Using dummies to estimate a fixed effects model is not recomended in a
non-linear model (especially if you have few observations in each
category, and 17 does not seem large enough to me). This point is
discussed here:
http://www.stata.com/statalist/archive/2003-09/msg00103.html .
So that precludes models 2 and 3. Model 1 is not without problems
either: it assumes that the city and time specific intercepts are
uncorrelated with the observed variables, which bothers quite a few
people. If I had to choose one model I would probably go for model 1,
as being the lesser of three evils.
However, you will probably want to look at what is called a
crossed-effects model, which can be estimated with -xtmelogit-. There
is a discussion on how to do that in the XT manual on pages 239-242.
Hope this helps,
Maarten
--- Noah Friedkin <[email protected]> wrote:
> The data I am analyzing are longitudinal, consisting of individuals
> located in 4 cities across 17 time periods. I.e., each individual
> contributes 17 observations, and each observation occurs in one of
> 4X17=68 settings. I am not concerned with the effects of these
> settings, but instead with a set of individual-level predictors x1,
> x2, x3 of the binary responses y of individuals in their city-time
> contexts. There are three credible approaches to the analysis of
> these
> observations:
>
> (1) xtmelogit y x1 x2 x3 || city: || time:
> (2) logit y x1 x2 x3 c2 c3 c4 t2 t3 ? t17, where c2-c4 and t2-t17
> are
> indicator variables with city 1 and time 1 as the omitted settings
> (3) logit y x1 x2 x3 c1t2 c1t3 ? c1t17 c2t1 c2t2 ? c2t17 ? c4t1 c4t2
>
> ?c4t17 where c_t_ are indicator variables with c1t1 as the omitted
> setting
>
> Model 2 is a popular approach. Model 3 appears closer to model 1.
> However, my sense is that models 2-3 are excessively conservative and
>
> potentially misleading in allowing (what may be viewed as)
> meaningless
> ?nuisance? associations to affect the estimates for x1, x2, and x3.
> Model 1 seems preferable.
>
> Question 1. What are the pros and cons of these three models?
>
> I am assuming that model 1 is equivalent to allowing different
> intercepts for each sity-time group of observations. I'm also
> assuming
> that the estimated random effects for the 68 city-time groups are not
>
> constrained, so that (for example) they might be found to be all
> positive values that monotonically increase with time.
>
> Question2. Are both of my assumptions about model 1 correct, or is my
>
> understanding of model 1 flawed?
>
> Comments on the above two questions would be much appreciated.
>
>
>
>
>
>
>
>
>
>
> *
> * For searches and help try:
> * http://www.stata.com/help.cgi?search
> * http://www.stata.com/support/statalist/faq
> * http://www.ats.ucla.edu/stat/stata/
>
-----------------------------------------
Maarten L. Buis
Department of Social Research Methodology
Vrije Universiteit Amsterdam
Boelelaan 1081
1081 HV Amsterdam
The Netherlands
visiting address:
Buitenveldertselaan 3 (Metropolitan), room N515
+31 20 5986715
http://home.fsw.vu.nl/m.buis/
-----------------------------------------
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
