2
$\begingroup$

Can anyone give me some pointers of how to get around the following error messages? I'm trying to run mixed models for significance using lme4 and got these messages:


Does anyone know what I should do about the following messages that I get when trying to run mixed models for significance one lme4?

Model failed to converge with max|grad| = 0.00448103 (tol = 0.001, component 1)

Model failed to converge: degenerate Hessian with 2 negative eigenvalues

fit warnings: fixed-effect model matrix is rank deficient so dropping 18 columns / coefficients convergence code: 0 unable to evaluate scaled gradient Model failed to converge: degenerate Hessian with 2 negative eigenvalues failure to converge in 10000 evaluations

Warning messages: 1: In vcov.merMod(object, use.hessian = use.hessian) : variance-covariance matrix computed from finite-difference Hessian is not positive definite or contains NA values: falling back to var-cov estimated from RX 2: In vcov.merMod(object, correlation = correlation, sigm = sig) : variance-covariance matrix computed from finite-difference Hessian is not positive definite or contains NA values: falling back to var-cov estimated from RX


I'm new to statistics and I have no idea what these mean or what to do about them. Could someone at least explain to me what the problem is? I understand that is is something about the model fit but other than that I have no idea what the problem is.

Many thanks in advance.

$\endgroup$
  • $\begingroup$ Possible to post a data set? $\endgroup$ – SmallChess Oct 28 '17 at 10:56
  • $\begingroup$ I suspect your data-set has insufficient information to fit the model you have tried. What happens if you fit a simpler model? $\endgroup$ – mdewey Oct 28 '17 at 11:09
  • $\begingroup$ Some simpler models do work but some I get the messages. I have a feeling that either my data might be too sparse. I had 70 replies in total but I split them by age and I think one of the age groups isn't represented. What should I do about this? I could try running them without the age column but i lose some information this way. Also my experiment was randomised and I had 3 levels for one of my variable so I essentially made 3 surveys and combined them later. Each survey had a different set of stimuli so maybe one of the surveys has too few respondants. $\endgroup$ – Helen Nov 10 '17 at 16:37
  • $\begingroup$ I have to go physically to a computer elsewhere to be able to post my actual data, but I can outline my experiment: I had 39 sentences recorded twice, once as statements and once as questions. The original soundclips were 3 words long and then were split into files containing just the first word, the first and second word and then all three words. The point of my experiment is try to find out how much of the sentences the participant needs to identify the 'mode' of the utterance. The statements have a rise at the end like Australians do so it's a harder task than it sounds. My groups 1/2 $\endgroup$ – Helen Nov 10 '17 at 16:44
  • $\begingroup$ My groups are divided into their varied experiences with these phenomena, self selected 'i use it myself', 'i know someone who uses it', 'i've seen it on tv' or 'none'. I want to see if experience makes a difference. I could take things like gender and age out, I thought i should collect the data and that they might be interesting but they're not central to the hypothesis. I'll try running it without considering those but what if it still doesnt work? Would bootstrapping be an option? This is my first statistical experience so I appreciate any informed replies that might come. 2/2 $\endgroup$ – Helen Nov 10 '17 at 16:47

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.