Bayesian Nonparametric Methods for Modelling Heterogeneous Treatment Effects in Randomised Controlled Trials in West African Health Settings

Randomised controlled trial analysis in West African health research typically reports average treatment effects that mask substantial individual and subgroup heterogeneity, and Bayesian nonparametric methods provide a theoretically coherent, flexible framework for recovering this heterogeneity without prespecifying effect modification structure. This dissertation develops Bayesian nonparametric methods for modelling heterogeneous treatment effects in RCTs conducted in West African health settings, with primary applications to malaria prevention and nutrition trials in Nigeria. Gaussian process priors for treatment effect surfaces, Dirichlet process mixture models for response subgroup discovery, and Bayesian Additive Regression Trees (BART) were extended to incorporate West African cluster randomisation designs and informative missing data patterns. Formal theoretical analysis establishes posterior contraction rates for the proposed nonparametric estimators under West African trial sample size constraints. The proposed methods were applied to reanalyse an ITN distribution RCT from Borno State (n=1,800 households) and a CMAM nutrition trial from Kano State (n=620 children). BART analysis of the ITN trial revealed that bed net effect on malaria incidence was three times larger in households with no prior ITN access than in prior-access households, an interaction undetectable by standard analysis. Dirichlet process mixture analysis of the nutrition trial identified a child subgroup with metabolic risk profiles showing significantly attenuated treatment response. The dissertation contributes novel West African-adapted nonparametric trial analysis methods. Keywords: Bayesian nonparametric, heterogeneous treatment effects, BART, Dirichlet process, West African RCT