RE: st: Marginal Effect

[Maarten buis <maartenbuis@yahoo.co.uk>]

 --- On Wed, 5/1/11, sazas  wrote:
> > > I estimated a random effect probit model using xtprobit in
> > > stata 11.
> >
> > > Marginal effects using margins, dydx(*) predict(pu0)
> > > assumes u_i =0 which l think implies rho =  0
> > > since rho = 0. Please is there any command that l can use
> > > to change this assumption?

--- I answered: 
> > It is not true that rho (proportion of variance due to
> > group level variance) is assumed to be 0. The way to think about
> > it is that there is an (unobserved) group level variable added
> > to your model and that you compute your marginal effects for
> > individuals that have an average value on this variable.

--- On Wed, 5/1/11, s_azagba@live.concordia.ca wrote:
> I thought if the unobserved individual effect i. e u_i = 0
> implies that rho which is the relative importance of the
> unobserved effect  is 0. I think it is same as the
> neglected heterogeneity in the estimation of partial
> effect  Wooldridge was refering to in his book on
> chapter 15. 

If you want average marginal effects you will actually get a sort
of mixture between average marginal effects over the observed 
variables and marginal effects at the average value of the 
unobserved group level variable. That is not very pretty, but it
is not strictly speaking wrong (as long as you interpret it 
correctly). This "un-pretty-ness" is rather typical for the 
friction you can expect when using marginal effects for this type
of models, as marginal effects will always give you somewhat 
"wrong" results: They are one-number-summaries of non-linear 
effects and  such a summary is always wrong and this tends to 
bite more in more complicated models. However, as long as you
interpret the marginal effects correctly, they can still be 
useful. Marginal effects are a model on top of a model, if used
correctly they are  wrong but still usefull summaries/
simplifications.

As I said before, if you really want to have a more pure 
representation of the effect then you should forget about 
marginal effects, and move to -(xt)logit- models and interpret 
odds ratios. Odds ratios follow directly from these models, and
do not need the "model on top of a model" approach of the marginal
effects. However the added precision you obtain by doing this 
only involves the description of your model results, typically 
the amount of random noise that occured during data collection, 
data preparation, modeling decisions, are many orders of 
magnitude bigger than the precision you obtain by moving from 
marginal effects to odds ratios.

Having said all that, my answer to your question is that you
can get fully correct descriptions of your model results by 
using -xtlogit- and interpret your results in terms of odds
ratios.

Hope this helps,
Maarten

-------------------------
Maarten L. Buis
Institut fuer Soziologie
Universitaet Tuebingen
Wilhelmstrasse 36
72074 Tuebingen
Germany

http://www.maartenbuis.nl
--------------------------



      

*
*   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/