No Data for MD and 1 other state?

[cschanck]

For the 8/6 9:02 run, "Last model run: 8/6 at 9:02PM", summary data here: https://github.com/covidestim/covidestim-products/raw/master/2020-08-06-allstates-ctp/summary.csv

You only show 48 states of data, and indeed in the summary data, there is no MD (no DC either, but I am unsure what the other missing state is).

Is there a reason? Transient? THe WashPo article that cites your data: https://www.washingtonpost.com/graphics/2020/health/coronavirus-herd-immunity-simulation-vaccine/?hpid=hp_visual-stories-8-12_no-name%3Ahomepage%2Fstory-ans&itid=hp_rhp__visual-stories-8-12_no-name%3Ahomepage%2Fstory-ans

includes all the states, naturally.

[marcusrussi]

Hey Chris,

Thanks for the note. Typically, one or two states per day have model runs that don't complete in the allotted 4 hours of time we give for each state to run. These failed states usually have an unusually high number of divergent transitions, if you let them run to completion. Our belief is that this is the result of unrealistic state data (caused by, for instance, huge data dumps of reclassified deaths), or bad RNG seeds used during MCMC, or a combination of the two issues.

Our current pipeline fails these states, and they are dropped from the summary data for that day. We are in the process of building a new pipeline, which retries these states automatically on timeout with a different seed. Then, if they still fail, we include the most recent successful run of the state in the summary data, as a failsafe.

We currently don't have a silver bullet for this problem, and are certainly open to ideas!

[cschanck]

Huh, that's interesting, and I can appreciate the difficulty. Just seems odd for MD in particular, when the data looks (intuition can lie) pretty clean. Data is hard!

[dmadeka]

Is the same true for counties like LA? Theyve had a lot of dumps recently -

[marcusrussi]

Hi Dhruv, Yes, we often have trouble fitting LA. They had a few clear inconsistencies with their data last month which may have been corrected by now, but in general the production version of the model sometimes has trouble fitting states/counties with large numbers of cases and deaths, and LA county alone has more cumulative cases and deaths than many states do. We hope to fix these issues in an upcoming package release.

…

-Marcus

On Mon, Jan 4, 2021 at 1:15 AM Dhruv Madeka ***@***.***> wrote: Is the same true for counties like LA? Theyve had a lot of dumps recently - — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#4 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAHA65LYXNYCSZYIHD2XUWDSYFMG7ANCNFSM4PXZQAEQ> .