A Novel Solution to Biased Data in COVID-19 Incidence Studies

Abstract

Complete novelty and uncertainty of the COVID-19 pandemic have created many challenging scientific problems, including biased data arising from a lack of randomized testing over the general population. We describe the bias problem and its solution from Econometrics literature, which seems to have been neglected by epidemiology experts. We study a large COVID-19 US data set, providing nationwide forecasts of deaths to illustrate the model's power. Our two-equation model overcomes the bias by using the inverse Mills ratio and improves forecasts of new deaths in all nine out of nine weekly out-of-sample comparisons. It can be applied to a variety of problems associated with the pandemic. A focused study of trends in deaths predicted by lagged cumulative infections reveals that forty-two states have negative trends and that seven of nine states with undesirable positive trends have Republican governors.

Note: Funding: None to declare

Declaration of Interest: None to declare

Keywords: Inverse Mills Ratio, Selection Models, Poisson probit model, outcome equation

JEL Classification: I18, C10, C33