Dear All, I fitted a linear mixed effects model using lmer() two weeks ago, and the model was able to converge. I rerun the model over the weekend (two days ago) and now there are warning messages about max|grad| and negative Herssian that the model failed to converge and the t values from the summary of the un-converged model are not the same as those I originally got from the originally converged one. Any idea why this could happen? Is it because of an update of the lme4 package or something? I saw a post here, http://r.789695.n4.nabble.com/Why-have-my-glmms-stopped-converging-lme4-td4691786.html, where Dr. Ben Bolker suggested that the max|grad| warning are just false positives, but since the un-converged one gives different t-values, I do not know which results I should rely on. E.maximal.model.RT<-lmer(RT~ 1+A*B*C+(1+A*C|Subject)+(1+A*B*C|Item), data=Edata, verbose=2,control=lmerControl(optCtrl=list(maxfun=50000))) The DV is a continuous one, and all the three IVs are categorical and sum-coded. Any suggestions would be greatly appreciated. Best, ZW
lmer model converged before, but couldn't now
2 messages · Zhaohong Wu, Ben Bolker
On 14-10-27 07:12 PM, Zhaohong Wu wrote:
Dear All, I fitted a linear mixed effects model using lmer() two weeks ago, and the model was able to converge. I rerun the model over the weekend (two days ago) and now there are warning messages about max|grad| and negative Herssian that the model failed to converge and the t values from the summary of the un-converged model are not the same as those I originally got from the originally converged one. Any idea why this could happen? Is it because of an update of the lme4 package or something? I saw a post here, http://r.789695.n4.nabble.com/Why-have-my-glmms-stopped-converging-lme4-td4691786.html, where Dr. Ben Bolker suggested that the max|grad| warning are just false positives, but since the un-converged one gives different t-values, I do not know which results I should rely on.
Do you know what version you were using before? What version are you
using now (packageVersion("lme4")) ?
What are the AIC and log-likelihoods of your previous and current models?
The standard error calculation
We certainly haven't updated the version on CRAN in the last two
weeks, but perhaps you had an older version and just now updated it?
The standard error calculation did change in version 1.1-4
<http://cran.r-project.org/web/packages/lme4/news.html>; you can check
whether this affects your results by comparing
sqrt(diag(vcov(fitted,use.hessian=TRUE)))
and
sqrt(diag(vcov(fitted,use.hessian=FALSE)))
(although I think if you are getting a negative-Hessian warning that the
program will have fallen back on use.hessian=FALSE anyway)
E.maximal.model.RT<-lmer(RT~ 1+A*B*C+(1+A*C|Subject)+(1+A*B*C|Item), data=Edata, verbose=2,control=lmerControl(optCtrl=list(maxfun=50000))) The DV is a continuous one, and all the three IVs are categorical and sum-coded. Any suggestions would be greatly appreciated. Best, ZW