A Retrospective Bayesian Model for Measuring Covariate Effects on Observed COVID-19 Test and Case Counts

As the COVID-19 outbreak progresses, increasing numbers of researchers are examining how an array of factors either hurt or help the spread of the disease. Unfortunately, the majority of available data, primarily confirmed cases of COVID-19, are widely known to be biased indicators of the spread of the disease. In this paper we present a retrospective Bayesian model that is much simpler than epidemiological models of disease progression but is still able to identify the effect of covariates on the historical infection rate. The model is validated by comparing our estimation of the count of infected to projections from expert surveys and extant disease forecasts. To apply the model, we show that as of April 20th, there are approximately 3 million infected people in the United States, and these people are increasingly concentrated in states with more wealth, better air quality, fewer smokers, more residents under the age of 18, more public health funding and a history of more cardiovascular deaths. On the other hand, the timing of state declarations of emergency and the proportion of people who voted for President Trump in 2016 are not clear predictors of COVID-19 trends. In addition, we find that the US states have increased testing at approximately the same level in line with infections, suggesting that testing has not yet increased significantly above infection trends.