search

epiforecasts / forecast.vocs

Contains models and tools to produce short-term forecasts for both case and sequence notifications assuming circulation of either one or two variants. Tools are also provided to allow the evaluation of the use of sequence data for short-term forecasts in both real-world settings and in user generated scenarios.
http://epiforecasts.io/forecast.vocs
Other
8 stars 3 forks source link

Divergent transitions #6

Open seabbs opened 3 years ago

seabbs commented 3 years ago

Both models currently have some issues with divergent transitions which can indicate model misspecification, non-optimal numeric implementations, or a difficult likelihood surface for NUTS to explore.

Explore this in more detail by looking at pairs plots, and divergent transitions by parameter plots. As in both models issue should first be looked at using the single strain model only.

seabbs commented 3 years ago

Looking at the pairs plot of different models/combinations (using inst/analysis/model-eval.R) it looks like modelling overdispersion is a major driver of fitting issues. This makes sense as there is a conflict/correlation between overdispersion and the allowed variation in the growth rate. The current suggestion is to explore this in terms of model fit and for this reason #49 has made this an explorable option.