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

• and total visitation counts for various points of interest aggregated by category, including grocery stores, hospitals, parks, restaurants, bars, colleges, etc. 3 These data are derived from GPS traces of tens of millions of mobile phones nationwide. "Home" and "work" locations for a device were inferred by Safe-Graph based on daytime and overnight locations over an extended period of time. The data were provided to us in an aggregated format; no device-level data or individual GPS traces were accessible by the research team. Figure 2 shows a selection of these social-distancing measures over time in both New York and Texas. We denote these D distancing metrics s it,1 , . . . , s it,D , observed each day in each state. To construct useful covariates out of this information, we proceed as follows. For each distancing metric s j , define a corresponding lagged versioñ s j as follows:s it,j = L ∑ l=1 w l s i,t−l,j where w = (w 1 , . . . , w L ) is a fixed vector of backward looking weights. The lagging is necessary to account for the time from infection to death. Social distancing actions taken on day t will not result in an immediate change in death rates; rather, the effect will show up many days in the future. We therefore tune this weight vector based on prior knowledge of the distribution of lags between infection with COVID-19 and death. Specifically, w is a Gaussian kernel centered 23.5 days in the past, with a standard deviation of 6 days. This is based on published estimates of time from contraction to illness onset [Lauer et al., 2020] and on time from illness onset to death [Zhou et al., 2020] .