[R-sig-ME] Underdispersion in multilevel logistic regression

[Douglas Bates]

You have a variance component for ClusterID that is being estimated as
"effectively zero".  You should refit the model with that random
effects term removed and check the estimated scale for the modified
model.

I don't have good guidelines for under- or over-dispersion.  When the
estimated scale is < 0.75 I get concerned and when it is < 0.5 I get
very concerned but I don't know if those are reasonable comparison
values or not.

On 9/9/07, Steven J. Pierce <pierces1 at msu.edu> wrote:
> Hi folks,
>
> I'm doing some multilevel logistic models with lmer() and I noticed that the
> estimated scale in my model (see code & results below) suggests the presence
> of under-dispersion. Are there any guidelines on when the scale is
> sufficiently far from 1 that one should conclude that underdispersion (or
> overdispersion) is serious enough to warrant switching from family =
> binomial(logit) to family = quasibinomial(logit)?
>
> > model.17f <- lmer(formula = EventTV ~ Period + OR1pc + OR1flyer + OR2pctv
> +
> +                             OR2flyertv + Spanish_Version + MUDwell + 0 +
> +                             (1 | ClusterID) + (1 | SurveyID),
> +                 data=RS05.Round1A, family = binomial(logit),
> method="Laplace")
> > summary(model.17f)
> Generalized linear mixed model fit using Laplace
> Formula: EventTV ~ Period + OR1pc + OR1flyer + OR2pctv + OR2flyertv +
> Spanish_Version + MUDwell + 0 + (1 | ClusterID) + (1 | SurveyID)
>    Data: RS05.Round1A
>  Family: binomial(logit link)
>   AIC  BIC logLik deviance
>  5324 5578  -2631     5262
> Random effects:
>  Groups    Name Variance   Std.Dev.
>  SurveyID       1.4455e+00 1.2023e+00
>  ClusterID      5.0000e-10 2.2361e-05
> number of obs: 27460, groups: SurveyID, 1787; ClusterID, 52
>
> Estimated scale (compare to  1 )  0.6785013
>
>
>
> Steven J. Pierce, M.S.
> Doctoral Student in Ecological/Community Psychology
> Department of Psychology
> Michigan State University
> 240B Psychology Building
> East Lansing, MI 48824-1116
>
> E-mail: pierces1 at msu.edu
> Web: http://www.psychology.msu.edu/eco/
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>