Keith Reynolds asked that I post his reponse to SOCNET (he couldn't from his
Paul, Tom and others,
Representing the correlation structure between dyads in a parameterized model
be quite difficult, though I have not looked at the problem. If you could do it
then you might simultaneously estimate the parameters of your ordinal regression
model and your correlational structure, (they might overlap in a mixed model
formulation), using maximum likelihood. You might have to do this by writing
own code, and using for example Numerical Recipies or NAG, for the optimization,
since, rather oddly, statistical packages don't seem well developed for this
though I have not systemaatically reviewed the facilities recently. If you try
this, your problem will be in obtaining the (asymtopic) covariance matrix (and
the standard errors) for the parameter estimates since the Hessian matrix (the
matrix of second order derivatives of the log-likelihood) at the fitted
values (i.e. the observed information matrix) does not seem easily available,
Numerical Recipies, or NAG. In various studies (but not social nets) I have used
observed information matrix as an approximation to the expected information
and usually found it adequate.
The above process is not particularly easy to do. However, a student (Wang) and
myself have recently gone through it for a simple Poisson Regression example
the Solver facility in Excel combined with some VB macro programming, as a basis
future modelling efforts. It should be easily adaptable to the ML estimation of
model, though I do not know how it will perform if the number of parameters gets
large. I can send a copy of the Excel demo to those who might be interested in
adapting it for use in their own statistical mdodelling work.
"Thomas W. Valente" wrote:
> Paul and others,
> The advice I've been given by many statisticians is to use GEE (General
> Estimating Equations) with no constraints on the correlation matrix
> (expectations of the degree of correlation within dyads). The other strategy
> has been to use the Sandwhich Estimator (Huber-White), both return similar
> results. In our paper on syringe sharing among needle exchange participants
> had a cohort sample with egocentric network data. The cohort was uneven in
> respondents varied in the number of followup interviews they had completed. We
> reshaped the data to be dyadic giving non-independence at 2 levels, the number
> of interviews and network. (Theoretically we might have been able to specify
> more covariation within respondents compared to within survey times, but
> mathematically this has not yet been implemented in any statistical package
> I know of.) You can also use a general Multi- level model framework (also
> as random effects model) specificing co-variation within respondents.
> on the statistical package, someone can provide model examples (I use STATA
> mostly now).
> Statisticians may provide better and more complete answers. - Tom V.
> Paul Chung wrote:
> > Hi! As a network novice, I've run into a problem that I imagine most
> > socnetters have already successfully handled.
> > I performed an analysis of the survey responses of 52 subjects, whom I
> > assorted into N*(N-1)/2 = 1326 unique dyads. My response variable was the
> > dyadic agreement in survey answers (measured on a scale), which I put into
> > an ordered logit regression.
> > The problem, of course, is that these dyads, while unique, are not
> > independent, so my standard errors are wrong. Does anyone have an easy
> > solution to this problem?
> > My e-mail address is below. If you feel that the question is of general
> > interest, please feel free to post your response.
> > Thanks! Any help at all would be appreciated. I look forward to hearing from
> > you.
> > Sincerely,
> > Paul Chung
> > Email: [log in to unmask]
> To learn more about my evaluation book go to:
> Thomas W. Valente, PhD
> Director, Master of Public Health Program
> Department of Preventive Medicine
> School of Medicine
> University of Southern California
> 1000 Fremont Ave.
> Building A Room 5133
> Alhambra CA 91803
> phone: (626) 457-6678
> fax: (626) 457-6699
> email: [log in to unmask]
To learn more about my evaluation book go to:
Thomas W. Valente, PhD
Director, Master of Public Health Program
Department of Preventive Medicine
School of Medicine
University of Southern California
1000 Fremont Ave.
Building A Room 5133
Alhambra CA 91803
phone: (626) 457-6678
fax: (626) 457-6699
email: [log in to unmask]