# RE: FW: st: mfx after xtnbreg and how to compute predicted Y

 From Maarten buis To statalist@hsphsun2.harvard.edu Subject RE: FW: st: mfx after xtnbreg and how to compute predicted Y Date Tue, 21 Oct 2008 10:00:20 +0100 (BST)

```--- Pek-Hooi Soh <phsoh22@gmail.com> wrote:
> Hi Maarten, Thanks so much for your detailed explanations, which help
> me understand stata and the technique of interpreting coefficients in
> nonlinear models with interaction terms. To my limited knowledge, few
> published papers discuss such technique in details, so I really

There are actually a lot of papers on that subject, there just isn't
(yet) a concensus on how best to handle interaction terms in a model
that is non-linear in the parameters. There are at least two issues,
both of which I ignored till now: the first is how best to represent
the interaction term. A good starting point there is (Norton, Wang and
Ai 2004) and the references there in. the second issue is that these
effects are likely to be biased. For this issue I recommend the various
working papers, handouts and references in
http://www.nd.edu/~rwilliam/oglm/index.html . However, I would
recommend you will continue to ignore these issues untill you fully
understand how to interpret interaction terms in a linear model. A book
I can absolutely recommend concerning the interpretation of these
models in general is (Long and Freese 2006). It doesn't do to much

> But according to your model x2prime x2primeXx3, you stated a3 as the
> effect of x3, do you mean a4 according to this exp(a1 + a2 x2 + a3 x3
> + a4 x2Xx3 )?

No, the effect of x3 on the linear predictor (ln[count]) is (a3 + a4
x2prime). If x2 has it's average value, then x2prime equals 0. So, the
effect of x3 when x2 has it's mean value is (a3 + a4 0) = a3.

> With regard to your suggestion about finding the change in predicted
> y for a standard deviation change in x3 , I want to try diving x3 by
> its standard deviation, do I do it with x3prime/(s.d. of x3prime) in
> both single and interaction term?

What I would do is the following:

// create a variable touse that is one when all variables are observed:
gen byte touse = !missing(y, x2, x3)

// standardize x2 and x3:
sum x2 if touse == 1
gen zx2 = (x2 - r(mean))/r(sd)
sum x3 if touse == 1
gen zx3 = (x3 - r(mean))/r(sd)

// create the interction term
gen zx2_X_zx3 = zx2 * zx3

Hope this helps,
Maarten

References:
Long, J. Scott and Jeremy Freese (2006) Regression Models for
Categorical Dependent Variables Using Stata, 2nd Edition, College
Station: Stata Press.
http://www.stata-press.com/books/regmodcdvs.html

Norton, Edward C., Hua Wang, Chunrong Ai (2004) Computing interaction
effects and standard errors in logit and probit models, The Stata
Journal, 4(2): 154--167.
http://www.stata-journal.com/article.html?article=st0063

-----------------------------------------
Maarten L. Buis
Department of Social Research Methodology
Vrije Universiteit Amsterdam
Boelelaan 1081
1081 HV Amsterdam
The Netherlands

Buitenveldertselaan 3 (Metropolitan), room N515

+31 20 5986715

http://home.fsw.vu.nl/m.buis/
-----------------------------------------

*
*   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/
```