What if we try to correct for delays by imputing race? We test the effectiveness of Bayesian Improved First Name, Surname, and Geocoding (BIFSG), a common method for race imputation. BIFSG does not lead to substantial improvement as it often overestimates prevalence for non-white race groups.
(8/n)

We observe similar impacts at the national level and across other health outcomes, though with less variation.
(7/n)

Consider a particularly striking example: prevalence of hypertension diagnoses in Arkansas. Failing to account for delays would not represent the true disparity between Asian and Pacific Islander patients vs. Hispanic patients.
(6/n)

We see differences in race reporting delays by race and ethnicity. Patients with delays are also older, tend to have more visits, and experience higher rates of adverse health outcomes such as clinical diagnoses, procedures, and measurements.
(4/n)

In our data, race information may not be reported or collected on a patient’s 1st health visit. However, over time, information on race may eventually be obtained. This is what we refer to as a "delay" in race reporting, which we observe via longitudinal EHR data.
(3/n)

We are increasingly aware that AI may exacerbate disparities, particularly in 🏥healthcare. Our work demonstrates that it is important to audit AI systems over time. We show in a large dataset of electronic health records that ~73% of patients experience race reporting delays.
(2/n)

I am presenting a new 📝 “Bias Delayed is Bias Denied? Assessing the Effect of Reporting Delays on Disparity Assessments” at @facct.bsky.social on Thursday, with @aparnabee.bsky.social, Derek Ouyang, @allisonkoe.bsky.social, @marzyehghassemi.bsky.social, and Dan Ho. 🔗: arxiv.org/abs/2506.13735
(1/n)