A simulation study comparing weighted estimation equations with multiple imputation based estimating equations for longitudinal binary data

Abstract

Introduction: Missingness has become an inevitable scenario in longitudinal data and this often complicates the proposed analyses. Different methods have been proposed, among which are likelihood-based methods where analysts rest assured that the missingness is taken care of, when, for instance, Linear, Generalized Linear, or Non- Linear Models are considered due to their validity under the missing at random assumption. Statistical methods which take no notice of the mechanism for dropout will show the way to biased inference. Likelihood-based methods have computational complexity when taking into consideration longitudinal binary data. Weighted Generalized Estimating Equations (WGEE) is one of the common methods for handling dropouts that is MAR and is more usually used in marginal models for discrete longitudinal data. Alternatively, multiple imputations can be used to pre-process incomplete data, after which standard GEE is applied (MI-GEE). Objective: The objective of this thesis was to compare weighted estimating equations with multiple imputation based estimating equations for longitudinal binary data. Method: In this study, both approaches WGEE and MI-GEE were compared for incomplete binary data, through so-called asymptotic simulation study as well as smallsample simulation. Bias, variances and mean square error (MSE) were the bases for the comparison between those two approaches. Results and conclusion: - The results provide evidence for the fact that MI-GEE is less biased and more accurate in small and moderate samples sizes, while WGEE is asymptotically unbiased and has only shown better performance for data having more percentage of dropout.