Impact of Propensity Score-Based Methods on Effect Estimates in Two Large Observational Endovascular Stroke Cohorts.

Abstract

Introduction

Observational studies in acute ischemic stroke are vulnerable to confounding by indication and treatment-selection bias, potentially distorting treatment-effect estimates. Propensity score methods can improve comparability between exposure groups and support more reliable interpretation of registry-based evidence. The Italian Registry of Endovascular Treatment in Acute Stroke (IRETAS) is an ongoing multicenter registry of patients undergoing endovascular thrombectomy (EVT).

Objective

To assess the impact of propensity score-based methods on exposure–outcome effect estimates in two large observational EVT cohorts.

Methods

Data were obtained from the IRETAS registry. Two independent cohorts were analyzed: a tandem occlusion cohort (n=2,148), comparing dissection-related versus atherosclerotic etiology, and an antithrombotic therapy cohort (n=10,662), comparing patients receiving any antithrombotic therapy before stroke onset with untreated patients. Associations were first estimated in unmatched cohorts using regression models appropriate for binary and ordinal outcomes.

Propensity score approaches were tailored to the exposure-group structure of each cohort. In the tandem cohort, dissection-related cases represented approximately 11% of patients; therefore, different propensity score matching strategies were evaluated to identify the approach best suited to the case-mix, balancing covariate comparability, sample-size retention, and precision. In the antithrombotic cohort, the smaller exposure group represented approximately 37% of patients; therefore, full matching was used to retain the entire study population while improving covariate balance and preserving precision. In this cohort, a hierarchical testing strategy was applied, first evaluating any pre-stroke antithrombotic therapy versus no therapy and subsequently assessing whether associations differed between antiplatelet therapy and oral anticoagulants. Balance was assessed using standardized mean differences and Love plots. After propensity score adjustment, associations were re-estimated using weighted generalized estimating equation models, incorporating propensity score-derived weights and accounting for the matched or weighted design.

Results

Substantial baseline imbalances were observed in both cohorts before adjustment. In the tandem cohort, propensity score matching generated a balanced matched population of 713 patients: 224 with dissection-related and 489 with atherosclerotic etiology. Before matching, dissection-related etiology was associated with higher odds of functional independence at 90 days, defined as mRS 0–2, compared with atherosclerotic etiology (OR 2.31, 95% CI 1.75–3.06). After matching, the association was attenuated and no longer statistically significant (OR 1.33, 95% CI 0.94–1.83; p=0.115).

In the antithrombotic cohort, full matching retained all patients while substantially improving covariate balance. Before matching, any pre-stroke antithrombotic therapy was associated with a less favorable 90-day mRS distribution compared with no therapy (OR 0.85, 95% CI 0.80–0.92). After matching, this association disappeared (OR 1.03, 95% CI 0.96–1.11). In the hierarchical analysis by antithrombotic class, APLT changed from a negative association before matching (OR 0.90, 95% CI 0.83–0.98; p=0.011) to a positive association after matching (OR 1.10, 95% CI 1.01–1.19; p=0.038), whereas the association for OAC was attenuated after matching.

Conclusions

In two EVT registry applications, propensity score-based methods materially changed the interpretation of crude exposure–outcome associations. Apparent clinically relevant differences were markedly attenuated or reversed after balancing baseline characteristics, highlighting the role of confounding and treatment-selection bias in observational stroke research.