1

Maybe my question will fail to be specific but when fitting a glme model (using lme4 package in R) I get for one of the parameters SE=1000, with the estimated parameter as high as 16. The variable is a dichotomous variable. My question is if there might be an explanation for such a result, considering that the other parameters have parameters and SE that seem ok

user2520918
  • 21
  • 2
  • WHich variable is dichotomous, predictor or response? An SE can be any number. I could have an effect of 1 and an SE of 1e6 and it be perfectly fine without some more context. Answering my first question just might provide it. – John Aug 20 '13 at 17:27
  • 6
    This question appears to be off-topic because it is about interpreting a statistical result, so belongs on stats.stackexchange.com. – Aaron left Stack Overflow Aug 20 '13 at 17:48
  • Both the predictor and the covariate are dichotomous. The problem is that such a big SE didn't make too much sense and neither the high estimated parameter – user2520918 Aug 20 '13 at 18:05

2 Answers2

4

That's a sign that you have complete separation. You should re-run the model without that covariate. Since its an ME model you may need to do a tabulation of outcome by covariate by levels to see what is happening. More details would allow greater specificity in our answers.

This is a link to a posting by Jarrod Hadfield, one of the guRus on the R mixed model mailing list. It demonstrates how complete separation leads to the Hauck-Donner effect, and it offers some further approaches to attempt dealing with it.

Ben Bolker
  • 211,554
  • 25
  • 370
  • 453
IRTFM
  • 258,963
  • 21
  • 364
  • 487
  • Thanks for your response. What happens is that at each level there is only 1 person having "1" on my covariate and that person always has "1" for the dependent as well. But there are cases with "0" for the covariate which have "1" on the dependent. Can I do anything with this variable or I have to run the models without it? (When I exclude it the model seems to fit ok) – user2520918 Aug 20 '13 at 18:01
3

You may be seeing a case of the Hauck-Donner effect. Here is one post that discusses it, you can read the original paper or search the web for additional discussions.

Greg Snow
  • 48,497
  • 6
  • 83
  • 110
  • Thanks for the reference. I will look into it but from the post it seems that this is the case: my covariate suggests a probability of 1 of the dependent – user2520918 Aug 20 '13 at 18:09
  • I don't have access to the full paper. Do you have a suggestion (personal or statistical ) on what to do with such a covariate? Can you say it's a good predictor, or it totally ruins the model? – user2520918 Aug 20 '13 at 18:18
  • 1
    @user2520918, test that term using the log-likelihood test on the full and reduced model and ignore the Wald standard error. Use profiling of the likelihood if you need a confidence interval. – Greg Snow Aug 20 '13 at 18:24