PROC MIXED vs. lmer()

This post was kindly contributed by Gregor Gorjanc - go there to comment and to read the full post.

I am using R for a few years now. I like it a lot. Previously I used SAS (notably its modules SAS/BASE, SAS/STAT, SAS/IML, …), because that is the statistical package of choice at my department. I had a hard time when I switched to R, mainly because I was thinking in the SAS way. However, I made a switch because I liked Rs’ open-source nature and its vibrant community. Up to now I did not want to make any comparisons between them. I simply use R because I like it and get along well with using it. Recently I helped a student to fit a quite big model. She used SAS and PROC MIXED failed due to “out of memory”. I was surprised, since SAS is known for its great stability and performance with big datasets. I was almoast sure that function lmer() in lme4 package in R will fail also, but could not resist to try it out. To my surprise, the model was fitted without problems. Perhaps I could tweak SAS to use more memory, but I did not invest time to study that option. Maybe new versions of SAS would perform better.

I used the following code in SAS 8.02:


proc mixed data=podatkiratio;
class genotype litter eage sex type year month herd hys;
model bw = genotype litter eage sex type year month;
random herd hys;
run;
quit;

This is the code for R 2.6.0, lme4 0.99875-9:


lmer(bw ~ genotype + litter + eage + sex + type +
         year + month +
         (1|herd) + (1|hys),
         data=podatki)

And the results, which include also the description of the model size:


Linear mixed-effects model fit by REML
Formula: bw ~ genotype + litter + eage + sex + type + year + month + (1|herd)  (1|hys)
  Data: podatki
   AIC    BIC logLik MLdeviance REMLdeviance
152403 152899 -76149     152011       152297
Random effects:
Groups   Name        Variance Std.Dev.
herd     (Intercept) 0.313    0.559  
hys      (Intercept) 0.147    0.383  
Residual             0.297    0.545  
number of obs: 85035, groups: herd, 346; hys, 9653

Fixed effects:
...
## genotype, 13 levels
## litter, 3 levels
## eage, 9 levels
## sex, 2 levels
## type, 2 levels
## year, 16 levels
## month, 12 levels

I am using SAS and R under Windows XP, running on Intel(R) Core(TM)2 CPU 6300 @ 2×1.86GHz with 2 GB RAM.

This post was kindly contributed by Gregor Gorjanc - go there to comment and to read the full post.

Post navigation