Racial Treatment Disparities after Machine Learning Surgical-Appropriateness Adjustment

Significant differences in inpatient surgery rates between black and non-black patients suggest a
racial treatment disparity. However, these rates must be adjusted for patient surgical
appropriateness to increase patient comparability. Inpatient data contains many potential control
variables, but only certain variables are valid controls in this regression framework. Researchers
then must select the proper controls from the large list of potential variables. In this paper, I
focus on the method of this appropriateness adjustment. The standard adjustment approach is to
include indicator variables for certain predetermined diagnosis groups to adjust for a form of
patient health severity. I propose an alternative data-driven approach that borrows the variable
selection tools of machine learning to choose as covariates the important predictors of invasive
surgery for acute myocardial infarctions (AMIs). The chosen control variables must be analyzed
to assess their validity as confounders between race and treatment. Using data from the
Nationwide Inpatient Sample, I show that the machine learning variable selection reveals surgery
appropriateness controls from diagnosis codes that decrease the standard adjusted treatment
disparity in AMIs by up to 50%. Nonetheless, a statistically and practically significant treatment
disparity of 6.5 percentage points remains after adjusting for many predictive controls, providing
further evidence of differential AMI treatment beyond that explained by health status
differences. Similar adjustments for patient-level health status and treatment appropriateness are
important and commonly used aspects of health research. The proposed approach can be used in
different contexts where empirical health adjustment is necessary to make patients more
comparable.