This is a big PR that does a bunch of different things. It:
Replaces the log–negative binomial model of read viral counts with a logit–binomial model. Resolves #62. (This also removes one of two non-identifiable overdispersion parameters.)
Fits rather than specifies the variance in the "true" value of the predictor for each sample.
Adds a hierarchical model of P2RA coefficients, with different fine_locations having different coefficients. (2) and (3) resolve #116 .
Fixes a bug in handling viruses with no observed reads.
Fits all viruses (except for hep b and c, which are missing some data). Resolves #115.
Adds a generator for all the fits we want to do: every virus, every taxid, every predictor type. This is used in fitting and for tests.
Adds a text file summarizing the fits, fit_summary.txt. (Planning to make this summary machine readable in a future PR)
Converts the model's coefficients (which are on a logit scale) to the more interpretable "relative abundance at 1 in 1,000 {prevalence,incidence}"
Converts annual incidence to weekly incidence to put it on a more similar scale to prevalence.
This is a big PR that does a bunch of different things. It:
fine_locationshaving different coefficients. (2) and (3) resolve #116 .fit_summary.txt. (Planning to make this summary machine readable in a future PR)