From Probability to Populations A Statistician’s Journey in Disentangling Health Disparities 

Racial disparities in health outcomes occur across a range of diseases. Mechanisms underlying such disparities are not fully understood, and often represent a combination of biology, access to care, and health system factors. Quantifying how much of the disparity is due to healthcare system versus individual patient differences is critical for directing policy and clinical interventions effectively. In this talk, we propose a variance partitioning approach for semiparametric Cox regression model for multilevel time-to-event data. Our model includes a random effect to account for correlation among patients within healthcare systems. While modelbased decomposition utilizing proportion of explained variation (PEV) can be obtained in this setting, we argue that PEV fails to appropriately quantify disparity. We propose using null martingale residuals from the model, and partitioning the variance by applying Blinder-Oaxaca decomposition, an econometric method used to quantify explained difference in outcomes attributable to observable characteristics versus unexplained difference attributable to discrimination or unobserved factors. Our methods can also be applied to parametric regression for time-to-event data such as the accelerated failure time model. Finally, we propose an alternative counterfactual approach using methods from causal inference. We illustrate our methodology using colorectal cancer data from the National Cancer Database in the United States to examine role of patient and hospital factors in survival differences between African Americans and Caucasians