st: margeff/margins discrepancy

["Joanne W. Hsu" <jwhsu@umich.edu>]

Hi Statalist-ers--

I've recently started using margins to compute probit average marginal 
effects in Stata 11 where I had previously used margeff.  I have found 
that although the AME coefficients are very close or identical, the 
standard errors reported by margins are vastly larger than those 
originally reported by margeff (z score of -2.21 by margins, versus 
-6.92).

I haven't been able to replicate discrepancies with such large 
magnitudes using the Stata example datasets, but from my own data 
(approx 230 obs), I've found that running the command

keep if e(sample)==1

prior to running margeff brings the standard error estimates in line 
with those reported by margins.  (Using

margeff if e(sample)==1

has not yielded anything different than just running an unconditional 
margeff).

Does anyone have any insight on what's driving the discrepancy? My 
understanding is that margeff numerically computes the AME while margins 
does it analytically, but this shouldn't yield zscores that are 3 times 
as large as one another.

Thanks,

Joanne
*
*   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/