TL;DR
This paper introduces a novel Bayesian statistical model to accurately estimate COVID-19 incidence across racial and ethnic groups, even when crucial demographic data (like race/ethnicity) is missing and not missing at random. Their method outperforms standard approaches like complete-case analysis and multiple imputation, revealing that previous estimates likely understated risks for non-white populations.
What Problem Does This Paper Address?
Understanding how diseases like COVID-19 affect different demographic groups is crucial for effective public health interventions. However, a significant challenge in analyzing this data is the widespread missingness of important covariates, such as race and ethnicity. This missingness is often not random; it's more likely to occur in groups that are already at higher risk or in communities with fewer resources for data collection. This "not-missing-at-random" (NMAR) phenomenon can lead to biased estimates if not properly accounted for. Standard methods like simply dropping cases with missing data (complete-case analysis) or imputing them assuming randomness (multiple imputation) can fail to correct for this bias, leading to potentially misleading conclusions about disparities in disease burden.
Methods
The authors propose a Bayesian parametric model that tackles NMAR missingness by jointly modeling two processes: the disease incidence and the missingness of the race/ethnicity covariate. This approach leverages spatial information and known population data (from census) to improve the estimation of both the disease rates and the probability of missing data. They utilize a selection model framework, which is a probabilistic approach to handling missing data. The model is designed to be locally identifiable, meaning that the parameters can be uniquely estimated under certain conditions. They also conduct extensive simulation studies to evaluate their model's performance against common alternatives like complete-case analysis and multiple imputation (both ad-hoc and Gibbs sampling).
Datasets
The study uses two main sources of data:
- Wayne County, Michigan, population data: This data, derived from the 2010 Decennial Census, provides population counts stratified by age, sex, and race/ethnicity at the Public Use Microdata Area (PUMA) level. This is crucial for identifying the model parameters.
- COVID-19 case data from Wayne County, Michigan: This data, obtained from the Michigan Disease Surveillance System (MDSS) for the period of March 2020 through June 2020, includes information on confirmed COVID-19 cases, including race/ethnicity where available.
Key Insights and Ideas
- NMAR missingness is a critical issue: The paper strongly emphasizes that assuming missing demographic data is "missing at random" (MAR) is often inappropriate in public health surveillance and can lead to significant biases.
- Joint modeling is key: By simultaneously modeling disease incidence and the missingness process, the model can learn about the complex relationships between them, leading to more accurate estimates.
- Spatial information matters: Incorporating spatial variation in both disease incidence and missingness helps to improve parameter identifiability and estimation.
- Standard methods can be misleading: The simulation results clearly demonstrate that complete-case analysis and multiple imputation (especially when assuming MAR for NMAR data) can produce biased estimates and underestimate uncertainty.
Key Results
- Underestimation of risk for non-white populations: When applied to COVID-19 data from Wayne County, Michigan, the proposed joint model revealed that previous estimates using complete-case analysis or multiple imputation likely understated the relative risk of COVID-19 for non-white residents compared to white residents.
- Improved performance in simulations: The joint model consistently outperformed complete-case analysis and multiple imputation methods in simulation studies, particularly in scenarios with higher proportions of missing data. While its performance slightly dipped in extreme missingness scenarios (20% observed), it generally provided more accurate and less biased estimates.
- Robustness to prior misspecification: While prior sensitivity was explored, the model's inferences for incidence were found to be relatively robust to misspecification of population hyperparameters, though care is needed for priors centered on missing-at-random scenarios.
Limitations and Future Work
- Dependence on census data: The model's identifiability relies on census data, which can be outdated (e.g., 2010 Census for 2020 data) and may not perfectly reflect current population distributions.
- Categorical covariate limitations: The study was limited by the available race/ethnicity categories in the census data, preventing a more nuanced analysis of ethnicity separate from race.
- Poisson assumption: The model assumes a Poisson distribution for disease counts, which might be less appropriate as cumulative incidence grows over time. A binomial model might be more suitable in such cases.
- Future work: The authors suggest extending the model to incorporate multi-level health system information, temporal and spatial autocorrelation (e.g., SIR models), finer spatial resolutions, and more sophisticated priors for hyperparameters. They also highlight the need for methods to account for uncertainty in population denominator data.
Summary updated