search

fhui28 / CBFM

Spatio-temporal joint species distribution modeling using community-level basis functions
13 stars 2 forks source link

More (flexible count) distributions? #13

Closed chrishaak closed 2 years ago

chrishaak commented 3 years ago

Wish list - to better deal with overdispersion; maybe poisson-tweedie as in bonat 2018 (maybe with covariate-dependent dispersion as in petterle 2019). R code and examples from both papers are attached in zip files...

bonat_2018_poisson-tweedie.zip petterle_2019_double-poisson-tweedie.zip

fhui28 commented 3 years ago

On a semi-completely different note, add ordered categorical responses well to the list. mgcv has ocat, but not sure if there is an easy way to do updates in TMB.

fhui28 commented 3 years ago

Wish list - to better deal with overdispersion; maybe poisson-tweedie as in bonat 2018 (maybe with covariate-dependent dispersion as in petterle 2019). R code and examples from both papers are attached in zip files...

On the issue of these flexible count distributions, I am not sure if the Poisson-Tweedie [at least without modeling the dispersion parameter] will solve the crazy overdispered counts your are seeing say in bottom trawl survey. Maybe modeling the dispersion parameter will help, but you can also do that with negbin, say.

Actually, I am curious how well VAST and its vast collection of response distributions [pun intended haha] goes for your data. Does it cope with the crazy overdispersion well? What distributions are recommended here? This would definitely be worth checking as that would provide a guide rather than blindly stabbing distributions in the dark

chrishaak commented 3 years ago

Totally agree that covariate-dependent dispersion (or ZI) seem like they are worth looking into, whether in an NB or PT context?

The overwhelming majority of work modeling counts (or biomass) from this dataset (or others with similar sampling methodology) uses the delta approach - modeling pres/abs (i.e., "encounter probability") first, and then modeling nonzero counts (using a zero-truncated NB ) and/or biomass (via gamma or log-transformed response with a gaussian error). So hurdles are pretty much the standard - however given that we both find the "two-stage" approach to be aesthetically unpleasing, it would be nice to find something more parsimonious...

On Mon, Sep 13, 2021 at 8:30 AM Francis KC Hui @.***> wrote:

Wish list - to better deal with overdispersion; maybe poisson-tweedie as in bonat 2018 (maybe with covariate-dependent dispersion as in petterle 2019). R code and examples from both papers are attached in zip files...

On the issue of these flexible count distributions, I am not sure if the Poisson-Tweedie [at least without modeling the dispersion parameter] will solve the crazy overdispered counts your are seeing say in bottom trawl survey. Maybe modeling the dispersion parameter will help, but you can also do that with negbin, say.

Actually, I am curious how well VAST and its vast collection of response distributions [pun intended haha] goes for your data. Does it cope with the crazy overdispersion well? What distributions are recommended here? This would definitely be worth checking as that would provide a guide rather than blindly stabbing distributions in the dark

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/fhui28/CBFM/issues/13#issuecomment-918143541, or unsubscribe https://github.com/notifications/unsubscribe-auth/AORLQXHJPKZQTIBGBVKGXWDUBXVGNANCNFSM5DXKDLCA . Triage notifications on the go with GitHub Mobile for iOS https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675 or Android https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub.

fhui28 commented 3 years ago

Currently in talks with Simon Wood to see if mgcv can be extended to fit zero-truncated Poisson/NB GAMs. The jump from this to hurdle + assumption that spatio-temporal fields are independent between the PA and non-zero part is then not too difficult.

It is possible to use gamlss to do fit truncated GAMs. But this would involve quite a bit of changes to CBFM to make it compatible with gamlss, given it was built with mgcv in mind. Also, initial exploration suggest gamlss is really slow!

fhui28 commented 2 years ago

Hurdle NB and Poisson now implemented. ZI with covariates in zero-inflation parts could be done, but leave that for a future me...especially if someone requests it.