Introduction to mixed model and missing data issues in longitudinal studies

Size: px

Start display at page:

Download "Introduction to mixed model and missing data issues in longitudinal studies"

Geraldine Meagan Mason
9 years ago
Views:

1 Introduction to mixed model and missing data issues in longitudinal studies Hélène Jacqmin-Gadda INSERM, U897, Bordeaux, France Inserm workshop, St Raphael

2 Outline of the talk I Introduction Mixed models Typology of missing data Exploring incomplete data Methods MAR data Conclusion

3 Longitudinal data : definition Definition : Variables measured at several times on the same subjects Examples : repeated measures of biological markers (CD4, HIV RNA) in HIV patients repeated measures of neuropsychological tests to study cognitive aging Repeated events : dental caries, absences from school or job,...

HIV RNA) in HIV patients repeated measures of neuropsychological tests to study

4 Longitudinal data analysis Objective : Describe change of the variable with time Identify factors associated with change Problem : Intra-subject correlation

5 Example : HIV clinical trial X i =1 if treatment A, X i =0 if treatment B Criterion : Change over time of CD4 Repeated measures of CD4 over the follow-up period. t = 0 at initiation of treatment. Y ij = CD4 measure for subject i at time t ij, i = 1,..., N, j = 1,..., n i.

of CD4 over the follow-up period. t = 0 at initiation of treatment.

6 Analysis assuming independence Y ij = β 0 + β 1 t ij + β 2 X i + β 3 X i t ij + ǫ ij with ǫ ij N(O, σ 2 ) and ǫ ij ǫ ij Intra-subject correlation ˆ Var( ˆβ) biased Tests for β biased For time-independent covariate : var(ˆβ 2 ) under-estimated Tests for H 0 : β 2 = 0 anti-conservative (p value too small)

Var( ˆβ) biased Tests for β biased For time-independent covariate : var(ˆβ

7 Linear mixed model with random intercept Y ij = (β 0 + γ 0i ) + β 1 t ij + β 2 X i + β 3 X i t ij + ǫ ij with γ 0i N(O, σ 2 0 ), and ǫ ij N(O, σ 2 ) and ǫ ij ǫ ij γ 0i are random variables Only one additional parameter : σ 2 0

8 Linear mixed model with random intercept (2) Population (marginal) mean : E(Y ij ) = β 0 + β 1 t ij + β 2 X i + β 3 X i t ij Subject-specific (conditional) mean : E(Y ij γ 0i ) = (β 0 + γ 0i ) + β 1 t ij + β 2 X i + β 3 X i t ij Assume common correlation between all the repeated measures

(conditional) mean : E(Y ij γ 0i ) = (β 0 + γ 0i ) + β 1 t ij + β 2 X i

9 Linear mixed model with random intercept and slope Y ij = (β 0 + γ 0i ) + (β 1 + γ 1i )t ij + β 2 X i + β 3 X i t ij + ǫ ij, γ 0i N(O, σ 2 0 ), γ 1i N(O, σ 2 1 ), ǫ ij N(O, σ 2 ), ǫ ij ǫ ij Population (marginal) mean : E(Y ij ) = β 0 + β 1 t ij + β 2 X i + β 3 X i t ij Subject-specific (conditional) mean : E(Y ij γ i ) = (β 0 + γ 0i ) + (β 1 + γ 1i )t ij + β 2 X i + β 3 X i t ij The correlation between repeated measures depend on measurement times

ij ) = β 0 + β 1 t ij + β 2 X i + β 3 X i t ij Subject-specific (conditional) mean : E(Y ij γ i ) = (β 0 + γ 0i

10 Linear mixed model : general formulation Y ij = X T ijβ + Z T ijγ i + ǫ ij γ i N(0, B) and ǫ i N(0, R i ). X ij : vector of explanatory variables β : vector of fixed effects Z ij : sub-vector of X ij (including functions of time) γ i : vector of random effects. Population (marginal) mean : E(Y ij ) = X T ij β Subject-specific (conditional) mean : E(Y ij γ i ) = X T ij β + ZT ij γ i

X ij : vector of explanatory variables β : vector of fixed effects Z ij : sub-vector of X ij

11 Linear mixed model : example Linear mixed model with AR Gaussian error Y ij = (β 0 + γ 0i ) + (β 1 + γ 1i )t ij + β 2 X i + β 3 X i t ij + w ij + e ij with γ t i = (γ 0i, γ 1i ) N(0, B), e ij N(O, σ 2 ), e ij e ij, w ij N(O, σ 2 w) and Corr(w ij, w ij ) = exp( δ t ij t ij )

12 Linear mixed model : Estimation Maximum likelihood estimator Y i = (Y i1,..., Y ij,..., Y ini ) T multivariate Gaussian with mean X i β and covariance matrix V i = Z i BZ T i + R i Softwares : SAS Proc mixed, R lme, stata

13 Generalized linear mixed model Y ij exponential family of distribution and g(e(y ij γ i )) = X T ijβ + Z T ijγ i with γ i N(O, B). Example : Logistic mixed model logit(pr(y ij = 1 γ i )) = Xijβ T + Zijγ T i with γ i N(0, B). Maximum likelihood estimation : Numerical integration Softwares : SAS Proc nlmixed, R nlme, stata

Example : Logistic mixed model logit(pr(y ij = 1 γ i )) = Xijβ T + Zijγ T i with

14 Typology of missing data in longitudinal studies Notation : Y i = (Y obs,i, Y mis,i ) with Y obs,i the observed part of Y i and Y mis,i the missing part, R ij = 1 if Y ij is observed and R ij = 0 if Y ij is missing R i = (R i1,..., R ij,..., R ini ) X i explanatory variables completely observed

missing part, R ij = 1 if Y ij is observed and R ij = 0 if Y ij is

15 Typology of missing data (2) Monotone missing data = dropout : P(R ij = 0 R ij 1 = 0) = 1 R i may be summarized by the time to dropout T i and an indicator for dropout δ i Intermittent missing data : P(R ij = 0 R ij 1 = 0) < 1

16 Typology of missing data (3) Missing Completely at random (MCAR) : P(R ij = 1) is constant The observed sample is representative of the whole sample. Loss of precision, no bias Covariate-dependent missingness process : P(R ij = 1) = f(x i ) Loss of precision, no bias if analyses are adjusted on X i

Loss of precision, no bias Covariate-dependent missingness process : P(R

17 Typology of missing data (4) Missing at random (MAR) : P(R ij = 1) = f(y obs,i, X i ) Example : Probability of dropout depends on past observed values Loss of precision, no bias with appropriate statistical methods Informatives or MNAR : P(R ij = 1) = f(y mis,i, Y obs,i, X i ) Example : Probability that Y be observed depends on current Y value Loss of precision, biases Sensitivity analyses

appropriate statistical methods Informatives or MNAR : P(R ij = 1) = f(y mis,i, Y obs,i, X i )

18 Exploring incomplete data Describe missing data frequency Cross classify missing data patterns with covariates Compare mean evolution for available data and complete cases Compare mean evolution until time t given observation status at time t + 1 Logistic regression for P(R ij = 1) given covariates and Y ik, k < j Cox regression for time to dropout given covariates Impossible to distinguish MAR from MNAR

time t given observation status at time t + 1 Logistic regression for P(R ij = 1) given covariates

19 An example : Paquid data set The Paquid Cohort in Gironde 2792 subjects of 65 years and older at baseline Living at home at the beginning of the study (1988) in Gironde (France) Seen at home at 1, 3, 5, 8, and 10 years after the baseline visit Cognitive measure : Digit Symbol Substitution Test of Wechsler (attention, limited time to 90s) Sample : 2026 subjects without diagnosis of dementia between T0 and T10 with the test completed at least once (at T0)

the baseline visit Cognitive measure : Digit Symbol Substitution Test of Wechsler (attention, limited time to 90s)

20 Description of dropout : Kaplan-Meyer Dropout time (=event) : first visit with missing score Probability to be in the cohort 1 95% confidence interval Kaplan-Meyer estimate Probability Follow-up time

cohort 1 95% confidence interval Kaplan-Meyer estimate 0.9 0.

21 Observed means of the DSST score given time 40 Available data Score years and + Age

22 Observed means of the DSST score given time 40 Complete data Available data Score years and + Age

23 Logistic regression model for dropout in the first 5 years Covariates OR 95% CI of the OR T T age age T age T previous MMSE score men Education (vs university level) No education no diploma CEP high school level

24 Methods for MCAR or MAR data Complete case analysis (loss of precision, require MCAR) Imputation (require MCAR or MAR) Maximum likelihood using available data (require MAR)

25 Maximum likelihood for MAR data (1) Objective : Estimate θ from the distribution f(y θ) Likelihood of the observed data : Y obs, R f(y obs, R θ, ψ) = f(y obs, Y mis θ)f(r Y obs, Y mis, ψ)dy mis

26 Maximum likelihood for MAR data (2) If the data are MAR : f(y obs, R θ, ψ) = f(y obs, Y mis θ)f(r Y obs, ψ)dy mis = f(r Y obs, ψ) f(y obs, Y mis θ)dy mis Log-likelihood : = f(r Y obs, ψ)f(y obs θ) l(θ, ψ Y obs, R) = l(θ Y obs ) + l(ψ R, Y obs ) If ψ and θ are distinct : the missing data are ignorable θ is estimated by maximisation of l(θ Y obs ) using only available reponses.

27 Example : MAR analysis of Paquid data Mixed effect model Y ij test score for subject i at time t ij Y ij = (β 0 + age iγ 0 + α 0i ) + (β 1 + age iγ 1 + α 1i ) t ij + β 3 I {tij =0} + e ij with α i = (α 0i α 1i ) T N(0, G), e ij N ( 0, σe 2 ) age i vector of indicators for baseline age classes (70-74, 75-79, 80 years and older, ref= 65-69) I {tij =0} indicator of the baseline visit

28 Observed and predicted means of the score given time Complete data Available data Mixed model (MAR) 30 Score years and + Age

29 Advantages of mixed models Conclusion use all the available information (repeated measures) Flexibly handle intra-subject correlation (unbiased inference) Any number and times of measurements Robust to missing at random data Available in most softwares Limits of mixed models Assume homogeneous population extended models included latent classes(mixture) As the MAR assumption is uncheckable, complete the study by a sensitivity analysis extended models for MNAR data

30 References Chavance, M. et Manfredi R. Modélisation d observation incomplètes. Revue d Epidémiologie et Santé Publique 2000,48, Diggle PJ, Heagerty P, Liang KY, Zeger SL. Analysis of Longitudinal Data.2nd Edition. Oxford Statistical Science series 2002, Oxford University Press. Jacqmin-Gadda H, Commenges D, Dartigues JF. Analyse de données longitudinales gaussiennes comportant des données manquantes sur la variable à expliquer. Revue d Epidémiologie et Santé Publique 1999, 47, Little R.J.A. et Rubin D.B. Statistical Analysis with Missing Data, New York : John Wiley & Sons, Verbeke G and Molenberghs G Linear mixed models for longitudinal data. Springer Series in Statistics, Springer-Verlag,2000, New-York.

Problem of Missing Data

Problem of Missing Data VASA Mission of VA Statisticians Association (VASA) Promote & disseminate statistical methodological research relevant to VA studies; Facilitate communication & collaboration among VA-affiliated statisticians;