[R] glmmPQL in R and S-PLUS 6 - differing results

[Andrew Robinson]

Professor Ripley,

> These are optimization problems with multiple local maxima, and 
> like any complex statistical fitting problem you should not expect all 
> programs to give the same answer.  

Examining the log-likelihoods from the anova statements, while fit1 and m1 are similar, the maximum for fit2 (due to R) is higher than the maximum for m2 (due to S-plus).

R:

>  anova(fit, fit2)
      Model df      AIC      BIC    logLik   Test  L.Ratio p-value
fit      1  4 87.79085 94.12493 -39.89543                       
fit2     2  3 86.41927 91.16983 -40.20964 1 vs 2 0.628421  0.4279

S-PLUS:

>  anova(m1, m2)
   Model df      AIC      BIC    logLik   Test L.Ratio p-value
m1     1  4 87.00400 93.33808 -39.50200                     
m2     2  3 91.75447 96.50503 -42.87724 1 vs 2 6.75047  0.0094


Can we interpret that in any way? For example, is it reasonable to infer that in this case the R results, suggesting no significant difference between the model, are preferable?

Andrew

Associate Professor, Forest Biometrics
University of Idaho
Moscow ID 83843