search

grantbrown / ABSEIR

A Flexible and Computationally Efficient SEIRS Modeling Framework
6 stars 3 forks source link

2 questions: p_ir static or dynamic? & option to use something besides Bayesian/MCMC to make the model? #9

Open nkniep888 opened 4 years ago

nkniep888 commented 4 years ago

Hello, We are trying to use ABSEIR to analyze local COVID-19 data and had two questions about the package: 1) For the variable p_ir, or the probability of transition from infected to removed, we plan to use the disease's CFR (case fatality rate). This value changes depending on the time point at which the analysis is being performed. Is this variable intended to be static or dynamic in the package? 2) Does ABSEIR have an option to use something besides Bayesian/MCMC to make the model, like a main field model without the posterior distributions?

As a side note, I see reference to "The ABSEIR documentation" on the vignettes page but I'm unsure if there is a separate page/document that encompasses this?

Thank you!

grantbrown commented 4 years ago

Hi @nkniep888,

Right now, p_ir is indeed static over time and space. It can be modeled as exponential or path-specific, but the software wasn't originally designed to deal with epidemics where the CFR changed. This is one of the features I'm actually currently working on adding to the package, though I've first needed to undertake some cleanup of how the simulations are initiated. The structure of passing in a huge list of parameters has grown in complexity over time, and led to some fragility, so I'm moving towards having the model component objects play a larger role. You can see the initial work here, but it's not yet ready. I'm hoping to get this up and running in the next week or so, and can keep you posted.

As far as alternate methods go, it's currently set up just to do ABC (not MCMC, which is often extremely slow for these problems). It is, however, possible to just do straight up simulations without posterior estimation if you want, followed by post processing in R. I can post some example code for that if such an approach would meet your needs.

Edit to clarify: the CFR is more naturally incorporated into the data model than into p_ir. R encompasses people recovering as well as passing away, and if the probability is limited to just deaths, then that would inflate I and the corresponding exposure probability. The new features will be at the data model level - of the new removed, what proportion are observed via various mechanisms (hospitalization/death/ventilation etc.)

nkniep888 commented 4 years ago

Thanks for the response! If you could provide some example code that would be great. I think we are trying to avoid the MCMC computations, as you mentioned they are time and resource intense.