Evaluating Treatment Effectiveness Under Model Misspecification: A Comparison of Targeted Maximum Likelihood Estimation with Bias-Corrected Matching

• Activity Number: 338
• Type: Topic Contributed
• Date/Time: Tuesday, August 6, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Noemi Kreif*+ and Susan Gruber and Rosalba Radice and Richard Grieve and Jasjeet S. Sekhon
• Companies: London School of Hygiene and Tropical Medicine and Harvard School of Public Health and Birkbeck, University of London and London School of Hygiene and Tropical Medicine and University of California, Berkeley
• Keywords: targeted maximum likelihood estimation ; bias-corrected matching ; treatment effectiveness ; machine learning ; double-robustness ; model misspecification

Abstract

This paper compares two approaches that combine estimates of the propensity score (PS) and the endpoint regression, for estimating treatment effectiveness. Targeted maximum likelihood estimation (TMLE) is a double-robust method designed to reduce bias in the estimate of the parameter of interest. Bias-corrected matching (BCM) reduces bias due to covariate imbalance between matched pairs using regression predictions. For both methods, estimates of the regression and the PS can be obtained using machine learning approaches, such as boosted classification and regression trees and "super learning", which can weaken the assumptions of correct model specification.

We contrast TMLE and BCM in a case study evaluating the effect of alternative types of hip replacement for patients' health related quality of life. The related simulation study compares bias and efficiency of the methods, for data generating processes including nonlinear functional form relationships, and good and poor overlap of the PS.

We demonstrate that incorporating machine learning into TMLE and BCM can reduce bias under misspecification. The relative performance of the methods varied by the complexity of the setting.