ABSTRACT
Objectives: The SARS-CoV-2 pandemic has been characterized by sharp, rapid increases in disease incidence, following by a relative long and slow decrease in new cases. This unusual curve was particularly surprising as governments instituted drastic measures to stop the disease progression. This study was designed to model the post-peak decline in SARS-CoV-2 infection cases as a function of social distancing scores, daily tests, population density and average family sizes in select US counties. Methods: Data for SARS-CoV-2 cases and daily testing counts was obtained from The COVID Tracking Project. The family sizes and population density were obtained from the US census. Social distancing scores were purchased from UnaCast. Family size, population density and social distancing scores were categorized by quartile, with lowest quartiles used as reference in the models. Two forecasting models, an exponential smoothing and auto-regressive integrated moving average (ARIMA) model, were built on data from New York Queens, New York Kings and Illinois Cook counties. Root mean square error (RMSE) and Akaike information criterion (AIC) were evaluate for both model types. Results: The forecast for infection rates post-peak using the exponential smoothing method produced models with AIC and RMSE of 340.7 and 12.6 for New York/Queens, 184.7 and 6.45 for Illinois Cook and 243.9 and 14.3 for New York/Kings, respectively. ARIMA models for all three areas resulted in AIC and RMSE values of: 242.12 and 7.31 for New York/Queens, 138.45 and 5.75 for Illinois Cook and 476.8 and 4.63 for New York/Kings, respectively. The calculated R squared value was greater for the exponential smoothing model versus the ARIMA for all counties and ranged from 0.45 (Illinois Cook) to 0.56 (New York Kings). Conclusions: The exponential smoothing method was more reliable than the ARIMA method for predicting the downwards trend following a COVID-19 peak, despite relative low R squared values.
ABSTRACT
Objectives: With the spread of the SARS-CoV-2 virus worldwide, governments have adopted stringent measures to prevent disease spread. As lockdowns are being eased, models to evaluate potential resurgence of disease are increasingly important. The aim of this study is to compare methodologies to predict incidence of COVID-19 for US counties. Methods: Reported number of COVID-19 positive cases were obtained from CDC, Social distancing scores (SDS) from Unacast, Population Density from the US Census data and testing rates obtained from the CDC website. The data assessed was during the period February 28, 2020 to May 28, 2020. Poisson and linear regression models were built to predict the number of reported cases using 1-week lagged SDS, tests per day and population density. Damped Holt linear trend (DHLT) coefficients and moving averages were calculated by using the daily number of cases in the latest 14 days. All the models were built at a county level. The following 4 methodologies were compared: Poisson Regression, Linear Regression, DHLT and simple moving average (SMA). Data from the month of June was used to validate the results. Results: US Counties were ranked in terms of annualized incidence of disease from highest to lowest and the top 100 counties were identified for each methodology. Counties that were predicted to be within the top 100 were compared to those that ended up being in the top 100, as per reported counts. The Poisson and linear regressions both correctly identified 45 out of top 100 counties. Whereas SMA and DHLT only identified 36 and 29 counties, respectively. Conclusions: Linear Regression and Poisson regression were the most accurate in predicting high incidence. In our study, confounding factors like usage of masks or changes in behaviors were not included. Further research on these different factors are needed to improve prediction accuracy.
ABSTRACT
Objectives: Social distancing (SocD) has been rapidly adopted as SARS-CoV-2 infection cases increased. This study’s aim was to evaluate the association between the SocD stringency and mortality from SARS-CoV-2 infections. Methods: Number of new COVID-19 fatalities and tests performed per day per state was obtained from The COVID Tracking Project. Social distancing activity was obtained from Unacast. Unacast creates SocD scores using an average of three dimensions based on cellular location data for 15-17 million people per day: the extent of person-to-person encounters, changes in average distance traveled and visitation to non-essential venues compared to a pre-COVID-19 period. SocD scores were categorized in this study as follows: <2 (reference), 2-2.9 (lower), 3-3.9 (medium), 4-5 (high). Poisson regression models were built to evaluate the impact of SocD score category on COVID-19 daily mortality, adjusting for population density and daily testing volume. The association between SocD_score category and new fatalities were modeled using a 1- to 10-week lag time. Results: Models with only 1-week lag showed positive associations between all SocD_score categories and increasing fatalities, the highest mortality proportion ratio (MPR) being observed for the median SocD_score category (MPR: 3.80, 95%CI: 2.07-7.00), suggesting that populations likely adopted higher SocD measures with increasing fatalities in the prior week. For the most stringent SocD_score category: models using 2- to 10-week lags showed SocD to be protective for mortality (MPR at 2 weeks: 0.79, 95%CI: 0.46-1.35, minimum MPR observed at week 11: 0.20, 95%CI: 0.10-0.40). For the lower and medium category of SocD_scores, it took 11 weeks to see a protective effect between SocD and fatalities. Conclusions: A protective effect of social distancing versus COVID-19 fatalities was shown after 2 weeks for the most stringent, and after 11 weeks, for the more relaxed, SocD measures. Further research on other confounders (eg. mask usage) is required.