Adjusted Bias-free Forecasts of Covid-19 Deaths for July 13 and July 20
12 Pages Posted: 23 Jul 2020
Date Written: July 12, 2020
Since testing for COVID-19 infections is not at all randomized over the general population, most epidemiological model forecasts of deaths are subject to ‘selection bias.’ This paper updates and supplements Vinod and Theiss (2020b), where we describe our bias correction methods using generalized linear models (GLM) and inverse mills ratio (IMR). Vinod and Theiss (2020a) describe our preliminary state-by-state forecasts using Poisson regression. This paper describes a localized adjustment to preliminary forecasts of new deaths by using an autoregressive distributed lag ARDL(1,0) model. The adjustment method is claimed to be relatively new with potential applications in improving disaggregate forecasts from Poisson regressions. This document focuses on ARDL-adjusted forecasts of new deaths for the week ending on July 13 and 20. The July 20 tables provide some guidance on using model errors for estimating extra costs of recent relaxing of restriction. We hope that our detailed forecasts with maximum entropy confidence intervals will help local governors and mayors in their opening up decisions.
Note: Funding: None to declare.
Declaration of Interest: None to declare.
Keywords: sample selection bias, autoregression, distribute lags, Poisson regression, maximum entropy bootstrap
JEL Classification: C10, C33, I18
Suggested Citation: Suggested Citation