re: st: does -margins- work after -menbreg- or -mepoisson- in Stata 13? ((out of office))

["Emma Slaymaker" <Emma.Slaymaker@lshtm.ac.uk>]

I'm on leave until the 9th September, looking at my e-mail but maybe not responding very fast.

>>> "Ariel Linden, DrPH" <ariel.linden@gmail.com> 08/13/13 00:36 >>>

Thank you, Rafal. This is extremely helpful. I intuited that the error was
related to differentiating between the fixed and random effects.

It would be helpful if the relevant help files and manual entries, were to
state what you wrote below. Even more specifically, I can't seem to find the
post-estimation -margins- for these commands. I am not sure if they are
there and I am not finding them, or they are just not available? In any
case, the issue that you raise (and the solution) should be made easy to
find...

Thanks again!

Ariel  



________________________________________
From
  rraciborski@stata.com (Rafal Raciborski, StataCorp)
To
  statalist@hsphsun2.harvard.edu
Subject
  st: does -margins- work after -menbreg- or -mepoisson- in Stata 13?
Date
  Mon, 12 Aug 2013 15:53:25 -0500
________________________________________
Steve Rothenberg <drlead@prodigy.net.mx> is trying to use -margins- after
-menbreg- but -margins- complains:

> I'm trying to use -margins- after -menbreg-.  No matter how I set up
> the -margins- command, I receive the following error message:
> 
> default prediction is a function of possibly stochastic quantities other
> than e(b)
> r(498);

Ariel Linden <ariel.linden@gmail.com> notes that -margins- works after
-mixed- but not after a linear mixed model estimated by -meglm-.

> ... I receive the same error when using -meglm- ... Substituting -mixed-
for
> -meglm- works fine.

-margins- requires that all quantities needed to produce the requested
margins
are available for the prediction calculation.  Namely, it requires that the
prediction is only a function of -e(b)- and the independent variables in the
model.  The default prediction for the new -me- commands is the mean
response
-mu- which is calculated as

        g^{-1}(X*b + Z*u)

where X in X*b is the list of fixed predictors (independent variables),
Z in Z*u is the list of random predictors (also independent variables),
and g^{-1} is the inverse link function that maps the linear prediction to
the
mean response.

The only unknown thing here is 'u'.  The random effects 'u' are not
variables
in the dataset; however, they can be substituted with empirical Bayes'
estimates.  The empirical Bayes' estimates are computed using the X, b, Z,
and
the dependent variable from the original estimation sample.

-margins- will exit with an error if any prediction it is working on depends
on anything other than X, b, or Z.

-margins- can be used to calculate margins of predictions that are functions
of only the fixed effects.

  . webuse melanoma, clear
  . menbreg deaths c.uv##c.uv, exposure(expected) || nation: || region:
  . margins, predict(mu fixedonly)

The reason -margins- works after -mixed- without complaining is that the
default prediction after -mixed- is -xb- which is a function of only the
fixed
effects.

-- Rafal
rraciborski@stata.com

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/



*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/


*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/