Evaluating a predictive model of avoidable hospital events for race- and sex-based bias.

Abstract

Objective: To evaluate whether race- and sex-based biases are present in a predictive model of avoidable hospital (AH) events.

Study setting and design: We examined whether Medicare fee-for-service (FFS) beneficiaries in Maryland with similar risk scores differed in true AH event risk on the basis of race or sex (n = 324,834). This was operationalized as a logistic regression of true AH events on race or sex with fixed effects for risk score percentile.

Data sources and analytic sample: Beneficiary-level risk scores were derived from 36 months of Medicare FFS claims (April 2019-March 2022) and generated in May 2022. True AH events were observed in claims from June 2022.

Principal findings: Black patients had higher average risk scores than White patients; however, the likelihood of experiencing an AH event did not differ by race when controlling for predicted risk (Marginal Effect [ME] = 0.0003, 95%CI -0.0003 to 0.0009). AH event likelihood was lower in males when controlling for risk level; however, the effect was small (ME = -0.0008, 95% CI -0.0013 to -0.0003) and it did not differ by sex for the target group for intervention (ME = 0.0002, 95% CI -0.0031 to 0.0036).

Conclusions: We implemented a simple bias assessment methodology and found no evidence of meaningful race- or sex-based bias in this model. We encourage the incorporation of bias checks into predictive model development and monitoring processes.