📊 statistics

Assessing Racial Disparities in Healthcare Expenditures via Mediator Distribution Shifts

This study employs a novel framework combining machine learning and influence-function techniques to decompose racial healthcare expenditure disparities using Medical Expenditures Panel Survey data, revealing that differences in socioeconomic status and health status are the primary drivers of gaps between racial groups, particularly between non-Hispanic Whites and Hispanics, while significant residual disparities persist due to unmeasured or structural factors.

Original authors: Xiaxian Ou, Xinwei He, David Benkeser, Razieh Nabi

Published 2026-06-02

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Xiaxian Ou, Xinwei He, David Benkeser, Razieh Nabi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the healthcare system as a massive, complex plumbing network. For decades, we've known that water (money spent on care) flows differently depending on which neighborhood (racial group) you live in. Some neighborhoods get a flood of water, while others get a trickle.

This paper asks a crucial question: Why is the water flowing so differently? Is it because the pipes in some neighborhoods are just wider (better health), or is it because the water pressure coming from the main source is different (income, insurance)?

The authors, researchers from Emory University, built a new "plumbing map" to figure this out. Instead of blaming the neighborhood itself (which is a fixed identity, not something you can easily "fix" like a broken pipe), they looked at the specific factors inside the neighborhood that change how much water flows.

Here is a breakdown of their study using simple analogies:

1. The Problem: The "Blind" Comparison

Usually, when people compare healthcare spending, they just look at the total bill.

The Old Way: "Group A spends more than Group B."
The Flaw: This is like comparing two cars where one is a heavy truck and the other is a tiny scooter, and saying the truck uses more gas because it's a "truck." It ignores that the truck is heavier and carries more cargo. In healthcare, different groups have different ages, incomes, and health needs. A simple comparison doesn't tell us why the spending is different.

2. The New Tool: The "Equalizer"

The researchers created a statistical "equalizer." Imagine you could magically take everyone in a disadvantaged group and give them the exact same Socioeconomic Status (SES), Insurance, Health Behaviors, and Health Status as the advantaged group, without changing their race.

They asked: "If we leveled the playing field on these specific factors, would the spending gap disappear?"

They broke the gap down into four specific "pipes" that carry the water:

SES (The Wallet & Education): Income, jobs, and schooling.
Insurance (The Ticket): Having coverage or not.
Health Behaviors (The Habits): Smoking, exercise, etc.
Health Status (The Engine): How sick or healthy the person actually is (chronic diseases, self-rated health).

3. The Findings: What the "Plumbing" Revealed

Using data from the Medical Expenditures Panel Survey (MEPS) from 2009 and 2016, they looked at how these pipes contributed to the gap between White, Black, Asian, and Hispanic populations.

The Heavy Hitters (SES and Health Status):
The biggest leaks in the system were Socioeconomic Status and Health Status.
- Analogy: If you gave the disadvantaged groups the same income and education levels as the advantaged groups, a huge chunk of the spending gap would vanish. Similarly, if you equalized their actual health conditions (like diabetes or heart disease), another massive chunk would disappear.
- Surprise: Even though marginalized groups often have worse health, the study found that if they had the same health status as White groups, their spending would actually go up. This suggests that White groups might have better access to doctors who diagnose and treat conditions more aggressively, leading to higher bills.
The Ticket (Insurance):
Insurance access was a major factor, especially for Hispanic populations.
- Analogy: In 2009, many Hispanic people didn't have tickets (insurance). When the researchers "gave" them the same insurance distribution as White people, the spending gap changed significantly. This highlights that lack of coverage is a huge driver of inequality. By 2016, after laws like the Affordable Care Act, this gap narrowed but didn't disappear.
The Small Leak (Health Behaviors):
Surprisingly, Health Behaviors (smoking, exercise) contributed very little to the spending gap.
- Analogy: You might think that if one group smoked more or exercised less, that would explain the cost difference. But the study found these habits were like a tiny drip in a massive pipe. They didn't explain much of the overall gap. The bigger issues were the wallet (SES) and the ticket (Insurance).
The Mystery Box (The Residual Gap):
Even after equalizing all four pipes (Money, Insurance, Habits, and Health), a significant gap remained, especially between White and Hispanic/Black groups.
- Analogy: This is the "ghost in the machine." It represents things the study couldn't measure, like structural racism, bias from doctors, or unmeasured life experiences. It suggests that even if two people have the same money, insurance, and health, they might still be treated differently or face different barriers in the healthcare system.

4. The Method: The "Super-Scanner"

To do this math, the researchers couldn't use simple formulas because healthcare spending is messy.

The Mess: Many people spend $0 (they don't go to the doctor), and those who do spend huge amounts (right-skewed).
The Solution: They used a "Super Learner" (a type of advanced computer algorithm that combines many different models) and a "Two-Part Model."
- Analogy: Instead of trying to guess the weather with one thermometer, they used a whole weather station with satellites, radar, and barometers all working together. This allowed them to handle the "zero" spenders and the "million-dollar" spenders without breaking the math.

5. The Bottom Line

The study concludes that racial disparities in healthcare spending are real and stubborn.

The Main Culprits: The gap is mostly driven by Socioeconomic Status (money and education) and Health Status (actual illness), followed closely by Insurance.
The Policy Takeaway: To fix the spending gap, we can't just tell people to "exercise more" (since behaviors didn't matter much). We need to fix the wallet (economic inequality) and the ticket (insurance access).
The Unfinished Job: Even if we fix the wallet and the ticket, a gap remains. This points to deeper, structural issues in how the healthcare system treats different groups that we haven't fully measured yet.

In short: The water flows differently not because of the neighborhood's name, but because of the pipes leading into it (money, insurance, and health). Fix those pipes, and you fix most of the leak. But there's still a hidden leak in the system itself that we need to find.

Technical Summary: Assessing Racial Disparities in Healthcare Expenditures via Mediator Distribution Shifts

Problem Statement
Racial disparities in healthcare expenditures are well-documented, yet existing methods for decomposing these disparities often face conceptual and methodological limitations. Traditional regression-based approaches and Oaxaca-Blinder decompositions frequently mischaracterize mediators as confounders or rely on rigid linear specifications that fail to capture complex, nonlinear relationships. Furthermore, standard causal mediation frameworks often treat race as a manipulable exposure, an interpretation that is ill-defined given race's nature as a socially constructed attribute shaped by historical and structural processes rather than a direct treatment. Consequently, there is a need for a framework that decomposes disparities without positing counterfactual interventions on race, while robustly handling the complex, zero-inflated, and right-skewed nature of healthcare expenditure data.

Methodology
The authors propose a nonparametric framework that defines racial disparities as differences in covariate-adjusted outcome distributions. Rather than estimating a "total effect of race," the study decomposes the total disparity into components attributable to shifts in the distributions of specific mediating variables (Socioeconomic Status [SES], insurance access, health behaviors, and health status) and a residual component.

Key methodological innovations include:

Distribution-Shift Decomposition: The framework defines the total disparity ( $\rho_{total}$ ) as a covariate-standardized difference in expected expenditures. It then isolates the mediator-attributable disparity ( $\rho_{R \to M_k \leadsto Y}$ ) by quantifying how the outcome distribution for a disadvantaged group would change if the distribution of a specific mediator $M_k$ were shifted to match that of an advantaged group, while holding the outcome-generating process and other mediators fixed at the disadvantaged group's levels. This avoids treating race as a manipulable cause.
Influence-Function-Based Estimation: To address challenges in model specification and the high dimensionality of mediator densities, the authors derive asymptotically linear estimators based on nonparametric influence functions. These estimators utilize a one-step correction to debias plug-in estimators, ensuring valid root- $n$ inference even when nuisance functions are estimated using flexible machine learning methods.
Multiply Robustness: The estimators exhibit a multiply robust structure, meaning consistency is achieved if any one of several distinct combinations of nuisance functions (e.g., propensity scores, mediator models, outcome regressions, and sequential regressions) is correctly specified. This protects against model misspecification in complex settings.
Handling Expenditure Data: Recognizing that healthcare expenditures are zero-inflated and right-skewed, the study employs a two-part modeling strategy within the super learner ensemble. The outcome regression models the probability of any expenditure and the conditional mean of positive expenditures (on a log-transformed scale) separately, then combines them to estimate geometric means.
Machine Learning Integration: The framework integrates super learners (ensemble methods combining GLMs, random forests, gradient boosting, etc.) to estimate nuisance parameters, enhancing robustness and predictive accuracy without relying on parametric assumptions.

Key Contributions

Conceptual Framework: The paper advances a perspective that treats race as an index of racialized social position rather than a manipulable treatment. It provides a formal statistical decomposition of disparities into contributions from specific mediating pathways and a residual component representing unmeasured structural factors.
Estimation Theory: The authors derive new asymptotically linear estimators for mediator-attributable disparities in settings with multiple, ordered mediators. These estimators are proven to be consistent and asymptotically normal under weaker convergence rate conditions than standard parametric approaches, provided specific pairs of nuisance functions are estimated with sufficient accuracy.
Software Implementation: The study introduces the flexPaths R package, extending the methodological toolkit for mediator-focused disparity analysis.
Empirical Application: The framework is applied to Medical Expenditures Panel Survey (MEPS) data from 2009 and 2016, analyzing disparities across non-Hispanic White, Black, Asian, and Hispanic populations.

Results
Applying the framework to MEPS data reveals substantial racial disparities in healthcare expenditures that persist across both years, with the largest gaps observed between non-Hispanic Whites and Hispanics.

Primary Drivers: Socioeconomic status (SES) and health status were identified as the largest contributors to expenditure disparities. Aligning SES distributions across racial groups would significantly reduce disparities, particularly for Hispanic and Black populations. Similarly, differences in health status distributions accounted for substantial portions of the disparity.
Insurance Access: Insurance access played a meaningful role, particularly for Hispanic populations. However, the disparity component attributed to insurance narrowed between 2009 and 2016, likely reflecting the impact of the Affordable Care Act, though Hispanics continued to face higher uninsurance rates.
Health Behaviors: Differences in measured health behaviors (smoking and physical activity) contributed minimally to expenditure disparities, with estimates often close to the null.
Residual Disparities: A significant residual disparity persisted, especially in comparisons involving non-Hispanic Whites. This residual, which captures differences not explained by the observed mediators, suggests the influence of unmeasured factors such as systemic racism, provider bias, or structural inequities in care delivery. Notably, the outcome-attributed disparity (the portion remaining after equalizing all mediators) was significant for White vs. marginalized group comparisons but often non-significant for comparisons between marginalized groups.

Significance and Claims
The paper claims that its framework offers a robust, interpretable, and policy-relevant approach to understanding healthcare disparities. By shifting the focus from the "effect of race" to the "effect of modifiable mediators," the study provides a data-driven basis for targeting interventions. The results suggest that reducing disparities requires addressing upstream structural determinants, specifically socioeconomic inequalities and barriers to insurance coverage, rather than focusing solely on individual health behaviors.

The authors emphasize that their findings highlight the persistence of structural inequities that are not captured by observed variables. The residual disparities indicate that even if SES, insurance, health behaviors, and health status were equalized, significant gaps would remain, pointing to the need for institutional reforms to address systemic bias and discrimination in healthcare. The study concludes that while methodological tools like influence-function-based estimators and machine learning are essential for valid inference in complex observational settings, advancing healthcare equity ultimately requires sustained policy efforts targeting the structural determinants of health.

1. The Problem: The "Blind" Comparison

2. The New Tool: The "Equalizer"

3. The Findings: What the "Plumbing" Revealed

4. The Method: The "Super-Scanner"

5. The Bottom Line

More like this