Notice: On March 31, it was announced that Statalist is moving from an email list to a forum. The old list will shut down on April 23, and its replacement, statalist.org is already up and running.
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: st: Fixed effects logit model
Maarten buis <email@example.com>
Re: st: Fixed effects logit model
Mon, 19 Jul 2010 13:50:11 +0000 (GMT)
--- On Mon, 19/7/10, Marc Michelsen wrote:
> I am estimating a logit model for a panel style data set.
> In order to guarantee unbiased estimation, I have used company,
> industry and/or offer year clusters (per Petersen, 2009). For
> my linear regressions I have made positive experience with
> fixed-effects models. Their application for binary outcome
> models is not as straightforward because the models rely solely
> on within-variance.
> more than 50% of my observations get lost in the regression
> because of zero within variance. Is it consistent to show also
> a fixed effects logit model beside standard logit models
> clustered by the above mentioned characteristics.
I would not do that, these two estimators just measure different
things, the fixed effects estimator controls for every
characteristic that remains constant, while your model with
clustered standard errors does not. I don't see how you can
compare the results of these two models. The point of presenting
two models side by side is that (it implies that) you can
compare models. If you can't compare those models, than
presenting the models side by side will just result in confusion.
The problem with a large proportion of dropped observations is
that you may need to think again about to what population you
are trying to generalize. For that reason I would look at
wether those that drop out of your analysis analysis are in
some sense different from those that are in the analysis in
terms of your observed variables. If you are lucky there isn't
much difference, and you can, with some arm waving, argue that
it doesn't matter. If there are considerable differences, than
I would just mention that, and at the very end of your paper
discuss some hypotheses of how this may influence your estimates.
Remember that you are trying to do something that is by
definition impossible: get an empricial estimate of an effect
while controlling for stuff that you haven't seen. So do not
expect to get the right answer. What you should aim at is to
look at your data as containing some information on the effect
that you are interested in; it is not enough, but it is not
zero either. There are now a variety of strategies you can
follow to extract that information. Pick one, and do that
one right. There are two reasons for that. First, using these
strategies right is hard (not surprising as they try to solve
an unsolvable problem...), so it really pays to focuss on one
strategy. Second, it is much easier this way to write your
paper in a way that it helps the reader to follow what data you
have used and what information it contains that help you get
an idea of what the effect of interest is (and what "information"
comes from the (untestable) assumptions underlying your strategy).
Others (or you in a different paper) can later use other
strategies. After a sufficient body of literature has been
assembled on this question, someone can try to summarize the
Hope this helps,
Maarten L. Buis
Institut fuer Soziologie
* For searches and help try: