Approaches for Developing Treatment Rules

Abstract

The availability of scientific knowledge and the strength of supporting statistical evidence for efficacy of available treatments varies considerably across clinical settings. Nonetheless, the goal of clinicians remains largely unchanged: to recommend patients the most beneficial course of treatment available. In this dissertation, we examine the problem of estimating treatment efficacy for heterogeneous individuals in a target population with varied levels of abstraction and with an eye toward varied study designs. In Chapter 2, we restrict ourselves to RCT data and frame the problem of identifying individual characteristics that affect treatment response as a global hypothesis test for qualitative interaction using a convex optimization problem that is solved either with or without the constraint of qualitative interaction under limited modeling assumptions. We also present a permutation-based testing procedure that yields a p-value or false discovery rate. In Chapter 3, we move away from the RCT setting and instead focus on observational study designs that introduce the significant complication of treatment not being assigned independently of patient characteristics. At the core of Chapter 3 is a principled framework and user-friendly R implementation in the DevTreatRules package that allow practitioners to develop a function (known as a "treatment rule") to recommend treatment based on individual characteristics, while also obtaining a trustworthy estimate of the treatment rule's benefit in the target population. We also introduces a four-category classification of characteristics collected in a given observational study based on whether each variable might influence treatment assignment and whether it is expected to be observed in independent clinical settings. Our framework and R implementation emphasize the distinct roles these variable types should play in a principled analysis to ensure that an estimated treatment rule is applicable in clinical settings and that the estimate of the rule benefit is reliable. We begin Chapter 4 by exploring the popular outcome-weighted learning (OWL) method that takes a "direct'' approach to estimating a treatment rule rather than the "indirect" approach taken by the split-regression procedure in Chapter 3. We present a simple Bayesian interpretation of OWL that offers a clear equivalence with split-regression when the outcome is binary and a more nuanced connection when the outcome is continuous. We show how OWL fits into the principled framework of Chapter 3 and we accordingly expand the R package DevTreatRules to accommodate OWL. We then conduct a simulation study that uses DevTreatRules to develop and compare the performance of treatment rules from OWL and split-regression under a range of scenarios. We also implement another promising direct approach to estimating treatment rules, referred to as direct-interactions, in DevTreatRules and include it in the simulation study. We share our proposed remedies for a few subtle but critical computational issues we encountered during our simulation study that have a substantial impact on the performance of OWL and direct-interactions in practice.