Re: st: Is rho the right indicator in a multilevel analysis?

[Austin Nichols <austinnichols@gmail.com>]

KERSTEN Sarah <sarah.kersten@unifr.ch>:
No, rho tells you little. Regress dummies for any paid (unpaid) work
on welfare policy variables, using -logit- or -probit- or -glm- and
cluster-robust standard errors, then -test- whether those policy
variables are jointly significant. Then regress hours of paid (unpaid)
work on welfare policy variables, using -glm- with a log link and
cluster-robust standard errors, then -test- whether those policy
variables are jointly significant.  These pairs of models are similar
to a two-part model and do not require the distributional assumptions
of -xttobit-.  Are you using -xt- because you have repeated
observations on people, or because you want to include effects for 26
political entities, and if you included fixed effects, policy
variables would drop out?  That is not a justification for -xttobit-
or the like; you must believe the distributional assumptions.

On Mon, Nov 19, 2012 at 6:47 AM, KERSTEN Sarah <sarah.kersten@unifr.ch> wrote:
> Dear Statalist,
> I am currently doing research about gender differences regarding the time use of paid and unpaid work, the dependent variable is therefore hours worked per week. This makes two analyses for paid and unpaid work, plus two for men and women. Important to say also that it is a multilevel analysis, as I have 26 unities and for each different variables coming from policies, economies etc. The dependent variable (lets just look at employed work) is censored, as I integrate also unemployed people, so there are a lot of zero's. This is why I use xttobit, I read quite a lot studies using this with no panel data. Now my problem is, that I am mostly interested in the level 2 variables, because my hypotheses is that the individual time use is influenced by the different welfare regimes etc. The only indicator the stata output gives me about the relevance of the second level is rho, in my understanding. It is so small (depending on the model from 0.02% to 0.005%) that I think the secon!
 d !
>  level is not relevant, contradicting a little bit actual research.
> My questions now are: is rho the only indicator, since I cannot calculate a null model (why?) with xttobit and calculate R square, about the influence of the second level variables? Is this method the right one?
>
> Thanks a lot,
> Sarah

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/