Understanding racial bias in health using the Medical Expenditure Panel Survey data

Moninder Singh, Karthikeyan Natesan RamamurthyarXiv 2019 · arXiv

Abstract

Machine learning models are increasingly used in healthcare to identify "high-risk" patients for care management programs. However, these models often rely on healthcare spending as a proxy for medical need. This paper demonstrates that such models inherit systemic racial biases because Black patients typically generate lower healthcare costs than White patients with similar health needs. Using Medical Expenditure Panel Survey (MEPS) data, we quantify these disparities and show that Black patients are significantly sicker than White patients at the same spending levels. We provide a framework for mitigating this bias using data-reweighing techniques, ensuring more equitable patient identification without sacrificing predictive accuracy.

Introduction

In the United States, healthcare organizations leverage predictive algorithms to allocate resources to patients who would benefit most from intensive care management. These algorithms frequently aim to predict which individuals will incur the highest costs in the subsequent year. The underlying assumption is that healthcare expenditure serves as a reliable proxy for healthcare need.

However, this assumption is fundamentally flawed. Systemic barriers—including differences in income, education, and access to care—result in Black patients generating lower healthcare costs than White patients, even when they suffer from equivalent or more severe health conditions. When AI models are trained on this biased spending data, they systematically under-identify Black patients for critical interventions, effectively automating and scaling existing health disparities.

Methods and Data

The study utilizes the Medical Expenditure Panel Survey (MEPS), a nationally representative survey from the US Department of Health and Human Services. The researchers focused on non-Hispanic White and non-Hispanic Black populations, using Panel 19 (2014-2015) for training and Panel 20 (2015-2016) for validation.

The research was conducted in three stages:

  1. Data Analysis: Quantifying racial disparities by comparing health indicators against spending deciles.
  2. Model Building: Training a baseline Logistic Regression model to predict the top 10% of healthcare spenders. Socio-economic proxies like prior spending and income were intentionally excluded to test if the model would still inherit bias from health-related attributes.
  3. Bias Mitigation: Applying a Reweighing preprocessing technique. This method assigns higher weights to high-expense Black patients and lower weights to high-expense White patients during training, encouraging the model to treat groups more equitably.

The "Sicker but Cheaper" Paradox

Analysis of the raw MEPS data reveals a stark disparity in health status within the same spending brackets. As shown in the table below, Black patients in the top 10% of spenders are significantly sicker than their White counterparts.

Comparison of health utilization metrics between White and Black populations in the entire sample versus the top expenditure decile.

Specifically, Black patients in the top decile averaged 0.83 ER visits compared to 0.62 for White patients, and nearly double the inpatient (IP) nights (4.91 vs 2.61). This confirms that for Black patients, a high spending level corresponds to a much higher degree of medical severity.

Model Bias and Mitigation

The baseline model, trained on spending data, inherited these systemic patterns. Even without explicit socio-economic inputs, the model predicted that only 6.8% of Black patients would be "high-risk" compared to 10.7% of White patients, as illustrated in the following table.

Model predictions for high-expense risk by race, illustrating the disparity in health status among those predicted as high-risk.

Furthermore, the Black patients the model did identify as high-risk were significantly sicker than the identified White patients (e.g., 39.9% vs 28.6% had ER visits). This indicates the model was applying a higher "sickness threshold" for Black patients to be classified as high-risk.

Applying the Reweighing technique successfully shifted the model’s output. The predicted high-risk group moved to 9.9% White and 11% Black, narrowing the health gap and selecting patients based on a more equitable standard of need. Importantly, the model's balanced accuracy remained stable at approximately 71%, suggesting that fairness does not necessarily require a trade-off in predictive power.

Discussion and Future Work

This study highlights that bias in healthcare AI is often a reflection of biased labels (spending) rather than just biased features. While reweighing offers a simple and effective mitigation strategy, several avenues for future work remain:

  • Data Volume: Pooling multiple years of MEPS data could enhance statistical robustness.
  • Granularity: Incorporating event-level data (specific diagnoses and prescriptions) could reveal more nuanced manifestations of bias.
  • Alternative Proxies: Future models should explore predicting high utilization (e.g., ER visits or hospitalizations) directly, as these may serve as better proxies for medical need than financial expenditure.

Summary updated

Abstract

Over the years, several studies have demonstrated that there exist significant disparities in health indicators in the United States population across various groups. Healthcare expense is used as a proxy for health in algorithms that drive healthcare systems and this exacerbates the existing bias. In this work, we focus on the presence of racial bias in health indicators in the publicly available, and nationally representative Medical Expenditure Panel Survey (MEPS) data. We show that predictive models for care management trained using this data inherit this bias. Finally, we demonstrate that this inherited bias can be reduced significantly using simple mitigation techniques.
BibTeX
@misc{singh2019understanding,
  author = {Moninder Singh and Karthikeyan Natesan Ramamurthy},
  title = {{Understanding racial bias in health using the Medical Expenditure Panel Survey data}},
  year = {2019},
  eprint = {1911.01509},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1911.01509},
}
Figure