Experimental Designs for Precision Health & Medicine

Randomized trials investigating the efficacy of a single treatment compared to a placebo,quantified by the average treatment effect (ATE), have long been the “gold standard” inevidence-based medicine (EBM). While profoundly successful, these trials provide guidancethat is better suited for regulatory rather than clinical decision making because the ATEdoesn’t consider the characteristics of individual patients. Precision medicine attempts tobridge this gap through a data-driven approach to the long-standing medical practice oftailoring treatment according to patient characteristics. Experimental designs can offer hopeto provide evidence supporting causally-valid decision making without sacrificing EBM’srigor by replacing the need for strong assumptions with randomization. In this dissertation,we detail the design of two such trials, the Biomarkers for Evaluating Spine Treatments (BEST)and the North Carolina Works for Health (NCW4H) studies, and present preliminary workon a third experimental design. We first present the statistical design of the BEST trial, a two-stage sequential multipleassignment randomized trial (SMART) comparing four active treatments that will inform aprecision medicine approach to chronic lower back pain (cLBP). We extend the minimiza-tion method for balancing covariates across treatments to accommodate contraindications,propose an analog for “power” when the primary aim is to estimate a dynamic treatmentregime (DTR), and conduct sample size simulations. We then present the NCW4H study investigating a multi-level intervention (MLI) toreduce chronic disease risks in socioeconomically disadvantaged, unemployed populations.Existing designs for MLIs focus on individually randomized factorial designs, which may beinfeasible due to logistical or other concerns, and cluster-randomized experiments whichcan pose identifiability challenges and be less efficient. We propose a design that combinesindividual-level randomization with a stepped wedge design at the cluster level. We derive aGEE-based power formula and provide an R package for sample size calculation. Finally, we present a sequential design for when the goal is to estimate a DTR and theexpected treatment response is a linear function of a patient’s characteristics . We proposeremoving suboptimal treatments based on changes in the value function thereby implicitlyaddressing potential subgroups with higher treatment response.Doctor of Philosoph