(209) Optimization-based Stable Balancing Weights vs. Propensity Score Weighting for Samples with High Covariate Imbalance

Publication Number: 1201

Background: Stable balancing weights (SBW; Zubizarreta, 2015) are an optimization methods-based alternative to propensity score weighting (PSW) to balance confounding covariates. SBW seeks to find the weights of minimum variance for a control group subject to prespecified covariate balancing constraints, thereby maximizing effective sample size while meeting covariate balance requirements.

Objectives: To compare performance (covariate balance, effective sample size) of SBW vs. PSW where propensity scores are computed via a parametric generalized linear model (GLM).

Methods: Observational, retrospective, comparative cohort study using the Premier Healthcare Database, which included patients who underwent a surgical procedure with one of two different bipolar forceps (treated, N=9,059 and control, N=2,474) between 2000 and 2020. Average treatment effect on the treated (ATT) weights were generated based on the following covariates: age (0-17, 18-44, 45-64, ≥65 years), sex, inpatient vs. outpatient procedure, surgical approach (open, laparoscopic, unknown), elective vs. non-elective surgery, Charlson comorbidity index score, and hospital size (small/medium, large). PSW was implemented with trim at 99th percentile using GLM and applied to patients in the control group. SBW was implemented using two methods: (1) “grid search” to find weights of minimum variance at the lowest target standardized mean difference (TSMD) an optimization solution can be achieved (TSMD=0.01, 0.05, and 0.1 (oSBW)); (2) finding the weights of minimum variance at prespecified tolerance of TSMD≤0.10 (pSBW). PSW and SBW methods were compared on post-weighting covariate TSMDs, number of imbalanced covariates (indicated by TSMD>0.10), and effective sample size of the ATT-weighted control group.

Results: Pre-weighting, covariate balance was poor between the treated and control groups, with mean (range) reference group TSMD for covariates of 0.555 (0.158-1.424) and no balanced covariates. After weighting, the effective sample size for the PSW-weighted control group (N=389) was lower than SBW methods (oSBW: N=530; pSBW: N=811). Covariate balance was poorer for PSW vs. SBW methods (mean TSMD=0.116 with 6/12 imbalanced covariate categories for PSW vs. mean TSMD=0.010 with 0 imbalanced covariates for oSBW and mean TSMD=0.088 with 0 imbalanced covariates for pSBW).

Conclusions: Optimization-based SBW methods provide ample flexibility with respect to pre-specification of covariate balance goals, resulting in better post-weighting covariate balance and larger effective sample size as compared with PSW.