How to reduce bias in the estimates of count data regression? · 2017. 10. 8. · • Count Data...

24
How to reduce bias in the estimates of count data regression? Ashwini Joshi Sumit Singh PhUSE 2015, Vienna

Transcript of How to reduce bias in the estimates of count data regression? · 2017. 10. 8. · • Count Data...

Page 1: How to reduce bias in the estimates of count data regression? · 2017. 10. 8. · • Count Data • Poisson Regression • Maximum Likelihood Estimate and Bias Reduction Method •

Howtoreducebiasintheestimatesofcountdataregression?

AshwiniJoshiSumitSingh

PhUSE2015,Vienna

Page 2: How to reduce bias in the estimates of count data regression? · 2017. 10. 8. · • Count Data • Poisson Regression • Maximum Likelihood Estimate and Bias Reduction Method •

13-Oct-15 2PhUSE2015:SP03

PrecisionProblem

bias

more less

more less

Page 3: How to reduce bias in the estimates of count data regression? · 2017. 10. 8. · • Count Data • Poisson Regression • Maximum Likelihood Estimate and Bias Reduction Method •

• CountData• PoissonRegression• MaximumLikelihoodEstimateandBiasReductionMethod

• ProfileLikelihoodBasedConfidenceInterval• NegativeBinomialRegression

Agenda

13-Oct-15 PhUSE2015:SP03 3

Page 4: How to reduce bias in the estimates of count data regression? · 2017. 10. 8. · • Count Data • Poisson Regression • Maximum Likelihood Estimate and Bias Reduction Method •

• Numberofadverseeventsoccurringduringafollowupperiod

• Numberoflesionsornumberofrelapsesinmultiplesclerosispatients

• Numberofhospitalizations• Numberofseizuresinepileptics• Andmanymore…

13-Oct-15 4PhUSE2015:SP03

Count Data in Clinical Trials

Page 5: How to reduce bias in the estimates of count data regression? · 2017. 10. 8. · • Count Data • Poisson Regression • Maximum Likelihood Estimate and Bias Reduction Method •

• ParameterEstimation• MaximumLikelihoodEstimate(MLE)

– Biasinthecaseofsmallsampledata– AsymmetricWaldConfidenceInterval

13-Oct-15 5PhUSE2015:SP03

Poisson Regression

Page 6: How to reduce bias in the estimates of count data regression? · 2017. 10. 8. · • Count Data • Poisson Regression • Maximum Likelihood Estimate and Bias Reduction Method •

• BiasCorrection:MLEcorrectedforbias

• BiasReduction:Likelihoodfunctionorscoreequationsaremodified

• Firth(1993):penalizationofthelikelihoodorscorefunctioninoppositedirectionofbias

13-Oct-15 6PhUSE2015:SP03

Bias correction and Reduction Methods

Page 7: How to reduce bias in the estimates of count data regression? · 2017. 10. 8. · • Count Data • Poisson Regression • Maximum Likelihood Estimate and Bias Reduction Method •

13-Oct-15 7PhUSE2015:SP03

Firth Method

:FisherInformation

Page 8: How to reduce bias in the estimates of count data regression? · 2017. 10. 8. · • Count Data • Poisson Regression • Maximum Likelihood Estimate and Bias Reduction Method •

• ForSingleCovariate

FindCIlimitswherelikelihoodsatisfy

• Incaseofmultiplecovariates– Fixparameterfordesiredcovariate– maximizethelikelihoodfunctionoverallotherparameters– Checktheabovecondition– Keepchangingparametervaluetillaboveconditionissatisfied

13-Oct-15 8PhUSE2015:SP03

Profile likelihood based Confidence Interval

21,1max0 5.0 αχ −−= ll

maxl

0l0l

Page 9: How to reduce bias in the estimates of count data regression? · 2017. 10. 8. · • Count Data • Poisson Regression • Maximum Likelihood Estimate and Bias Reduction Method •

• Aspectsofbiasreductionmethodforsmallsampledata

• ExistenceinthecaseofQuasi-separation

• Performanceincaseoflargedata

13-Oct-15 9PhUSE2015:SP03

Scenarios

Page 10: How to reduce bias in the estimates of count data regression? · 2017. 10. 8. · • Count Data • Poisson Regression • Maximum Likelihood Estimate and Bias Reduction Method •

13-Oct-15 10PhUSE2015:SP03

Example 1: Multiple Sclerosis• CommonlyusedEndPoints:

– NumberofRelapses– Annualizedrelapserate(ARR)

• Relapseistheappearanceofanewneurologicalabnormality

FitPoissonModel:numberofrelapses~BaselineEDSSEDSS:ExpandedDisabilityStatusScore

Page 11: How to reduce bias in the estimates of count data regression? · 2017. 10. 8. · • Count Data • Poisson Regression • Maximum Likelihood Estimate and Bias Reduction Method •

data MultScler;inputSubID EDSSAgeSexCode RaceCode ArmCountTime;cards;1 1 19 2 1 1 0 7582 1 54 1 2 2 0 57...31 2 53 2 1 1 0 75132 2 46 1 1 0 1 449;PROC LOGXACT DATA=MultScler;MODELCount=EDSS/link=Poisson;ES/ASEDSS;RATETime;RUN;

PROC LOGXACT DATA=MultScler;MODELCount=EDSS/link=Poisson;ES/ASPMLEEDSS;RATETime;RUN;

13-Oct-15 11PhUSE2015:SP03

Example 1 (cont.):Proc LogXact

Page 12: How to reduce bias in the estimates of count data regression? · 2017. 10. 8. · • Count Data • Poisson Regression • Maximum Likelihood Estimate and Bias Reduction Method •

TheSASSystem16:15Saturday,August1,20154

Outputfrom LogXact(r)(v11.0)PROCs _SAS9_1Copyright(c)1997-2015 CytelInc.,Cambridge,MA,USA.

--------------------------------------------------------------------------------------------COUNTREGRESSION--------------------------------------------------------------------------------------------BASICINFORMATION:Datafile:MULTSCLERModel:Count=Intercept+EDSSLinktype:PoissonRateMultiplier:TimeStratumvariable:<Unstratified>Analysistype:Estimate::AsymptoticNumber oftermsinmodel:2Number ofterm(s)dropped: 0Sumof RateMultiplier:20654Number ofrecordsrejected:0Number ofgroups:32--------------------------------------------------------------------------------------------SUMMARYSTATISTICS:--------------------------------------------------------------------------------------------StatisticValueDFP-valueDeviance32.8859300.3275--------------------------------------------------------------------------------------------PARAMETERESTIMATE:--------------------------------------------------------------------------------------------

PointEstimateConfidenceIntervalandPValueforBetaModelTermTypeBetaSEType95.0%C.I.P-Value

LowerUpper2*1-sided--------------------------------------------------------------------------------------------InterceptMLE-5.23581.1002Asymptotic-7.3922-3.07930.0000EDSSMLE-1.53661.0260Asymptotic-3.54750.47430.1342--------------------------------------------------------------------------------------------AnalysisTime=00:00:00

13-Oct-15 12PhUSE2015:SP03

Example 1 (cont.): Output

Page 13: How to reduce bias in the estimates of count data regression? · 2017. 10. 8. · • Count Data • Poisson Regression • Maximum Likelihood Estimate and Bias Reduction Method •

PARAMETERESTIMATE:--------------------------------------------------------------------------------------------

PointEstimate ConfidenceIntervalandPValueforBetaModelTermType BetaSEType 95.0%C.I.P-Value

LowerUpper2*1-sided-------------------------------------------------------------------------------------------------------------------------------InterceptMLE-5.23 Asymptotic -7.3922-3.07930.0000EDSSMLE-1.54 Asymptotic -3.54750.47430.1342-------------------------------------------------------------------------------------------------------------------------------

PARAMETERESTIMATE:-------------------------------------------------------------------------------------------------------------------------------

PointEstimate ConfidenceIntervalandPValueforBetaModelTermType BetaSEType 95.0%C.I.P-Value

LowerUpper2*1-sided-------------------------------------------------------------------------------------------------------------------------------InterceptPMLE-5.58 Asymptotic -7.4193-3.75932.706e-009EDSSPMLE-1.16 Asymptotic -2.81780.50360.1721-------------------------------------------------------------------------------------------------------------------------------

13-Oct-15 13PhUSE2015:SP03

Example 1 (cont.): Comparison

1.10

0.93

1.03

0.85

1.10

0.930.85

1.03

Page 14: How to reduce bias in the estimates of count data regression? · 2017. 10. 8. · • Count Data • Poisson Regression • Maximum Likelihood Estimate and Bias Reduction Method •

13-Oct-15 14PhUSE2015:SP03

Example 1 (cont.)

Page 15: How to reduce bias in the estimates of count data regression? · 2017. 10. 8. · • Count Data • Poisson Regression • Maximum Likelihood Estimate and Bias Reduction Method •

• ProblemofSeparation(PerfectFit)occurswhen– Responsevaluedividessetofcovariateorcombinationsofcovariates

– Acovariatevaluepredictsvalueoftheresponse

ForAge<60,Events=0:Separation

ForGender:Quasi-separation

13-Oct-15 15PhUSE2015:SP03

Separation, Quasi-separation

Age Gender Eventsold Male 1old Male 1old Female 1

young Female 0young Female 0young Female 0

Page 16: How to reduce bias in the estimates of count data regression? · 2017. 10. 8. · • Count Data • Poisson Regression • Maximum Likelihood Estimate and Bias Reduction Method •

AnimalData:DatafromHeinze andPuhr (2010).Study:Toinvestigateeffectofheparinized vs non-heparinized,

vascularsubstituteinratsonaneurysmformation.

Theevent ofinterestisaneurysmformation.Thecovariate isnon-heparinized implant.Stratumisfollowuptime.

• Model:Aneur ~Nohep• CaseofQuasi-separation• MLEcannotbeobtained

13-Oct-15 16PhUSE2015:SP03

Example 2

Page 17: How to reduce bias in the estimates of count data regression? · 2017. 10. 8. · • Count Data • Poisson Regression • Maximum Likelihood Estimate and Bias Reduction Method •

--------------------------------------------------------------------------------------------PARAMETERESTIMATE:--------------------------------------------------------------------------------------------

PointEstimateModelTerm Type Beta SE--------------------------------------------------------------------------------------------NoHep MLE ? ?--------------------------------------------------------------------------------------------NoHep PMLE 2.1970 1.6665--------------------------------------------------------------------------------------------

?:non-convergence

13-Oct-15 17PhUSE2015:SP03

Example 2 (cont.)

Page 18: How to reduce bias in the estimates of count data regression? · 2017. 10. 8. · • Count Data • Poisson Regression • Maximum Likelihood Estimate and Bias Reduction Method •

13-Oct-15 18PhUSE2015:SP03

Example 2 (cont.)

Page 19: How to reduce bias in the estimates of count data regression? · 2017. 10. 8. · • Count Data • Poisson Regression • Maximum Likelihood Estimate and Bias Reduction Method •

• Poisson:Variance=mean

• OverdispersioninPoisson:Variance>mean

• NB-2variance=μ +a.μ2

13-Oct-15 19PhUSE2015:SP03

Negative Binomial Regression

Page 20: How to reduce bias in the estimates of count data regression? · 2017. 10. 8. · • Count Data • Poisson Regression • Maximum Likelihood Estimate and Bias Reduction Method •

• Data:FromtheUSnationalMedicareinpatienthospitaldatabase(Medpar)forthe1991MedicarefilesforthestateofArizona(Hilbe2011).

• RESPONSE

los:lengthofstayinthehospital

• PREDICTORS

hmo:PatientbelongstoaHealthMaintenanceOrganization(1),orprivatepay(0)

white:PatientidentifiesthemselvesasprimarilyCaucasian(1)incomparisontonon-white(0)

type:Athree-levelfactorpredictorrelatedtothetypeofadmission.1=elective(referent),2=urgent,3=emergency

• FitNB-2 model:los~factor(type)13-Oct-15 20PhUSE2015:SP03

Example 3 – Medpar Data

Page 21: How to reduce bias in the estimates of count data regression? · 2017. 10. 8. · • Count Data • Poisson Regression • Maximum Likelihood Estimate and Bias Reduction Method •

---------------------------------------------------------------------------------------------------------------------------------------PARAMETERESTIMATE:---------------------------------------------------------------------------------------------------------------------------------------

PointEstimate ConfidenceIntervalandPValueforBetaModelTermType BetaSE Type 95.0%C.I. P-Value

LowerUpper2*1-sided---------------------------------------------------------------------------------------------------------------------------------------

InterceptMLE2.1783 Asymptotic2.13472.22190.0000type_2MLE0.2379 Asymptotic0.13940.33630.0000type_3MLE0.7252 Asymptotic0.57680.87362.04e-019------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------InterceptPMLE2.1781 Asymptotic2.13452.22170.0000type_2PMLE0.2368Asymptotic0.13840.33520.0000type_3PMLE0.7235 Asymptotic0.57520.87172.278e-019---------------------------------------------------------------------------------------------------------------------------------------

13-Oct-15 21PhUSE2015:SP03

Example 3 (cont.)

0.02220.0502

0.0757

0.02220.0502

0.0756

Page 22: How to reduce bias in the estimates of count data regression? · 2017. 10. 8. · • Count Data • Poisson Regression • Maximum Likelihood Estimate and Bias Reduction Method •

• PropertiesofBiasreducedEstimatesforsmallsampledata:– Smallerstandarderrors– Shorterconfidenceintervals– Existenceinthecaseofquasi-separation

• ProfileLikelihoodbasedCI– betterthanWaldCIforasymmetriclikelihoodfunction

13-Oct-15 22PhUSE2015:SP03

Concluding remarks

Page 23: How to reduce bias in the estimates of count data regression? · 2017. 10. 8. · • Count Data • Poisson Regression • Maximum Likelihood Estimate and Bias Reduction Method •

• FirthD.(1993).“Biasreductionofmaximumlikelihoodestimates”.Biometrika (1993),80,1,pp.27-38

• Heinze G.,Puhr R.(2010).“Bias-reducedandseparation-proofconditionallogisticregressionwithsmallorsparsedatasets”.StatisticsinMedicine

• HilbeJ.(2011).“NegativeBinomialRegression”.CambridgeUniversityPress

• Kosmidis I.(2007).“BiasReductioninExponentialFamilyNonlinearModels”.

• Kosmidis I.(2010).“Agenericalgorithmforreducingbiasinparametricestimation”.ElectronicJournalofStatistics.Vol.4(2010)1097–1112

• LogXact11SoftwareandUserManual13-Oct-15 23PhUSE2015:SP03

References

Page 24: How to reduce bias in the estimates of count data regression? · 2017. 10. 8. · • Count Data • Poisson Regression • Maximum Likelihood Estimate and Bias Reduction Method •

13-Oct-15 24PhUSE2015:SP03

Thankyou

[email protected]@cytel.com