A 1:1:1 propensity score matching design to account for center-specific antimicrobial treatment or prophylaxis in observational studies: the case of the POSACOVID study

Propensity score matching (PSM) is employed in observational studies to reduce confounding by baseline characteristics. We considered a scenario where one center administered a given treatment/prophylaxis, and two other centers provided standard-of-care. We aimed to develop a three-group matching method balancing measured confounders between treated and control subjects while preserving center-effects, since they represent the effect of treatment/prophylaxis in the given scenario. Using multinomial logistic regression, we estimated generalized propensity scores and matched patient triplets by minimizing the perimeter of the triangle formed by their scores. We compared this method to classical 1:1 PSM in both a simulation and a case study (the POSACOVID study). While 1:1 PSM showed better balance, only 1:1:1 PSM balanced the two control groups and preserved center-effects. Overall, 1:1:1 PSM may be preferable when is essential to preserve center-specific characteristics influencing the treatment allocation and then to balance multiple control groups despite increased computational cost.