karimn / covid-19-transmission Goto Github PK

Running the model with just the US data takes a very long time and some chains finish before others even reach 10% of their iterations.

Drop subnational treatment effects in mobility model and R0.

https://www.apple.com/covid19/mobility

https://mc-stan.org/loo/articles/loo2-lfo.html

Test with synthetic data.

For example, the Poland report produces:

## Warning in grid.Call(C_textBounds, as.graphicsAnnot(x$label), x$x, x$y, :+1: 
## conversion failure on 'Łódź' in 'mbcsToSbcs': dot substituted for <c5>

Labels in plots end up with missing characters.

Some solution I found Googling:

• used cairo_pdf
• pdf.options(encoding='ISOLatin2')

Need to make sure this works for all languages.

@wwiecek found that we are hitting convergence problems because of non-identifiability of R0 and imputed cases, so setting a more informative (restrictive) prior on imputed cases eliminates some of these issues.

We also discussed using a prior for imputed cases that is conditional of the population size of sub-regions.

@wwiecek I'm opening this issue so we can keep track what we discussed as a solution. Please add anything you think is relevant here. I'm assigning to you, but feel free to re-assign back to me if you have a clear implementation of these priors you want me to do. I'm going to go ahead and do a cluster run with the new restricted priors you checked in.

Plot mobility data time series.

In the spread data for France the variable new_confirmed is NA for dates before 5/13. We need this data for the "contact rate" trend of the mobility model.

This might be the cause of bug #28

if (!is_null(params$country_code)) {
results %<>%
mutate(run_data = map(run_data, filter, fct_match(country_code, params$country_code)))
}

Error in map(run_data, filter, fct_match(country_code, params$country_code)) :
object 'run_data' not found

SE, FR, and IT are having problems. Need to investigate this.

Investigate job 64349420

Hey, so you pushed PR and PL runs (Pushed pl.RData and pe_.RData to data/mobility/results on main-karimn_) and I wanted to review them this morning, but am not sure what is the fastest way to get a national report because I'd need results object (use_subnat_data)

I can run the chunks of run_mob.R that go before & after the model, but maybe there a quicker/cleaner way to get this?

Problem appears to be drive by three smaller sub-regions.

mobility_report.pdf

https://mc-stan.org/users/documentation/case-studies/qr_regression.html

All hierarchical parameters are currently uncentered. I was reading this discussion https://discourse.mc-stan.org/t/low-ess-and-high-rhat-for-random-intercept-slope-simulation-rstan-and-rstanarm/9985/6 and thought I should try adding the option to have these param centered and see what happens to our runs.

Run high adapt delta to see if we can eliminate all divergent transitions.

Job 64395309 problem countries: DO (singleton), SE, ES, MY, NL, FR (high divergence but Rhat < 1.01)

A run including AR AU CA PT runs fine and a separate run for PL also runs fine, but combining results in high Rhat and very low ESS measures. @wwiecek suspects that the root of the problem is the beta coefficients of the mobility model.

Sorry, we discussed this before but I forgot. S(t) (1-cumulative_cases) is a transformed parameter, so we can't directly put a probabilistic statement on it. But do you know of any other simple method to do it, @karimn? Looking at another country where S(t) tends to 0 made me think of how nice it would be to have a control over S(t) at particular t's. And we even have data for this!

Attached are the reports for the model with contact rate trend

it_trend_report.pdf

@wwiecek I'm starting to get this all the time now. Now sure if it's the latest data.

When I add the stringency_index to the model using the new dataset (which can be found in the Dropbox), it reports the following:

Error in new_CppObject_xp(fields$.module, fields$.pointer, ...) :
Exception: mismatch in dimension declared and found in context; processing stage=data initialization; variable name=design_matrix; position=0; dims declared=(2361,4); dims found=(2250,4) (in 'model359444027750_mobility' at line 50)

See #28 and #31

On the side, I ran with only two countries (with "--no-pooling"), PY and PT, and the model has some problems. Same run without "--no-pooling" -- so it's hierarchical over countries ---- worked well. I wonder if with no pooling we end up with the same problem we see when we have a completely uninformative prior on the SD param that was causing divergences before.

This is a good case to investigate to identify non-identification or weak identification problems. It's relatively small and quick to run, and has no sub-regions so we can run out the hierarchical parameters as the cause of divergence.