Julie,
1. You might want to check that structure is
affected by gender, experience, and admin position type by running an
autocorrelation test. This would tell you, for example, whether people tend to
interact more within gender than between, or certain admin position types talk
more to certain other types.
2. An n of 36 is not in itself a big problem for a
t-test. (Yes, you will need large effects to be significant with such small n,
but that is as it should be -- no scientific misconduct there, just
non-significance). There may be other reasons for distrusting a classical
t-test/anova however. E.g., your data were not obtained from a random sample,
distribution of pop variables may be different from what is required,
observations may not be independent, etc etc. Plus, you may not be interested in
whether the means are different in "the" population, which the classical test
answers among other things, but only what are the chances of obtaining a
difference in means as large as observed if centrality values are assigned
independently of gender. For all of these reasons you might
consider permutation/randomization analogs of t-test, anova,
etc.
steve.
----- Original Message -----
Sent: Tuesday, October 15, 2002 9:03
AM
Subject: Actor Attributes & Network
Structure
The issue of actor
attributes and network structure is the question of one of my current research
papers. I am examining the structure of a network of public school
administrators and, based on the graphical mapping, it looks clear that the
structure is affected by gender, administrator's experience and the type of
administrative position.
However, I only have n=36
administrators. At the dyadic level of the ties, I have a large enough
set of cases to import network data to SPSS and use dyadic variables in
correlations, etc. However, at the actor level, what types of tests can
I run to determine if there are group differences? I've run t-test for
gender and ANOVA for type of school (e.g. elementary, middle, high school) on
actor centrality or other positional variables as well as
characteristics of their egocentric networks. However, it occurs to me
that the small n might a problem. If so, what small group statistics
would work best for this population data to support the finding of group
differences? Are there any network specific tests (in UCINet) that can
address this question?
I'd appreciate any direction.
Julie
Julie M. Hite
Brigham Young University
Dept. of
Educational Leadership & Foundations
Provo, UT 84602
801-422-5039
[log in to unmask]