R help - May 2002 - Multilevel model with dichotomous dependent variable

Greetings-

I'm working with data that are multilevel in nature and have a dichotomous
outcome variable (presence or absence of an attribute). As far as I can
tell from reading archives of the R and S lists, as well as Pinheiro and
Bates and Venables and Ripley,

- nlme does not have the facility to do what amounts to a mixed-effects
logistic regression.
- The canonical alternative is GLMMgibbs, but there are concerns about
this as well.
- There has been some talk of wrappers around nlme that would add PQL (a
technique about which I know nothing) as a way of estimating such
equations.

Does this accurately summarize the state of software availability? If not,
what updates should I know about? If so, what would be the costs and
benefits of the following courses of action:

1.) Use nlme, violating the assumption of continuous outcomes, simply
assigning 0 and 1 as the outcome values. (A crude option, but one that has
the distinct advantage of increased comprehensibility to sociologists, the
research's target audience.)

2.) Use glmmGibbs, which introduces some new assumptions and requirements
of the data, but is perhaps the closest to a "correct" approach

3.) Find a PQL implementation and learn enough about the technique to use
it

Any comments or advice will be much appreciated.

----------------------------------------------------------------------
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
_._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._

Andrew Perrin <clists at perrin.socsci.unc.edu> writes:
> Greetings-
> 
> I'm working with data that are multilevel in nature and have a
dichotomous
> outcome variable (presence or absence of an attribute). As far as I can
> tell from reading archives of the R and S lists, as well as Pinheiro and
> Bates and Venables and Ripley,
> 
> - nlme does not have the facility to do what amounts to a mixed-effects
> logistic regression.
> - The canonical alternative is GLMMgibbs, but there are concerns about
> this as well.
> - There has been some talk of wrappers around nlme that would add PQL (a
> technique about which I know nothing) as a way of estimating such
> equations.
> 
> Does this accurately summarize the state of software availability? If not,
> what updates should I know about? If so, what would be the costs and
> benefits of the following courses of action:
> 
> 1.) Use nlme, violating the assumption of continuous outcomes, simply
> assigning 0 and 1 as the outcome values. (A crude option, but one that has
> the distinct advantage of increased comprehensibility to sociologists, the
> research's target audience.)
> 
> 2.) Use glmmGibbs, which introduces some new assumptions and requirements
> of the data, but is perhaps the closest to a "correct" approach
> 
> 3.) Find a PQL implementation and learn enough about the technique to use
> it
Recent versions of the MASS package in the VR bundle have a glmmPQL
function that does exactly this.  Penalized Quasi-Likelihood (PQL) is
a method of fitting these models.  The name refers to the
approximation to the likelihood that is used for fitting.

I would recommend starting with glmmPQL.

-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-
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
_._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._