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

One note: we have phrased this model as a raw polynomial. But we actually fit the model using orthogonal polynomials to stabilize estimation; one may convert between the raw and orthogonal parameterizations easily. We fit the model using Markov Chain Monte Carlo, sampling from the posterior distribution of all model parameters. We use weakly informative priors on the fixed effects β, the second-stage regression coefficients Γ, and the covariance matrix Σ of the random effects. We also explored the possibility of using more informative priors based on daily death rates from European countries. But this had almost no effect on the fit of the model. Model fitting was carried out using the rstanarm package in R [Goodrich et al., 2020] . Forward-looking model projections are based on draws from the posterior predictive distribution. Because of the way that our covariates are constructed (see below), most of the x values for these forward-looking projections correspond to social-distancing behavior that has already been observed. However, at a longer horizon some extrapolation of x is necessary to generate projections; here we assume that socialdistancing behavior in a state remains unchanged from the average behavior over the seven most recent days of available data.