It seems that unless you have a very large data set you will have
degraded your sample size to point that it is useless by relying
solely on within subject change for predictors like marital status.
Marital status likely does not change much in 5 years.
More important, why not take advantage of the between subject effects?
Unless you have some theoretical or statistical reason to believe that
any between subject effects should be excluded (a classic example
would be drinking during pregnancy and birth weight, since drinking is
likely confounded with other predictors that cannot be controlled for
by including additional covariates), it is more appropriate to use a
random effects model that relies on both within and between subject
estimation since you bring to bear more data points for your estimate,
you are also afforded the opportunity to evaluate some time invariant
predictors like gender, and the interaction of time invariant with
time varying predictors.
We can test the hypothesis that between and within effects are equal
using the hausman test in Stata: 1) Estimate a random effects model
using the re option and store the results, 2) Estimate a fixed effects
model with the same covariates and store the results, and 3) Compare
the results "hausman 'model1' 'model2'." If the estimates for marriage
are statistically different you have justification to look for
additional covariates that may explain the confounding or if not
possible use a fixed effects model. The first is theoretically more
interesting.
As for interpretation of the estimates: A fixed effects model relies
solely on within subject change and can be interpreted as such.
However, a random effects using both within and between subject
effects under the assumption they are equal may be treated the same
way (a change in X results in a change in Y) in your discussion.
The coding you used appears to be fine.
You should become intimately familiar with the work of Sophia
Rabe-Hesketh and her colleagues if you continue in your efforts.
http://www.gllamm.org/sophia.html
Best of luck,
john
On Mon, Jan 19, 2009 at 2:50 AM, Anna Reimondos <[email protected]> wrote:
> Hello,
> I am attempting to use a fixed-effects model to examine attitudes in a
> longitudinal setting.
> I have a panel dataset and I would like to explain the attitudes of
> individuals (independent variable) at time t. The attitudes are
> measured on a score from 0-10 and I am using a fixed-effects model
> (using xtreg, fe and treating the attitude variable as an interval
> level variable) as I think this is most appropriate model for me to
> study within-person change over time.
>
> I have included a set of time-varying predictors such as marital
> status (coded as 0 =not married, 1 = married). I was hoping to get
> some advice on if I am coding and interpreting the effect of the
> time-varying predictors correctly.
>
> Example.
> Subject #1 is included for 5 waves of the data, and has married by
> the second time period.
>
> Id Wave Attitude Mar_status
> 1 1 6 0
> 1 2 8 1
> 1 3 9 1
> 1 4 8 1
> 1 5 9 1
>
>
> Would this be the correct way to set up the time-varying mar_status variable?
>
> If the coefficient for mar_status (marital status) in the
> fixed-effects model is positive, could this be interpreted as meaning
> that a change from not being married to getting married has a positive
> effect on the attitudes.
>
> I am very new to using longitudinal models and would appreciate
> knowing if I am going in the right direction or not.
>
> Thanks very much in advance
>
> Anna
> *
> * 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/
>
*
* 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/
