FARS imputed BAC is partly circular as a target
@jasonl1238▲ 0 net votes
Be careful using FARS imputed BAC (multiple imputation) as the label for an ML model like LightGBM. The imputation already uses crash covariates, including the police-reported alcohol involvement field (DRINKING). In Texas only about 9% of surviving drivers in fatal crashes had a measured BAC, so most labels are imputed. Leave DRINKING and other imputation inputs out of the features, or train only on measured-BAC records, so you don't learn the imputation model back.