Original Statistical Methodology for Causal Mediation Analysis with Time-Varying Confounding in Longitudinal Cohort Studies in Sub-Saharan Africa

Causal mediation analysis in longitudinal cohort studies in Sub-Saharan Africa requires methods that handle time-varying confounders affected by prior exposure, a setting where conventional regression-based mediation estimators are biased and for which no methodological framework grounded in African study design characteristics has been developed. This dissertation develops original statistical methodology for causal mediation analysis with time-varying confounding in longitudinal cohort studies in Sub-Saharan Africa, with primary applications to Nigerian maternal and child health data. New identification conditions for natural direct and indirect effects under time-varying treatment-induced confounding are established using nonparametric structural equation models. G-computation estimators, inverse probability weighted marginal structural models, and doubly robust estimators based on efficient influence functions are derived theoretically for the time-varying confounding setting. Semiparametric efficiency bounds are established. Data-adaptive targeted minimum loss-based estimation (TMLE) extensions are implemented using SuperLearner ensemble methods. Simulation experiments calibrated to the Enugu HDSS cohort data structures confirmed double robustness and efficiency of the proposed estimators. The methodology was applied to decompose the effect of socioeconomic status on under-five mortality in the IITA-Ibadan longitudinal cohort into direct effects and effects mediated through maternal nutrition status over pregnancy, with time-varying postnatal care as a time-varying confounder. Natural indirect effect through nutrition mediation was 38 percent of total effect, substantially different from naive regression mediation estimates of 22 percent. The dissertation contributes original causal mediation methodology with African cohort study applications. Keywords: causal mediation analysis, time-varying confounding, TMLE, longitudinal cohort, Sub-Saharan Africa