using smoothing spline anova to examine the relation of - Deep Blue

20 Pages·1997·1.15 MB·English

STATISTICSINMEDICINE,VOL.16,1357—1376(1997)
USING SMOOTHING SPLINE ANOVA TO EXAMINE THE
RELATION OF RISK FACTORS TO THE INCIDENCE
AND PROGRESSION OF DIABETIC RETINOPATHY
YUEDONGWANG(cid:49)*,GRACEWAHBA(cid:50),CHONGGU(cid:51),RONALDKLEIN(cid:52)
ANDBARBARAKLEIN(cid:52)
(cid:49)(cid:68)(cid:101)(cid:112)(cid:97)(cid:114)(cid:116)(cid:109)(cid:101)(cid:110)(cid:116)(cid:111)(cid:102)(cid:66)(cid:105)(cid:111)(cid:115)(cid:116)(cid:97)(cid:116)(cid:105)(cid:115)(cid:116)(cid:105)(cid:99)(cid:115)(cid:44)(cid:85)(cid:110)(cid:105)v(cid:101)(cid:114)(cid:115)(cid:105)(cid:116)(cid:121)(cid:111)(cid:102)(cid:77)(cid:105)(cid:99)(cid:104)(cid:105)(cid:103)(cid:97)(cid:110)(cid:44)(cid:49)(cid:52)(cid:50)(cid:48)(cid:87)(cid:97)(cid:115)(cid:104)(cid:105)(cid:110)(cid:103)(cid:116)(cid:111)(cid:110)(cid:72)(cid:101)(cid:105)(cid:103)(cid:104)(cid:116)(cid:115)(cid:44)(cid:65)(cid:110)(cid:110)(cid:65)(cid:114)(cid:98)(cid:111)(cid:114)(cid:44)(cid:77)(cid:73)(cid:52)(cid:56)(cid:49)(cid:48)(cid:57)(cid:44)(cid:85)(cid:46)(cid:83)(cid:46)(cid:65)(cid:46)
(cid:50)(cid:68)(cid:101)(cid:112)(cid:97)(cid:114)(cid:116)(cid:109)(cid:101)(cid:110)(cid:116)(cid:111)(cid:102)(cid:83)(cid:116)(cid:97)(cid:116)(cid:105)(cid:115)(cid:116)(cid:105)(cid:99)(cid:115)(cid:44)(cid:85)(cid:110)(cid:105)v(cid:101)(cid:114)(cid:115)(cid:105)(cid:116)(cid:121)(cid:111)(cid:102)(cid:87)(cid:105)(cid:115)(cid:99)(cid:111)(cid:110)(cid:115)(cid:105)(cid:110)(cid:44)(cid:49)(cid:50)(cid:49)(cid:48)(cid:87)(cid:46)(cid:68)(cid:97)(cid:121)(cid:116)(cid:111)(cid:110)(cid:83)(cid:116)(cid:46)(cid:77)(cid:97)(cid:100)(cid:105)(cid:115)(cid:111)(cid:110)(cid:44)(cid:87)(cid:73)(cid:53)(cid:51)(cid:55)(cid:48)(cid:54)(cid:44)(cid:85)(cid:46)(cid:83)(cid:46)(cid:65)(cid:46)
(cid:51)(cid:68)(cid:101)(cid:112)(cid:97)(cid:114)(cid:116)(cid:109)(cid:101)(cid:110)(cid:116)(cid:111)(cid:102)(cid:83)(cid:116)(cid:97)(cid:116)(cid:105)(cid:115)(cid:116)(cid:105)(cid:99)(cid:115)(cid:44)(cid:85)(cid:110)(cid:105)v(cid:101)(cid:114)(cid:115)(cid:105)(cid:116)(cid:121)(cid:111)(cid:102)(cid:77)(cid:105)(cid:99)(cid:104)(cid:105)(cid:103)(cid:97)(cid:110)(cid:44)(cid:77)(cid:97)(cid:115)(cid:111)(cid:110)(cid:72)(cid:97)(cid:108)(cid:108)(cid:44)(cid:65)(cid:110)(cid:110)(cid:65)(cid:114)(cid:98)(cid:111)(cid:114)(cid:44)(cid:77)(cid:73)(cid:53)(cid:56)(cid:49)(cid:48)(cid:57)(cid:44)(cid:85)(cid:46)(cid:83)(cid:46)(cid:65)(cid:46)
(cid:52)(cid:68)(cid:101)(cid:112)(cid:97)(cid:114)(cid:116)(cid:109)(cid:101)(cid:110)(cid:116)(cid:111)(cid:102)(cid:79)(cid:112)(cid:104)(cid:116)(cid:104)(cid:97)(cid:108)(cid:109)(cid:97)(cid:108)(cid:111)(cid:103)(cid:121)(cid:44)(cid:85)(cid:110)(cid:105)v(cid:101)(cid:114)(cid:115)(cid:105)(cid:116)(cid:121)(cid:111)(cid:102)(cid:87)(cid:105)(cid:115)(cid:99)(cid:111)(cid:110)(cid:115)(cid:105)(cid:110)(cid:44)(cid:54)(cid:49)(cid:48)(cid:87)(cid:97)(cid:108)(cid:110)(cid:117)(cid:116)(cid:83)(cid:116)(cid:46)(cid:77)(cid:97)(cid:100)(cid:105)(cid:115)(cid:111)(cid:110)(cid:44)(cid:87)(cid:73)(cid:53)(cid:51)(cid:55)(cid:48)(cid:54)(cid:44)(cid:85)(cid:46)(cid:83)(cid:46)(cid:65)(cid:46)
SUMMARY
SmoothingsplineANOVA(ANalysisOfVAriance)methodsprovideaflexiblealternativetothestandard
parametricGLIM(generalizedlinearmodels)methodsforanalysingtherelationshipofpredictorvariables
tooutcomeswithdatafromlargeepidemiologicstudies.Thesemethodsallowthevisualizationofrelation-
shipsnotreadilyfitbysimpleGLIMmodels,andprovidefortheabilitytovisualizeinteractionsbetweenthe
variables.Atthe sametime, theyreduceto GLIMmodelsif the data suggestthat the addedflexibility is
unwarranted. Using this method, we investigate risk factors for incidence and progression of diabetic
retinopathyinagroupofpatientswitholderonsetdiabetesfromtheWisconsinEpidemiologicalStudyof
DiabeticRetinopathy.Wecarryoutfouranalysestoillustratevariouspropertiesofthisclassofmethods.
Some of the results confirm previous findings with use of standard methods, while others allow the
visualizationofmorecomplexrelationshipsnotevidentfromtheapplicationofparametricmethods.(cid:40)1997
byJohnWiley&Sons,Ltd.
Statist.Med.,16,1357—1376(1997)
No.ofFigures:10 No.ofTables:1 No.ofReferences:38
1. INTRODUCTION
Manydemographicmedicalstudiescollectdataoftheform(cid:77)y,t(i),i"1,2n(cid:78),whereiindexes
(cid:105)
theithstudyparticipant,y is1or0,indicatingwhetheramedicalconditionofinterestispresent
(cid:105)
orabsentatfollow-up,andt (i)"(t (i),2,t (i))isavectorof dpredictorvariablesatbaseline,
(cid:49) (cid:100)
thatmayormaynotrelatetothelikelihoodthaty isa1.Fromthesedatawewishtoestimate
(cid:105)
p(t)(cid:34) Prob(cid:77)y"1(cid:68)t(cid:78), the probability that a person with predictor vector t presents with the
condition of interest at follow-up. We use such estimates for the prevalence of the condition of
interestin generalpopulations and to study the sensitivity of p to the predictor variables,or, if
* Correspondenceto:Y.Wang,DepartmentofBiostatistics,UniversityofMichigan,1420WashingtonHeights,Ann
Arbor,MI48109,U.S.A.
Contractgrantsponsor:NIH
Contractgrantnumber:EY09946,P60DK20572,P30HD18258,U54HD29184,EY03083
Contractgrantsponsor:NSF
Contractgrantnumber:DMS9121003,DMS9301511
CCC 0277—6715/97/121357—20$17.50 (cid:82)(cid:101)(cid:99)(cid:101)(cid:105)v(cid:101)(cid:100) (cid:68)(cid:101)(cid:99)(cid:101)(cid:109)(cid:98)(cid:101)(cid:114) (cid:49)(cid:57)(cid:57)(cid:53)
(cid:40) 1997 by John Wiley & Sons, Ltd. (cid:82)(cid:101)v(cid:105)(cid:115)(cid:101)(cid:100) (cid:65)(cid:117)(cid:103)(cid:117)(cid:115)(cid:116) (cid:49)(cid:57)(cid:57)(cid:54)
1358 Y.WANGE„A‚.
possible, to combinations of them. The traditional GLIM models(cid:49) do this by defining f(t), the
logit, as
f(t)"log[p(t)/(1!p(t))] (1)
and assuming that f is a simple parametric function of the components of t. When the t are
(cid:97)
continuous variables, the most commonly used model is linear in the components of t
(cid:100)
f(t)"f(t ,2,t )"(cid:107)#(cid:43) (cid:98) t (2)
(cid:49) (cid:100) (cid:97) (cid:97)
(cid:97)(cid:47)(cid:49)
butsometimesoneusessecondoreventhirddegreepolynomialsifitappearsthatalinearmodel
isinadequate.Ifsomeofthet arecategoricalvariables,thenwecanuseindicatorfunctions.More
(cid:97)
generally, given M parametric functions (cid:47) ,(cid:108)"1,2,M subject to some identifiability condi-
(cid:108)
tions, f is modelled parametrically as
(cid:77)
f(t)"(cid:43) (cid:98) (cid:47) (t). (3)
(cid:108) (cid:108)
(cid:108)(cid:47)(cid:49)
Then ((cid:107) and) the (cid:98)’s are obtained by minimizing the negative log-likelihood(cid:76)(y,f) given by
(cid:110)
(cid:76)(y,f)"(cid:43) [y(cid:105)f(t(i))!log(1#e(cid:102)(cid:40)(cid:116)(cid:40)(cid:105)(cid:41)(cid:41))] (4)
(cid:105)(cid:47)(cid:49)
with(2)or,moregenerally,(3),substitutedforf.Ifthe‘truebutunknown’fisactuallysomelinear
combinationofthespecified(cid:47) thenthisis,ofcoursethecorrectprocedure.Unfortunately,aswe
(cid:108)
examine large data sets more closely, it becomes clear that linear models, or even quadratic or
cubic models, are often inadequate. What happens if, for example, the dependence on t is ‘J’
(cid:97)
shaped,or,evenhastwopeaks,possiblyrepresentingtwodistinctsubpopulations?Ifthe‘true’log
oddsratiofis notwellapproximatedbysomefunctionofthespecifiedform,thenwe mayhave
largebiasesintheestimatesoff.Furthermore,thecommonstatisticalhypothesistests,confidence
intervals,P-values,etc.arenotnecessarilyvalidiffisnotofthespecifiedform.Inthispaperwe
describe and demonstrate the use of smoothing spline ANOVA (SS-ANOVA) methods to
estimate f. These methods avoid many of the above difficulties, yet they provide the ability to
visualize some of the relationships between the variables not easily observed with use of more
traditional methods. We demonstrate their use to examine the relation of risk factors to the
incidenceandprogressionofdiabeticretinopathy,usingdatafromtheWisconsinEpidemiologi-
cal Study of Diabetic Retinopathy (WESDR).
The analysis here is based on (a special case of) the smoothing spline ANOVA method for
modelling and estimating f that appears in Wahba, Wang, Gu, Klein and Klein(cid:50) (WWGKK),
andwe haveimplementedthe analysisviathe publiclyavailablecodeGRKPACKdescribedin
Wang.(cid:51)Thecodeitselfisavailablefromnetlib(cid:64)research.att.cominthegcvdirectorythere,
and through statlib(cid:64)lib.stat.cmu.edu.
An SS-ANOVA estimate is a form of penalized likelihood estimate. We first note two
importantpenalizedlikelihoodestimatesthatappearintheliterature.Then,intheremainderof
thisintroduction,wedescribethemainfeaturesofSS-ANOVAingeneraland thedetailsofthe
SS-ANOVAmodels in particular that we apply to the WESDR data.
MajorprecursorsoftheworkunderdiscussionincludeO’Sullivan(cid:52)andO’Sullivanetal.,(cid:53)who
proposedapenalizedlog-likelihoodestimateforf based onthin platesplines.Thesesplinesare
useful in many contexts,and one can incorporatethem in an SS-ANOVAmodel;(cid:54)(cid:44)(cid:55) we do not
employ them in the present work.
Statist.Med.,Vol.16,1357—1376(1997) (cid:40)1997byJohnWiley&Sons,Ltd.
SMOOTHINGSPLINEANOVA 1359
Hastie and Tibshirani(cid:56) (see also other references there) discussed estimates of f of the form
(cid:100)
f(t)"f(t ,2,t )"(cid:107)#(cid:43) f (t ) (5)
(cid:49) (cid:100) (cid:97) (cid:97)
(cid:97)(cid:47)(cid:49)
where the f are ‘smooth’ functions obtained by some smoothing process, including the use of
(cid:97)
cubic smoothing splines. The S code (Chambers and Hastie(cid:57)) provides the facility for fitting
models of the form (5). They also note that some of their work extends to the SS-ANOVA
methods that we consider here. The varying-coefficient models of Hastie and Tibshirani(cid:49)(cid:48) are
averyinterestingsub-familyoftheSS-ANOVAmodels.Cubicsmoothingsplinesforthef in(5)
(cid:97)
are the solution to the minimization problem: find f (in an appropriate function space), and
(cid:97)
subject to some identifiability criteria such as (cid:58)f (t )dt "0, to minimize the penalized log-
(cid:97) (cid:97) (cid:97)
likelihood functional
(cid:100)
(cid:73)(cid:106)(y,f)"(cid:76)(y,f)#(cid:43) (cid:106)(cid:97)J(cid:97)(f(cid:97)) (6)
(cid:97)(cid:47)(cid:49)
where we define the penalty functionals J by
(cid:97)
(cid:80)(cid:49)
J (f )" (f(cid:64)(cid:64)(t ))(cid:50)dt . (7)
(cid:97) (cid:97) (cid:97) (cid:97) (cid:97)
(cid:48)
As the smoothing parameter (cid:106) tends to infinity, f tends to a linear function in t , so that the
(cid:97) (cid:97) (cid:97)
minimizer of (6) tends to the linear function as in (2) if all the (cid:106) ’s become large. Penalized
(cid:97)
likelihoodmethodswithmoregeneralpenaltyfunctionalsandunpenalizedtermsarediscussedin
reference 11.
Recentresearchhasfocusedonmoregeneralmodelsforfthatallowtheexplicitmodellingand
visualization of possible interactions between variables, via functional analysis of variance
decompositions(to be described)and SS-ANOVAestimationmethods.Given a fairlyarbitrary
functionf(t ,2,t )ofseveralvariables,wecandefinea(functional)ANOVAdecompositionof
(cid:49) (cid:100)
f (generalizing ideas from parametric ANOVA) as
(cid:100)
f(t ,2,t )"(cid:107)#(cid:43) f (t )#(cid:43) f (t ,t )#2#f (t ,2,t ) (8)
(cid:49) (cid:100) (cid:97) (cid:97) (cid:97)(cid:98) (cid:97) (cid:98) (cid:49)(cid:44)2(cid:44)(cid:100) (cid:49) (cid:100)
(cid:97)(cid:47)(cid:49) (cid:97)(cid:58)(cid:98)
where the f are the main effects, f are the two factor interactions, etc. The components are
(cid:97) (cid:97)(cid:98)
uniquelydeterminedgivenasetofaveragingoperators(cid:69)(cid:97),whichaveragefunctionsoverthet(cid:97)in
some specified way. For example, if t is a continuous variable in the interval [0,1], then
(cid:97)
a possible choice for (cid:69)(cid:97) is
(cid:80)(cid:49)
((cid:69)(cid:97)f) (t(cid:49),2,t(cid:97)(cid:126)(cid:49),t(cid:97)(cid:96)(cid:49),2,t(cid:100))" f(t(cid:49),2,t(cid:100))dt(cid:97). (9)
(cid:48)
Then the mean is (cid:107)"(cid:60)(cid:100)(cid:97)(cid:47)(cid:49)(cid:69)(cid:97)f, the main effects are f(cid:97)"(I!(cid:69)(cid:97))(cid:60)(cid:98)O(cid:97)(cid:69)(cid:98)f, the two factor
interactionsaref(cid:97)(cid:98)"(I!(cid:69)(cid:97))(I!(cid:69)(cid:98))(cid:60)(cid:99)(cid:47)(cid:97)(cid:44)(cid:98)(cid:69)(cid:99)f,etc.Since(cid:69)(cid:97)(cid:69)(cid:97)"(cid:69)(cid:97),wecanseethattheterms
inthefunctionalANOVAdecompositionsatisfysideconditionsanalogoustothoseinordinary
parametricANOVA,forexample, (cid:69)(cid:97)f(cid:97)"0. In general,for model fittingpurposes,we eliminate
higherordertermsinthefunctionalANOVAdecompositionandweestimatesomeorallofthe
lower order terms by finding (cid:107), and (some of) the f ,f , etc. to minimize a penalized log-
(cid:97) (cid:97)(cid:98)
likelihoodfunctional(cid:73)(cid:106)(y,f)givenbyanexpressionthatgeneralizes(6),withaseparatepenalty
functionalfor eachindependentlysmoothedterm in the model. We can also incorporatein the
(cid:40) 1997byJohnWiley&Sons,Ltd. Statist.Med.,Vol.16,1357—1376(1997)
1360 Y.WANGE„A‚.
modelindicatorfunctionsforcategoricalvariables.Otheraveragingoperatorsincludeweighted
sums over possible or observed values of t .
(cid:97)
It is well known(cid:49)(cid:49) that solutions to variational problems like the minimizer of (cid:73)(cid:106)(y,f) and
generalizationsto (cid:73)(cid:106)(y,f) with f containing interactionterms as in (8) havea representation as
alinearcombinationoftheunpenalizedtermsinthemodelplusn(basis)functions,thatwecan
constructfromthet(i),thereproducingkernelsassociatedwiththepenaltyfunctionals,andthe
smoothingparameters,seereferences3,4,11—14.Giventhesmoothingparameters,thenumerical
problemofcomputingtheminimizerf reducestofindingthecoefficientsofarepresentationforit
(cid:106)
asa linearcombinationofthe unpenalizedtermsand the abovementionedbasis functions.We
can solve the numerical problem for the coefficients with a Newton—Raphson iteration in
conjunction with matrix decompositions. The history of these numerical methods includes
references4,5,11,15—18.Therearevariousapproximatenumericalmethodsavailableandunder
developmentfordatasetsthataretoolargeformatrixdecompositionmethods;these,however,
are beyond the scope of this article.
Objective methods for choosing the smoothing parameters are desirable. This paper and the
code GRKPACK employ the iterative unbiased risk method given in Gu(cid:49)(cid:53)(cid:44)(cid:49)(cid:54) for the non-
Gaussiancaseandextendedtothepsmoothingparametercase(cid:106)"((cid:106) ,2,(cid:106) )inWang(cid:49)(cid:57)and
(cid:49) (cid:112)
WWGKK. This method is a computable proxy for the Kullback—Liebler information distance
from the estimate to the ‘truth’, see references 2, 16, 20 and 21. Thus, the method attempts to
choosethe smoothingparametersto minimizethe Kullback—Lieblerinformationdistancefrom
the estimate f and the unknown ‘true’ f. Other related references are 22—24.
(cid:106)
As with any estimation method, it is important to have some indication of its accuracy.
It is particularly important in the case here of non-parametric regression with Bernoulli data
derivedfrommedicalrecordsbecausethesedatatendtobeirregular,andmaycontaininfluential
outliers.Therigidityofparametricmodelsmaymasktheeffectofoutliers,aswellasobscurethe
fact that we sometimes make inferences in fairly data-sparse regions. The non-parametric
estimatesmaybemoresensitivetooutliers,anditisimportanttobeabletodelineatearegionin
thepredictorvariablespacewherewecanrelyupontheestimate,aswellaswherewecanprovide
some sort of confidence statement. In this work we use the Bayesian ‘confidence intervals’
proposedinreference25,adapatedtothecomponent-wisemultiplesmoothingparametercasein
reference 7, to non-Gaussian data in reference 26, and to the multiple smoothing parameter
non-Gaussian case in reference 19 and WWGKK. We use the Bayesian ‘confidence interval’
forf to delineatethe regioninpredictorvariablespace forwhichwedeemtheoverallestimate
(cid:106)
as reliable, by computing an appropriate level curve in a contour plot of the width of the
confidence interval, and using this to enclose a ‘reliable’ region. We also use it in conjunction
with cross-sectional plots so that we can visualize the accuracy estimates. We have used the
component-wise confidence intervals to eliminate some terms that we cannot distinguish from
noise, by deleting terms whose confidence intervals contain 0 over most of their domain. We
remarkthatwe base theseconfidenceintervals onan across-the-functionproperty,that is, a 95
per cent confidence interval has the property that (approximately) we expect the confidence
intervals to cover the true curve at about 95 per cent of the n data points, see references 3, 21
and 26.
We now proceed to the details of the particular models employed in the analysis of the
WESDRdata.Wecarryoutfouranalyses,eachofwhichillustratessomeparticularfeatureofthis
classof models.In eachcasewehave rescaledthe(continuous)predictorvariablest ,2, t to
(cid:49) (cid:100)
[0,1], andweusetheaveragingoperator(cid:69)"(cid:69)(cid:97)definedin(9).Weconsideronlymaineffectsand
twofactorinteractions.WeemploythepenaltyfunctionalJ of(7)andatwo-dimensionalrelative
(cid:97)
J defined below.If we eliminatedall of the twofactorinteractions, thenthe modelsdescribed
(cid:97)(cid:98)
Statist.Med.,Vol.16,1357—1376(1997) (cid:40)1997byJohnWiley&Sons,Ltd.
SMOOTHINGSPLINEANOVA 1361
immediately below reduce to the form (5) as studied by Hastie and Tibshirani,(cid:56) although our
estimation method is different.
We require somedefinitions to define the terms in the particular ANOVAdecomposition(8)
thatweuse.Definethe(linear)‘trend’function(cid:47)(u)"u!1/2foru3[0,1],and,forcontinuous
functions g defined on [0,1] let (cid:66)g stand for (cid:66)g"g(1)!g(0) . For future reference note that
(cid:69)(cid:47),(cid:58)(cid:49)(cid:47)(u)du"0 and (cid:66)(cid:47)"1. We use (cid:69) and (cid:66) to define ANOVA and identifiability side
conditio(cid:48)ns.Asubscript(cid:97)on(cid:69)or(cid:66)meansthatitappliestowhatfollowsconsideredasafunction
of t . The typical main effect term f(t ) in this set up has a decomposition of the form
(cid:97) (cid:97) (cid:97)
f (t )"d (cid:47)(t )#f (t ) (10)
(cid:97) (cid:97) (cid:97) (cid:97) (cid:52)(cid:97) (cid:97)
whered (cid:47)(t )isthelinearandunpenalizedpartoff andf isthe(detrended)‘smooth’partoff .
(cid:97) (cid:97) (cid:97) (cid:52)(cid:97) (cid:97)
f(cid:52)(cid:97)satisfiesthesideconditions(cid:69)(cid:97)f(cid:52)(cid:97)"(cid:66)(cid:97)f(cid:52)(cid:97)"0andappearsinsideapenaltyfunctionalasJ(cid:97)(f(cid:52)(cid:97)).
Due to the side conditions on f ,J (f )"0 implies that f "0.
(cid:97) (cid:97) (cid:52)(cid:97) (cid:52)(cid:97)
The two factor interaction f (t ,t ) has an analogous decomposition into four interaction
(cid:97)(cid:98) (cid:97) (cid:98)
terms, namely
f (t ,t )"d (cid:47)(t )(cid:47)(t )#(cid:47)(t ) f(cid:40)(cid:97)(cid:41)(t )#(cid:47)(t ) f(cid:40)(cid:98)(cid:41)(t )#f (t ,t ) (11)
(cid:97)(cid:98) (cid:97) (cid:98) (cid:97)(cid:98) (cid:97) (cid:98) (cid:97) (cid:52)(cid:98) (cid:98) (cid:98) (cid:52)(cid:97) (cid:97) (cid:52)(cid:97)(cid:98) (cid:97) (cid:98)
where the ‘smooth’ factors of the trend(cid:93)smooth interactions satisfy the side conditions
(cid:69)(cid:97)f(cid:52)(cid:40)(cid:97)(cid:98)(cid:41)"(cid:66)(cid:97)f(cid:52)(cid:40)(cid:97)(cid:98)(cid:41)"(cid:69)(cid:98)f(cid:52)(cid:40)(cid:98)(cid:97)(cid:41) "(cid:66)(cid:98)f(cid:52)(cid:40)(cid:98)(cid:97)(cid:41)"0 and the ‘smooth—smooth’ interaction term satisfies
((cid:69)(cid:97)f(cid:52)(cid:97)(cid:98))(t(cid:98))"((cid:66)(cid:97)f(cid:52)(cid:97)(cid:98))(t(cid:98))"((cid:69)(cid:98)f(cid:52)(cid:97)(cid:98))(t(cid:97))"((cid:66)(cid:98)f(cid:52)(cid:97)(cid:98))(t(cid:97))"0, for all t(cid:97),t(cid:98). Letting J(cid:97)(cid:98)(f(cid:52)(cid:97)(cid:98))"
(cid:58)(cid:49)(cid:58)(cid:49)( (cid:76)(cid:52) f (t ,t ))(cid:50)dt dt ,thenonecanshowthatthesideconditionsinsurethatJ (f )"0
im(cid:48)p(cid:48)lies(cid:76)t(cid:50)(cid:97)th(cid:76)ta(cid:98)(cid:50)t(cid:52)(cid:97)f(cid:98) "(cid:97) 0(cid:98). Lett(cid:97)ing(cid:98)(cid:106)"((cid:106) ,(cid:106) ,(cid:106)(cid:40)(cid:98)(cid:41),(cid:106)(cid:40)(cid:97)(cid:41),(cid:106) ) and (cid:97)(cid:98) (cid:52)(cid:97)(cid:98)
(cid:52)(cid:97)(cid:98) (cid:97) (cid:98) (cid:97) (cid:98) (cid:97)(cid:98)
(cid:74)(cid:106)(f)"(cid:106)(cid:97)J(cid:97)(f(cid:52)(cid:97))#(cid:106)(cid:98)J(cid:98)(f(cid:52)(cid:98))#(cid:106)(cid:40)(cid:97)(cid:98)(cid:41)J(cid:97)(f(cid:52)(cid:40)(cid:97)(cid:98)(cid:41))#(cid:106)(cid:98)(cid:40)(cid:97)(cid:41)J(cid:98)(f(cid:52)(cid:40)(cid:98)(cid:97)(cid:41))#(cid:106)(cid:97)(cid:98)J(cid:97)(cid:98)(f(cid:52)(cid:97)(cid:98)) (12)
we estimate f by finding (cid:107),d ,d ,d , and f ,f ,f(cid:40)(cid:98)(cid:41),f(cid:40)(cid:97)(cid:41) and f to minimize
(cid:97) (cid:98) (cid:97)(cid:98) (cid:52)(cid:97) (cid:52)(cid:98) (cid:52)(cid:97) (cid:52)(cid:98) (cid:52)(cid:97)(cid:98)
(cid:73)(cid:106)(y,f)"(cid:76)(cid:106)(y,f)#(cid:74)(cid:106)(f). (13)
Duetothesideconditions,asacomponentof(cid:106)becomeslarge,theestimateofthe‘smooth’term
thatitmultipliesbecomessmall.Thus,agoodmethodforchoosingthecomponentsof(cid:106)fromthe
data effectively eliminates unneeded terms in the expansion if one estimates their companion
smoothing parameters as large.
We can include categorical variables in the model, by, for example, letting
(cid:75)
f(t ,t ,z)"(cid:107)#f (t )#f (t )#f (t )#(cid:43) (cid:99) I (z) (14)
(cid:97) (cid:98) (cid:97) (cid:97) (cid:98) (cid:98) (cid:97)(cid:98) (cid:97)(cid:44)(cid:98) (cid:107) (cid:107)
(cid:107)(cid:47)(cid:50)
wherezisavariablewithKpossiblevaluesz ,2,z andI (z)"1ifz"z and0otherwise.We
(cid:49) (cid:75) (cid:107) (cid:107)
then include the (cid:99)’s in the minimization of (13).
Themathematicsbehindtherepresentationoff,thenumericalmethodfortheminimizationof
(cid:73)(cid:106),thedata-basedchoiceof(cid:106)accordingtotheiterativeunbiasedriskmethod,andthecalculation
oftheconfidenceintervalsaredescribedinWWGKK.Furtherdetailsappearinreference3and
the documentation for GRKPACK. In the four analyses that we carry out below, there are
betweend"2andd"4predictorvariables.Inthed"4case,consideringthepenaltyfunctional
in(12),andincludingall ofthepossiblemaineffectsandtwofactorinteractiontermswiththeir
own smoothing parameters, the result is 4 main effects smoothing parameters ((cid:106) ), 12
(cid:97)
trend(cid:93)smoothsmoothingparameters((cid:106)(cid:40)(cid:98)(cid:41))and6smooth(cid:93)smoothparameters((cid:106) ),morethan
(cid:97) (cid:97)(cid:98)
wewouldwanttofitsimultaneouslywiththesamplesizeshere(alllessthan500).Wereducedthe
(cid:40) 1997byJohnWiley&Sons,Ltd. Statist.Med.,Vol.16,1357—1376(1997)
1362 Y.WANGE„A‚.
numberoftermswithsmoothingparameterstoamaximumof7basedonpreviouslypublished
analyses of these data by more traditional methods, by extensive pre-screening, by fitting
parametricpolynomialGLIMmodelswithtermsashighascubicordertodetectanddeleteterms
thatfailedtogiveanyevidenceofsignificance,andbyfittingsomemarginalSS-ANOVAmodels
asdemonstratedinSection6.Explorationofmoresystematicpre-screeningmethodsisanareaof
activeresearch.Oncewehadreducedthenumberofsmoothingparametersto7orless,weused
informal model selection methods as discussed in WWGKK. These included the deletion of
component terms too small to have an observable effect on cross-sectional plots of f, and the
deletion of terms whose componentwise confidence intervals included the 0 function. Further
detailsspecifictoeachanalysisappearbelow.Leaving-out-one-thirdmodelselectionprocedures
have been discussed in reference 20, but we have not used them here.
Section 2 discusses the Wisconsin Epidemiologic Study of Diabetic Retinopathy, and
Sections3,4,5and6discussfouranalysesofthisstudy.Section7isasummaryandconclusion.
2. WISCONSIN EPIDEMIOLOGICSTUDY OF DIABETIC RETINOPATHY(WESDR)
TheWESDRisanongoingepidemiologicstudyofacohortofpatientswhoreceivetheirmedical
care in an 11-county area in southern Wisconsin, who were first examined in 1980—1982, then
again in 1984—1986 and 1990—1992. At this writing a third follow-up is currently in progress.
Detaileddescriptionsof thedatahaveappearedinreferences27and28andreferencesthere.In
brief,asampleof2990diabeticpatientswasselectedinan11-countyareainsouthernWisconsin.
This sample consisted of two groups. The first group was 1210 patients diagnosed as diabetic
beforetheywere30yearsofageandwhotookinsulin(‘youngeronsetgroup’).Thesecondgroup
consistedof1780patientsdiagnosedwithdiabetesafter30yearsofage.Ofthese,824weretaking
insulin (‘older onset group taking insulin’) and 956 were not (‘older onset group not taking
insulin’).Ofthe2990eligiblepatients,2366participatedinthebaselineexaminationfrom1980to
1982.
A large number of medical, demographic, ocular and other covariates were recorded at the
baseline and later examinations along with a retinopathy score for each eye (to be described).
Relationsbetweenvariouscovariatesandtheretinopathyscoreshavehadextensiveanalysiswith
standardstatisticalmethodsincludingcategoricaldataanalysis andlinearlogistic models,with
resultsreportedinaseriesof WESDRmanuscripts29—35.WeappliedSS-ANOVAmethodsto
asubsetofthedatafromtheyoungeronsetgroupinWWGKKinconjunctionwithanextensive
technicalaccountofthemathematicaltheoryandnumericalmethodsbehindthemethod.Itisthe
purpose of this account to carry out four SS-ANOVA analyses on data from the older onset
WESDR group, and in the process, to illustrate some of the more important features of the
method. Our goal is to explain the possible results, advantages and disadvantages in a less
technical manner, aimed at clinicians and medical researchers.
We limited the present study to the construction of predictive models for incidence and for
progression(tobedefined)ofdiabeticretinopathyatthefirstfollow-up,asafunctionofsomeof
thecovariatesavailableatbaseline.Wedothisbothforthe‘olderonsetgrouptakinginsulin’(ID
group) and the ‘older onset group not taking insulin’ (NID group). We analyse both ‘incident
events’and‘progressionevents’forboththeIDandtheNIDgroups.Weonlylistthecovariates
pertinent to our analysis:
1. Age: age at the examination (years);
2. Duration: duration of diabetes at the examination (years);
3. Glycosylated haemoglobin: a measure of hyperglycaemia (%);(cid:56)
Statist.Med.,Vol.16,1357—1376(1997) (cid:40)1997byJohnWiley&Sons,Ltd.
SMOOTHINGSPLINEANOVA 1363
4. Body mass index (bmi): weight in kg/(height in m)(cid:50);
5. Pulse rate counted for 30 seconds;
6. Baseline retinopathy severity levels (base-retinopathy-level).See explanation below.
At the baseline and follow-up examinations, each eye was graded as one of the 6 levels:
10,21,31,41,51 and 60#, in order of increasing retinopathy severity with 10 indicating no
retinopathyand60#indicatingthemostseverestage,proliferativeretinopathy.Wederivedthe
retinopathy level for a participant by giving the eye with the higher level (more severe re-
tinopathy) greater weight. For example, we specify the level for a participant with level 31
retinopathyineacheyewiththenotation‘level31/31’;foraparticipantwithlevel31inoneeye
andlesssevereretinopathyintheothereyethenotationis‘level31/(31’.Thisschemeprovided
an 11-step scale: 10/10; 21/(21; 21/21; 31/(31; 31/31; 41/(41; 41/41; 51/(51; 51/51;
60#/(60#,and60#/60#.Participantsintheanalysisforincidenceconsistedofsubjectswith
level10/10atthebaselineandnomissingdata.Themodelprovidesanestimateoftheincidence
rate,definedastheprobabilitythatsuchaparticipanthaslevel21/(21orworseatthefollow-up
examination. Participants in the analysis for progression consisted of subjects with no or
non-proliferativeretinopathyatbaselineandnomissingdata.Themodelprovidesanestimateof
theprogressionrate,withprogressiondefinedasanincreaseinbaselinelevelbytwostepsormore
(10/10 to 21/21 or greater, or 21/(21 to 31/(31 or greater, for instance). These analyses
correspond to analyses previously carried out by traditional methods in some of the WESDR
referencescited.Notethatthisallowssubjectsincludedintheincidenceanalysisalsotobepartof
the progression analysis.
3. INCIDENCE IN THE OLDER ONSET GROUP NOT TAKING INSULIN
After excluding participants with missing values, there were 297 participants in the older onset
NID group with a baseline score of 10/10, thus qualifying them for inclusion in the NID
‘Incidence’analysis.Therewasoneobservationofglycosylatedhaemoglobinrecordedas23·6
per cent, much greater than the others. We decided to delete this influential observation. Our
conclusions would remain the same if we included this observation in the analysis.
Kleinetal.(cid:50)(cid:55)found,usinglinearlogisticmodels,thatglycosylatedhaemoglobinistheonly
statisticallysignificantpredictorofincidenceofretinopathyinolderonsetpatients.Usinglinear
logisticmodelsasascreeningtool,wefoundthattheeffectof ageis notlinearand thatthereis
a strong interactionbetween age and glycosylated haemoglobin.
We first fit an SS-ANOVA model with the main effects of age and glycosylated
haemoglobinandallthreeinteractionterms.Themaineffectofglycosylatedhaemoglobin
islinear.Allinteractiontermsexcepttrend(glycosylatedhaemoglobin)(cid:93)smooth(age)arenear
zero. The final model is
f(age, glycosylated haemoglobin)
"(cid:107)#f (age)#a (cid:93)glycosylated haemoglobin
(cid:49) (cid:49)
#trend(glycosylated haemoglobin) (cid:93)smooth(age). (15)
We plot age versus glycosylated haemoglobin on Figure 1(a). We have marked those
participants with incident retinopathy as solid circles and those without as open circles. We
superimposethecontourlinesofestimatedposteriorstandarddeviations.Thesecontoursagree
wellwiththedistributionoftheobservations.Thuswecanusethemtodelineatearegioninwhich
(cid:40) 1997byJohnWiley&Sons,Ltd. Statist.Med.,Vol.16,1357—1376(1997)
1364 Y.WANGE„A‚.
Figure1. NID‘Incidence’analysis:(a)dataandcontoursofconstantposteriorstandarddeviation.Solidcirclesindicate
incidenceandopencirclesindicatenon-incidence;(b)estimatedprobabilityofincidenceinthedefinedregion,asafunction
ofageandglycosylatedhaemoglobin
Statist.Med.,Vol.16,1357—1376(1997) (cid:40)1997byJohnWiley&Sons,Ltd.
SMOOTHINGSPLINEANOVA 1365
Figure2. NID‘Incidence’analysis.Cross-sectionsofestimatedprobabilityofincidenceasafunctionofage,withtheir90
per cent Bayesian confidence intervals, at four quantiles of glycosylated haemoglobin. q1, q2, q3 and q4 are the
quantilesat0·125,0·375,0·625and0·875
wedeem the estimateof the probabilityfunctionas reliable.We decidedto use the region with
estimated posterior standard deviations less than or equal to 1, in the logit scale.
We plot the probability functionestimate on Figure 1(b). Figure 2 gives cross-sectionsof the
estimatedprobabilityofincidenceasafunctionofagewiththeir90percentBayesianconfidence
intervals at the cross-sections, at four quantiles of glycosylated haemoglobin. The width of
theconfidenceintervalssuggeststhatthesmallbumpsareprobablyanartifact.Thegeneralshape
of the response, however, is evident, and it appears that the age effect is not particularly well
modelled with a second or even a third degree polynomial.
The risk for incidence of retinopathy increases with increasing glycosylated haemoglobin. The
risk increases with increasing age for lower glycosylated haemoglobin. This increase might result
fromthedevelopmentofotherconditionssuchashypertensionandatheroscleroticvasculardisease
inoldercomparedtoyoungersubjectsandthatleadstoincreasedriskofretinopathy,evenwhen
the glycosylated haemoglobin level is relatively low. The risk decreases with increasing age for
higherglycosylatedhaemoglobin.Thisdecreasemayresultfrommortalitysinceanoldparticipant
with high glycosylated haemoglobin has a higher mortality risk during the observation period.(cid:51)(cid:54)
(cid:40) 1997byJohnWiley&Sons,Ltd. Statist.Med.,Vol.16,1357—1376(1997)
1366 Y.WANGE„A‚.
Figure3. NID‘Progression’analysis.Maineffectsestimatesofglycosylatedhaemoglobinandbmialongwiththeir
component-wise90percentBayesianconfidenceintervals,logitscale
4. PROGRESSION OF THE OLDER ONSET GROUP NOT TAKING INSULIN
After excluding participants with missing values, there were 432 participants in the older onset
NID group qualified for inclusion in the ‘Progression’ analysis. Klein et al.(cid:50)(cid:55) found that age,
duration, and glycosylated haemoglobin were statistically significant in a linear logistic
model.Usinglinearlogisticmodels,wefoundthatthedurationmaineffectofpolynomialsupto
thecubicissignificantandthebmimaineffectofpolynomialsuptothequarticissignificant.We
also found that the multiplication interaction between age and polynomials up to cubic of
durationissignificant.Theseindicatethatalinearlogisticmodelwithlowerorderpolynomials
may be inadequate. We decided to fit the SS-ANOVA model:
f(age, duration, glycosylated haemoglobin, bmi)
"(cid:107)#f (age)#f (duration)#f (age, duration)
(cid:49) (cid:50) (cid:49)(cid:44)(cid:50)
#f (glycosylated haemoglobin)#f (bmi). (16)
(cid:51) (cid:52)
Figure3plotsthemaineffectsestimatesandtheir90percentBayesianconfidenceintervalsfor
glycosylatedhaemoglobinandbmi.Weseethattheeffectofglycosylatedhaemoglobinis
strong.Theeffectofbmiissmall,and0iscontainedwithinitsconfidenceinterval.Theestimates
areeffectivelylinearonalogitscaleeventhoughwealloweda‘smooth’termforeachofthemin
the model. This illustrates the ability of the method to reduce to a partially parametric form if
warranted by the data.
The effects of glycosylated haemoglobin and bmi are additive in the logit scale. In the
remaining plots, we fix glycosylated haemoglobin and bmi at their median values.
WeplotageversusdurationonFigure4(a).Participantswhohadprogressionofretinopathy
appear as solid circles and those with no progressionas open circles. We superimposecontour
lines of the estimated posterior standard deviations as a function of age and duration with
glycosylatedhaemoglobin and bmi fixed attheir medianvalues. Thesecontoursagree well
with the distribution of the observations. We decided to use the region with estimated
Statist.Med.,Vol.16,1357—1376(1997) (cid:40)1997byJohnWiley&Sons,Ltd.