Dear Roger and Florian, Thanks so much for your comments. A model with random slopes but no random intercepts (e.g. (0+type|spk)) also seems to lead to singularity. As I said in my previous message, this happens too for a model with both random intercept and slop (e.g. (1+type|spk)). I understand Roger's suggestion to merge levels 'e' and 'i'. However, if I am fitting the model, it's precisely to compare the level means :-)
I have therefore fitted a model with a random intercept and calculated the CI for the level means using Baayen's pvals.fnc(). I suppose that the CI obtained this way are not equivalent to the ones obtained with post-hoc comparison procedures (e.g. TukeyHSD). Does anyone have an idea how to do this with a mixed model? For the moment I am satisfied with these intervals though. Along with some lattice graphs I think I can already get a good idea of what's going on in my data. Francisco On 7/31/07, T. Florian Jaeger <[EMAIL PROTECTED]> wrote: > Dear Francisco, > > as Roger said, too strong correlations between the variances of the random > effects can lead to a singularity in the estimation of the > variance-covariance matrix for the random effects. This can also happen, if > any of the variances are indistinguishable from zero. Like Roger, I do not > have a clear understanding of the underlying fitting procedure, but too the > best of my knowledge the singularity is due to one of the underlying > parameters determining the random effects whose value is being optimized is > too close to zero. > > I suggest the following: look at the a couple of different models. I would > start by comparing a model with only a random intercept vs. a model with > only the random slopes (the "type | spk" part). If a model with only the > random slopes does not converge, the singularity due to some of the levels > of "type" being indistinguishable with regard to their random effects and > you should do what Roger suggested. If the model with only random slopes > DOES converge, you can compare it against a model with only the random > intercept. Too a first approximation, you may use the model fit measures, > e.g. AIC to compare the two models. When you compare these models, keep in > mind that the slopes have more DFs than just the intercept. If a model with > only the random intercept has basically the same model fit quality as a > model with only the random slopes, than it seems that (given the fixed > effects that you are considering) the random slopes do not seem to do much ( > i.e. the different types do not seem to affect your dependent variable, at > least not under the assumption that their effect is normally distributed). > Have a look at Baayen, Davidson, & Bates, 07 for more detail on how to > compare different models based on their random effects. > > Florian > > > On 6/23/07, Francisco Torreira <[EMAIL PROTECTED]> wrote: > > > > Hello, > > > > I am fitting a mixed model that prompts the following warning messages: > > > > Warning messages: > > 1: Estimated variance-covariance for factor 'spk' is singular > > in: `LMEoptimize<-`(`*tmp*`, value = list(maxIter = 200L, tolerance = > > 1.49011611938477e-08, > > 2: nlminb returned message function evaluation limit reached without > > convergence (9) > > in: `LMEoptimize<-`(`*tmp*`, value = list(maxIter = 200L, tolerance = > > 1.49011611938477e-08, > > > > Although the model is fitted, R does not let me run simulations on it > > with mcmcamp(). This is the error message I get: > > > > > mcmcsamp(full, n=10000) > > Error: inconsistent degrees of freedom and dimension > > Error in t(.Call(mer_MCMCsamp, object, saveb, n, trans, verbose, > deviance)) : > > error in evaluating the argument 'x' in selecting a method for > function 't > > > > The model was: > > full <- lmer(an ~ type + (1 + type | spk) - 1) > > > > My design included 5 speakers (spk) and 5 utterance types (type). For > > each combination of speaker and utterance type there were > > approximately 20 repetitions. If I fit a more reduced model with no > > random effect for type within speakers, as in lmer(an~type+(1|spk)), > > no warning appears. Here is the summary of my full model: > > > > Linear mixed-effects model fit by REML > > Formula: an ~ 1 + type + (1 + type | spk) - 1 > > AIC BIC logLik MLdeviance REMLdeviance > > 4663 4747 -2311 4653 4623 > > Random effects: > > Groups Name Variance Std.Dev. Corr > > spk (Intercept) 1568.3 39.601 > > typee 1037.3 32.208 -0.745 > > typeg 1303.7 36.107 -0.659 0.946 > > typei 1780.9 42.200 -0.778 0.976 0.864 > > typel 757.4 27.521 -0.725 0.839 0.826 0.865 > > Residual 598.8 24.470 > > number of obs: 498, groups: spk, 5 > > > > Fixed effects: > > Estimate Std. Error t value > > typea 78.87 17.88 4.410 > > typee 18.49 12.13 1.524 > > typeg 50.86 14.26 3.566 > > typei 11.42 12.48 0.915 > > typel 14.94 12.46 1.199 > > > > Correlation of Fixed Effects: > > typea typee typeg typei > > typee 0.570 > > typeg 0.491 0.898 > > typei 0.240 0.851 0.722 > > typel 0.699 0.758 0.739 0.622 > > > > I wonder if the high correlations correlations between several > > utterance types and the intercept in the random part of the model > > aren't causing all this trouble. I would appreciate any comment on the > > warnings. > > > > Thanks in advance, > > Francisco Torreira > > > > -- > > Francisco Torreira > > PhD Candidate in Hispanic Linguistics > > University of Illinois at Urbana-Champaign > > > > > https://netfiles.uiuc.edu/ftorrei2/www/index.html > > tel: (+1) 217 - 778 8510 > > _______________________________________________ > > R-lang mailing list > > [email protected] > > > https://ling.ucsd.edu/mailman/listinfo.cgi/r-lang > > > > -- Francisco Torreira PhD Candidate in Hispanic Linguistics University of Illinois at Urbana-Champaign https://netfiles.uiuc.edu/ftorrei2/www/index.html tel: (+1) 217 - 778 8510 _______________________________________________ R-lang mailing list [email protected] https://ling.ucsd.edu/mailman/listinfo.cgi/r-lang
