Causal Inference and Prediction in Health Studies: Environmental Exposures and Schistosomiasis, HIV-1 Genotypic Susceptibility Scores and Virologic Suppression, and Risk of Hospital Readmission for Heart Failure Patients

摘要: Causal inference-inspired semi-parametric methods of measuring variable importance are well designed to answer questions interest in health settings. Unlike traditional regression approaches, such measures based on causal parameters that have straightforward real-world definitions, regardless the approach used estimate them. Parameters models, contrast, not at all interpret settings, because their definition relies completely correctness pre-specified model. Prediction-focused machine learning can avoid issues model pre-specification, but still do provide estimates be easily interpreted; set predictors chosen also highly variable. Semi-parametric combine best both and able utilize data-adaptive estimation algorithms while returning a parameter is meaningful simply understood. In this dissertation, assess applied three applications: relationship between types water contact prevalence schistosomiasis infection rural China; HIV-1 treatment regimen genotype susceptibility scores with rate virologic suppression; impact telemanagement program association multiple risk factors rates hospital readmission for heart failure patients. Emphasized (1) choice as motivated by research question, (2) estimator consideration theoretical properties performance under non-ideal conditions, (3) use during process candidate models. Four different defined described, estimators considered. Each data analysis presents opportunities investigate aspects inference-based methods. analysis, compared Estimator HIV particularly context observed extreme violations experimental assignment (ETA) assumption. The G-computation estimator, inverse-probability-of-censoring-weighted (IPCW), its double-robust counterpart (DR-IPCW), targeted maximum likelihood (TMLE), included comparison. addresses differences community-level treatment, related assumptions must added typical framework. Also comparison super terms predictive performance.