[R] GML with tweedie: AIC=NA

Gordon Smyth smyth at wehi.EDU.AU
Tue Aug 14 23:15:06 CEST 2007


Dear Catarina,

As I've already indicated, I have no plans to include AIC for tweedie 
families, because the AIC computation is much more computationally 
intensive and less reliable than the actually fitting of the glm. 
That fact that dtweedie() is not currently available for Windows is a 
demonstration of my point.

I'm sorry that this causes you a problem. I don't pretend to know 
your problem but, in my experience, it can often be an advantage to 
do your model selection more hands on rather than relying on an 
automatic algorithm.

Best wishes
Gordon

At 09:06 PM 14/08/2007, Catarina Miranda wrote:
>Thanks for your reply, but I still couldn't solve the problem.
>
>I am using the package statmod, and I need the AIC because I want to 
>use the step function (I am modelling many species, so I would 
>prefer to do the step automatically).
>I can't find the tweedie  package in the R packages list and I don't 
>know how to download it from the package source available in 
><http://cran.r-project.org/src/contrib/Descriptions/tweedie.html>http://cran.r-project.org/src/contrib/Descriptions/tweedie.html 
>.
>However, after running the  tweedie.R file from that package source 
>I am able to use the dtweedie() function, but still I didn't figure 
>out a way to do the step (or to get the AIC from the glm command).
>
>Thank you again for your help;
>
>Catarina
>
>
>On 13/08/07, Gordon Smyth 
><<mailto:smyth at wehi.edu.au>smyth at wehi.edu.au > wrote:
>Dear Catarina,
>
>I prefer to leave the AIC value as NA for the tweedie GLM family
>because it takes extra time to compute and is only occasionally
>wanted. It's easy to compute the AIC yourself using the dtweedie()
>function of the tweedie package.
>
>Best wishes
>Gordon
>
>At 03:05 AM 14/08/2007, Catarina Miranda wrote:
> >Dear Gordon;
> >
> >I have also sent this email to R help mailing list, so I apologize
> >for duplicated mailing.
> >I am modelling densities of some species of birds, and I have a
> >problem with a great amount of zeros.
> >I have decided to try GLMs with the tweedie family, but in all the
> >models I have tried  I got an NA for the AIC value.
> >Just  to check the problem I've compared the a glm using the
> >Gaussian family with the identity link and a glm using the tweedie
> >family with var.power=0 and link.power=1. These are equal, as
> >expected, except the fact that the tweedie output gives me an NA 
> for the AIC.
> >Could you help me with this problem?
> >Below you can find the two outputs I refer.
> >
> >Best Wishes;
> >
> >Catarina
> >
> > > summary(glm(formula=ACIN~DIST_REF+DIST_H2O+DIST_OST+
> > COTA+H2O_SUP+vasa,family=gaussian(link="identity")))
> >Call:glm(formula = ACIN ~ DIST_REF + DIST_H2O + DIST_OST + COTA
> >+     H2O_SUP + vasa, family = gaussian(link = "identity"))
> >Deviance
> >Residuals:       Min         1Q     Median         3Q        Max
> >-0.112792   -0.042860  -0.021113  -0.006311   1.551824
> >Coefficients:              Estimate Std. Error t value
> >Pr(>|t|)  (Intercept)
> >-6.625e-02  5.454e-02  -1.215   0.2256  DIST_REF     3.581e-06
> >1.336e-05   0.268   0.7889  DIST_H2O    -
> > 3.168e-05  1.527e-05  -2.074   0.0391
> >*DIST_OST    -1.799e-05  1.953e-05  -0.921   0.3579  COTA
> >5.648e-04  2.470e-04   2.287   0.0230
> >*H2O_SUP     -2.172e-04  3.994e-04  -0.544   0.5870  vasa
> >3.695e-02   4.573e-02    0.808   0.4199  ---Signif. codes:  0 '***'
> >0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
> >(Dispersion parameter for gaussian family taken to be 0.02151985)
> >     Null deviance: 5.6028   on 257  degrees of freedomResidual
> > deviance: 5.4015  on 251  degrees of freedomAIC: -249.33
> >Number of Fisher Scoring iterations: 2
> >
> >
> > > summary(glm(formula=ACIN~DIST_REF+DIST_H2O+DIST_OST+
> > COTA+H2O_SUP+vasa,control=
> > glm.control(maxit=750),family=tweedie(var.power=0, link.power=1)))
> >Call:glm(formula = ACIN ~ DIST_REF + DIST_H2O + DIST_OST + COTA
> >+     H2O_SUP + vasa, family = tweedie( var.power = 0, link.power =
> >1),     control = glm.control (maxit = 750))
> >Deviance
> >Residuals:       Min         1Q     Median         3Q        Max
> >-0.112792  -0.042860  -0.021113  -0.006311   1.551824
> >Coefficients:              Estimate Std. Error t value
> >Pr(>|t|)  (Intercept) -
> >6.625e-02  5.454e-02  -1.215   0.2256  DIST_REF     3.581e-06
> >1.336e-05   0.268   0.7889  DIST_H2O    -3.168e-05   1.527e-05
> >-2.074   0.0391
> >*DIST_OST    -1.799e-05  1.953e-05  -0.921   0.3579  COTA
> >5.648e-04  2.470e-04    2.287   0.0230
> >*H2O_SUP     -2.172e-04  3.994e-04  -0.544   0.5870  vasa
> >3.695e-02   4.573e-02   0.808   0.4199  ---Signif. codes:  0 '***'
> >0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
> >(Dispersion parameter for Tweedie family taken to be 0.02151985)
> >     Null deviance: 5.6028  on 257  degrees of freedomResidual
> > deviance: 5.4015  on 251  degrees of freedomAIC: NA
> >Number of Fisher Scoring iterations: 2
> >
> >
>



More information about the R-help mailing list