search

CDCgov / ww-inference-model

An in-development R package and a Bayesian hierarchical model jointly fitting multiple "local" wastewater data streams and "global" case count data to produce nowcasts and forecasts of both observations
https://cdcgov.github.io/ww-inference-model/
Apache License 2.0
17 stars 2 forks source link

Modify infection feedback prior #228

Closed kaitejohnson closed 1 week ago

kaitejohnson commented 4 weeks ago

Originally was N(500,200) so lognormal(6, 0.4) which was a VERY strong infection feedback forcing term, especially at high incidence.

This PR proposes changing this prior to be centered around 1 but in the range of o to 400 (exp(6) ~ 400).

We plan to test this our in the wweval pipeline, look at performance, and then consider adding this to a version update.

github-actions[bot] commented 4 weeks ago

Thank you for your contribution, @kaitejohnson :rocket:! Your page is ready to preview here

seabbs commented 4 weeks ago

Have you dug in to the original motivation for this? As this is likely to not be very informed by the length of data we are fitting to having a very weak prior means widening the prediction intervals essentially.

Was the motivation flawed or has the change in model over time changed the definition?

kaitejohnson commented 4 weeks ago

Have you dug in to the original motivation for this? As this is likely to not be very informed by the length of data we are fitting to having a very weak prior means widening the prediction intervals essentially.

Was the motivation flawed or has the change in model over time changed the definition?

I tried my very hardest to find where I had documented a bunch of posterior estimates of the infection feedback posterior from 2022-2023s data fits in cfa-forecast-renewal-ww but I wasn't able to find them. In that instance, I think the mistake was mine that in order to see a meaningful (e.g. large) reduction in R(t) the magnitude of the term needed to be around 1e2 bc incidence was being estimated as peaking around 1e-3 -- however this is likely unrealistically low, and I probably was just trying too tight of priors that were already forcing incidence to be artificially low.

Basically I think my strategy at guessing these priors was flawed, and we had an open issue to revisit it but didn't get to it

kaitejohnson commented 1 week ago

Closing this in favor of #236