R help - Aug 2007 - (Ben Bolker) AIC and logLik for logistic regression in R and S-PLUS

Leandra Desousa <sousa <at> ims.uaf.edu>
writes:
>> > I am using 'R' version 2.2.1 and 'S-PLUS' version
6.0; and I loaded the
>> > MASS library in 'S-PLUS'.
>> > 
>> > I am running a logistic regression using glm:
>> > 
>> >  >summary(mydata.glm)
>> > Call:
>> > glm(formula = COMU ~ MeanPycUpT + MeanPycUpS, family =
binomial,data =
>> > mydata)
  [snip]
>> >     Null deviance: 30.316  on 21  degrees of freedom
>> > Residual deviance: 23.900  on 19  degrees of freedom
>> > AIC: 29.9
>> >
--------------------------------------------------------------------
>> > 'R'
>> > -----------------
>> >  > AIC(mydata.glm)
>> > [1] 29.89986
>> > 
>> >  > logLik(mydata.glm)
>> > 'log Lik.' -11.94993 (df=3)
>> > -----------------
>> > 
>> > 'S-PLUS'
>> > -----------------
>> >  > AIC(mydata.glm)
>> > [1] 71.03222
>> > 
>> >  > logLik(mydata.glm)
>> > [1] -31.51611
>> > -----------------
>> > 
>> >
>> > 1) Which AIC value is the correct one?
>> > 2) Which log-likelihood value is the correct one?
>    AIC and log-likelihood are often defined differently
> in software packages -- specifically, additive constants
> can be included or excluded as long as they are done consistently,
> without affecting inferences from the model.  The absolute
> values of AIC and logLik aren't that important; the only thing
> that really matters are differences among models.
Dear Ben Bolker,
Thank for clarifying the details on how AIC and Log-likelihood may be 
defined differently in statistical packages.
I did not know that.
>   Have you
> tried comparing models within R and within S-PLUS to establish
> whether they give the same inferences (I would guess they do)?
Yes, I did compare models within R and within S-PLUS and the inferences 
were the same.

>> > 3) If  'extractAIC' in 'S-PLUS' and all values in
'R' are the correct
>> > ones, and the 'AIC' and 'logLik' in
'S-PLUS' values are wrong then:
>> >    Why 'S-PLUS' cannot retrieve a log-likelihood value
from my glm
>> > object('mydata.glm'), even though it is using
log-likelihood to
>> > calculate its residual deviance?
>>     
>    It's very hard for us to debug S-PLUS!  Some (most?) of
> us on the list don't even have S-PLUS installed any more ...
> Perhaps you should ask this question on an S-PLUS mailing list,
> or of the (paid) S-PLUS technical support team ...
>  
>    And the obligatory R-list nags:
>
>  * it would help if you upgraded your R version -- R 2.3.0
> came out in April 2006, and we're now up to 2.5.1
>   * it's the MASS "package", not the MASS "library"
>
>   cheers
>     Ben Bolker
I just upgraded my R version thank you.
I may just use R for the analysis instead, since I only have an older version of
S-PLUS (2002 version 6).
Thank you for your comments.
Leandra de Sousa


------------------------------