Likely responder analysis and tests of model misspecification in randomized controlled trials of treatments for Alcohol Use Disorder

点击翻译按钮获取中文摘要

Project Summary/ Abstract We have developed a strategy for the analysis of randomized clinical trials (RCTs) using a potential outcomes causal framework. Likely responders (LRs) to a test treatment T are identified at the end of the trial and a statistical test of the difference between T and placebo, P in this enriched sample is performed. LRs are identified at the end of the trial by fitting a model, called a prognostic score function, that estimates the expected response to T as a function of baseline features. The LR subset comprises individuals whose expected response exceeds a pre-specified clinically defined minimum. Identifying LR achieves an important goal of precision medicine. The causal effect of T compared to P among LRs is appraised based on the observed outcomes within strata of samples matched on their prognostic score. It is well known that, especially for subsets of a random sample, misspecification of the model can lead to spurious conclusions. To protect against this possibility in the estimation of the prognostic score, we have adapted an approach, novel to RCTs, that we call the RCT dry run (DRrct) diagnostic. It formally evaluates the potential for model misspecification. The value of the LR method has been demonstrated in a reanalysis of a large multisite 26-week long double-blind RCT of extended release gabapentin enacarbil (GE-XR) compared to placebo for the treatment of alcohol use disorder (AUD). Substantial benefits of treatment with GE-XR were found for the subset of patients predicted to be LRs based on their clinical features. In this research project, we will explore new statistical and machine learning modeling strategies for the prognostic score function and expand our knowledge of the statistical properties of the LR and DRrct methods. The goal is to minimize bias and increase precision in estimation of the prognostic score model and increasing power to test treatment effects in the LR subpopulation. To accomplish this we will use three strategies: analytic/theoretical methods where possible, simulation of RCTs and the reanalysis of six NIAAA RCTs comparing treatments for AUD. Although in most of the six trials, no treatment differences were found, it may be that LR subgroups can be identified whose members obtain substantial clinical benefit. Each reanalysis will utilize the DRrct method to appraise the possibility of model misspecification. The LR method has the potential to change standard practice for the analysis of RCTs, reduce the rate of failure caused by analyses limited to whole sample mean differences, and facilitate personalized medicine; the DRrct method has the potential to reduce the rate of irreproducible RCTs; and the reanalysis of the six NIAAA studies has the possibility of uncovering clinically meaningful relationships between patient characteristics and likely responders to previously studied candidate AUD treatments.