Abstract
The COVID-19 pandemic has necessitated high-resolution surveillance to understand mortality trends across different demographics. This study introduces a hierarchical Bayesian model utilizing a novel non-parametric spatial approach: a low-rank Gaussian Process (GP) projected by regularised B-splines. This method addresses the computational challenges of standard GPs while maintaining smoothness. Applied to censored and biased CDC data, the model reconstructs longitudinal trends in COVID-19 deaths by 1-year age bands. The findings reveal that lower vaccination rates among adults aged 18-64 were significantly associated with death resurgences in states like Florida and Texas, highlighting the critical need for broad-age vaccination coverage.
The Challenge of Messy Public Health Data
Public health data is rarely "clean." In the United States, the CDC reports COVID-19 deaths in age bands that change over time or across states. Furthermore, to protect privacy, the CDC censors low death counts (between 1 and 9), leaving significant gaps in the record.
Standard epidemiological models are often too computationally expensive to run at a 1-year age resolution over long periods. Conversely, standard smoothing techniques like Gaussian Processes (GPs) struggle with the dimensionality of a 2D surface (age × time), while basic splines often overfit. The researchers sought a "Goldilocks" solution: a model flexible enough to handle censoring and reporting delays, precise enough to estimate deaths by single-year age increments, and fast enough for practical Bayesian inference.
Methodology: The Projected Gaussian Process
The core innovation is the Regularised B-splines projected Gaussian Process.
- GP and Splines Integration: Gaussian Processes are prized for their smoothness but are computationally heavy. B-splines are fast but can be prone to overfitting. By placing a GP prior on the B-spline coefficients, the authors created a low-rank approximation that inherits the GP’s smoothness while drastically reducing the number of parameters to estimate.
- 2D Latent Surface: The model treats the age-time distribution as a 2D surface. It uses a "softmax" transformation to ensure that the age contributions at any given week sum to 100%.
- Handling Censoring: The model uses a Negative Binomial likelihood to account for overdispersion and incorporates a cumulative density function (CDF) to "probabilistically fill" the censored gaps based on surrounding trends.
- Calibration: To fix reporting delays, the model is calibrated against all-age death data from Johns Hopkins University, which is generally more up-to-date than age-stratified records.
Key Results and Vaccination Impact
The projected GP approach proved to be approximately 74% faster than a standard 2D GP while maintaining nearly identical predictive accuracy. This efficiency allowed the researchers to couple the mortality model with age-specific vaccination rates to analyze the 2021 resurgences.
While the elderly were vaccinated first, the study highlights that the resurgence of deaths in 2021 was heavily tied to the 18-64 demographic. Lower vaccine uptake in this group didn't just lead to more deaths among younger adults; it likely sustained transmission chains that eventually reached the remaining vulnerable elderly.
A counterfactual analysis estimated that if every state had reached the 46% vaccination rate seen in New York for 18-64-year-olds before the Summer 2021 wave, 41.96% of all deaths in that age group across the ten most populous states could have been avoided. In Florida, an estimated 57% of deaths in the 18-64 group were potentially avoidable under this higher-vaccination scenario.
Conclusion and Future Work
The B-splines projected GP framework offers a robust tool for high-dimensional public health challenges where data is sparse, censored, or inconsistently reported. By effectively reconstructing trends, it underscores the critical importance of vaccination across all adult age groups to limit fatal outcomes. Future work may apply this framework to other challenges, such as mapping opioid overdoses or cancer clusters.
Summary updated