Assessing racial disparities in healthcare expenditure using generalized propensity score weighting

This paper extends current propensity score weighting methods for causal inference to better understand disparities in healthcare access across multiple racial groups. By treating each racial group as a distinct entity (or "treatment") in the causal inference framework, we can assess and evaluate heterogeneity in healthcare outcomes across various racial or ethnic categories. Furthermore, we leverage modern propensity score weighting techniques to address the challenges inherent to multiple group evaluations, such as violations of the positivity assumption, and compare the performance of different propensity score weights.

We use generalized propensity score methods to assess racial disparities across 4 specific racial or ethnic groups: Whites, Hispanics, Asians, and Blacks. We first calculate weights that standardize the participants' characteristics and then compare their weighted outcomes. We consider four distinct measures (i.e., causal estimands) and estimation methods: the conventional average treatment effect on the treated (ATT), the ATT trimming, the ATT truncation, and the overlap weighted ATT (OWATT). These estimands are applied under a multi-valued "treatment" framework, where the said "treatment" is defined by non-manipulable racial or ethnic group memberships. Using data from the Medical Expenditure Panel Survey (MEPS), we assess disparities in healthcare expenditures across the 4 racial and ethnic groups.

We found significant disparities in healthcare expenditure between White participants and all the other racial or ethnic groups when using OWATT and ATT truncation. Conventional ATT and ATT trimming could indicate non-significant difference due to larger variance estimates. Moreover, the conventional ATT was found to be the least efficient estimation method, even when its variance was estimated via non-parametric bootstrapping. Overall, the OWATT emerges as a promising estimation method; it retains the available information from all samples, avoids subjectivity (inherent to choosing thresholds by its competitors) and mitigates judiciously pernicious inferential effects (such as the inflated variance estimates) by extreme propensity score weights.

We found that generalized propensity score weighting (GPSW) methods are valuable quantitative tools to standardize and compare characteristics as well as outcomes of non-manipulable groups. This helps assess disparities across multiple racial and ethnic groups, as demonstrated in this study. These methods offer flexible and semi-parametric analysis on the primary causal parameters of interest (such as the racial disparities), with straightforward and intuitive interpretations. In addition, when there is violation of the positivity assumption, OWATT serves as an excellent alternative due to its greater efficiency, evidenced by relatively smaller variance. More importantly, the OWATT uses the entire dataset by assigning weights to all participants, regardless of their propensity score values. This feature of OWATT circumvents the need to specify user-defined thresholds, as required in ATT trimming or truncation, and retains as much data information as possible, leading to more reliable estimation results.