs a large sample to show iL .. this a Dumber of Kialively Dl1IIOW' ap bands. A ClUcIe rate is the:
_wness
case an 1IIe-n:Jated sWlI1IIIII)' slillistic for the has to Dumber or ewmls occurriq in a papulation duriDg a spec-
be estilDlllecl, alGIII with the ....relalal mean and SO. The iftecllime period divided by an eSlilllllac of' Ihc: size of the:
I
AGRESMENT ________________________________________________________________

populatiOD. However. when comparinl rates belween 00la: of NaIi..aI Statistics (ONS) in Ensland and Wales
populations willa diffen:nt !lie dislributions. it is necessary (ONS. 2002). Ase-spc:cific disease incideace rates are also
to consider IDles at specific ages separately. published in wrious cauntries, mosI DOIably cllllCer inci-
In the table:.. clealh rales an: pn:sc:naed for COIla Rica and dence. for which inlclnalional data an: compiled by Ihe
Ibe UDib:d Kinplana for 1999. derived fna dala flUID the Inlemalionai Asency for Research OD Canter (Parkin
United Nations (2002). The final colunm lives the SIC- el QI•• 2(03). Age-spccilic prevalence ndes for exposures
specific rales far bmad !lie bands and the crude (lotal) rate. such as smokinl can also be derived. but ~ mon: usually
The QC-specific rate is calculated as the number ofdeaths in obtained from specific surveys such as Ihe Oenn Hause:-
Ibe palticular DIe lroup.ln Costa Rica. thc death rate aI ages hold Survey (Walker etlll•• 2(01). HI
0-5 is calculated as 129611 070000. 'I1Ie rate isexpmlSCd per (See also CAUSE-SPEC'mC DEA1H RA1E.. STANDARDISED . . . .AIJJ'Y
1000 penons so the nile is multiplied by 1000 to live the rate RATIOI
of 1.2 pel' 1000 in the final column of the table.
0.... far N...... staIIItIcI 2002: MorlDli" "D'islier mII.ff'.
R~,~,,· oflM bgalr. G~I'tIIIIII_11u ilyeDuse, .x_IIg~. in
.....pecHic rates Population, number of deaths and ErtglDRJ _ WDks. 200/. I.oadoa: Ofticc for NIIIioDaI Statistics•
deaIh I8Ies from all causes lor Costa RIca and tire PmtdD, M., WIIeIaII, So, FIrIa)", J., TIppO, L -1'IIaIaaI, D. B.
2003: Olmw iItritImt:r iIr ftl'e ftHIliIIt"ls. ~l. VOl. Lyan: IARC
United Kingdom for the YfHU 1999
Scientific Publicldioas. Valid Na_ 2002: 2000 tltmogrtlphic
CD~III RiclI )WlrbooIc. New VOlt: UnilCd NIIIions. WaI... A., ........ J.,
CaaBard, .... GaddanI, Eo ... n-. M. 2001: liI'ing in
BrilDiR: re.ll~from lilt 200D c;,lItrtll HoaseIroIti SlInyY. Landan:
A,e PtlplllllliDIr '96 in.II,e DmIJu DeIIIIr
11Ie Stationery Offtce..
11'0"11 (IOOOOD.s) gl'tlllp Nle//ODD
0-15 10.7 32'1\ 1296 1.2 8gl'88lllani Apeement in repeated assessments is a
15-49 17.4 52~ 2766 1.6 funclameatal crilcrion for quality of usessmenlS on mtins
50-69 3.9 12 '1\ 3447 8.8 scales. The use or ratinl seales and ather kinds of anIercd
70+ 1.3 4., 7523 56.6 cllllSi6calions of complex qualitative variables is inlel'disci-
Tala) 33.4 15032 4.' plinuy and unlimilc:cl. RatiRl scale assessments produce
Uniled Kirrgtlom tlrr/ina/.'11. the anIemI caleCOrics JqHaentinl only a rank
onIc:r of thc intensily of a particular variable and IIDI a
Age PtlpIIlllliorr '96 in lI,e DmIJu DeIIIIr numerical value in a IDIIthemalicai sense. althqh the use
,,.D"II (IOOOOD.s) gRlllp Nle//ODD of numerical labelliag could give a false impression of
0-15 113.9 19 'I. S8SO 0.'
quantilalive data. (sec RANK INVARWICE). The main qualilY
15-49 288.0 48~ 31228 1.1 concepts of scale assessments are reliabilily and "alidiIY.
50-69 126.1 21 '1\ 120759 9.6 Reliability (sec ME.o\SUItEMENT PIECISION AND REl.lABlury)
70+ 67.0 11'1\ 474225 70.1 refers 10 the extenlto which repelllecl measun:ments of dae
TaIaI 59'~0 632062 10.6 same: abject yield thc same raula. which means IIIR'Cment
in n:peated usessments of various designs. In intell1lk:r
n:liabilily (see MEASUJlfJ.I!NI' PRECISION AND REUAllurY)
The: crude: (talal) rate forCosta Rica is less than halflhat for studies an: madeofthc level oragn:emenl belween obscm:rs
Ibe UK. However. at DO age is the nile in the UK double that duat classify the SIIIIIC object or individual. and inlnlraler
for Costa Rica and far some qe groups the rate is higher in reliabilily (see INrRACUSS CORREl.ATION COEFFICIENT) slUdies
Costa Rica than in the UK. Note that the pe~nlQc:s of the refer to ap:cment in lesl-relesl scale usc:ssmc:nts by dae
populalion in each age group (third column) diller markedly. same: nIIcr.
The UK papulati.. is much older (II .., of thc population the f""lucDC')' disbibution of pain of ordinal data is
~ oyer 70 compan:d with 4'1\ in Costa Rica). The different described in a square COJrnHOEI'D' TABLE (sec the figure wilh
. . st~ explaias the misleadins mmparison between pans I, II and III on page II). and in the cue of continuous
the crude rateL usessmenls on a visual analogue scale. VAS. by a scllllcr
Agc-spec:iftc rates are cwnbenomc: to mml'lR across a plot. 'I1Ie percentage apecment (PA) is a basic agn:emc:at
number of populalions. Standardisation mc:lhacIs are often meas~. When theqreemena is unsatisfactar)' small RUDDS
used to provide an qe-adjusted sunurtal)' nile far each f.disqn:ement ca bce'Vllluatc:cl by a &IaIisticai meIhod lhal
population. labs Kalant of the rank-invariant propc:dic:s or antinal clata
Many countries publish qe-specific ntc:s for all cause and and thai makes it possiblc 10 identify and measure systelDlllic
specific causes of death. c.l. tile annual publications of the disasn:ement. when 1B1CDt. separately from disagn:ement
________________________________________________________________ AGREEMENT

caused by individual variability in assessments. SySlemalic Systematic diSllJl'CCmcDl is evident by the lIIBIIinai
disa&Rcment is population based and maIs a sySlemalic helUogencity. and by pairilll otT abc two sets or lIIBIIinai
duuage in conditions or memory bias bcIwccn 1CSl-n:1CSt rrequencies. the so-called nnIt-lnInsrannabie pattem of
assc:ssmenlS. or bclWccn ndcI5 who interpn:t the seale caI- &p:ClDent (RTPA) is cODstnlclcd. The RTPA clc:scribcs the
elDries ditTeratly.l..arJe individual wriabilily. on the aIhcr expected paltera in abc case of systematic disqn:emc:nt
haad. is a sip orpoorqualily ora ralilll scale as it allows for only. All pain or observations or the RTPA will have tile
uaccrlainty in ~1Ilcd assessmc:atl. TIle pracnc:e or sys- same rank onicrilll in lhe two assessments provided
tematic disagn:c:mcnt in Ihe use of the scale calCpric:s dial the ranks am lied to the cells. which is the clcfinilion
between the two usc:ssmcnts is ~ed by ditTcrcat fRe or the augmcnled ranking pnJCCdun: (aug-nmb)
quency dislribulions. which means IlUll'linal distributions (sec IAXKINO).
(paris I and II). A systematic disagr1:clDenl n:pnIiac the Part U in the figun: is the RTPA orthe pallem in part I. The
categorical levels and in the way or concc:allatin& the observed distribution or pairs in part J deviates from this
asseamenlS on Ihe categories ~ mellllUml by the n:lali~ RTPA. which means that some or the pain or aug-ranks
position (RP) ..... the n:lali~ conccnlralion (RC) n:spcc- given to Ihe observations diO'er. The relalive rank yariance
tively. The RP expn:sscs the extent to which the IIIBIJinal (RY) is a rank-based IDCDSUK or this observed individual
distribution of assessments Y is shined lOWanIs hillier variability. i.e. unexplained by the measures or SyslCmalie
categories than the lIIIIIIinai distribution or X. nather ..... diS81n:cmcnl:
the opposite. A lheon:ticaJ clcscripti_ is the: diR'c:n:nc:e
bclwecn the probabiUtics P(X < JI) - pcY < Xl. Possible
values or RP ranie from (-1) 10 1. and RP is positive
whcll higher seale calcJories ~ more frequently used in
the assessments Y than in X when CXIIIIp8I'Cd with the
opposite. Com:spondingly. Ihe RC expn:sscs Ihe extent to when: n is the number or paiRd assessments and (.di,)1 is
whieh the IIIBIJinai distribution or Y asscssmen15 is II'ICR Ihe squsn: or the mean 8U&-lDllk diR'en:ac:e of the Qlh cell.
COIICCnlndcd to central seale catcgoric:s than is the rnaqinal ad the summation is made over all cells ij or the m x n,
distribution or X. thc:oIclicaily clcscribecl by Ihe diR'c:n:acc square table, 0 ~ RV ~ I (Swnsson Itl DI., 1996; Svensson,
in probabilities P(X,< r,,<x,)-P(Y,<X,, < lj). Possible 1998&). TIle Cohen's coelltcienl kappa (,,) is a commonly
values mnge from (-1) to 1. and a posilive RC indicates used mcasun: or 8IrecmeDt adjuslccl ror lIIe chance
thai Ihe assessmcnlS r an: man: CXIIICCntratcd than x. Zao expected qn:c:mc:nl (sec KAIIPA AND WEIOII1'ED KAPM).
or wry small values or both RP and RC mean lhat the the calculations of Cronbach's alra and other so-called
systematic part of an observed clisqrccmenl .-ireel assess- n:liability CXlCl1icicn15 an: baed on the lIlISumplion or quan-
meats is ncglipblc. titative, normally clisbibutcd daIa. whieh is not achievable in

I RatcrX II RalcrX OJ RatcrX

A B C D tat A B C D tat A B C D tat
RaIcr D I I 2 IJ 'Z Z D I 4 14 .9
r
c 2 '·Z. 14 18 C t 17' II C 1 2 10 4 17

B 1 1 II 3 16 B 16" 16 B 1 6 3 1 II
A 2 8 3 1 J4 A 3 11 14 A 1 2 3
tal 3 II 17 19 !50 tal 3 II 17 19 50 tal 3 II 17 19 !50

PA.12'1 PA.12" PA,62~

RP. -0.49 RC.O.l6 RP,-G.49 R~0.16 RP=RC.O
RV,o.OB RV.O RV,o.OS
agraement EJtIJmpIes of psifed ordinal data from Intemtter assessments on 8 four-poinl SCIIle with the ORIered
calegodesl8belledA < B< C < D. 7beranlc-transfotmablepaltem oIlJ111'f18f11f11(RTPA) Is sIuIded. Themeasul8SoI
percsnlllge agreement (PA), lhelflllltive position (RP), Iherel8t1ve concenItaIion (RC) 8IId the 18Ia1lve mnIc VBIfance
(RV) 818 given

11
AKAIKFS INFORMATION CRITERION _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ __

data rrom nding scales. There is also a widespread misuse IKJIU1CSledmodels when likelihood ratio tc:sts cannoI be
or the correlation coefficient as a reliability measure. The applied.
eorrclalion coemcienl (see CORRELATION) mcasun:s Let z denote the data and' the com:spondi~ maximum
the degree of associalion between two variables and does likelihood estimates (MLEs) or the pararnclcl'5. Then. the
nOi measure the level of agn:ement. In part I of the figure AIC for a given model is denoted by:
the PA is 12~. and the observed disagreemeat is mainly
explained by a systematic disagreement in position. The AlC = -2 log L(i; x) + 2p
negative RP value (-0.49) and the RTPA (parlll) shows where p denotes the number of parameters in the given
that the assessmenls r systematically used lower catqo- model being filled to the data and log L(i; z) the corre-
ries than X. A slight additional individual variability. sponding log-likelihood evaluated at the MLEs of the
RV =0.08 is observed. SPEARMAN'S RANK CORRELAll0N parameters. The AIC statistic is calculated ror each
COEfFICIENT. r~. is 0.66 in part I or the figure and 0.97 in possible model being considcrc:d. The model deemed
part II. ignoring the ract that the assessments an: system- optimal is the one with the smallest AIC value. i.e. a
atically biased and unreliable. The same holds for the model with a relatively small number or panunelers that
eoemcienl kappa (-0.14). In pari III the marginal homo- adequately fits tbe data. The AIC is generally easy to
geneity and the zero RP and RC values confirm that calculate given the maximum or the likelihood function
the disagreement (39~) is entirely explained by slight and is vcr)' versatile. allOWing us to compa~. for exam-
individual dispersion (RV =0.05) from the RTPA. which ple. nonnesleci models. We note that c:orrec:tions have
is the main diagonal in this case. The ra is 0.61 and the It been suggestc:d to the AIC statislic to allow for data with
is O.4S. ovenlispersion (denoted by QAIC) and small sample sizes
Besides reliability studies. the level of disagreement is or (AIC,.). See. for example. Burnham and Anderson (2002),
main inteRst in paired asscssmenls ·berore and after' Sections 2.4-5.
lrealmeat for analysi~ change in outcome or treatment 1be AIC statistic has also been used to eompare Ihe
effect. In this application or the disagn:ement measures, performance of difl"cmlt models. relative to each othu
nonzero RP and RC values indicate the level or eommon (Buckland. Burnham and Augustin, 1997: Burnham and
group change in outeomes. and the hcterogcaeity in Anderson. 2002. Section 2.6). It is not the absolute values
changes among the individuals is measun:d by the RV of the Ale statistics thal are important but their relative
(Svensson. 1998b). ES values, in parliculartheir difference. Foreach model the tcnn
Swauoa, E.. 1998&: AppIicatiCIII of a llIIIk-invariant medaod to .dAIC = AIC - min AlC is calculated. where min AlC is the
C\'IIIuale n:liability of ontcmI catepD:al asscs:smcnlS.. Jorulftll of value of the AlC slalistic ror the model deemed optimal.
Epidemiology I11III Bioslalillirs 3. 403-9. SYIIISIOII, It. I998b: Clearly. AAIC = 0 ror the model deemed optimal: the largCl"
OrdiDDl invariant measures for indiridual and group cbanp:s the value or .dAiC the poorer the model. The relati\'e
in ordmd categorical data. SlaliNlits in Meditbre 17, 2923-36- penalised likelihood weights n', can also be calculated ror
S........, ...... sr.m.tc, J ..E., BboIm, S.. YOB . . . . . c. aad each model i = 1•. , '. m. where:
Me....... A. 1996: Analysis of'inter-Clbsen'Udisagrmnc:n1 in die
asscssment of subarachnoid blood . . acute hydrocephalus on cr
exp( -.dAlC;/2)
scans. Nftlrological Remuth 18. 487-94. W; = -=",r~----';':"""";'-

Eexp(-AAlCj/2}
AIcaIke'slnformaUon criterion Akaike's infonna- j=1
lion criterion (AIC) is an index used to discriminale between
compcti~ models. It is widely used when then: is the issue or and AICI denotes the carrcsponding AlC value associated
model choice where we wish to find the most parsimaaious with model i. The weights provide a scale to interpn:t Ihe
model (see Akaikc. 1974). Often there may be a number or difference in values ror the models. Finally. these model
possible models that can be Hlted 10 the cIaIa. from which weights can be used to obtain a (weighted) model-averaged
parameters can be estimated using. rorexample. the MAXIMUM estimate or parameters of interesL RK
UKaJIIDOD ESllMATION. Generally. complex models an: mo~ [See also DEVIANCE. UKflJIIOOD RATIo)
ftexible. but contain a n:lalively large number of paramclers.
Ablb, H. 1974: A DCW look atlhc Slatistical model identifica-
whereas simpler models with rewer parameters may com-
tion. IEEE TrtmJGt:lioRs on Alliomalit COIItrol AC 19. 716-72,
promise the fit or the model to the data. Eueatially. the AIC
Backlaad. S. T., BIIlIIIuua. K. P. ad Aapstta, N. H. 1997:
statistic compan:s competing models by CXJllsideri~ the Modcl selection: an integral part of inrc~nce. Biomelrirs 53.
trade-off between the complexity of Ihe model and the 603-11. Bu........... K. P. &ad Andenoa, D. R. 2002: Model
carrcspanding fit of the model to the clata. The AlC stalistic sc/~clion aM mullimotlel i"J~rDlte. 2nd editiCIII. Heidelbell:
is widely used. particularly as it can be used to compare a'en Springer Verlag.
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ ALL SUBSETS REGRESSION

allelic association 1'bis is an association between .Classical epidemioloJical designs (CASE-coNI'IDL SnJDl!S.
two alleles (at two dilTermtlaci). ar betwcea an allele and COHORI' STUDIES. CROSS-5EC11ClNA1. SnJDlES) are mldily appli-
a phenotypic bait. in Ihc population. Since humans IR cable to lhe study or diseasc>oallele associations. as IR the
diploid a morelc:dmical definition of the ranner is nca:ssary: Slatistieal methods developed ror Ihcse designs (e.l. LOOIS11C
two alleles are associated iflhcir frequency orCO-OCCUJ'RDC."C REORESSION. SURVIVAL ANALym). These designs are polCDliaily
in the same: haplolype (i.e. Ihe genetic maIeriaI transmitted susceptible to Ihe problem ofhidclen population stratiftcation.
from one pan:nt) is gmder Ihan the product of the l1III'Iinal which can lead 10 spurious associations or mask true a~
frequencies of the lwo alleles. cialions. Family-based association designs an: robuSl to
Association belween two alleles is also known as lilrJcoge population stnliftcaliaa and usually consist of the use of
dlseqrtili"'iuIII. 1bc reason is thai. in a lillie papulation under eilller parenlal ar sibling controls. Melhads forlhe analysis of
mncIom mating.1hc extent of association between two alleles matched samples. such as the McNBIAR·S 11!ST (also called
(as me&lllRd by the difference between Ihc rraaucncy of the transmission disequilibrium lesl in the contexl of'pamatal
the haplotype containinr: Ihc two alleles and the product or conlmls) and CONDITIONAL LOOJS11C RBJUSSIDN an: applicable
the frequencies of'the two alleles) clc:c:reases by a radar equal 10 these dc:sips.
to one minus the n:combinalion fraction (see OENETIC UNIt. The study of diseasD-8llele associations is a complemcn-
AGE) between the two loci. per leneration. "Ibus allelic IaIy slndegy 10 link. analysis. in the localisation and
association n:pn:sents a slate of disequilibrium that k:ads to identification or genes thDl incn:ase the risk or disease. In
dissipate at a nile determined by the sln:nlth or linkage general. allelic association is unlikely to be detected when Ihc
belwcen the lwo alleles. lOW'" the stale of equilibrimn
whea the: frequency oflhe haplotype is equal to the produc:cor
marker locus is quite far (>1 mclabase) from the disease
locus. but can be much mon: powelf'ul than linkage whea Ihe
the hqucncies of the two constituenl alleles. marker locus is close enouP to the disease locus to be in
AssociDlions between lWO alleles can arise in a population substantial IiDkqe disequililxium with it. particularly when
for a number of R:8SDRs. The mutation that pve rise to the the effecl size of the disease locus is small. For this reason.
more rc:ccnt allele may ha'Ve occuned on a chromosome allelic assaciation is particularly appealiq for sc:archillJ
tlud happened to contain the other allele. Random genetic n:gions that demonstrate linkage to Ihc disease or to Ihe
drift duriq a population botdeneck may have led to the inveSlilation of specific candidate aenes. However. techno-
oVcm:pR:SCntalion or some haplotypes. The mixiq of two Iopcal developments have enabled lhe efficient lenolypiq
populations with differenl allele frequencies may havc of up 10 I million SNPs in a single 8II'8y. and this has led
resulted in associations between alleles in the ovmall to association sludies on the whole-genome scale (called
population. When. for any of' these ~asons. such allelic gellOl11C-widc assoc:iation studies. or GWAS) that ha'Ve cov-
associations arose many lenemtions BlO. OIIly those occur- erage of over 90 fJt of common variants (allele frequency
ring between lightly linked loci are likely to havc penisted > S CJ.) in the lenome. PS
10 the Clll'l'CDt ICDeraliaa. We would thcrefce expect an
impelf'ed inverse rdationship belween Ihc extent of asso-
cialion between two alleles and the distance belwecn 1hc1D. all ..bsets regresSion A form of n:grasion in
An assocwion between an allele and a disease may be lhe which ull possible models an: compared usillJ some appro-
result Oradi~cl causal Rlationship.ln other words. the allele priale criterion for indicatinllhe "best' models. If there an: p
is a causal varianllhal is fuactional and increases Ihc risk of explanatory variables in the data. there are a total of r- I
the disease. However. il could also be indin:cl. with the allele possible reIRssioD models because each explanatory vari-
beiDg in linkage disequilibrium with a causal variant. The able can be in oroul ofthe madel and the model containiq no
pI1:lICIICe of' link. disequilibrium belween tightly linked explanatory variables is excluded. One possible criterion rar
loci means that it is possible to seRen a chromosomal n:gion comparinl models is the MAu..aNs C,. STA11STIC' and 10 iIIur
for a causal varianl without eumininJ all the alleles. only lI1Ile ils use we will apply it to data that arise from a study of
a sufticienl number to ensure Ihat any causal variant in the 2S patients with cystic ftbrosis Rported in O'Neill el til.
~gion is likely 10 be in linkage disequililximn with one or (1983). and also gi\'en in Altman (1991). Data for Ihc first
mon: of tile alleles examined. The poIymorphisms chasen 10 tine patienls an: Jiven in the lint table. The dependenl
JqJn:5Cnl itself and associated polymorphisms in ils vicinily variable in this case is a mcasu~ or malnutrition (PE_).
in an association study are called TAG polymorphi&ms. The Some of Ihe models consideml in the all subsets n:gression
International HapMap Projecl (www.hapmap.DII) has char- of Ihcse data are shown in the second lable. Iogcther with
acterised the pauem of allelic associations among over their assaciated C,. wlues. where p n:fers to the number of
3 million sinlle nucleolide poJymorphisms (SNPs) in the paramelcls in a particular model. Le. a model that includes
human lenome in line major populDlions (Europeans. a subsca of p - I of the explanalaly variables plus an inter-
Arricans and Asians). cepl. If Cp is plouecl against p. Ihc subsets of explanaIDry

13
ALTERNA~HYPOTH8S~ _______________________________________________________

all .uba. regression Cystic flJrosis data; fbI thl8fJ subjecIs

Sub Se.T. Height Weight BMP FIN RV FRC TLC

I 7 o 109 13.1 68 32 2S1 113 137 95

2 7 1 112 12.9 6S 19 449 24S 134 85
3 8 o 124 14.1 6S 22 441 268 147 100
Sub: subjcc:t number
Sex: O=maIe. 1=female
BMP: bady mass (ftiJhIillc~) as • pcrcen1age or Ihc agwpcc:ill: median in IICInII8l indiriduals
FEY: fon:ed cxpiratary volume in CIIIC secand
RV: raiduaJ wIume
FRC: filDclional residual capacity
TLC: lotaIlunc capacity
Pf..u: maximal statistic clpinllOl)' pn:SS1R (cmH:O)

alternative hypothesis See IIYIVl1IESIS nsrs

all ....... regie_on Some of ths models fitted in
~ths~ _ _~~m~~~.~ AMOS See STRurnJRAL EQUATION MDOEWND sm:rwARE
data (size is one mote than ths numberofvllliables in.
model, 1o Include the Intercept) a..lysls of covariance (ANCOYA, ANOCOVA)
This is an extensiao of the analysis or variance (ANOVA)
Model Size Tnnu Cp
that incorporales a eantinuous cxplanalOry YDriable. When:
7 2 Sex 17.24 ANOVA aims to ddc:ct if then: is a change in the mean value
14 3 Sex, weight 4.63 or a wriable across two ar IIICR puups, ANCOVA (or rarely
21 4 Age. FEV. RV 2.62 ANOCOVA) docs the same but adjusts for a mntinuous
~ 4 Age. BMP. FEV 4.5 covariate.
3S 5 Sex, weight, BMP. FEV 2.95 Most commonly this cowrialc will be a baseline ~~
42 6 Ale. wei&ht. BMP. FEV, RV 2.8 menl, a1lowinl the analysis 10 adjust ror initial variation
49 6 Age, sex. height. FEV. TLC 6.99 between participants and isolate the etrects due to the lRat-
56- 7 Age. sex. height. FEV. RV, n..c 7.06 ment factor. However. sometimes a dift'en:nt covariate is
63 I Sex, weight. BM.,. FEV~ RV. 6.49 used. Far example. Xarhune el QI. (1994) consider the
FRC. TLC association' bc:twccn alcohol intake (divided into rour cllle--
70 9 Age, height. weight. BMP. FEV. 8.06 garies) and numbers of Purtinje cells. In doing so they
RV.FRC, TLC introduce DIe as a continuous covariate in order to 'control'
77 9 Ale, sex. height. BMP. FEV. 10.29 or 'adjust' for the effects of age on cell numbers.
RV.FRC,TLC Under other cirannstanc:es the aulhan 'could have been
inten:stcd in the etrects of age and wantinl to adjust far
• Models close to abc line e,. =p.
alcohol intake. Despite being the same analysis computa-
tionally. this is not typically what is lhought or as analysis of
variables mast wonh cODsidcrini in trying to ftnd a parsi- covariance and might mo~ commonly be IR5Cnled as a
monious model am Ihase Iyilll close 10 the line C,,=p. '~grasion' . Indeed the various analysis of wriance methods
All subsets IqIasion has been foUDd to be particularly can all be Yiewed from within a repession fnIInewoItt. which
userul in applicaliOM or 'COX"s REGIlBSION MODEL (sec demonstrates that ANCOVA can be extended to cope with
Kuk. 1914). SSE much m~ than ODe continuous mvariate.
(See also MUL11FLE LINEAR RBJRESSION] Malhematically. ANCOVA follows a similar path to thal
far ANOVA and the output is usually summarised in a similar
AltmaD , D. O. 1991: PrortimJ stillwiea for rrwdiml remut:b.
table. aJlhoUlh the details may vary.
London: CRCJChapman & Hall. Kale, A. Y. c. 1984: All subsets
~ion in a pruportioDaI hazuds model. BioIrwtrim. 71~.587-92.
The promised beneftls of the analysis of eovariance an:
O'NtIU. s., ....." r., PasCerIcaInp, H. aad Tal, A. 1983: The clear. If one has an unbalanced obserwtional study. then
dfects of chronic bypmunclion. nutriliaaaJ staha and postuK on ANCOVA can adjust for dift"en:nces in baseline 'Values and
rapiratar)' muscle sbmIIh in cystic fibrosis. AlfWriftlll Rft'ieM' of ~movc a potential bias from the ~sults. By the same token, if
Respiratory DiJortlers 128. 1051-4. one has a randomiscd biaI thai is naturally balanced. then
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ ANALYSISOFVARIANCE (ANOVA)

ANCOVA mluces the amount of unexplained variation in the the populations have the same mean). If the two estimalCs
data and Ibus increases the power of the test. of the variance an: different. then Ibis is evidence Ibat our
However. ANCOVA can only be employed if the appro- assumption of equality failed and. therefore.lbat the popula-
priate assumptions are met. 'lbcse include those of ANOVA tions do not all have the same mean.
(i.e. normality of n:siduals. holllOSC."Cciaslicity) as well as the Note that the variance of a single sample is eslimated as Ihe
appropriateness of the ANCOVA model. Is the n:lationship sum of squared ditTen:nces from the mean divided by the
with the covariate IrUly linear? Docs the etTc:ct of the COWl'- sample size minus one. 1be sum of squared differences tenD
iate vary between groups? Failing to meet these assumptions is interpretable as a measure of the total variation in the
can lead to the introduction of important but subtle biases. It sample. In the analysis of variance. by combining all groups
is a frequent c:onccm that medical n:scarch papers report together. one can calculate this measure for all the data. This
a covariate as having been "conlrOlled' or 'adjusacd' for, with is termed the 'total sum of squares' or 'total SS·.
no evidence that the conlrOl 01' adjuslment was appropriate. Variation in the data is either- "between' or 'within' the
For further- details sec Allman (1991), Owen and Froman samples. The "wilhin poups sum of squares' or 'within sse
( 1998). Miller and Chapman (2001) and Vickers and Allman can be calculatc:d as the sum of squared difren:nces rrom the
(2001). AGL individual sample means (mther Ihan the ditTen:aces from
(See also GIlNER.wSED LINEAR Mooa.l the overall mean Ihat produced the tcMal SS). 'Between
groups sum of squares' or "between SS' can be calculatc:d
AItawI. D.O. 1991: Pmt:tiCill statistic's lor mttliml researm.
London: ChIpman &: Hall. KariIuDe, P. J., ErldDJutti, T. &ad directly. but is most easily calculated by subtraction of the
LalJIPaIat P. 1994: Moderate alcohol ccaumplion aod loa of' within SS (rom the total SS.
cmbcllar Prikinjc cells. British M«Iim/ Joumal 301. 1663-7. 11ae two estimates of the variance (or "mean square' as it is
MOler, O. A. and CIIa......,J. P. 2001: Misunderstanding anal)'lls often termed in this context) can then be calculated. 1bc:
of covariaDc:e. Journal ofAbnormtJI PS)~lIoIogy 110.~. OweD, between groups mean square is equal to the between SS
S. V. aDd Ji'I'oawa, R. D. 1998: Uses and abuses oflhc analysis or di\'ided by the number of groups minus one. 11ae within
CO\'lIriancc. Rt.seorm in NlII'sing and Hm/11r 21, 557-62. Vkken, groups mean square is equal to the wilhin SS divided by the
A. J. aDd Altman. D. O. 2001: AnaI)'s~ coatrolled trials with number of observations minus the number of groups.
basdiae and follcM'-up measun:mcnts. Britislr Medical JDllT1ItIl323. An F-slatistic is then calculated as the betwe:en
123-4.
groups variance divided by the within groups variance.
Under the assumptions of normality and homoscedasti-
analysis of variance (ANOVA) Often referring to city (common vananee) Ibis statistic will be an
the one-way analysis of "'ariancc. it is a test for a common obsen'ation from an F-DISTRIB~ ir Ibe groups come
!.lEAN in multiple groups that we describe in detail here. from populations with a common mean. The DEOREES OF
Analysis of variance frequently arises in the comparison of FREEDOM of the F -distribution are the number of groups
more c:omplicaled models, but the same logical argwnenlS minus one and the number of observations minus the
apply. In all eases. the undcrIyiRl concept is to partition number of groups.
the observed wriance into quanlities attributable to specific From the F-disbibulion. we: can calculate: the probability
explanatory soun:cs. and then consider important those of observiRl such an extreme value of the F-statistic if the
SOUK'es thai explain 'more than their rair share' of populations have a common mean. This is a one-tailed test. If
the variance. the value is unusually small. this suggests the between groups
Despite the confusion sometimes caused by the name. the variance is unusually small and so is not evidence of variation
one-way analysis of wriance is a method for testing to sec between the groups. Therefore. the test is to find the prob-
whether multiple samples come from populations that share ability. ir the populations do have a common mean, or
the same mean.. In this mspcct it can be viewed as an observing a value greater than that observed.
exlcnsion to the '-test, which assesses whether samples rrom A natural way ofpn:scnting ANOVA is the ANOVA table.
two populations share a common mean. An analysis of Given Nobsc:rvalions that fall into k groups. it is necessary to
variance performed on two samples is equivalent to perform- calculate the total SS and the within SS as described earlier
ing a l-lest. and then the analysis can be completed as presented in the
ANOVA assumes that all the samples come from popula- first table. Murphy el aI. (1994) conducted an analysis of
tions with a NORMAL DI5TRIBU11O.~ that share the same variance to see if milk consumption before the age or 25
VARIANCE. It can be viewed in a number of ways. but essen- affects bone density of the hip in later life. A total of 248
tially allDpare5 the estimate of the variance obtained within women palticipalc:d in this part of their study (N =248) and
samples (that makes no assumption that the populations have wen: divided intogroupslbat reprcscntlow. mediwn and high
a common mean) with an estimate of the variance rrom the milk consumptions (k =3). The samples had similar var-
sample means (which will requin: the assumption that iances aDd so atlcast one of the assumptions for ANOVA was
15
ANALYSIS OF VARIANCE (ANOVA) _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ __

analysis of variance (ANOVA) 7he anaJys;s of valiance table

Soune of Degree:lof SIIIfI:I of Mf!tJR aqlltlTe:l F P

l'Qrilllrce freedo", :lqutlnS
8etwa:a graups It-I Between SS= Between MS = Between SS/(It - I) BetweenMS p
Total SS - Within SS WilhinMS
Within groups N-k Within SS Within MS = Within SSI(N - It)
TaIaI N-I Total SS

analysis of variance (ANOVA) Approximate reconstruction of the analysis of variance table from MulPhyet a/.
(1994).
Source of var;lIIIce Degree:l offreedom Sunu of :lquares F p

8etwa:a graups 2 0.15 0.01 3.8 0.23

Within groups 245 4A 0.02
Talal 247 4.6
(Entries in bold weft: infemd from the paper. lite rat simply follow from the cU:ulalioas)

salisftai. As is common for reasons of space. the ANOVA If the assumptions do not hold. thea "I1W5RItMA'I1DN of abe
table was not preseated in the published paper. just the P- daIa mighl comx:t dU. Otherwise a number of nan)Jlll'1llDClric
value. but enough dala wm: pn:scnted for an approximate altemativc:s to ANOVA exist. the most commonly used beins
R:Consb'UctiOll. the KlUsKAL-W~ 1BI' and the FREDMAN 1I5SJ'.
We can infel' btllle within SS is approximately 4.4 and 1bc one-way analysis of variance is appropriate when our
the between SS is approximately O.IS. nus leads 10 an F- data an: simply divided into a number of groups. 'Ibm: an:
statistic or approximately 4. From the rcpaned P-value many other forms of analysiS of ,·arianc:c. 'nIe TW~WAY
(0.023). it can be calculated from the F-distribution (with ANALYSIS CE VARIANCE should be used when the IJOIIPS are
2 and 245 respcclively for numerator and denominator definc:d by two factors. S~ for example. we had six
degrees of fR:Cdom) thai the F-statistic was 3.1. The aJO- groups: the three groups of women in Murphy el QI. (1994)
clusion then is thatthc:n: is evidence Ihal these samples cIonot and tlee gruups of men at die same levels of milk atnsump-
come from papulations thai share a common mean. n.c lion. Rathel' than a on~way analysis of variance. a two-way
rcconsb'Uctcd table is presealed in the second table (entries analysis oharianee with gender and milk consumption as Ihe
in bold iD this table were infem:d from the papeI'. the rest two factors would be appropriate in this instance.
simply follow from the calculations). Ir the data are multiple observations from the same
It is preferable to conduct an analysis of varillDlX: rather subjects, perhaps measurements· of cholesterol levels O.
than 10 conducl/-tesls betwc:c:n all pairs of groups. ANOVA 7, 14.21 and 21 days after slarting a new diet 011 several
awills problems of multiple testing and Ihus keeps CX1IIlIUI individuals. then a REFEATED ).IEASURES ANALYSIS CE VARIANCE
oflhe SIDNlflCANCE LEVIiL Having amduclc:d an ANOVA and would be appropriate. This is a special case of the two-way
rejected Ihc hypothesis of common means. it may then be ANOVA and can be viewed as an extension of the paired
clcsin:d to lest to see which graups ~ raponsible (although sample I-tell.
a plot of the data might be as infonnalivc). In this case, care If lhere ~ observations of ~ than one characteristic
must be takea to comd for the problems of making 1ftILDPLE from the individuals in several poops. i.e. measures of bulb
~. the diastolic and systolil: blood prasun:. then a multivariate
It is imporlanttotakc: note or the assumptions being made. analysis of variance (MANOVA) can be used. If. however. it
rather than simply ignoring them. ANOVA can be quite is desired to correct for a measw-ed baseline atwriate. such
robust to variatio. from nonnality. but hcterosccdasticity as body mass index. in the analysis. then aD ANALYSIS OF
can be a serious problem. Residual plots can be used to help COYARIANa. (ANt'OVA) may be used.
assess the normality and IOXPLOIS can be used to help All these techniques could be implemented through a
assess the hctcmsc:edaslicity. Passible formallcsts for the regRssion framework. in most cases MUIlIPLE lINEAR REORES-
assumptions ~ the KOLMOOOItO\L-SMIRNOY 1BI' and I..EVENES SION. TheadvanlqeS ofdoing so would be the transition from
1BI' rcspeclivcly. the usc of a IIYIIOI'HESIS TEST to an actual estimate of etrcct
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ ARTIFICIAL INTEWGENCE (AI)

sizes. This approach would also allow IIICR ftcxibility; for diseased subjects by a fair coin toss. Consicicr the example
inslanc:e in Ihc case ofMurpily el 01. ( 1994) we could account discussed in the enlly ror the ROC curve. The points on the
for lhe naluml ordering oflhe 1c~ls of milk consumption that curve IR given in the table.
ANOVA ignores. As a gencnl priDc:iple. eSlimDlian and
modelling IR usually pn:fenm to testing of h)'potheses. area uncler the cwve SUmmary dIIIa ussd in an ROC
For fUJthcr details sec Allman (1991) and Altman and CUIVB
Bland (1996). AGL
Speciftcit)' ()'I) 0 0.56 0.14 0.94 0.98 1.00
.AJtmu. D. G. 1991: ProdimJ Jla,istk, foT medical l'esearcl,. I-Sensitivit)' (I,) 0 0.04 0.12 0.32 0.60 1.00
London: Chapman a Hall. AIIDwI. 0. G. aad BIaDd, J. M.
1996: Slalistic:s nota: oompariag several poups usia: analysis of
variaJtI::e. British MedkolJoumol312. 1472-3. M1II'pIJy. s., Ka.w,
K.-T.. Ma1. H. aDd Compsloa. J.IL 1994: Milk CIODS1UIIpIion and 11Ic data pn::scnted ~ult in an AUe as follows:
bone mineral densil)' in middle ~ and elderty WOIDeD. BrilM
Met/iral JOIITtIa/ 308, 939-41. AUe - 0.5(0.04-0) x (0+0.56) + (0.12-0.04)
x(0.S6 + 0.84) + (0.32-0.12) X (0.84
area uncler the curve (AUC) This is a simple and +0.94) + (0.60-0.32) x (0.94 + 0.98)
useful mediad of obiainiDg a summary measu~ from plotted
+( 1.00-0.60) x (0.98 + 1.00») - 0.91
cIaIa. Medical research is frequently concerned with serial
data. as in repeated measurcments (sec REPEATm MEASURB AD area or 0.9 I indicates the high discriminatory ability or
ANALYSIS OF VARIANCE) an a subject oyer time. e.g. blood the marker.
aspirin c:onL'lCntralion mcasun:d at various times over a 2- For the ROC curve. estimating the area by the trapezium
hour interval (Matthews el m., IWO). Sa)' we have n mea- rule is equivalent to computing the Wilcoxon or Mann-
sun:ments y, laken at times II (i-I, ..., n). Such data arc Whitney stalislic divided b)' Ihc products of the sample
tRqucntl)' cxhibited by plotting )'1 yersus I, and joining the sizes on the healthy and diseased populations. For smoothed
resulting points by straight-line segments n:sulting in a ROC curves.. allemaliYe eslimates of the AUe arc available
·curve'. The n:sulting an:a under Ihc curve (AUe) is often (faraggi and Reiser. 20(2). The effectiveness of aitemative
used as a single-numb« summary measu~ for Ihc indiYiduai diagnostic marten is usually studied by comparing their
subject. Further analysis of Ihc subjects or comparison or AUes (Wieand el a/., 1989). Adjustments of these ~u ror
groups of subjects is carried out based on the summary covariate infonnalion, selection bias and pooling effects an:
measures. The AUe for the SCI of points (y,. I,) i - I ..... discussed in Ihe ~fe~nClCs gi~n in Ihc entry for the ROC
n is t),picall), calculalcd by the lnpCZium rule: curve. Sc:histerman elol. (200 I) consider com:clions or the
lolli-I
AUe for measurement error. For rurthel'details scc Hanle)'
and McNeil (1982). DFIBR
AUC - 2~(/i+I-li)tl'; + )";..... )
........, D. and RII. ., B. 2002: Eslimalioa of the lR:a under the
'11tc AUe is used as a suml118l)' measu~ in many amIS or ROC CUI\'e., Slalulit'$ in Medicine 21.3093-106. HuIe)"J. A. and
medical ~search. including bioc:quivalClltlC and pharmaco- McNeIl, B. J. 1912: The meaning and use or the IRa under the
n:ceiver opending clwKteristk (ROC) cun~ RlMJiolol)' 143,
kinetics. It pla),s an especially important role in the analysis
29-36. ~........ J. N. s., A...... D. G .. Call1pIIeII. 1\1. J. and
of RECEIYER 0PERA1IN0 CHARAC'JERISIlC (ROC) CURVES. The
ItoJstaa. P. 1990: AIIalysis of suiaI melSU~meDts in medical
area under lhe ROC curve of a diagnostic: marker (1Csl) rescuda.. BrilU/r Medicol JtHll"lJQl 300. 230-5. SdIIIt--. E..
measures the ability of Ibc I1UIdcr 10 discriminate between ........, o.,Reiler, B.andTrmsan.M.200I: Statistical infamce
health)' and diseased subjects. II is the most commonly used for the ma under Ibc ROC CUJVe in die pmlCDCC of random
measure of perfonnance or a lII8.Iler. We use the convention measumaenl CI'RIf'. Am~rit.Ylft Journal of Epitkmiolol)' 154.
thDIlarger marlccr values ~ IIICR indicative ordisease. 11Ien 174-9. WIeaIId, s.,Gal, l\L H., J ..... B... and J..... K. L
if we randomly pick one subject rlOm the health)' population 1989: A family of non-parametric stali.stics for comparinc diagnostic
and one flOm the diseased population we would 'cxpect' that marken with paimI or uapaimt daIa. Biometriko 76. S8S-92.
lhe value of the muter far the healthy subject would be
smaller than thec:om:sponding yalue ror the diseased subject. artificial Intelligence (AI) This branch of computei'
AUC is the probabilit)' that this. in fact, occurs. The larger the scielK'e is devoted 10 the simulDiion of intelligent behaviour
AUC. Ihc bencr Ihc overall discriminator)' accurac)' of the in machines. Traditional focus an:as of AI an: machine
marker. An Bra or I ~presents a perfect test while an IRa or vision. MAOIINE lEARNING. natural-Ianguqe processing
112 rqRSCnts a worthless test having a discriminatory ability. and speech n:cognilion. Historically an interdisciplinary
which is the equiyalent ofdiffen:ntialing between healthy and field. and helK'e characlerised by the pn:scnce of several

17
ASSOCIATION __________________________________________________________________

competing paradigms and approaches. recently AI has summer of 1956, although many key ideas had already bc:cn
staJtcd developing a more unified conccptual rramcwodt. debated befon:, during lhc early years of cybcmc:tics.
based largely on the convergence of statistical and Modem techniques of artificial intelligence include
algorithmic ideas. Baycsian belief networks. part of lhc more general ftcld of
A constant theme of AI Ihroughout its history has been probabilistic graphical models: pattern-recognition algo-
'pauem m:opition', Ihc cruciallask of delCcting "pa1lems' rithms such as SUPPORT \'EtTOR MACHINES. which represenl
(regularities. relations. laws) within daIa. This task has the con\'ergence of ideas from classical stalistics and from
elllCJ'Fli as a roadblock in aUIhc lladilional areas mentioned neural networks analysis; statistical analysis of natural
earlier and hence has aunctcd significant aucntion. Since languagc tcxt and machine vision algorithms: reinforce-
most cum::nt approaches to pallern R:COgnition involve sig- me:ntlearning algCll'ithms which represent a connection with
niftcant usc of statistics. this has bc:come an important ~I in control theory: and many other methods. !VenDS
AI in general. BIIbop. c. 1996: Nellral nelK--orks jor pQllern recogni-
Recently. AI has bc:cn applied to a new series or important lion. Oxford: Oxford University Press. Mllcbell, T. 1995:
problems and this. in tum. has hc:avily aJTc:ck:d general AI Mamine learning. Maidenhead: McGraw-Hili. R...., S.
~. Important applications of modem AI include: and NOI'YIat P. 2002: Artijit'ia/ inlelligen~e: a ",odem
intelligcnt data analysis (sec also DATA MININO IN MEDICINE): approach. 2nd edition. Harlow: Pn:ntice Hall. Sbawe-
information retrieval and filtc:ri~ from the web; bioinfar- Taylor. J. aDd CrlsUanlnl, N. 2004: Kemel methods for
matics; and computational biology. Tmditional application pallerrr allQlysu. Cambridge: Cambnqe Univcnity Press.
areas. by way of contrast. included lhc design of EXPER1'
SYSTDIS for medical or indusbiDI diagnosis. me:thods for
scheduling in logistics and creation ofoIhcr decision-making
association This is the statistical depcndc:noe between
two variables. Measura of association, unlilce descriptive
assistant software. statistics of a single variable. summarise thc extent to which
The imprecise definition of what AI actually is has made it one variable: inm:ascs ar dccrc:ascs in relation to a change
harder in time: to gauge the impact of this n:sc:an:h ftcld on in a sc:cond variablc. The basic graphical analysis of two
everyday applications. A number of widely used computer variables is the SCA1TERPLGr. wbk:h provides evidence of
programs would have met early definitions of artificial association in the shape and direction of the seanc:r or points.
intelligence.. e.g. popular wcb-bascd n:commcndation sys- In the example given here. there appears to be an association
tems or air travel planning advisors. between bady mass index and systolic blood pn:ssure values
Popular techniques for pallc:m n:cognilion such as NEI1RAL in a samplc of a few thousand middle-aged men and women:
NETWORKS, clc:cision ~ and cluSlu analysis (see nUSI"ER
higher values of body mass index lc:ad to be associated
ANALYSIS IN MmICINE) have made lhcir way into the standard
with higher \'alues of systolic blood pn:ssure.. suggesting
toolbox of data analysis and are commonly found in lhc a "positivc' association. A ·neplive' association. in cantrast,
toolbox of any biology lab. Machinc vision methods are would dc:smbe a situation where an increase in one wriable:
routinely used in analysing medical images. as wcll as parts tc:nds to be n:latcd to a dc:cn:asc in lhc second variable.
of systems such as microanay machines far collecti~ gene Various statistical measwa can be lRd to inlCrpn:t Ihc
expression data. Web retrieval and email filtering software degn:e of association.
also incorponde several ideas from natural-language proccs-
si~ and paIteIn m:ognition and Ihc modem sequence
analysis of genomic data heavily relics on techniques orig- Correlatiorr c.oeffi~ienl. This specifically measures lhc
inally developed for spccc:h nxvgnition. Intelligent web degree of lirrear association between lWO quantitative
agents exist to find, assess and rclricve relevant infonnalion variables on a scalc from negative one to positive one. A
for the user and spc:cch-n:cognition systems are routinely value ofzero indicates a total absence orlincar association.
used in automatic pbanc information systems. 111e field of whilc a value of positive or nqativc one indicates a perfecl
artificial intelligc:nee has clearly produced a number of linear relationship. The correlalion coemcient between
pnc:tical applications, but-lhc critics say - these have been body mass index and systolic blood pressure in our
achieved without solVing lhc general problem of building example was 0.25, indicating a positive association that
intelligent madUncs. Maybe for this n:ason, gc:nc:rally the is less than perfectly linear. Howevc:l'. adherencc to a
main suc:ccss story of AI is n:ported to be the defeat of the linear relationship is only one form of association and it
chess world champion Gary Kasparov by an IBM algorithm is easy to imagine other plausiblc patterns of association.
in 1997. such as a parabolic scatter. in which the change in one
The origin of Ihc field or AI is often idcntified with a paper variable: may be perfectly reftected in thc change in
by A. M.1Uring. whichappean:d in 1950 in thejournal Mintl, the second variable, but the correlation coemcient might
and with a workshop held at Danmouth College in the be close to zero.
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ ATTENUATION DUE TO MEASUREMENT ERROR

Regreuion t:oejJicienl. In the cascofsimple lincarreg~ssion. multiple regression models to estimate adjusted regression
then: is a complc:tc c;orrcspondcnce between the (."C)IKlation coefficients and partial com:lation coeflicients by including
c:oeflicient and the regression c:oeflicient for the slope (/J). all relevant variables in the model. However. even aIlc:r
1'11c: regression coefficient. thCRforc. also measmes allowing for such intc:rdependencies. Ihc much strongc:r
association. but its value is interpreted as the magnitude of claim of CAUSALllY between two variables would generally
change in the dependent variable thDl arises. on average. from require cxaminalion of more stringent CriteriL
a unit chang~ in the independent variable. In our example. an Third. an observc:d association may be speciftc 10 Ihc
estimate of fJ = 1.37 indicated Ihat a I kglm2 increase in the chosen range of the variables or to Ihc particular group
body mass index was associated with an average in~ of of subjects studied and any inrerence beyond the range of
1.37 mmHg sylllolic blood pn:ssurc. However. in more the data to hand would require careful consideration of lite
complex regression models, the regression coefficient can method of sample selc:ction. Various forms of selection bias
measure other forms of association beyond linear may limit the ,eneralisabilily of the association. JGW
dependence. For example. either the dependent or
independent variable may be mathematically transfonnc:d. as treated See INTENTJON-TO-TREAT
such as raising to a higher power. taking logarithms, etc., and
the association measun:d by the regression coefficient would
express a nonlinear change in one variable in response to a attenuation due to measurement error This is a
bias reducing Ihc size of a correlation or a regression coef-
change in the second \·ariable.
ftcic:at due to imprecision of data measurc:mcnt. Consider an
Relllli.'e riJlc. In the special case of two binary variables. analytical epidemiological study in which the aim is to
estimate lite CClItJI.EU.TKJN bc:twc:en true average consumption
various nlio measun:s arc often used to quantify the degree
of alcohol (mg per day) and true a'lmlle systolic blood
of association. forexample. one variable might be a measure
IRS~ (mmHg). Blood pressure measurc:mcnts arc wc:lI-
ofdisease OCCUI'ICncc:.the other a biolOgical or environmental
known to be variable within individuals and a single mea-
quantity. Most commonly the ratio would compare
surement is likely to be rather imprecise (see J.lEASUREMfJO'
probabilily of disease expressed as an odds. a risk or some
FREaSJON AND RELLO\BWTY). Such a statement is even more true
olhcr relevant approximalion to the risk. A relative risk value
of a single day's intake or alcohol as a measure of Ihc true
of 1. indicDling equal risks in both groups. suggc:5Is that no
a\'CnIIC daily intake of alcohol (even if that day's intake wen:
association exists between the biological or environmental
found to be measured without cnur). Now. in the c:pidemi-
quantity and disease.
If a statistical measure suggests posilive or negalive ological study we chose. ror each participant. to measure
systolic blood pressure once and then ask them to recalllhcir
association. this should not immediately be taken to imply
alcohol intake the previous day. If we now calculate lite
thal the association is valid and gc:ncraJisable. Several ~
siderations mighl lead us to question the importance of an Pearson product-moment correlation bet\WCn Ihc two mc:a-
SUfCS we are likely 10 get a positive value that may be
observc:d statistical association.
statistically signiftcanl (assuming we hayc a large enough
Firsl. considc:ntion of the STANIlo\RD ERRDR of the measure
of association. generally re8ccting the size of the sample. sample) but will not be particularly high (i.e. not far above
zero). Suppose, ror the sake of argument that we have found
places Ihc magnitude of association in perspective with the
a value of lhis correlation 10 be 0.20. It should be fairly
magnitude of random c:rror. Apparently shollg associations
obvious that as Ihc measures or systolic blood prcs5ure and
may in fact be poorty estimalc:d and fall short or SlAtislical
signiflcance. alcohol get less pteCisc (equivalenl for a nx.c:d population to
lowering their reliabilities) the correlation will tend to zero.
Second. an apparent association may be entirely spurious
This is alIcnuation due to measuremenl erTOr.
(i.e. 'confoundc:d') due to the inftuenc:e of Glher measuraL or
unmeasun:d. variables that hayc not been accounted for in the l.etlhc observed measurement of blood pressure for the 4h
participant be Y, and the corrcsponding true average blood
analysis. For example. in II IRliminary statistical enquiry.
prcssa=be t'.. Similarly.lct the measured alcohol intake bcXt
risk of coronary hc:an disease may appear to be associated
with watching television. although consideration of the wilh a true 8\OCrage of",. We have estimated the correlation
underlying relalionship with obesity and physical exercise between Y and X. PYX. when we are really interested in Ihc
would probably suggest that Ihc preliminary Onding was corrclDlion between the lnIe w1ues. PII9' If the c:rrors of
measurement for blood prcssa= are uncorn:lated with those
spurious. An association may alter after adjustment for the
ror alcohol consumption then it can be shown thDl Ihc
interdependence oroeher variables and Ihc gcnc:nl validity of
rollowing relationship holds:
a measa= of association would often depend on the extent to
which such poICntial interdependencies have been taken into (1)
K-aJUnt. Studies measuring sevc:nl variables often utilise
19
ATTRIBUT~~SK ____________________________________________________________

HcM. "r and "Jt ~ the reliabilitics of the blood pleSAR as good an cstimale as possible. especially when one employs
and alcohol a1Dsumption mc:asuremenls respcc:tivcly. It anobscrvational study. Using BAYESTHE(IlEM and rearranging
follows that: the cqualiDD. we can obtain an expreSsion exprasc:cl in tenDs
of the relative risk (RR):

Provided wc know the reliabililies fCJl'the two measuremenls. 1 = Pr{EHRR-l)

this equation can be used 10 adjust Ihe observed atJRlalion
between Yand X to obtain the required com:lation between where Pr(£} is the pn:valence of exposun: in the population
their true average values. If we bow that "Jt I : 0~3 and at large. This is a COIlYCnicnl way of expressing the measure
ItJt=0.7. for example. the n:quin:d comdalion is 0.21 of association, because RR is often cslimalc:d usine alterna-
,J(0.3 x 0.7) = 0.44. tive study designs. including C.~SE-CON1ROL. COHOIn' AND
If. instead ofa correlation. the lincarregn:ssion coefficient CIlOSS-SECIlOIW. SlUDIES.
for the effecl of blood presSIR on alcohol CODSumption wcre Attributable risk is most easily inlClpretc:d when the
of key intelallhcn: factor of interest increases risk. i.e. RR > I. and in these
cases the possible range of the measure is from 0 10 I. An
(3)
altributable risk of zero can occur when no individuals in the:
and. again.1he n:quiml adjustmenl isstraightforwani. Equa- popuIalion ~ exposed 10 the factor of inlc:rcst. or if the:
tion (3) also holds approximately if we were 10 use a logistic faclor is not relaled to risk of disease. RR = I. The measure
rqrasion to pn:dict the presencclabllClltlC of hypertension. is nol easily interpreted when the expos~ is proteclive.
These calculalions are One as 10Rl as we ha\'C valid RR < I. so it isgenemllynot used in thisca5e. By n:cieftning
cstimates of the reliabilities. Howcver. Ihcy ~ only valid the: reference graup. one can always cxpress the results of a
in these veJY simple situations as described. Epiclcmiologisls study in a fonn in which RR is greater than I. so this is not a
almost always wish to adjust their estimates to allow for serious limitation. In addition. the lDCasure is oftcn ex-
confounding and some of these confounders an: inevitably pressed as a percent. As RR became large.).. goes 10 I. but A.
goiRl to be prone 10 MEASUREMENT ERROR. Undel' these goes 10 zero either as the proportion exposed. Pr{EJ.
cimlmstances life is considembly more complicated! We becomes small or as the relative risk. RR. appraached the:
cannot even be ccrtain that the estimate of the required null value of 1. If an enlim population is cxposed to a
parameter will be allenualc:d, never mind heine altenuated particular faclor. Pr{ E} = 1. then the: sc:cond equalion
in a way described by equation (3). Readers ~ refClRCi (above) reduces to A. =(RR -1)lRR.
clsewhere to these much mom challeRling but more realistic The lable: shows a typical 2 x 2 tablc that can he usc:cllo
situations (Carroll. Ruppert and Stefanski. 1995; Cheng and display Ihe results from an cpidemiological study. In a case-
Van Ness. 1999; Ouslafson~ 20(3). GO control study. the column totals arc generally regarded as
being fixed by design and the odds ratioorcross-produci mtio
Carrol, R. J., RuppeJ1, D. and S......... L A. I99S: MI!tUUI'r-
is usc:cl as a good approximation 10 the estimate of RR when
,.nl ~rTtJf' in IIIHflinetlr motkls. London: Oapman " Hall. a..a.
Co-L ..... Va Ness. J. W. 1999: StQI&tkal re,rrssiolr M'ilh the disease is rare. In addition. the exPOSlR distribution in Ihc
IMtISII1'tRfMt e"OI'. London: Amold. CiuIIIIfIaa. P. 2003: MetlSUTr!-
conbols. Pr{E) = Pr{EtO}. is conside=llo be representa-
,.nl ~"or and miNmsijitaliDII ;" stalistia tmtl epidemiology. tive of the exposure distribution in Ihc overall population.
London: Cbapmm" HalIlCRC. Substituting in the samplc estimatcs ofthcse quantities gives
rise to whal is the maximum likelihood cstimate of A:
attributable risk As a mcaslR of the public health • ad-be
signiftcance of exposure to a risk factor for disease. the 1=----
d(a+c)
attributable risk provides an estimate of the proportion of
diseased subjcc:tslhat may be allributc:d to the exposure. It is
defined by: attributable risk Results from an epidemioIogIcalstudy
A = Pr{D}-Pr{DIE} with two lwels 01 exposure and disease status

where PrID) is the probability that an individual develops Disetue slahU

disease and E and E reprellCllt whclhcr an individual is
cxposed CJI' not exposed to Ihc factor of intcn:st (Levin. Expo:mi D fJ Tolal
1953). Ideally. one would like 10 know bath Prt DJ and
£ a b a+b
Pr{ DIE} fCJl' Ihc population under study. bul for some study E c tl c+tl
designs this is not possible. so if one wishes to use the
Total a+c b+d N
measure, can: is needed to design a study that will proviclc
Exploring the Variety of Random
Documents with Different Content
passed near here before when out in the woods, and
knew of many darkys who befriended them. Had a
surfeit of food. Stayed at the huts until after midnight,
and then a woman brought us to this place. To-night
we go to Jocko’s hut, across the river. A darky will
row us across the Little Ogechee to Jocco’s hut, and
then he will take us in tow. It is a rice country about
here, with canals running every way. Negroes all
tickled to death because Yankees coming. I am
feeling better than yesterday, but difficult to travel.
Tell the boys they had better leave me with the
friendly blacks and go ahead to our lines, but they
won’t. Plenty to eat and milk to drink, which is just
what I want. The whites now are all away from their
homes and most of the negroes. Imagine we can
hear the booming of cannon, but guess we are
mistaken. Dave is very entertaining and good
company. Don’t get tired of him and his talk. Both of
them are in rebel dress throughout, and can talk and
act just like rebels. Know the commanders of
different rebel regiments. They say that when out
before they on different occasions mixed with the
Southern army, without detection. Said they didn’t
wonder the widow woman knew I was a Yankee.
Ain’t up to that kind of thing.
Dec. 15.—Jocko’s hut was not across the river as I
supposed and wrote yesterday, but on the same side
we were on. At about ten o’clock last night we went
to his abiding place as directed and knocked. After a
long time an old black head was stuck out of the
window with a nightcap on. The owner of the head
didn’t know Jocko or anything about him; was short
and crusty; said: “Go way from dar!” Kept talking to
him and he scolding at being disturbed. Said he had
rheumatics and couldn’t get out to let us in. After a
long time opened the door and we set down on the
door step. Told him we were yankees and wanted
help. Was the funniest darky we have met yet. Would
give something for his picture as he was framed in
his window in the moonlight talking to us, with the
picturesque surroundings, and us yankees trying to
win him over to aid us. Finally owned up that he was
Jocko, but said he couldn’t row us across the river.
He was lame and could not walk, had no boat, and if
he had the river was so swift he couldn’t get us
across, and if it wasn’t swift, the rebels would catch
him at it and hang him. Talked a long time and with
much teasing. By degrees his scruples gave way,
one at a time. Didn’t know but he might row us
across if he only had a boat, and finally didn’t know
but he could find a boat. To get thus far into his good
graces took at least three hours. Went looking
around and found an old scow, fixed up some old
oars, and we got in; before doing so however, he had
warmed up enough to give us some boiled sweet
potatoes and cold baked fish. Rowed us way down
the river and landed us on the noted Miller plantation
and a mile in rear of the negro houses. Jocko, after
we forced our acquaintance on him with all kind of
argument, proved to be a smart able-bodied old
negro, but awful afraid of being caught helping
runaways. Would give something for his picture as
he appeared to us looking out of his cabin window.
Just an old fashioned, genuine negro, and so black
that charcoal would make a white mark on him. Took
us probably three miles from his hut, two miles of
water and one of land, and then started back home
after shaking us a dozen times by the hand, and
“God blessing us.” Said “Ole Massa Miller’s niggers
all Union niggers,” and to go up to the huts in broad
daylight and they would help us. No whites at home
on the plantation. We arrived where Jocko left us an
hour or so before daylight, and lay down to sleep
until light. I woke up after a while feeling wet, and
found the tide had risen and we were surrounded
with water; woke up the boys and scrambled out of
that in a hurry, going through two feet of water in
some places. The spot where we had laid down was
a higher piece of ground than that adjoining. Got on
to dry land and proceeded to get dry. At about ten
o’clock Dave went up to the negro huts and made
himself known, which was hard work. The negroes
are all afraid that we are rebels and trying to get
them into a scrape, but after we once get them
thoroughly satisfied that we are genuine Yanks they
are all right, and will do anything for us. The negroes
have shown us the big house, there being no whites
around, they having left to escape the coming
Yankee army. We went up into the cupola and looked
way off on the ocean, and saw our own noble
gunboats. What would we give to be aboard of them?
Their close proximity makes us discuss the feasibility
of going down the river and out to them, but the
negroes say there are chain boats across the river
farther down, and picketed. Still it makes us anxious,
our being so near, and we have decided to go down
the river to-night in a boat and see if we can’t reach
them. It is now the middle of the afternoon and we
lay off from the huts eighty rods, and the negroes are
about to bring us some dinner. During the night we
traveled over oyster beds by the acre, artificial ones,
and they cut our feet. Negroes say there are two
other runaways hid a mile off and they are going to
bring them to our abiding place. Later.—Negroes
have just fed us with corn bread and a kind of fish
about the size of sardines, boiled by the kettle full,
and they are nice. Fully as good as sardines. Think I
know now where nearly all the imported sardines
come from. Negroes catch them by the thousand, in
nets, put them in kettles, and cook them a few
minutes, when they are ready to eat. Scoop them out
of the creeks. The two other runaways are here with
us. They are out of the 3d Ohio Cavalry. Have been
out in the woods for two weeks. Escaped from
Blackshear and traveled this far. I used to know one
of them in Savannah. We do not take to them at all,
as they are not of our kind. Shall separate to-night,
they going their way and we going ours. Have
secured a dug-out boat to go down the Ogechee
River with to-night. The negroes tell us of a Mr.
Kimball, a white man, living up the country fifteen
miles, who is a Union man, and helps runaways, or
any one of Union proclivities. He lays up the river,
and our gunboats lay down the river. Both have
wonderful charms for us, and shall decide before
night which route to take. Are on rice plantation, and
a valuable one. Before the “wah” there were over
fifteen hundred negroes on this place. Cotton is also
part of the production. Have decided to go down the
river and try to reach our gunboats. It’s a very
hazardous undertaking, and I have my doubts as to
its successful termination.
Dec. 16.—Another adventure, and a red hot one.
Started down the river in our dug-out boat
somewhere near midnight. Ran down all right for an
hour, frequently seeing rebel pickets and camp fires.
Saw we were going right into the lion’s mouth, as the
farther down the more rebels. All at once our boat
gave a lurch and landed in a tree top which was
sticking out of the water, and there we were, swaying
around in the cold water in the middle or near the
middle of the Ogechee. Dave went ashore and to a
negro hut, woke up the inmates, and narrated our
troubles. A negro got up, and with another boat came
to the rescue. Were about froze with the cold and
wet. Said not more than a mile farther down we
would have run right into a chain boat, with pickets
posted on it. It really seems as if a Divine providence
were guiding us. After getting a breakfast of good
things started off toward the Big Ogechee River, and
have traveled three or four miles. Are now
encamped, or rather laying down, on a little hillock
waiting for evening, to get out of this vicinity which is
a dangerous one. In our river escapade lost many of
our things, but still hang to my coverlid and diary.
There are three or four houses in view, and
principally white residences, those of the poor white
trash order, and they are the very ones we must
avoid. Have caught cold and am fearfully out of
traveling condition, but must go it now. A mistake in
coming down the river. Am resting up, preparatory to
traveling all night up the country. No chance of
getting out by the coast. Have enough food to last all
day and night, and that is a good deal. Can’t carry
more than a day’s supply. Have now been out in the
woods, this is the fourth day, and every day has been
fresh adventures thick and fast. If I could only travel
like my comrades, would get along. Bucks praise me
up and encourage me to work away, and I do. For
breakfast had more of those imported sardines.
Storm brewing of some sort and quite chilly. Saw
rebel infantry marching along the highway not more
than eighty rods off. Hugged the ground very close.
Dogs came very near us, and if they had seen us
would have attracted the rebels’ attention. Am writing
with a pencil less than an inch long. Shall print this
diary and make my everlasting fortune, and when
wealthy will visit this country and make every negro
who has helped us millionaires. Could not move from
here half a mile by daylight without being seen, and
as a consequence we are feeling very sore on the
situation. Don’t know but I shall be so lame to-night
that I cannot walk at all, and then the boys must
leave me and go ahead for themselves. However,
they say I am worth a hundred dead men yet, and
will prod me along like a tired ox. Dave goes now
bareheaded, or not quite so bad as that, as he has a
handkerchief tied over his head. The programme
now is to go as straight to Mr. Kimball’s as we can.
He is probably twenty miles away; is a white Union
man I spoke of a day or so ago in this same diary.
Will stick to him like a brother. Can hear wagons go
along the road toward Savannah, which is only
thirteen or fourteen miles away. Later.—Most dark
enough to travel and I have straightened up and am
taking an inventory of myself. Find I can walk with the
greatest difficulty. The boys argue that after I get
warmed up I will go like a top, and we will see.
Dec. 17.—And another day of vicissitudes. We
traveled last night about four miles, piloted by a
young negro. It was a terrible walk to me; slow and
painful. Were fed, and have food for to-day. Are now
about three miles from a canal which we must cross
before another morning. Negroes say “Sherman
most here” and “Bress de Lord!” Mr. Kimball lives
nine miles away and we must reach him some way,
but it seems an impossibility for me to go so far. Are
now in a high and fine country, but too open for us.
Have to lay down all day in the bushes. David is a
thorough scout. Goes crawling around on his hands
and knees taking in his bearings. Troops are
encamped on the main road. Every cross road has
its pickets, and it is slow business to escape running
into them. Eli S. Buck has a sore throat and is
hoarse. Pretty good jaunt for him, tough as he is.
Shall have no guide to-night, as Dave thinks he can
engineer us all right in the right direction. Some
thinks he will leave us both and reach Kimball’s to-
night, and then come back and see us through.
Guess I will be on hand to go along however.
Dec. 18.—Six days of freedom and what a sight of
hardship, sweetened by kind treatment and the
satisfaction of being out from under guard. We
traveled last night some four miles and now are in a
very precarious position. When almost daylight we
came to the canal, and found cavalry pickets all
along the tow-path; walked along until we came to a
lock. A cavalryman was riding his horse up and down
by the lock. At the lock there was a smouldering fire.
It was absolutely necessary that we get across
before daylight. As the mounted picket turned his
horse’s head to go from us, Dave slid across the tow-
path and went across the timbers which formed the
lock, and by the time the picket turned around to
come back Dave was hid on the opposite shore. At
the next trip of the rebel Eli went the same as Dave.
The third one to go was myself, and I expected to get
caught, sure. Could not go as quiet as the rest, and
was slower. Thought the picket saw me when half
way across but kept right on going, and for a wonder
made it all right. Was thoroughly scared for the first
time since jumping off the train. Am very nervous. All
shook hands when the picket turned about to go
back the fourth time. Getting light in the east and we
must move on, as the country is very open. Dare not
travel over half a mile, and here we are hid almost in
a woman’s door yard, not over thirty rods from her
very door. Are in some evergreen bushes and
shrubs. It’s now most noon, and have seen a rather
elderly lady go out and in the house a number of
times. The intrepid Dave is going up to the house to
interview the lady soon. Later.—Dave crawled along
from our hiding place until he came to the open
ground, and then straightened boldly up and walked
to the house. In fifteen minutes he came back with
some bread and dried beef, and said the woman was
a Union woman and would help us. Her daughter
slept at her uncle’s a mile off last night, and expected
her back soon, and perhaps the uncle, who is a
violent Secesh, with her. Said for us to lay low.
Later.—The daughter came home on horseback
and alone. Could see the old lady telling the daughter
about us and pointing our way. About the middle of
the afternoon the old lady started out toward us.
Behind her came a young darky, and behind the
darky came another darky; then a dog, then a white
boy, then a darky, and then the daughter. Old lady
peeked in, and so did the rest except the grown-up
girl, who was too afraid. Finally came closer, and as
she got a good view of us she says: “Why, mother,
they look just like anybody else.” She had never seen
a Yankee before. Brought us some more food, and
after dark will set a table for us to come to the house
and eat. Her name is Mrs. Dickinson. They went
back to the house and we proceeded to shake hands
with one another. During the afternoon five rebel
soldiers came to the house, one at a time. It is now
most dark and we are about ready to go to the house
and eat. Mr. Kimball lives only four miles away.
Dec. 19.—We are now less than half a mile from
Mr. Kimball’s. After dark last night we went to Mrs.
Dickinson’s house and partook of a splendid supper.
I wrote a paper directed to the officer commanding
the first Yankee troops that should arrive here telling
what she had done for us runaway Yankees. She
talked a great deal, and I thought was careless
leaving the front door open. Three or four times I got
up and shut that door. We had taken off our blankets
and other wraps and left them in a sort of a kitchen,
and were talking in the best room. I heard the gate
click, and on looking out saw two rebel officers
coming to the house and not six rods off. We jumped
into the other room and out of the back door and
behind a corn house, bare headed. The officers were
asked into the front room by the daughter. They
asked who the parties were who ran out of the back
way. She said she reckoned no one. They kept at her
and jokingly intimated that some of her skulking
lovers had been to see her. She kept talking back
and finally said: “Mother, did any one just go away?”
And the old lady said: “Why, yes, brother Sam and
his ‘boy’ just went off home.” Them confounded
rebels had come to see the girl and spend the
evening, and we shivering out in the cold. Joked her
for an hour and a half about her lovers and we
hearing every word. Finally they got up and bid her
good night, saying they would send back some men
to guard the house and keep her lovers away. Just
as soon as they were down the road a ways, the
daughter came out very frightened and said for us to
hurry off, as they would send back troops to look for
us. Hurried into the house, got our things and some
dried beef, and started off toward Mr. Kimball’s
house. Reached here just before daylight and lay
down back of the house about eighty rods, in the
corner of the fence, to sleep a little before morning.
Just at break of day heard some one calling hogs.
David got up and went toward an old man whom we
knew was our friend Kimball. Came to us, and was
glad to shake hands with genuine Yankees. Said one
of his neighbors was coming over early to go with
him to hunt some hogs, and for us to go farther off
and stay until night, and he would think up during the
day what to do with us. Did not want anything to eat.
Came to this place where we now are, and feeling
that our journey was most ended. Mr. Kimball said
that Sherman was not over fifty miles off, and coming
right along twenty miles per day, and our plan was to
hide and await coming events. Mr. Kimball is an old
man, probably sixty years old, white haired and stoop
shouldered. He had five sons, all drafted into the
rebel army. All refused to serve. Two have been shot
by the rebels, one is in some prison for his Union
proclivities, and two are refugees. The old man has
been imprisoned time and again, his stock
confiscated, property destroyed, and all together had
a hard time of it. Still he is true blue, a Union man to
the back bone. Really think our troubles coming to an
end. Kimball said: “Glory to God, the old Stars and
Stripes shall float over my house in less than a
week!” It’s a noble man who will stand out through all
that he has, for his principles, when his interests are
all here. Is not only willing, but glad to help us, and
says anything he has is ours, if it will help us toward
our escape. Later.—Have been laying all day
watching Kimball’s house. Along in the morning the
neighbor spoken of came to Kimball’s, and they both
went off on horseback to shoot hogs. The swine here
roam over a large territory and become most wild,
and when they want fresh pork they have to go after
it with a gun. You may be sure the hunters did not
come near us with Mr. Kimball for a guide. A negro
boy went with them with a light wagon and mule
attached. Near noon they returned with some killed
hogs in the wagon. At three or four o’clock the old
man came down where we were “to look after his
boys,” he said. Is in the best of spirits. Says we are to
hide to-night where he tells us, and stay until our
troops reach us. That is jolly good news for me, as I
hate to travel. Said come to the house after dark and
he would have a supper prepared for us, and has just
left us. Later.—Have just eaten a splendid supper at
Kimball’s and getting ready to travel three miles to a
safe hiding place.
Dec. 20.—Well, we are just well fixed and happy.
After partaking of a royal repast last night, served in
an out-building near the main building of the Kimball
home, we were directed to this place which is on the
banks of the Big Ogechee river, in a most delightful
spot. While we were at Kimball’s he had negro
sentinels stationed at different points on the
plantation to announce the coming of any rebel
soldiers or citizens that might see fit to come near.
He gave us an axe, a quart of salt, a ham too big to
carry conveniently, and all the sweet potatoes we
could drag along; also a butcher knife. Went with us
a mile as guide and then told us so we found the
place pointed out. Also gave us some shelled corn to
bait hogs and told Dave how to make a deadfall to
catch them. We left the main road going directly West
until we came to a fence, then turned to the left and
followed the line of the fence, and when we had got
to the end of it kept straight ahead going through a
swampy low section. After a while came to higher
and dry land and to the banks of the river. Is a sort of
an island, and as I said before, a very pretty and
pleasant spot. Out in the river grows tall canebrake
which effectually hides us from any one going either
up or down the river. Tall pines are here in
abundance and nice grass plats, with as handsome
palm clusters as ever I saw. Are going to build us a
house to keep off the cold and rain. Have matches
and a rousing fire cooked our breakfast of nice ham
and sweet potatoes. We also roasted some corn and
had corn coffee. Any quantity of hogs running around
and Dave is already thinking of a trap to catch them.
It will be necessary for we are making that ham look
sick. Eat so much breakfast that we can hardly walk
and don’t know but will commit suicide by eating.
Buzzards fly around attracted by the cooking. Are as
large and look like turkeys. Our government should
give to Mr. Kimball a fortune for his patriotism and
sacrifices to the Union cause. About eight miles
above is a long bridge across the river and there it is
thought a big fight will take place when Sherman
attempts to cross, and so we will know when they
approach, as we could hear a battle that distance.
Night.—We have built the cosyest and nicest little
house to lay in. Cut poles with the axe and made a
frame, and then covered the top with palm leaves
just like shingles on a house at the North, then fixed
three sides the same way, each leaf overlapping the
other, and the fourth side open to a fire and the river.
The water is cold and clear and nice to drink; just like
spring water. Have eaten the ham half up; ditto
potatoes. The increased prosperity makes me feel
well bodily, and mentally am more so. It is still the
“Astor House Mess.” We all cook, and we all eat.
Dave prays to-night as he does every night and
morning, and I ain’t sure but all through the day. Is a
thorough Christian if ever there was one. I also wrote
a letter for Mr. Kimball to the commanding Union
officer who may first approach these parts. In it I told
how he had befriended us and others. We heard
boats going by on the river to-day. At such times all
we do is to keep still, as no one can see us. Rebels
are too busy to look for us or any one else. All they
can do now to take care of themselves. Eli is making
up our bed, getting ready to turn in. I have just
brought a tin pail of nice water and we all drink. Take
off our shoes for the first time in some days. A
beautiful night—clear and cold. And thus ends
another day, and we are in safety.
Dec. 21.—Got up bright and early. Never slept
better. Getting rested up. We talk continually. Both
Bucks are great talkers, especially David. Cooked
and ate our breakfast, and would you believe it the
ham is all gone. Incredible, the amount of food we
eat. Wonder it don’t make us all sick. Sweet potatoes
getting low. Dave fixing up his deadfall for hogs. Has
rolled some heavy logs together forty rods away from
our house, and fixed up a figure four spring trap, with
the logs for weight to hold down the animal which
may be enticed into it. Has scattered corn in and
around the trap, and we wait for developments. Hogs
are very shy of us and surroundings. Are apparently
fat and in good order. Plenty of roots and shack
which they eat, and thrive thereon. Buzzards are very
curious in regard to us. They light on the limbs in the
trees, and if their support is a dead limb it breaks and
makes a great noise in the still woods. Two or three
hundred all together make a terrible racket, and
scare us sometimes. The weather is very fine, and
this must be a healthy climate. Dave is going out to-
day to look around. As I have said before, he is a
scout and understands spying around, and won’t get
caught. If we had a fish hook and line or a net of
some sort could catch fish to eat. That would be a
grand sport as we can see nice large fish in the
water. The main road is away about one and a half
miles we think by the sound of the teams which
occasionally rumble along. Often hear shouting on
the road as if cattle were being driven along toward
Savannah. Once in a while we hear guns fired off,
but it is no doubt hogs being killed. We also hear
folks going up and down the river, but cannot see
them. After dark we have no fire as that would
expose us, it is so much plainer to be seen in the
night. The river is wide; should think a third of a mile,
as we can view it from away up the stream. The cane
that grows in the river is the same as we have for fish
poles at the North, and are shipped from the South.
Have added some repairs to the house and it is now
water tight, we think. Made a bed of soft boughs, and
with our three blankets have a good sleeping place.
Dave got a tall cane and fastened up on the house,
and for a flag fastened on a piece of black cloth—the
best we could do. That means no quarter; and it is
just about what we mean, too. Don’t believe we
would be taken very easy now. I am getting fat every
day, yet lame, and have come to the conclusion that
it will be a long time before I get over it. The cords
have contracted so in my right leg that they don’t
seem to stretch out again to their original length. That
scurvy business came very near killing me. Later.—I
also went out of our hiding place, and saw away out
in a field what I took to be a mound where sweet
potatoes were buried. Came back and got a pair of
drawers, tied the bottom of the legs together, and
sallied forth. The mound of potatoes was a good way
back from the house, although in plain sight. I
crawled up, and began digging into it with a piece of
canteen. Very soon had a hole in, and found some of
the nicest potatoes that you can imagine, of the red
variety, which I believe are the genuine Southern
yam. Filled the drawers cram full, filled my pockets
and got all I could possibly carry, then closed up the
hole and worked my way back to camp. Eli was
alone, Dave not having returned from his scouting
trip. Had a war dance around those potatoes. Believe
there is a bushel of them, and like to have killed
myself getting them here. After I got into the woods
and out of the field, straightened up and got the
drawers on my shoulders and picked the way to
head-quarters. We don’t any of us call any such thing
as that stealing. It’s one of the necessities of our lives
that we should have food, and if we have not got it,
must do the best we can. Now if we can catch a
porker will be fixed all right for some days to come.
Think it is about the time of year for butchering. We
don’t expect to be here more than two or three days
at fartherest, although I shall hate to leave this
beautiful spot, our nice house and all. Listen all the
time for the expected battle at the bridge, and at any
unusual sound of commotion in that direction we are
all excitement. Later.—Dave has returned. He went
to the main road and saw a negro. Was lucky enough
to get a Savannah paper three days old in which
there was nothing we did not know in regard to
Sherman’s coming. The negro said yankee scouts
had been seen just across the river near the bridge,
and the main army is expected every day. The rebels
will fall back across the river and contest the
crossing. Fortifications are built all along clear to
Savannah, and it may be reasonably expected that
some hard fighting will take place. Savannah is the
pride of the South and they will not easily give it up.
Dave did not tell the negro that he was a yankee, but
represented himself as a conscript hiding in the
woods to keep from fighting in the rebel army. Was
glad to see supply of potatoes and says I will do. Has
freshly baited his trap for hogs and thinks before
night we will have fresh pork to go with the potatoes.
Later.—We went around a drove of hogs and
gradually and carefully worked them up to the trap.
Pretty soon they began to pick up the corn and one
of them went under the figure four, sprung it and
down came the logs and such a squealing and
scrambling of those not caught. The axe had been
Welcome to our website – the ideal destination for book lovers and
knowledge seekers. With a mission to inspire endlessly, we offer a
vast collection of books, ranging from classic literary works to
specialized publications, self-development books, and children's
literature. Each book is a new journey of discovery, expanding
knowledge and enriching the soul of the reade

Our website is not just a platform for buying books, but a bridge
connecting readers to the timeless values of culture and wisdom. With
an elegant, user-friendly interface and an intelligent search system,
we are committed to providing a quick and convenient shopping
experience. Additionally, our special promotions and home delivery
services ensure that you save time and fully enjoy the joy of reading.

Let us accompany you on the journey of exploring knowledge and

personal growth!

ebookfinal.com