4G-02 - Application of the high-dimensional propensity score on medico administrative databases: an example in assessing comparative mortality in COPD patients on triple therapy.

Background: Observational studies, although using large databases, are often subject to confounding biases, particularly in terms of indication. The propensity score (PS) is commonly used to minimize these biases, but it is limited by a priori selection of covariates. The high-dimensional propensity score (HDPS) was developed to overcome this limitation by including a large number of covariates without a priori selection.

Objectives: This study aimed to compare the performance of PS and HDPS matching in the assessment of mortality of COPD patients on triple therapy in France (exacerbators vs. non-exacerbators).

Methods: We carried out a retrospective observational study using the French National Health Data System (SNDS). We included COPD patients who had initiated triple therapy in 2015. Two matching methods (PS and HDPS) were applied to compare patients with and without exacerbations. The PS was constructed with covariates selected by a committee of clinical experts. The HDPS was developed through several steps: identification of the dimensions of interest, selection of covariates in each dimension, evaluation of their recurrence and adjustment potential, estimation of the HDPS by logistic regression, and use as a matching variable (1:1 ratio). Sensitivity analyses were carried out by varying the number of covariates in the HDPS (k=100, 200, 300, 400, 500). Survival analyses were performed using Kaplan-Meier curves and Cox models, adjusted for the selected covariates.

Results: A total of 186,963 patients were included, of whom 39,847 had exacerbations and 147,116 were without exacerbations. The SP matched 79,654 exacerbators (with an equivalent number of non-exacerbators) and the HDPS 75,084. The quality of the matching did not increase beyond k=100. Survival analyses showed a significantly higher risk of mortality in patients with exacerbations, with a hazard ratio (HR) of 1.18 (95% CI: 1.16 – 1.21) for PS and 1.10 (95% CI: 1.07 – 1.12) for HDPS (p < 0.001). Interestingly, the dimensions of the covariates selected in the SP and the HDPS were different. The HDPS mainly selected covariates related to hospital diagnosis, outpatient drugs and biological tests.

Conclusions: HDPS matching produced results comparable to those obtained by PS, despite a different selection of covariates (including an increased importance of biological tests). The HDPS allowed an adjustment on additional covariates not retained by the committee of experts in PS, suggesting that the HDPS could be a valuable tool for sensitivity analyses, to confirm results obtained with more conventional methods.