Using unmarked's predict with QAIC
Javan Bauder
I was wondering if unmarked's 'predict' function can incorporate QAIC instead of AIC (bearing in mind earlier posts discussing the use/merits of QAIC!)? I have been working with a couple snake datasets, running single-season occupancy models, and testing GOF with AICcmodavg's 'mb.gof.test' function to estimate c-hat (the overdispersion parameter). I typically observe "moderate" overdispersion (c-hat: 1.5-2.5) so I use QAIC in my subsequent model selection and model-averaging. I would like to use unmarked's 'predict' function but it apparently only compares models using AIC. Is there a way to use QAIC instead? I have been using the AICcmodavg package's 'modavgpred' function which allows users to specify a value for c-hat. When comparing results from unmarked's 'predict function and AICcmodavg's 'modavgpred' function (with c-hat = 1), I get very similar and highly correlated (r = 0.99995) predicted values and the SE's are also very similar and highly correlated (r = 0.9864). I really appreciate having the 95% prediction intervals that unmarked provides so it would be great to be able to use QAIC.
Thank you very much,
Javan
Kery Marc
Dear Javan,
I think you confuse a couple of things. One is goodness of fit, the other is model selection and the third one is prediction, in particular the computation of the uncertainty around predictions. You seem to want to inflate the prediction standard errors or confidence intervals by the additional uncertainty represented by a c.hat, which you obtain in a Gof test. You can’t do that with unmarked’s own predict function. However, Marc Mazerolle’s AICcmodavg package has a number of very useful functions that work specifically for unmarked fitted objects and includes functions for prediction with inflated SEs/CIs. To be honest, I don’t know whether this also applies to occupancy models, but it does for N-mixture models. I attach a blurb from our upcoming AHM book that describes prediction for Nmix models fit with pcount() and assume that something similar works also for occupancy models fit with occu().
Kind regards --- Marc
--
You received this message because you are subscribed to the Google Groups "unmarked" group.
To unsubscribe from this group and stop receiving emails from it, send an email to
unmarked+u...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.
Kery Marc
Dear Javan,
Here is part 2 of my reply (I should have read your email till the end, sorry !). To get 95% prediction intervals that are inflated by some c.hat value, you can predict on the link scale, the code for and result of which may look something like this:
> modavgPred(cand.set = list(fm12), newdata=newData, parm.type = "detect",
+ type = "link", c.hat = 3.04) # Could be used to get 95% CIs
Model-averaged predictions based on entire model set:
mod.avg.pred uncond.se
1 0.58 0.56
2 0.53 0.52
3 0.49 0.48
4 0.45 0.45
5 0.41 0.43
6 0.37 0.41
Then you can take the mean plus/minus twice that SE and backtransform (apply the inverse logit) and this will give you approximate, c-hat inflated 95% prediction intervals.
Best regards --- Marc
MarcJ.M...@uqat.ca
To complement Marc Kéry's answer, modavgPred( ) indeed works with occupancy models with overdispersion. The current output includes model-averged predictions and unconditional SE's, but I'm working on including the CI's on model-averaged predictions in AICcmodavg for most models fit in unmarked. Now that I got some grant applications out of the way, I hope to include them in the next version of the package this fall.
Sincerely,
Marc
____________________________________
Marc J. Mazerolle
Département des sciences du bois et de la forêt
2405 rue de la Terrasse
Université Laval
Québec, Québec G1V 0A6, Canada
Tel: (418) 656-2131 ext. 7120
Email: marc.ma...@sbf.ulaval.ca
Envoyé : jeudi 29 octobre 2015 04:16
À : unma...@googlegroups.com
Objet : AW: [unmarked] Using unmarked's predict with QAIC
Kery Marc
Dear Marc,
This will be a very welcome addition to your highly useful package; thank you !
Best regards --- Marc
Javan Bauder
Javan
Kery Marc
I am also looking forwards to the new AICcmodavg, but the twice the SE up and down rule and then backtransforming is quite valid, too. No problem to publish that.
Regards -- Marc
Sent: 29 October 2015 19:20
To: unmarked
Subject: [unmarked] Re: Using unmarked's predict with QAIC