In general, one uses "clustered standard errors" if you think that there are
observations that don't fully contribute an observation's worth of
information. For example, in everything that we do, we assume that all our
observations are iid. However, it might be the case that some observations
are correlated with each other, so two correlated observations don't
contribute as much new information to the analysis as two fully independent
observations. Clustering on a variable means that that variable defines how
the observations are correlated. In general, doing the clustered standard
errors should increase your standard errors (you might think of it as
decreasing the number of independent n observations). However, in some
cases, the opposite might occur, which might be the case here. Maybe this
will help a little bit, although it is for the OLS case:
http://www.stata.com/support/faqs/stat/cluster.html
In terms of checking for overdispersion, remember that the poisson is nested
within the negative binomial. That is, a negative binomial with a
dispersion parameter of 1 is equal to a poisson. So you might want to
figure out what the dispersion parameter is on the negative binomial. Also,
you might think about making plots similar to the ones Gary has when
introducing the models. Also, remember that the dependence is not over the
observations, but rather over the event generating process for the negative
binomial.
2009/3/23 sparsha saha <sparshahoneysaha at gmail.com>
Hi everyone!
JeeHye and I are working on a paper that uses a negative binomial
model. The dependent variable is the number of uses force in a given period
of time (it might have been every three months or something like that...).
So, this seems like it might be pretty NOT iid. So that is why they use the
neg bin. So we ran in this R using zelig and did robust sandwich clustered
errors and pretty much did not have their key explanatory variable come out
significant (percent in congress from same political party as president
during that period of interest). So we thought okay let's try stata. So we
did and there was an option to cluster on a particular variable...they
clustered on "president." Got their results, their exp var of interest came
out significant. So, why was this? Why would you cluster your standard
errors on something? When is it okay to do this? When is it not? To be
honest, what does clustering on something for your standard errors even
mean?
Also, how would we go about checking for overdispersion...we want to try
the poisson model and then compare it to the neg bin...see if there is even
overdispersion (which there must be). But how do you check for it? And are
there other ways of controling for dependence between observations in event
count models? I mean with other regression types there are splices and time
dummy variables, could you do something similar with an event count model?
Sorry about all the questions...I guess it would be just be helpful to
hear back from some of you regarding whatever it is that you know about in
regards to a subset of them, all of them, none of them...whatever comments
you have.
Hope everyone is having a great break!
sparsha
_______________________________________________
gov2001-l mailing list
gov2001-l at
lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l
--
Patrick Lam
Department of Government and Institute for Quantitative Social Science,
Harvard University
http://www.people.fas.harvard.edu/~plam