OBJECTIVES : Appropriate adjustment for confounding in total hip arthroplasty (THA) requires controlling patient-, surgeon-, and hospital-level factors. A combination of direct matching and propensity score matching (PSM) has shown adequate confounder control. However, matching methods exclude a large proportion of comparison subjects and result in considerable loss of information. This study evaluated direct and PSM versus propensity score stratification (PSS) methods for confounding adjustment in THA by covariate balance and by sample size loss.

METHODS : Patients with THA between 2010-2016 were identified from Medicare database. The exposure variable was presence of dislocation diagnosis within one-year after THA. Three approaches to confounding adjustment were assessed: 1) direct matching and PSM (1:1), 2) PSS with 10 strata, 3) PSS with 100 strata. Area under the curve (AUC) was evaluated for the approaches as a measure of overall covariate balance between patients with and without dislocation. To assess the balance on individual covariates, absolute standardized differences (ASD) were calculated.

RESULTS : Of 450,355 patients identified with THA, 1.7% (N=7,680) had dislocation diagnosis. The AUCs for the matching approach, PSS with 10 strata and 100 strata were 0.54, 0.51 and 0.51 respectively. In matching method, two covariates, were found to have ASD >0.1, while no covariates from either PSS approach showed ASD >0.1. Though the matching method perfectly balanced the surgeon and hospital factors (ASD=0), it showed greater ASD for other covariates as compared to PSS approaches. When compared between the two PSS approaches, the 10 strata had higher ASD than 100 strata. With matching and PSS approaches 97.8% and 0.2% sample was lost, respectively.

CONCLUSIONS : Direct matching and PSM, and PSS provided adequate confounder control between patients with and without dislocation after THA. However, matching approach resulted in considerable loss of sample size and PSS with 100 strata showed tighter covariate balance as compared with the other approaches.