1

I just had a question as to how to identify the degree of correlation of the random effects in a mixed effects model using R's lme4 package.

I'm investigating a mixed model using lme4, which has the instructions to specify an option in the model specification as to whether the random effects of the model are correlated.

I've tried using the VarCorr and V function on my model after running it, but its returning unrealistically small values, and I'm running into trouble interpreting the values.

Any help/advice as to how to determine which option to use would be greatly appreciated.

kjetil b halvorsen
  • 63,378
  • 26
  • 142
  • 467
user124123
  • 280
  • 2
  • 9
  • 1
    You'll get some good information here: [R's lmer cheat sheet](http://stats.stackexchange.com/q/13166/7290). – gung - Reinstate Monica Jun 19 '14 at 13:30
  • 1
    Need more information please. What is your random-effects specification in the model formula (please specify whether variables appearing in it are numeric or categorical predictors). What is the `V` function? Can you show the results you're getting? – Ben Bolker Jun 19 '14 at 16:28
  • My model is pretty big, 20+ variables both categoric and continuous, the random effects are a subset of the fixed effects, the response is continuous, the VarCorr function is giving me results 10^-3 which I don't believe. – user124123 Jun 20 '14 at 08:46
  • 1
    It is very common for overfitted models to be singular (e.g., variances collapse to zero, correlations collapse to +/- 1): http://rpubs.com/bbolker/6226 shows some simple examples. If you don't like this behaviour you can regularize the model by adding priors, with the `blme` package ... – Ben Bolker Jun 21 '14 at 21:20
  • Hi Ben many thanks for the info I'll check out blme – user124123 Jun 24 '14 at 08:12
  • 1
    @user124123 Have you found the answer to your problem following Ben's suggestion? Could you consider posting it as an answer and accepting it to share it with the other users? – Charlotte R Sep 20 '16 at 15:21

0 Answers0