Alicia Dowd <[email protected]> asks:
> I wonder if someone can help me understand a warning message I am receiving
> after _mfx compute_.
> The warning is
> . mfx compute
> Warning: predict() expression unsuitable for standard error calculation
> option nose imposed
> I am running two logistic regression models, which are the same except with
> different dependent variables.
> After the first model, which has a predicted probability of the positive
> outcome of .81, the standard errors are issued. After the second model,
> which has a predicted probability of .55, the warning and the marginal
> effects are issued, and no standard errors are provided.
> When I simplify the model to just one explanatory variable to explore the
> problem, I get the same warning on the second model.
> I suspect this is a statistical issue having to do with the nearly equal
> probability of the positive and negative outcomes in the second model, but I
> haven't encountered this before. A brief explanation and a reference to a
> resource that would explain the issue would be much appreciated.
> I am running Stata 7 in Windows. Reference to the Users Manuals and online
> help didn't turn up an explanation of this warning message.
The standard error calculation performed by -mfx- will produce valid standard
errors when the only stochastic component of the prediction is the estimated
coefficient vector, e(b). If there are other stochastic components to the
prediction (such as some other stored e() result), -mfx- will refuse to
perform the standard error calculation. -mfx- for Stata 7 was updated on
08may2002. With this update, -mfx- now does a much better job of detecting
whether this assumption is true, but there do exist rare cases where this
check will produce a false positive. That is, -mfx- thinks there are random
components to the calculation other than e(b), but in fact there are not. As
a result, rather than sometimes calculate invalid standard errors, -mfx- will
sometimes refuse to calculate standard errors when they would have been valid
had they been computed. Better to err on the conserative side.
That been said, I am still puzzled that this check of assumption was tripped
by Alicia's model. If she could privately send me the data and a do-file
reproducing this result, I'll be glad to take a closer look. Once resolved,
I will report back to the list.
--Bobby
[email protected]
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
