How different propensity score methods change the estimated cost of healthcare-associated infections

Healthcare-associated infections (HAIs) are a major patient safety issue and a significant source of healthcare expenditure. Their economic burden is often estimated using routinely collected administrative data. Because infected and non-infected patients differ in complexity, simple comparisons may overestimate costs attributable to infection. Propensity score methods are widely used to improve comparability in observational data, yet different approaches may yield different estimates, particularly for economic outcomes. This study assessed whether different propensity score methods and estimands yield different estimates of the hospitalization costs attributable to HAIs using real-world administrative data. We analysed 12,033 hospitalizations (1,212 HAI cases) at a Rome university hospital in 2018. Costs were estimated using Diagnosis-Related Group (DRG) reimbursements. Propensity scores were calculated using logistic regression and generalized boosted models. Matching, inverse probability of treatment weighting (IPTW), and stratification were compared, with and without length of stay (LOS) in the propensity score model. Costs were analysed on the log scale and expressed as percentage increases attributable to infection. HAIs were associated with higher DRG- reimbursements, but estimates varied by analytical approach, covariate specification, and estimand. With LOS, estimates ranged from 2.85% with stratification to 84.24% with untrimmed IPTW. Without LOS, they ranged from 6.72% with stratification to 129.14% with matching. Balance improved, but residual imbalance and extreme weights remained in some analyses. These findings show that the estimates depend strongly on analytical choices. Considering estimates under different assumptions, alongside balance and weighting diagnostics, may enhance transparency and support cautious decision on prevention strategies.