Ed Baskerville, Lauren McGough, Timothy M Pollington & Katie Gostic
19 July 2020
This repository contains example code for deconvolving a known delay distribution from observed epidemic time series.
None of the code here is ready for prime time, but we’re sharing it in case it’s a useful starting point for other people.
There is code here for two methods, both with attendant warnings:
- The Richardson-Lucy-type deconvolution used in Goldstein et al. 2009 PNAS, implemented straightforwardly in unoptimized R code. Either a bug in this code or an inherent property of the algorithm makes it extremely sensitive to initial guess. Certainly the observed curve needs to have structure of some kind (i.e. look like a curve) to have a chance of deconvolving into the infection epidemic curve.
- An analogous Bayesian model, implemented straightforwardly in Stan. There are two versions of the model, one (
uncorrelated) where unobserved states are given identical independent priors, and one (randomwalk) where log(unobserved states) follow a Brownian motion. Neither of these implementations have been used for real work, but initial runs seem promising for doing things this way.