Hi, I wanted to raise a concern regarding how the compartmental models beyond SIR collect observations, found in:
SEIR
SEIRD
OverdispersedSEIR
SuperspreadingSEIR
My suggestion is that we change this observed value to be E2I, rather than S2E. If we were to condition our models on real world data such as to get these observations as accurate as possible, then we would not have access to the number of exposed individuals, but would have a fairly good idea of the number of new infected individuals, which would be a part of the E2I transition, not S2E.
Hi @nsk367,
thanks for the suggestion. Actually our early prototypes did observe E2I rather than S2E as you suggest. Lucy Li and I changed it to S2E because, in our infection data, each infection was back-dated to the estimated date of exposure. I agree that a downside of that sort of preprocessing is that it prevents end-to-end Bayesian models (where we could try to learn the date of testing), but that’s the data we had.
Note that those models are half intended as reusable and half intended as things you can fork. I hope it will be easy for you to fork one of the models to fit your own data.
Cheers!