Useful Links for COVID Forecasting

The Reich lab is compiling a list of teams to go after future COVID-19 numbers. Really nice place to see what others are doing:

https://github.com/reichlab/covid19-forecast-hub/tree/master/data-processed

https://reichlab.io/covid19-forecast-hub/

----


https://www.cdc.gov/coronavirus/2019-ncov/covid-data/forecasting-us.html

https://covid19.healthdata.org/united-states-of-america

https://covid-19.tacc.utexas.edu/projections/

https://github.com/reichlab/covid19-forecast-hub

https://github.com/shaman-lab/COVID-19Projection

https://www.covidanalytics.io/projections_documentation

COVID-19: Modeling Growth Rates and Acceleration Directly

A few weeks ago Arena posted our work on COVID. I probably wasn't keeping up closely enough with the different methods out there, but while reading Andrew Gelman's blog I came upon Michael Levitt's early analysis on COVID-19 projections, and the importance of modeling Daily Percent Growth rate as a function of Cumulative cases. This was such great work and very similar to what we ended up doing and open sourcing.

Here's an example plot of Levitt's regression of Growth rates against normalized China cases:

And here's a similar view but on some of the top US states by cumulative COVID cases per capita:

It's nice to know that other people came to this view point of "throwing" out first principle SIR modeling and looking instead at this "phase space" -- growth rate in cases vs. cases -- type view point. It's very accurate/reliable, despite noise in the data. We've used a Statespace model  / rolling regression to react to recent changes in acceleration -- seems like Levitt used a linear regression but we will definitely explore that.

Testing rates:

Also on Levitt's analysis: it's nice to hear another voice echoing the importance of modeling daily growth rates, and how that is a valid thing to track even in the presence of test censorship. We've actually put some effort into estimating and accounting for censorship with a Case observation error model. You can read about it here

The wider we go on nets, the more they look like GP's...

Like Rasmussen mentions in his book, seems like people are finally getting around to experimentally verifying:

https://github.com/thegregyang/GP4A