st: Dummy variables in longitudinal models

[Niels Schenk <niels@oudwijk.cistron.nl>]

Dear statalist,

I am having trouble getting my head around the following issue. It's 
not directly related to Stata but I'm hoping you are willing to help me 
out. The dataset I use is a panel of respondents (see example below). 
Some of these respondents transition into a certain state during the 
period of observation, some don't. The variable 'trans' denotes if a 
respondent has transitioned at a certain time point. In my example, 
respondents 1 and 2 have, respondent 3 hasn't. I'm trying to determine 
if and how the transition affects my dependent variable. The basic 
model I'm estimating is: 'xtmixed depvar trans || time: '. I know I can 
use xtreg, but I'm using a simplified representation of the model I'm 
estimating.

My issue is the following:
I have both continuous and categorical variables that I want to use in 
my estimation. Estimating the effect of my variable 'contvar', and 
distinguishing the effect between those that have and have not 
transitioned seems straightforward:
xtmixed depvar trans contvar trans#c.contvar || time:

However, the categorical variable I would like to use is only 
available for those that have made the transition. Creating dummy 
variables out of this categorical variable yields three dummy 
variables, one is the reference category. The problem I'm having with 
this is that for those respondents who have not transitioned, all dummy 
variables are zero. For those that have transitioned, one of them is 
one. When estimating the model xtmixed depvar trans dummy1 dummy2 || 
time:. From the results I want to be able to conclude that respondents 
that have transitioned where dummy1==1, are significantly different in 
the depvar from respondents that have transitioned where dummy3==1. I'm 
thinking though that the reference category here is blurred, I can be 
either dummy3, or the case where the categorical variable is zero (i.e. 
not applicable). My conclusion would be that this approach is simply 
not valid in that the coefficients do not represent what they are 
supposed to represent, but I have seen papers that use such an 
approach.

My question is if people on this list think that this use of dummy 
variables described above is appropriate or not? Estimating a model 
only for those that have transitioned is simply not an option. If it is 
not appropriate, perhaps some of you have suggestions on how to 
estimate the model I would like to estimate.

Another totally unrelated question I have is if it is possible to 
mean-center dummy variables over time (within respondents) to separate 
the fixed and time-varying effects of dummy variables as is possible 
with continuous variables? If so, any pointers to papers that do would 
be greatly appreciated.


Thanks very much for any help,

Niels Schenk



Simplified representation of my dataset:

id	time	trans	depvar	contvar	catvar	dummy1	dummy2	dummy3
1	1	0	4	34	0	0	0	0
1	2	0	5	55	0	0	0	0
1	3	1	6	43	1	0	1	0
1	4	1	3	34	2	1	0	0
2	1	0	7	54	0	0	0	0
2	2	1	6	23	1	0	1	0
2	3	1	5	32	2	1	0	0
2	4	1	6	34	3	0	0	1
3	1	0	3	54	0	0	0	0
3	2	0	4	43	0	0	0	0
3	3	0	7	23	0	0	0	0
3	4	0	6	23	0	0	0	0

*
*   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/