We show that a simple model of COVID-19 that incorporates feedback from disease prevalence to disease transmission through an endogenous response of human behavior does a remarkable job fitting the main features of the data on the growth rates of daily deaths observed across a large number countries and states of the United States from March to November of 2020. This finding, however, suggests a new empirical puzzle. Using an accounting procedure akin to that used for Business Cycle Accounting as in Chari et al. (2007), we show that when the parameters of the behavioral response of transmission to disease prevalence are estimated from the early phase of the epidemic, very large wedges that shift disease transmission rates holding disease prevalence fixed are required both across regions and within a region over time for the model to match the data on deaths from COVID-19 as an equilibrium outcome exactly. We show that these wedges correspond to large shifts in model forecasts for the long-run attack rate of COVID-19 both across locations and over time. Future research should focus on understanding the sources in these wedges in the relationship between disease prevalence and disease transmission.