Not sure whether this is still of interest. There was some interest about speed of lmer and lmer2. Attached are comparisons for seven rather complex models estimated from a set of eye fixations measured during reading isolated sentences (N=29114 reading fixations; 3 random partially crossed groups of 109 subjects, 144 sentences, and 550 words. Note that most words occur only in one sentence but some words (e.g., articles) can also occur more than once per sentence. Also, for all subjects some of the sentences are missing. In addition to the three intercept variances, I always estimated subject variances for three slopes with their covariances forced to zero. The models all use the same predictors; they differ in what subset of fixed effects of the full factorial is estimated. The number of model parameters (fixed and random effects) varies from 406 to 158. I am quite amazed that there were absolutely no problems with convergence, etc.. (It is also nice that the results make a lot of sense, too.) lmer2 was faster than lmer on the largest model (588 vs. 642). For the smaller models lmer was typically faster than lmer2 but not always. Given its overall speed even for such complex models, the timing differences are not important for these data. Most importantly, the estimates are incredibly similiar for all seven models. Details (fit statistcs and random effects for lmer and lmer2, but not fixed-effect estimates) are given in the attachment. R version 2.4.1 Patched (2007-01-26 r40579) i386-apple-darwin8.8.1, lme4 0.9975-13. Best Reinhold ?---- Reinhold Kliegl, Dept. of Psychology, University of Potsdam, Karl-Liebknecht-Strasse 24-25, 14476 Potsdam, Germany phone: +493319772868, fax: +493319772793 http://www.psych.uni-potsdam.de/people/kliegl/ ---- Reinhold Kliegl, Dept. of Psychology, University of Potsdam, Karl-Liebknecht-Strasse 24-25, 14476 Potsdam, Germany phone: +493319772868, fax: +493319772793 http://www.psych.uni-potsdam.de/people/kliegl/
lmer/lmer2 timing results
1 message · Reinhold Kliegl