search

athowes / multi-agyw

Spatio-temporal estimates of HIV risk group proportions for AGYW across 13 priority countries in sub-Saharan Africa
https://athowes.github.io/multi-agyw
MIT License
2 stars 2 forks source link

Very large local WAIC values #2

Closed athowes closed 2 years ago

athowes commented 3 years ago

Local WAIC very large for x_eff values of zero (this more often occurs for sexpaid12m). DIC output also has problems resulting in either NA or implausible values. Use the local values to find rows causing the problem, then try to resolve somehow. Could it be related to survey weights? Have added the diagnostic output but remains to say how to solve it

athowes commented 3 years ago

"Now there remains the question about the DIC infinite of my model that is still unresolved. Do you have any idea?"

This is usually an overflow issue, mainly because of an weakly indentified model. Like DIC would require an integral like \int exp(x) dx and if the marginal variance of x is large, then exp(x) at the upper limit might evaluate to infinity, giving an infinite DIC.

athowes commented 3 years ago

This is happening only for the spatial only models (just one DHS), not for the spatio-temporal models. I think this could be because it's a problem when you have zero counts in a category (and it's neighbours?) and there is more smoothing in the spatio-temporal. But typing that out I'm not sure how much sense it makes.

athowes commented 2 years ago

I think this was related to trying to fit four categories when there is no data in the fourth category. Haven't seen it recently so closing for now.