st: size of cluster robust SEs relative to regular SEs

[David Quinn <dxquinnx@gmail.com>]

Hello,

I am running a series of logit and multinomial logit models using
STATA and have very strong reason to suspect some clustering on a
particular indicator.  Hence, I ran the models with regular standard
errors and with cluster robust standard errors and compared the two.
As I understand the theory, the cluster robust standard errors should
be larger to account for serial correlation and heteroskedasticity,
thereby resulting in a more strict (since the SEs will be higher) and
less biased estimation of variation and test of statistical
significance.  But my results consistently show that the cluster
robust standard errors for half of the variables in my equation are
smaller than the regular standard errors, while for the other half of
the variables the cluster robust standard errors are larger as
expected.  Because of this, some variables that are only approaching
statistical significance in the regular models become significant (at
p<.05) in the cluster robust models, while others lose their
significance in the cluster robust models as one would expect.

I've tried to come up with an answer to what is going on, but I have
been unsuccessful.  Is this normal?  And if not, does anyone know what
could be driving these results?

I do have a small sample problem, but if that was the main issue,
wouldn't the size of ALL of the robust standard errors display the
same biased directionality relative to the size of the regular
standard errors?

My syntax is of the following ilk:

logit (or mlogit) Y X1 X2 etc. [regular models]
logit (or mlogit) Y X1 X2 etc., cluster(clustervar) [cluster robust models]

Thanks for any insight that you can provide.

--David
*
*   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/