The Results of Generalized Estimating Equations in the Presence of Monotone Missing Patterns

Objective: During the analysis of ordinal longitudinal datasets, the most frequent problem is missing or incomplete data. In this study, when comparing two independent groups for this type of datasets, the effects of different missing ratio and different sample size on Type I and Type II error rates were investigated. Material and Methods: The data for different missing ratio and sample sizes (n=50,100,200,400) using simulation technique were generated. Repeated measurements (at four time points) for each group were generated from multivariate normal distributions using SAS MVN macro and transformed ordinal structure with quintiles method. Completely random monotone missing data for predefined ratios were created. On the comparison of two independent groups using generalized estimating equations (GEE), Type I and Type II error rates were investigated. These simulations were each replicated 1000 times. Results: While the Type I error rate was not affected seriously from missing clusters, the Type II error rate was affected. In the complete dataset for both small and large datasets, Type I error rates were <0.0001. While Type II error rates were greater than 0.10 for small sample sizes (n<100), this value was greater than 50% in the presence of missing datasets. Conclusion: The sample size and missing ratio are still important tasks for ordinal longitudinal data both in the planning and analyzing stages of an experimental design.