Data-Science-Brigade / modelo-epidemiologico-sc

Bayesian epidemiological model, with a focus on the state of Santa Catarina and its micro and macroregions. https://arxiv.org/abs/2104.01133

MIT License

2 stars 1 forks source link

Revise cases likelihood #62

Closed jonjoncardoso closed 3 years ago

jonjoncardoso commented 3 years ago

What this PR does

Introduce a variable phi2 to model the scale of the cases likelihood independently of the deaths likelihood
Set infection_overestimate lower bound to 1, the intuition being that the number of estimated cases should be at least the number of reported cases
Comment figure saving parts of the code in script combine_independent_results.R. --> This does not change the steps required to reproduce the weekly model

Closes #47 Closes #56 (by simply commenting the parts of the code where figures are saved)

jonjoncardoso commented 3 years ago

These changes have slightly improved the fit on the B graph (reported deaths vs estimated deaths) for the two microregions in which the third peak was more pronouced:

SC_RSA_GRANDE_FLORIANOPOLIS SC_RSA_GRANDE_FLORIANOPOLIS_3panel

SC_RSA_OESTE SC_RSA_OESTE_3panel

jonjoncardoso commented 3 years ago

The only STAN model that led to a low number of effective sample size was SC_RSA_FOZ_DO_RIO_ITAJAI_3panel. This produced a nonsensical peak in graph A:

SC_RSA_FOZ_DO_RIO_ITAJAI_3panel

Suggestions for tackling this:

Tighten up the prior distributions of the y and tau

jonjoncardoso commented 3 years ago

Another side effect of this version is that it does allow for the number of estimated cases to be a bit lower than the reported number of cases (graph A).

SC_RSA_ALTO_URUGUAI_CATARINENSE

SC_RSA_ALTO_URUGUAI_CATARINENSE_3panel

This happens despite the lower bound to infection_overestimate=1 because our likelihood equation (negative binomial of cases ~ prediction allow for some variation.

jonjoncardoso commented 3 years ago

My conclusion is that this change of parameters are welcome as they slightly improve the fit of the third peak in graphs A and B for the most densely populated micro-regions of the state -- the ones that have been notoriously harder to fit -- even though it introduces some small drawbacks.

NataliaDelCoco commented 3 years ago

Suggestions for tackling this:

Tighten up the prior distributions of they and tau

Thinking about last week diagnostic analysis, this conclusion is another evidence that reparametrize y and/or tau could lead to significant improvement.

I believe the benefits overcome the drawbacks (=