Projections for first-wave COVID-19 deaths across the US using social-distancing measures derived from mobile phones

Spencer Woody; Mauricio Garcia Tec; Maytal Dahan; Kelly Gaither; Spencer Fox; Lauren Ancel Meyers; James G Scott

The copyright holder for this preprint (which was not peer-reviewed) is the . https://doi.org/10. 1101 /2020 in the original formulation of the model) can inform U.S. state-level forecasts. For a variety of reasons, we find this assumption problematic. Our forecasts therefore rely solely on U.S. data, with state-level parameters shrunk toward a common mean in a hierarchical Bayesian model. Difference 3: valid uncertainty quantification. We address a problem with the IHME model by relying on a fundamentally different statistical assumption about model errors. Briefly: the IHME model fits cumulative death rates using a least-squares-like procedure on the log scale and applying standard large-sample statistical theory to get confidence intervals. For this procedure to result in valid uncertainty quantification, one must assume that successive model errors are independent. But in the IHME fitting procedure, this assumption is violated: today's cumulative death rate is yesterday's plus an increment, so the two must be correlated. Our model repairs this problem by fitting daily (noncumulative) death rates using a mixed-effects negative-binomial generalized linear model.