1

Running R 3.1.0, and I'm having a lot of trouble with lme since the latest 1.0 update. I have a data set I was working with under an older version, getting fine parameter estimates, no issues with collinearity and no failures of convergence. The same analysis was then failing to converge after updating to 1.0, using both bobyqa and Nelder-Mead and playing around with increasing max iterations. I poked around and it seems that others have encountered the same problem and solved it by installing lme4.0 and running that instead. I did that, but am now getting the error typically associated with collinearity:

Error in mer_finalize(ans) : Downdated X'X is not positive definite, 1.

As far as I can tell, I have no issues with collinearity in my dataset (certainly not perfect collinearity) and none of the data set or my code has changed since the older version that used to run fine. It seems that a similar problem has been reported here but I don't really understand the solution, or whether it is actually the same issue.

I used this bit of code, from Florian Jaeger's blog, to install lme4.0: 

install.packages("lme4.0", type="both",
                 repos=c("http://lme4.r-forge.r-project.org/repos",
                         getOption("repos")[["CRAN"]]))

But don't know how to install the version suggested in that link, lme4 0.999902344-0, or if that might be problematic for other reasons. Any help would be greatly appreciated.

Roder
  • 53
  • 4
  • why are you not installing from cran? that webpage you linked to is over a year old – rawr May 14 '14 at 19:06
  • installing from CRAN gets me the latest 1.0 version of lme4 -- that version is apparently having some issues with models failing to converge, which used to converge fine in 0.9-X. Or is there something I should be doing through CRAN that I'm missing? – Roder May 14 '14 at 19:27
  • you can select the older version you were using from http://cran.r-project.org/src/contrib/Archive/lme4/ and see if it works again? I don't really keep up with the updates, sorry – rawr May 14 '14 at 19:32
  • are you sure you're not getting the [frequently-reported false positive convergence warnings](http://thread.gmane.org/gmane.comp.lang.r.lme4.devel/11674/)? Is there a chance you can send the data to `maintainer("lme4")` ? I confess I don't understand the problems with `lme4.0` -- I wonder if it might be due to a change in the `Matrix` package ... – Ben Bolker May 14 '14 at 19:46
  • Thanks for your reply, Ben. Just sent it. – Roder May 14 '14 at 20:03
  • one other thing to try: can you re-install lme4.0 *from source*, i.e. rebuild the binaries? It occurs to me there might be some binary incompatibility with newer versions of Matrix ... – Ben Bolker May 14 '14 at 20:57
  • @Roder, haven't seen data? – Ben Bolker May 15 '14 at 13:26
  • I sent it directly through gmail -- is there a better way? – Roder May 19 '14 at 21:29
  • I don't remember whether this fell through the cracks or not. I know it's a long time later, but if you're still interested in the answer could you re-send? – Ben Bolker Oct 23 '14 at 20:00

0 Answers0