[R-sig-ME] mean and variance of random effects in glmer

[ONKELINX, Thierry]

Dear Ken,

Very large variance for the random effect in a binomial glmer is an indication for (quasi-)complete separation. Here is some info on that issue: http://www.ats.ucla.edu/stat/mult_pkg/faq/general/complete_separation_logit_models.htm

If the values of Problem and Across are constant within each level of PID, I would aggregate the data (sum per PID) and then use a simple glm()

Best regards,

Thierry

ir. Thierry Onkelinx
Instituut voor natuur- en bosonderzoek / Research Institute for Nature and Forest
team Biometrie & Kwaliteitszorg / team Biometrics & Quality Assurance
Kliniekstraat 25
1070 Anderlecht
Belgium
+ 32 2 525 02 51
+ 32 54 43 61 85
Thierry.Onkelinx op inbo.be
www.inbo.be

To call in the statistician after the experiment is done may be no more than asking him to perform a post-mortem examination: he may be able to say what the experiment died of.
~ Sir Ronald Aylmer Fisher

The plural of anecdote is not data.
~ Roger Brinner

The combination of some data and an aching desire for an answer does not ensure that a reasonable answer can be extracted from a given body of data.
~ John Tukey

-----Oorspronkelijk bericht-----
Van: r-sig-mixed-models-bounces op r-project.org [mailto:r-sig-mixed-models-bounces op r-project.org] Namens Ken Kelley
Verzonden: dinsdag 24 juli 2012 8:29
Aan: r-sig-mixed-models op r-project.org
Onderwerp: [R-sig-ME] mean and variance of random effects in glmer

Hi everyone,

I'm fitting a straightforward glmer model with the family=binomial. I expected the mean of the random effect for the intercept to be near zero, but that isn't the case, as the mean is .91:

> (model.3 <- glmer(TA ~ 1 + Problem + Across + (1|PID), data=Data.Timed, family = binomial, nAGQ=100))
Generalized linear mixed model fit by the adaptive Gaussian Hermite approximation
Formula: TA ~ 1 + Problem + Across + (1 | PID)
   Data: Data.Timed
   AIC   BIC logLik deviance
 158.8 172.9 -75.38    150.8
Random effects:
 Groups Name        Variance Std.Dev.
 PID    (Intercept) 18.869   4.3439
Number of obs: 256, groups: PID, 64

Fixed effects:
            Estimate Std. Error z value Pr(>|z|)
(Intercept)  -1.1328     0.7304  -1.551   0.1209
Problem      -0.5864     0.2449  -2.394   0.0167 *
Across       -1.3768     0.4280  -3.217   0.0013 **
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Correlation of Fixed Effects:
        (Intr) Problm
Problem -0.390
Across  -0.208 -0.150

 That is, I'm doing a mixed effects logistic regression. The PID is the participant ID; there are 4 Problems (essentially timepoints: 0, 1, 2, 3) and Across is a time-varying covariate (0, 1, 2, or 3).

The mean of the random effects is:
> colMeans(ranef(model.3)$PID[])
(Intercept)
  0.9137307

Additionally, the variance of the random effect is in the model output as 18.869, yet when I calculate the variance of the random effects directly, I get a much smaller value:
> var(ranef(model.3)$PID[])
            (Intercept)
(Intercept)    7.806402

Should I be surprised by either of the issues I note above? My concern is that I was planning on plotting the model implied curves using the fixed effects (so that the curves would represent an individual specific trajectory for a participant with a random effect of 0). Yet, there are no individuals with a random effect of zero and the mean is not zero. Thus, such a plot doesn't seem as useful as I initially thought it would.

Thanks for any thoughts on this,
Ken



        [[alternative HTML version deleted]]

* * * * * * * * * * * * * D I S C L A I M E R * * * * * * * * * * * * *
Dit bericht en eventuele bijlagen geven enkel de visie van de schrijver weer en binden het INBO onder geen enkel beding, zolang dit bericht niet bevestigd is door een geldig ondertekend document.
The views expressed in this message and any annex are purely those of the writer and may not be regarded as stating an official position of INBO, as long as the message is not confirmed by a duly signed document.