Modeling racial/ethnic differences in COVID-19 incidence with covariates subject to non-random missingness

Rob Trangucci, Yang Chen, Jon Zelner

TL;DR

This paper introduces a novel Bayesian joint model designed to accurately estimate racial and ethnic disparities in disease incidence when data is "Not Missing At Random" (NMAR). By simultaneously modeling the disease process and the missingness process—and anchoring them to known census data—the authors demonstrate that traditional methods significantly underestimate COVID-19 risks for non-white populations.

What Problem Does This Paper Address?

Public health surveillance data is the backbone of pandemic response, yet it is often riddled with holes. During the COVID-19 pandemic, a substantial percentage of case reports lacked race and ethnicity information.

The core challenge is that this missingness is rarely random. In statistical terms, it is Not Missing At Random (NMAR). For example, resource-strapped health departments in marginalized neighborhoods might have higher rates of missing data, or certain groups might be less likely to provide demographic info due to systemic distrust.

Standard fixes—like "complete-case analysis" (simply dropping records with missing info) or "multiple imputation" (which usually assumes data is missing at random)—fail here. They often produce "overconfident and biased" estimates, typically understating the true burden of disease on the very communities most at risk.

Methods

The authors propose a Bayesian parametric joint model. Unlike standard approaches that treat missing data as a preprocessing step, this model handles the disease incidence and the "missingness mechanism" as two parts of a single mathematical system.

Key technical features include:

  • Joint Modeling: It defines a Poisson sampling model for disease counts and a conditional binomial model for whether a case's race is observed.
  • Spatial Anchoring: The model leverages known population data (from the U.S. Census) at a granular spatial level (Public Use Microdata Areas, or PUMAs). By knowing exactly how many people of each race live in a specific area, the model can "infer" the likely race of the missing cases.
  • Identifiability: The authors prove that their model is "locally identifiable," meaning there is enough information in the combined spatial and census data to uniquely determine the model's parameters, even with high levels of missingness.
  • Implementation: The model is implemented using Stan, a probabilistic programming language that utilizes Hamiltonian Monte Carlo for efficient Bayesian inference.

Datasets

The study evaluates the model using two primary sources:

  1. Simulation Study: 200 simulated datasets based on the geography of Wayne County, Michigan, varying the proportion of observed race data from 90% down to 20%.
  2. Real-World Application: COVID-19 case data from the Michigan Disease Surveillance System (MDSS) for Wayne County (including Detroit) from March to June 2020.
  3. Population Denominators: 2010 Decennial Census data, aggregated to PUMA levels and stratified by age, sex, and race/ethnicity.

Key Insights and Ideas

  • The Danger of "Standard" Imputation: The simulation showed that Multiple Imputation (MI) can actually increase bias if the data is NMAR. Because MI often assumes missingness is independent of the missing value, it tends to "smear" cases across groups incorrectly.
  • Spatial Variation as a Signal: The "aha moment" is that spatial clustering of both disease and missingness provides the necessary information to recover the true distribution. If a PUMA is 80% Black and has a high number of missing-race cases, the model can statistically assign those cases more accurately than a non-spatial model.
  • Sensitivity to Priors: In scenarios with extremely high missingness (e.g., only 20% of cases have race data), the model becomes sensitive to "priors" (initial assumptions). This highlights the need for researchers to have some baseline understanding of the communities they are studying.

Key Results

  • Underestimation of Risk: In the Wayne County analysis, the joint model revealed that relative risk estimates for non-white residents were significantly understated by traditional methods.
  • The "Other" Category: The most dramatic finding involved the "Other" race category (including mixed-race and Middle Eastern individuals). The joint model estimated their relative risk compared to white residents at nearly 14, whereas complete-case analysis suggested a much lower disparity.
  • Accuracy in Simulations: The joint model consistently outperformed complete-case analysis and MI in terms of Root Mean Squared Error (RMSE) and interval coverage, especially as the proportion of missing data increased.

Limitations and Future Work

  • Census Lag: The study relied on 2010 Census data for a 2020 pandemic. Demographic shifts over a decade can introduce errors in the "anchoring" population counts.
  • Poisson Assumption: The model assumes a Poisson distribution, which doesn't account for "overdispersion" (when cases cluster more than expected). Future versions could use a Negative Binomial distribution to better handle the "bursty" nature of infectious diseases.
  • Computational Intensity: Bayesian joint models are computationally "expensive" compared to simple imputation, requiring more time and specialized software (like Stan).
  • Future Directions: The authors suggest extending the model to include temporal autocorrelation (how risks change over weeks/months) and finer spatial resolutions to better capture neighborhood-level dynamics.

Summary updated

Abstract

Characterizing the cumulative burden of COVID-19 by race/ethnicity is of the utmost importance for public health researchers and policy makers in order to design effective mitigation measures. This analysis is hampered, however, by surveillance case data with substantial missingness in race and ethnicity covariates. Worse yet, this missingness likely depends on the values of these missing covariates, i.e. they are not missing at random (NMAR). We propose a Bayesian parametric model that leverages joint information on spatial variation in the disease and covariate missingness processes and can accommodate both MAR and NMAR missingness. We show that the model is locally identifiable when the spatial distribution of the population covariates is known and observed cases can be associated with a spatial unit of observation. We also use a simulation study to investigate the model's finite-sample performance. We compare our model's performance on NMAR data against complete-case analysis and multiple imputation (MI), both of which are commonly used by public health researchers when confronted with missing categorical covariates. Finally, we model spatial variation in cumulative COVID-19 incidence in Wayne County, Michigan using data from the Michigan Department and Health and Human Services. The analysis suggests that population relative risk estimates by race during the early part of the COVID-19 pandemic in Michigan were understated for non-white residents compared to white residents when cases missing race were dropped or had these values imputed using MI.
BibTeX
@article{trangucci2023modeling,
  author = {Rob Trangucci and Yang Chen and Jon Zelner},
  title = {{Modeling racial/ethnic differences in COVID-19 incidence with covariates subject to non-random missingness}},
  journal = {The Annals of Applied Statistics},
  year = {2023},
  doi = {10.1214/22-aoas1711},
}
Figure