[R] Interpreting output from glmmPQL

Andrew Perrin clists at perrin.socsci.unc.edu
Fri Jun 21 18:04:44 CEST 2002


I'm running some models under R using glmmPQL from MASS. These are
three-level models (two grouped levels and the individual level) with
dichotomous outcomes.  There are several statistics of interest; for the
moment, I have two specific questions:

1.) This question refers to the following model (I present first 
the call, then the output of summary():

morality.restr.1.pql<-glmmPQL(random = ~ 1 | groupid/participantid,
                              fixed  = r.logic.morality ~ 1,
                              data = fgdata.df[coded.logic,],

Linear mixed-effects model fit by maximum likelihood
 Data: fgdata.df[coded.logic, ] 
       AIC      BIC    logLik
  4427.735 4447.531 -2209.868

Random effects:
 Formula: ~1 | groupid
StdDev:   0.3312237

 Formula: ~1 | participantid %in% groupid
        (Intercept)  Residual
StdDev:   0.3651775 0.9765288

Variance function:
 Structure: fixed weights
 Formula: ~invwt 
Fixed effects: r.logic.morality ~ 1 
                 Value Std.Error  DF   t-value p-value
(Intercept) -0.1699931 0.1039887 905 -1.634727  0.1025

Standardized Within-Group Residuals:
       Min         Q1        Med         Q3        Max 
-1.2951648 -0.8865510 -0.7183326  1.0428044  1.6135857 

Number of Observations: 1042
Number of Groups: 
                   groupid participantid %in% groupid 
                        20                        137 

Raudenbush & Bryk (1992; 2002) suggest that the Intraclass Correlation is
a useful statistic for a hierarchical linear model. My understanding is
that this statistic is the proportion of the model's total variance that
is "explained" by each level of the model. I have calculated this for
level 2 as 0.3312237^2 / (0.3312237^2 + 0.3651775^2 + 0.9765288^2) and for
level 3 as 0.3651775^2 / (0.3312237^2 + 0.3651775^2 +
0.9765288^2). However, Guo and Zhao imply that the total variance for a
dichotomous-outcome (logistic) model should be a constant, specifically
pi^2/3.  Clearly pi^2/3 is a very different number from (0.3312237^2 +
0.3651775^2 + 0.9765288^2). Can anyone shed light on this? Does this
calculation make sense at all?

2.) There is the possibility in these models of using some
cross-classification. The lowest unit of analysis here is the
utterance: one statement made in a group discussion. Each statement is
(currently) nested within a speaker, who is in turn nested within a
group. The complication is that each statement is *also* nested within one
of four scenarios, and the scenarios are repeated across the 20
groups. Using the scenario as a fixed covariate in the model results (or
seems to) in erronenously assuming too many degrees of freedom, since
utterances are clustered within scenarios. But cross-classifying the
scenario * group into 80 clusters seems like it will seriously impede
intepretation. Any advice? Ultimately it may not be terribly important to
include the scenario as a covariate, but I would like to be able to do so
if necessary.

Thanks for any advice.

Andrew J Perrin - http://www.unc.edu/~aperrin
Assistant Professor of Sociology, U of North Carolina, Chapel Hill
clists at perrin.socsci.unc.edu * andrew_perrin (at) unc.edu

r-help mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html
Send "info", "help", or "[un]subscribe"
(in the "body", not the subject !)  To: r-help-request at stat.math.ethz.ch

More information about the R-help mailing list