Multivariate Regression of Current and N-day Future-looking Deaths and Cases per Million, per Day, aggregated by US State
Sage Betko, Rishabh Shetty, Jeff Morgan. Anne Gibbon, Prahlad G Menon, PhD
- Carnegie Mellon University; Columbia University; Catholic University of America; Matri Design; University of Pittsburgh [prm44@pitt.edu]
We develop a novel analytic approach to modeling current and future COVID-19 risk using the COVID-19 Symptom Survey data aggregated daily by State from CMU, joined with daily time-series data on confirmed cases and deaths from the covidtracking.com project, in order to augment situational awareness of the outbreak. Specifically, we model current and N-day forward-looking estimates for: a) deaths per million (DPM); and b) cases per million (CPM). We find that while Facebook survey evidenced behaviors and attitudes of individual respondents collectively estimates DPM from 0 to 20 days in advance with R^2 north of 80% in out-of-sample testing, whereas the same survey data optimally predicts CPM rates 20+ days in future of a given survey.
a) ReadCMU_StateWise_COVID19SymptomsSurveyData.ipynb : This script creates our dataset of daily aggregated CSV files of the COVID-19 Symptom Survey data by US State curated by CMU, joined with daily time-series data on confirmed cases and deaths from the covidtracking.com project [1,2] b) COVID19_MultivariateModelingAndFeatureSelection.ipynb : This script develops the models that created the results which we report.
Dataset references: [1] Symptoms survey data aggregated by state by day: https://cmu.app.box.com/s/ymnmu3i125go4aue0qxosi3rbcae20bj ; [2] Confirmed cases and deaths data: https://covidtracking.com/api/v1/states/daily.json https://covidtracking.com/api/v1/us/daily.json