Comparison of various modeling approaches in the analysis of longitudinal data with a binary outcome: The Ontario Mother and Infant Study (TOMIS) III

Longitudinal studies are often used to investigate the developmental trends of outcomes over time. Several modeling strategies can be applied for the analyses of longitudinal data. In this study, various statistical approaches were discussed and compared using data from The Ontario Mother and Infant Study (TOMIS) III. TOMIS III was a longitudinal cohort study that assessed the associations between the method of delivery and health outcomes and service utilizations. The primary outcome of postpartum depression was used as an example. Generalized estimating equations (GEE) assuming a serial correlation structure were used as the primary method of analysis to assess the association between the method of delivery and postpartum depression over 12 months. We performed sensitivity analyses using three other methods – namely, the (1) generalized linear mixed-effects model (GLMM), (2) hierarchical generalized linear model (HGLM), and (3) Bayesian hierarchical model (BHM), to compare the robustness of the results. The results from all four models indicated that the method of delivery had no significant effect on postpartum depression. However, GEE, GLMM, and BHM identified the following seven predictors of depression: annual household income; urinary incontinence (bladder problems); English or French (Canada’s official languages) spoken at home; a lower SF-12 mental component score; unmet learning needs in the hospital; lower social support; and a lower SF-12 physical component score. HGLM showed similar results to the above three models with the exception of language spoken at home, which was not significant. GEE provided the good fit statistics for the data. Method of delivery had no significant effect on postpartum depression, based on GEE analysis. This result remained robust under different methods of analyses. GEE demonstrated a good fit for the TOMIS III data.