Designed sampling from large databases for controlled trials

Abstract

ABSTRACT Controlled trials are ubiquitously used to investigate the effect of a medical treatment. The trial outcome can be dependent on a set of patient covariates. Traditional approaches have relied primarily on randomized patient sampling and allocation to treatment and control groups. However, when covariate data for a large set of patients are available and the dependence of the outcome on the covariates is of interest, one can potentially design treatment/control groups that provide better estimates of the covariate-dependent effects of the treatment or provide similarly accurate estimates with a smaller trial cohort size. In this article, we develop an approach that uses optimal Design Of Experiments (DOE) concepts to select the patients for the treatment and control groups upfront, based on their covariate values, in a manner that optimizes the information content in the data. For the optimal treatment and control groups selection, we develop simple guidelines and an optimization algorithm that achieves much more accurate estimates of the covariate-dependent effects of the treatment than random sampling. We demonstrate the advantage of our method through both theoretical and numerical performance comparisons. The advantages are more pronounced when the trial cohort size is smaller, relative to the number of records in the database. Moreover, our approach causes no sampling bias in the estimated effects, for the same reason that DOE principles do not bias estimated effects. Although we focus on medical treatment assessment, the approach has applicability in many analytics application domains where one wants to conduct a controlled experimental study to identify the covariate-dependent effects of a factor (e.g., a marketing sales promotion), based on a sample of study subjects selected optimally from a large database of covariates.