SDN 2016-20

[QLLZ]

Data pull: HASH: 6772ea9da4f43ddc3da11c494a573af2ca3dd3c8 config

[QLLZ]

Model run: HASH: 6772ea9da4f43ddc3da11c494a573af2ca3dd3c8

[QLLZ]

After double checking, SDN2016-2020 config was missing before.need to regenerate config and rerun from data pull

[QLLZ]

new config

[QLLZ]

rerun from data pull: data pull on dev: HASH c9a89d4ac5e4f63cb75fa5c006a900ece2839199

[QLLZ]

Model run: HASH 6772ea9da4f43ddc3da11c494a573af2ca3dd3c8

[QLLZ]

Country data report

[eclee25]

Convergence, fits, Rhats look good, but there is a strong effect of the adjacency matrix on the rate maps that seems unusual.

Opinion: Investigate adjacency matrix

[javierps]

Maybe same issue on population density as in NER. Large and sparsely populated places are hard for the model to fit?

[javierps]

Dagar fixed should solve this, waiting for ID 45 report.

[QLLZ]

Config

Data pull:

HASH: 211913e0eb7d570ed62c48bc1ae9ecd6671a6cbc

[QLLZ]

Failed stan model run log file

[QLLZ]

Rerun on dev_u_combs_fix

HASH: e56580fa5ddab00293e31ff90139351f80ae3f6c

[QLLZ]

country data report

[QLLZ]

Overall cases are overestimated. convergency looks good/

[eclee25]

Instability in sd_w traceplots but otherwise the model convergence looks good. There is overestimation in 2016 and 2019 and it's not clear why. Perhaps the 2016 overestimation is related to the multi-year observation from 2016-2018?

Investigate the 3-year multi-year obs in OCs 20189? 21107? It's hard to tell which OCs they are in based on the cdr -- may be better to verify which OCs have these observations in the stan input file.

I think we could perhaps try running it after removing or handling the 3-year multi-year obs in a different way?

[QLLZ]

double checked the 3-year observation in 21089. it looks correct based on the source doc

[javierps]

The 2016 data is an imputed observation, so may make sense it is hard for the model to fit.

For std_dev_w, given the very high values of rho I don't think this has a large impact (the w traceplots look very good).

I don't think the multi-year observations are breaking things here, but could try a run without it.

[eclee25]

Rerun with updated national od parameter

[eclee25]

@javierps determined that the initialization failures we're having with the new od admin0 parameter are caused by numerical stability issues in the evaluation of the censored observation probability -- added a possible fix to the Stan code that improves numerical stability of the probability is close to 0 or 1 --> implemented as dummy censoring variables

[javierps]

Sep 2023 Production run: chain differences in std_w, rest convergence Ok, Rhats Ok.

Suggestion: accept.

[eclee25]

Noting that Tab 6 2017 estimates are between the two observations and the PI does not include either obs, likely due to the slightly restricted overdispersion param here. Other diags look okay.

Discuss this but leaning to Approve. We are okay with the 2017 estimate but we want to rerun the cdr to show the 2016 imputed observations.

Action Required - rerun cdr

[eclee25]

2016 imputed obs is ok. Approve