Open jamesmbaazam opened 1 month ago
Oh interesting. I assume this is because of numerical overruns for very large numbers (i.e because the prior distribution is multiplier case values a incidence of 40 million might allow some prior values to be 10^3 * 40 million) I would guess we could improve this by improving the prior distribution
or usinng a different default model where you enforce some concept of population size.
Did you look at see if NUTs has the same scaling problems? I would guess it would do but maybe not? If not that is interesting
Sorry, I missed this message. I'll try with NUTS and see.
In my bid to get
pathfinder
running for the vignette in #695, I realised that it was quite likely failing because of the extremely large case reports in the simulated data (approx 40M at peak), so I decided to run a few data scenarios to see how the method performs with different case numbers. My suspicion has so far been confirmed, as the method performs well with case numbers between 100 and 10K, but starts to fail from there. Also notice that the credible intervals get narrower as the data magnitude increases. I am wondering why this is the case from the theoretical perspective, and if there is a way to improve the method to handle such large case numbers.The log is quite rich, so I've attached it here for further discussion. pathfinder_scenarios_log.txt
Created on 2024-07-24 with reprex v2.1.1