[ADMB Users] Fwd: glmmADMB predict with negative binomial

[Ben Bolker]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On 14-12-10 07:35 PM, jnancheta wrote:
> ---------- Forwarded message ---------- From:
> <sheryn.olson at maine.edu> Date: Fri, Nov 28, 2014 at 8:26 AM 
> Subject: glmmADMB predict with negative binomial To:
> admb-users+owners at googlegroups.com
> 
> 
> Hello, May I post a question?
> 
> I would like to plot a predictive graph of hare pellet count
> response vs. conifer saplings.
> 
> It is unclear to me what predict.glmmADMB extracts when the family
> is nbinom, with log-link.
> 
> The documentation says [type   Whether to return predictions on the
> scale of the linear predictor ("link") or the scale of the data
> ("response")]
> 
> Does that mean specifying      type = "link"    will extract
> exponentiated fits? Or do I exponentiate the fit AFTER it is
> extracted by predict?

  using type="response" will give you the exponentiated fits; using
type="link" will give you the results on the log scale
(I will say that in general you should be able to look at the results
and make a pretty good guess -- for example, any negative values
indicate for sure that you're still on the log scale ...)

> 
> R-code follows Call: glmmadmb(formula = pellets ~ season *
> t.con.splgs + offset(ln.days) + (1 | stand/plot) + (1 | hareyr),
> data = hv, family = "nbinom") AIC: 8391.6 Coefficients: Estimate
> Std. Error z value Pr(>|z|) (Intercept)               -4.67230
> 0.17840  -26.19  < 2e-16 *** season[T.smr]             -0.23872
> 0.09559   -2.50  0.01251 * t.con.splgs                0.01686
> 0.00447    3.77  0.00016 *** season[T.smr]:t.con.splgs -0.01653
> 0.00402   -4.12  3.8e-05 ***
> 
> 
> and ## prepare predictive dataframe: ## Conifer saplings, wi is
> winter intercept ##  and predict pellets based on square root
> transformed conifer saplings/0.1 ha
> 
> predctCSwi <- data.frame( t.con.splgs = rep(seq(from =
> min(hv$t.con.splgs), to = max(hv$t.con.splgs),length.out =
> 100),2), ln.days=rep(log(30.25),200),   ###### account for month 
> season = factor(rep(1:2, each = 100), levels = 1:2,
> labels=levels(hv$season)))
> 
> predctCSwi <- cbind(predctCSwi, predict(wi.tconsplgs, predctCSwi,
> type = "link", se.fit=TRUE)) head(predctCSwi) t.con.splgs  ln.days
> season       fit    se.fit 1   0.0000000 3.409496    wtr -1.262801
> 0.1784000 2   0.7142493 3.409496    wtr -1.250759 0.1784448 3
> 1.4284985 3.409496    wtr -1.238718 0.1785467 4   2.1427478
> 3.409496    wtr -1.226676 0.1787054 5   2.8569971 3.409496    wtr
> -1.214635 0.1789210 6   3.5712464 3.409496    wtr -1.202593
> 0.1791931
> 
> # Then scale up to pellets/ha/month (phm) #  and backtransform
> saplings to per 0.1 ha by squaring the square-root values 
> predctCSwi <- within(predctCSwi, { pellets <- exp(fit) phm <-
> pellets/1.5*10000 LLha <- (exp(fit - 1.96 * se.fit))/1.5*10000 ULha
> <- (exp(fit + 1.96 * se.fit))/1.5*10000 con.splgs.1ha <-
> (t.con.splgs^2) })
> 
> 
> Thank you, Sheryn Olson
> 
> 
> 
> _______________________________________________ Users mailing list 
> Users at admb-project.org 
> http://lists.admb-project.org/mailman/listinfo/users
> 

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.11 (GNU/Linux)

iQEcBAEBAgAGBQJUiOlwAAoJEOCV5YRblxUHD6kIANXL+5XZ8xoMmjaIvw24pCMt
gWbVvB9R8gMVAMgAxnknb/I9wtmxK8gPkSPRD/KlAKM/AY1SDKip2RQu8YDbwVvA
8WxxyVLUgT4jflVewwhQ1fGUCOyG3amaDX9mM4mBj+cSGM/GpyFLnmHlKOAYA8x+
TTwa2c+NkEDZMrpWAuIRAgiWCKr8zj0aLrkqVn0MSJ6lJSuRLreovWTOwtWiq8tq
w6b8rzy3zdIQ7d3mzpF92fQUel+uLPUUZuLsBtXCyCMPWDc/X+9s5THrIGgiTtGF
HZMvaS/HAg303MlQdu++iexNC9YRh7BsayI4VVGYh+VnJgXtoHN31nREeXPVGTA=
=g8RD
-----END PGP SIGNATURE-----