Covariate adjustment in subgroup analyses of randomized clinical trials: A propensity score approach

Abstract

Background: Subgroup analyses are frequently conducted in randomized clinical trials to assess evidence of heterogeneous treatment effect across patient subpopulations. Although randomization balances covariates within subgroups in expectation, chance imbalance may be amplified in small subgroups and adversely impact the precision of subgroup analyses. Covariate adjustment in overall analysis of randomized clinical trial is often conducted, via either analysis of covariance or propensity score weighting, but covariate adjustment for subgroup analysis has been rarely discussed. In this article, we develop propensity score weighting methodology for covariate adjustment to improve the precision and power of subgroup analyses in randomized clinical trials. Methods: We extend the propensity score weighting methodology to subgroup analyses by fitting a logistic regression propensity model with pre-specified covariate–subgroup interactions. We show that, by construction, overlap weighting exactly balances the covariates with interaction terms in each subgroup. Extensive simulations were performed to compare the operating characteristics of unadjusted estimator, different propensity score weighting estimators and the analysis of covariance estimator. We apply these methods to the Heart Failure: A Controlled Trial Investigating Outcomes of Exercise Training trial to evaluate the effect of exercise training on 6-min walk test in several pre-specified subgroups. Results: Standard errors of the adjusted estimators are smaller than those of the unadjusted estimator. The propensity score weighting estimator is as efficient as analysis of covariance, and is often more efficient when subgroup sample size is small (e.g. <125), and/or when outcome model is misspecified. The weighting estimators with full-interaction propensity model consistently outperform the standard main-effect propensity model. Conclusion: Propensity score weighting is a transparent and objective method to adjust chance imbalance of important covariates in subgroup analyses of randomized clinical trials. It is crucial to include the full covariate–subgroup interactions in the propensity score model.

Bootstrap vs asymptotic variance estimation when using propensity score weighting with continuous and binary outcomes

“…The current study was focused on the use of different sets of weights to estimate the effect of treatment in observational (or nonrandomized) studies. Both Zeng et al and Yang et al described the use of overlap weights in randomized controlled trials 29,30 . The use of weighting allows the analyst to reduce the effects of residual imbalance in measured baseline covariates between treatment groups.…”

“…Both Zeng et al and Yang et al described the use of overlap weights in randomized controlled trials. 29,30 The use of weighting allows the analyst to reduce the effects of residual imbalance in measured baseline covariates between treatment groups.…”

The Characteristics and Mortality of Chinese Herbal Medicine Users among Newly Diagnosed Inoperable Huge Hepatocellular Carcinoma (≥10 cm) Patients: A Retrospective Cohort Study with Exploration of Core Herbs

“…Furthermore, we used several PS-based models to correct for the imbalanced case numbers and estimated the risk of mortality under different model settings, including overlap weighting, average treatment effect on the treated (ATT), and inverse probability treatment weighting (IPTW). For example, the overlap weighting model added weights to subjects on the basis of demographic features between CHM users and nonusers, including age, gender, lifestyles, comorbidities, initial treatments, and medications [35]. These covariates were used to generate the probability of using CHM as PS, and the PS was assigned as a weight to CHM nonusers while 1 − PS was assigned to CHM users [36].…”

The Uses of Propensity Scores in Randomized Controlled Trials

“…OW completely removes imbalance, attaining the same semiparametric variance lower bound as the ANCOVA/IPTW estimator and consistently outperforming IPTW in finite samples. Yang, Li, Thomas, and Li (2021) extended OW to subgroup analyses and compared OW to a full-interaction ANCOVA-S model on the overall sample with baseline covariates, treatment indicator and their interactions with the subgroup variables. They showed that ANCOVA-S estimator is as efficient as weighting methods with full interaction, when the ANCOVA model is correctly specified, while OW estimator with propensity score estimated from a full-interaction model on the treatment indicator outperforms ANCOVA-S under small subgroup sample size, and/or when the ANCOVA is misspecified.…”