--- Gao Liu wrote
> I am examining the impacts of some state-imposed policies, measured as a set
> of dummy variables, on school districts. Some states adopt some of these
> policies, others not. But if a state adopts one policy, it would apply to
> all school districts in the state. In other words, a policy dummy variable
> has the same value for all school districts within the same state.
>
> The dataset is an unbalanced panel, containing a sample of school districts
> from different states in different periods. School district that appear in
> this period generally would not be in the dataset next period, although some
> exceptions exist. Policies were adopted before the start period I am
> looking at. And during my examining period, no states change their policies.
>
> The problem is: since the dummy variables of interest are linear combination
> of states, we can't include state dummy variables. Thus, the results of the
> impact of these policies on school districts may capture some unobserved
> state-wide effects. It is not so convincing to simply interpret the
> coefficients of those dummy variables as the policy impacts. Is there any
> way to solve this problem?
Your schools are nested in states, and you want to add a variable that
remains constant at the state level. In that case you can't use a
fixed effects model (add state dummies), but you can use a random
effects model. Downside is that now you have to make assumptions about
the unobserved state-wide effects (well, you could not expect it to be
otherwise: the information about unobserved effects has to come from
somewhere, if it is not from your observations (which it is not,
otherwise it wouldn't be unobserved) and it is not from your design
(as would be the case in a fixed effect model) than it has to come from
your assumptions). In this case the assumption that troubles the most
people is that the unobserved state wide effects are uncorrelated with
the observed variables. But given your design, you don't have much
choice. The Stata commands implementing these models are part of the
-xt- family of commands, which one you choose depends on your dependent
variable. See: -help xt- for a list of commands.
Hope this helps,
Maarten
-----------------------------------------
Maarten L. Buis
Department of Social Research Methodology
Vrije Universiteit Amsterdam
Boelelaan 1081
1081 HV Amsterdam
The Netherlands
visiting address:
Buitenveldertselaan 3 (Metropolitan), room Z434
+31 20 5986715
http://home.fsw.vu.nl/m.buis/
-----------------------------------------
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
