[R-sig-ME] lmer vs SAS results

[Douglas Bates]

Thanks very much for all your work on this question, Ben.  Did you
happen to check the convergence trace for the problematic model in
lme4?

I would look at it myself but I'm still a little out of it from a bout
of pneumonia.  About 3 a.m. on Wednesday with a raging fever
(sometimes close to 104 F) and piercing headaches for hours on end
plus the continuous cough, I managed to convince myself that the
headaches were caused by brain cancer. I felt I would need to write to
the list to apologize for never having gotten lme4 to version 1.0 and
for never having finished the book on lme4

On Wed, Nov 24, 2010 at 11:11 AM, Ben Bolker <bbolker at gmail.com> wrote:
>  I got a little carried away: see
> <http://www.math.mcmaster.ca/~bolker/misc/preyswim.pdf> for details.
>
>  My conclusions:
>
>  * lme and lme4a agree with SAS (and not lme4) in estimating the MLE of
> among-trial variance as >0.  There are a variety of differences between
> lme4 and lme4a, I don't really know why lme4 performs suboptimally in
> this case.
>
>  * lme4a on R-forge (r1088) does not build on my system [gives the
> 'drtrs' symbol missing error I posted yesterday]; if I revert to r1080 I
> can get it built.  (Doug, Martin?)  I was using a slightly older version
> here.
>
>  * If you're really trying to test a hypothesis here (rather than find
> the best predictive model), and if you're obeying the magic "p=0.05"
> rule, you may be out of luck; the p-value for the LRT between the model
> with (swim+light) and (swim) alone is 0.0537.  This should (?) be a
> fairly reliable number because the number of groups and data points is
> fairly large (lme gives a similar p-value for the F test, which it
> claims has 80 denominator df).  Looking at the pictures, I don't see
> much of an effect of light jumping out at me except (maybe) at the
> lowest light level in the "preyswim=N" group. Maybe the trial effect is
> blurring the picture, or ???
>
> On 10-11-23 07:16 PM, David Duffy wrote:
>> On Tue, 23 Nov 2010, Beth Holbrook wrote:
>>
>> lme4:
>>>       Df    AIC    BIC   logLik  Chisq Chi Df Pr(>Chisq)
>>> model3  4 214.21 229.98 -103.103
>>> model2  9 215.26 250.74  -98.628 8.9494      5     0.1111
>>> model1 14 219.15 274.35  -95.575 6.1062      5     0.2960
>>>
>> SAS:
>>>               AIC    BIC    -2 Log Likelihood
>>> model1   216.7  251.1   188.7
>>> model2   213.3  235.4   195.3
>>> model3   214.2  224.0   206.2
>>>
>>
>> Well, SAS agrees with lme here (with method="ML"):
>>
>> Model df      AIC      BIC     logLik   Test  L.Ratio p-value
>> m1    14 216.7262 271.9254  -94.36309 1 vs 2  6.59408  0.2526
>> m2     9 213.3203 248.8055  -97.66013 2 vs 3 10.88519  0.0537
>> m3     4 214.2055 229.9767 -103.10273
>>
>> Directly maximizing the likelihood (using AS319), I get the
>> m2 v. m3 LRTS to be 10.8852.
>>
>> I haven't evaluated likelihood at the lmer and SAS solutions yet, but
>> obviously the likelihood surface will be fairly flat.
>>
>> Cheers, David Duffy.
>>
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>