This study examines racial and gender disparities in the applications and awards of Social Security Disability Insurance (SSDI) and Supplemental Security Income (SSI) programs, using administrative records from the Social Security Administration and self-reported data from the Health and Retirement Study. Our research aims to enrich the toolset for analyzing disparities by adopting a more data-driven approach with fewer subjective assumptions. Our study highlights the sensitivity of disparity estimates obtained from the ordinary least squares (OLS) method, commonly used in disparity examinations. We find that OLS estimates are sensitive not only to the control variables in a regression model but also to the interaction terms between the group indicator (race or gender) and these control variables. These sensitivities are tied to the subjective assumptions researchers must make when estimating disparities using OLS.
To address these limitations, we employ a new method known as the double/debiased machine learning estimator, proposed by Chernozhukov et al. (2018). This alternative approach enables us to estimate racial and gender disparities with fewer subjective assumptions, and it flexibly accounts for numerous potential interaction effects arising from various characteristics like education and income. Using this estimator, we find minimal evidence of racial disparities (White vs. Black) in SS(D)I applications and awards. However, suggestive evidence points to a lower prevalence of SSDI applications among women. This contrast suggests the need for increased outreach efforts to facilitate SSDI applications for women. Meanwhile, racial disparities in SS(D)I applications and awards may be attributed to disparities in other socioeconomic dimensions