Abstract

Random effects (RE) models and marginal models based on generalised estimating equations (GEE) are frequently used to analyse longitudinal repeated measurements health studies where subject dropout is common. RE models require MAR assumption. Because marginal models are not based on likelihoods, they require data to be MCAR . When the data are Gaussian, the GEEs reduce to score equations and provided the correct correlation structure is applied the two types of models are equivalent and marginal models should then be robust to MAR . However, equivalent marginal and RE models for Gaussian data may not necessarily produce identical parameter estimates due to missing data as GEEs may not reduce to score equations in that situation, even in presence of MCAR . For binary data the marginal models require MCAR assumption. By definition neither RE or marginal models are robust to MNAR .

In practice, the extent to which missing observations cause bias to the parameter estimates of these models and affect their clinical and statistical significance is not clear. Limited simulation studies have been conducted. However, it is not clear what proportion of missingness leads to substantial bias or how sensitivity to missing data compares between cluster-level and cluster-varying covariates. It is not known to what extent the marginal model is robust to misspecification of the working correlation matrix in presence of missing data and whether the strength of the intracluster correlation coefficient affects the bias in parameter estimates caused by MNAR . The aim here is to explore the effects of dropout on parameter estimates of RE and marginal models for repeated measurements data for both Gaussian and binary outcome using simulation studies in order to make some practical recommendations regarding analyses.