Online Program

WITHDRAWN: Performance of Methods to Assess Heterogeneity in Treatment Effect

Diane Richardson, VA Center for Health Equity Research and Promotion 

Keywords: comparative effectiveness, heterogeneity in treatment effect, statistical learning

BACKGROUND In comparative effectiveness research, parametric regression models are the most commonly used methods for estimating average treatment effects. These analyses rely upon assumptions that may not be valid for the estimation of treatment heterogeneity even in randomized studies, including correct specification of the parametric statistical model. With the increasing use of administrative and electronic medical record databases from clinically diverse populations for comparative effectiveness studies, this heterogeneity provides valuable information and a challenge to develop and evaluate methods that can better inform patient-centered decisions of optimal treatment. Machine-learning approaches that do not rely upon the assumptions of parametric regression models may provide improved estimates of heterogeneous comparative effects, compared to standard methods. OBJECTIVE The objective of this work is to compare the performance of conventional parametric regression modeling with targeted learning for the evaluation of treatment heterogeneity in the comparative effectiveness setting. We estimate power to detect heterogeneity in treatment effect (HTE), and bias in the resulting estimates of HTE, using multivariable risk stratification and interaction-term subgroup approaches with both methods. METHODS We performed a series of Monte Carlo simulations to examine realistic settings based on recent comparative effectiveness studies: a randomized controlled trial comparing two non-pharmacologic interventions to improve diabetes control; a regional retrospective study of disparities in joint replacement outcomes, and a study of gender and homelessness risk. In each setting, we imposed varying degrees of treatment heterogeneity, applied conventional and statistical learning approaches, and then assessed power and bias in estimation of HTE. We will present representative results from these settings to demonstrate the performance of parametric regression and statistical learning approaches.