Refactor prior construction and make clearer the logic #566

SamuelBrand1 · 2024-12-20T12:55:48Z

This PR is (hopefully) an improvement to #565 aimed at addressing @seabbs concerns about the implementation. Concerns I share because I think the opaque implementation got in the way of me implementing the priors I wanted (and @seabbs reviewing them from a scientific point of view).

github-actions · 2024-12-20T12:56:02Z

Try this Pull Request!

Open Julia and type:

import Pkg
Pkg.activate(temp=true)
Pkg.add(url="https://github.com/CDCgov/Rt-without-renewal", rev="new-way-to-specify-priors", subdir="EpiAware")
using EpiAware

pipeline/src/constructors/remake_latent_model.jl

SamuelBrand1 · 2024-12-20T13:04:18Z

The idea of this approach is that the remake_latent_model function first calls down to an internal function that sets our expectation of the fluctuations (std) and autocorr based on the IGP and whether the latent process representation is stationary or not.

This passes to a function which then constructs the latent process in one of two ways:

If stationary, it aims to (approximately) get the stationary std and autocorr as prior expected.
if non-stationary its this but aimed at the step by step fluctuations.

seabbs · 2024-12-20T13:10:28Z

pipeline/src/constructors/remake_latent_model.jl

+The autocorrelation is expected to be 0.1.
+"""
+function _make_target_std_and_autocorr(igp; stationary::Bool)
+    if igp == Renewal


This much nicer but my training for you makes me want it to dispatch down based on igp vs ifelse it

Aha thing here is that igp is the type itself rather and an instance of a type so got to remember how to do that

And it was indeed nicer!

codecov-commenter · 2024-12-20T13:17:36Z

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 90.96%. Comparing base (1b3d397) to head (dd689bd).

Additional details and impacted files

@@               Coverage Diff                @@
##           specific-priors     #566   +/-   ##
================================================
  Coverage            90.96%   90.96%           
================================================
  Files                   60       60           
  Lines                  863      863           
================================================
  Hits                   785      785           
  Misses                  78       78

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

pipeline/src/constructors/remake_latent_model.jl

seabbs

Yeah this is much nicer. I think it could get even nicer but happy with it as it is

SamuelBrand1 · 2024-12-20T13:31:40Z

SamuelBrand1 · 2024-12-20T13:32:16Z

Some of those priors look too confident at no outbreak

SamuelBrand1 · 2024-12-20T14:01:14Z

As ever, Julia rewards thinking about dispatch. I think this version is getting there. Ready for re-review and sci review

seabbs · 2024-12-20T14:42:01Z

Nice:

Prior predictive review:

Direct infections looks too tight
AR for Expontial growth is a bit tight
Otherwise I think this is looking reasonable

seabbs

Implementation wise I think this is nice enough now (though I think there are a few alternative patterns that better match our more general approach and worth thinking about why/if they aren't easy to do).

So just the science to discuss now

SamuelBrand1 · 2024-12-20T15:08:51Z

Nice:

Prior predictive review:

Direct infections looks too tight

AR for Expontial growth is a bit tight

Otherwise I think this is looking reasonable

Agree with all this

SamuelBrand1 · 2024-12-20T15:24:39Z

Bumped the variance for models mentioned by @seabbs above. Think it looks ok... Obviously same $\log R_t$ gives very different dynamics with different mean GI in the Renewal model.