search

karimn / covid-19-transmission

0 stars 0 forks source link

Italy does not converge #45

Open karimn opened 4 years ago

karimn commented 4 years ago

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.

wwiecek commented 4 years ago

It didn't have this problem before, right? e.g. there is a model of Italian data with this extra logistic trend that works fine?

karimn commented 4 years ago

Yes, but probably not since the latest data update. I have no clear hypothesis of why this is happening. The model itself has changed very little since it was last run.

On Thu, Jul 30, 2020 at 5:22 AM Witold Wiecek notifications@github.com wrote:

It didn't have this problem before, right? e.g. there is a model of Italian data with this extra logistic trend that works fine?

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/karimn/covid-19-transmission/issues/45#issuecomment-666253810, or unsubscribe https://github.com/notifications/unsubscribe-auth/AABB552V5MUYOL7LP4LIH3DR6E3TXANCNFSM4PL2BA6A .

wwiecek commented 4 years ago

How about re-running with imputed_cases fixed to a particular value (e.g. eyeballing median in the "good" regions)? Or a strongly informative prior to force imputed_cases to be somewhere in that region.

karimn commented 4 years ago

Ok, I'll have to get to it next week.

On Thu, Jul 30, 2020 at 10:07 AM Witold Wiecek notifications@github.com wrote:

How about re-running with imputed_cases fixed to a particular value (e.g. eyeballing median in the "good" regions)? Or a strongly informative prior to force imputed_cases to be somewhere in that region.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/karimn/covid-19-transmission/issues/45#issuecomment-666386232, or unsubscribe https://github.com/notifications/unsubscribe-auth/AABB552QQNX4YDVMHRI5XPLR6F5ATANCNFSM4PL2BA6A .

karimn commented 4 years ago

Here's the latest with 10 imputed cases forced everywhere. Still not good.

mobility_report.pdf

wwiecek commented 4 years ago

I can see that Junyi has the same problem in his 2 new runs. This is a bit baffling, but the pattern is clear: it's the regions that have low prevalence that cannot fit.

Do you have a coverged run for Italy saved? Could we check the parameters that were used to run it? Didn't we increase hyperSD on R0 between when the old model was ran and now? Seems like R0 is just allowed to wander into very high regions. BTW couldn't we truncate R0 at e.g. 6? We know it's impossible for it to be this high.

There are also two problems, one with data and one with constraints:

1) @Junyi-Que I see negative number of cases for Basilicata and Calabria. @karimn I don't think this impacts fitting because cases are not an input, but double-checking in case this is significant in some other way

2) @karimn we get non-monotonic S(t):

image

I guess this is because of extreme values of R pushing cases into negative. I think you told me we can't constrain that easily, but flagging this in case I misremembered.

karimn commented 4 years ago

Yes, the first place I'll investigate is look at older versions of the data, model, etc. Separately, I'll see what reintroducing the constraints does.

On Fri, Aug 7, 2020 at 6:09 AM Witold Wiecek notifications@github.com wrote:

I can see that Junyi has the same problem in his 2 new runs. This is a bit baffling, but the pattern is clear: it's the regions that have low prevalence that cannot fit.

Do you have a coverged run for Italy saved? Could we check the parameters that were used to run it? Didn't we increase hyperSD on R0 between when the old model was ran and now? Seems like R0 is just allowed to wander into very high regions. BTW couldn't we truncate R0 at e.g. 6? We know it's impossible for it to be this high.

There are also two problems, one with data and one with constraints:

1.

@Junyi-Que https://github.com/Junyi-Que I see negative number of cases for Basilicata and 2.

@karimn https://github.com/karimn we get non-monotonic S(t):

[image: image] https://user-images.githubusercontent.com/20645798/89634927-1c553500-d89e-11ea-8a74-86cd57dd0655.png

I guess this is because of extreme values of R pushing cases into negative. I think you told me we can't constrain that easily, but flagging this in case I misremembered.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/karimn/covid-19-transmission/issues/45#issuecomment-670441667, or unsubscribe https://github.com/notifications/unsubscribe-auth/AABB553NOZOHAKCJURDXEZ3R7PHENANCNFSM4PL2BA6A .

karimn commented 4 years ago

From @wwiecek

Actually, one more question (I will need more time to write some of these things up anyway): Karim, do you think having upper constraints on R0 parameters would help with some of the worst convergence issues? I think national_R0 should not be more than 1 on log scale, ever. Same for subnational. Maybe we could use Italy as a testing model for this concept?

wwiecek commented 4 years ago

@karimn don't think if you're working on this today, but if yes, give me a shout, we could discuss live if that helps

karimn commented 4 years ago

I'm not working on this today. I'll try to come back around to it tomorrow. I'll email you in the morning.

On Sat, Aug 8, 2020 at 5:20 AM Witold Wiecek notifications@github.com wrote:

@karimn https://github.com/karimn don't think if you're working on this today, but if yes, give me a shout, we could discuss live if that helps

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/karimn/covid-19-transmission/issues/45#issuecomment-670849508, or unsubscribe https://github.com/notifications/unsubscribe-auth/AABB553IESA2O22X7WIMNHTR7UKEVANCNFSM4PL2BA6A .

wwiecek commented 4 years ago

@Junyi-Que -- we discussed this problem with @karimn and wanted to ask you to help out a bit. (It's simple and similar to running of your last model!) Could you take the Italian model with the latest version of the repo but choose only 1 region? (Imagine it's 1 country with only 1 region -- rest of data does not exist.) There is even an option in the script that should allow picking data for 1 region in 1 country only.

We want to see what happens to the few regions that have bad results. The main one is Calabria, but you will see there is a few problematic ones, e.g. Basilicata. It's all the regions that have very high R0's. If you could run that and post the .pdf of mobility results (like Karim does above), that would be great.

We do this because we want to find out if running the model with just 1 region will fix the MCMC convergence issues.