Re: st: problem using mfx, pred(pu0) after clogit

[Maarten Buis <maartenlbuis@gmail.com>]

On Wed, May 25, 2011 at 9:53 PM, Samia Costa wrote:
> I am using Stata 10.1 and I have been trying to obtain marginal effects after clogit using mfx with the predict(pu0) option. However, when I run the mfx command, I get an output as follows:
>
> . mfx, pred(pu0)
>
> warning: predict() expression pu0 unsuitable for standard-error calculation;
> option nose imposed
>
>
> Marginal effects after clogit
>      y  = Pr(response|fixed effect is 0) (predict, pu0)
>         =  2.575e-69
> -------------------------------------------------------------------------------
>                      variable |          dy/dx                 X
> ---------------------------------+---------------------------------------------
>                            x1 |               .            3.44053
>                            x2*|        1.99e-69            .108374
>                            x3 |               .            1.47502
> -------------------------------------------------------------------------------
> (*) dy/dx is for discrete change of dummy variable from 0 to 1
>
>
> I can easily obtain the marginal effects after logit or xtlogit, re. Is there any other way I could obtain the marginal effects? The equation I estimate is
>
> xi:clogit y x1 x2 x3 i.year i.dis, group(number) robust cluster(number)

I cannot reproduce that behavior using standard Stata example datasets
(I added the -cluster(pairid)- option to make the example as similar
to yours, not because I think it is a good idea):

*----------------------- begin example -----------------------------
use  http://www.stata-press.com/data/r11/lowbirth2.dta
xi: clogit low lwt smoke ptd ht ui i.race, ///
    group(pairid) robust cluster(pairid)
mfx, predict(pu0)
*------------------------ end example -------------------------------

One thing that strikes me as very suspicious is that Pr(response|fixed
effect is 0) in your model is de facto 0 (2.575e-69 =
0.0000000000000000000000000000000000000000000000000000000000000000000002575).
This suggest to me that your -clogit- did not converge to a plausible
solution. Once you have fixed your -clogit- model, you will probably
be able to get marginal effects.

It is hard for us to tell you how to fix your -clogit- model, as you
did not tell us anything about your -clogit- model, so it is hard for
us to diagnose what might be wrong there. The first thing I would do
is center all my continuous variables, i.e. subtract a constant such
that the value 0 falls within the range of the data. This constant can
be the mean, or any other meaningful value. This won't change the
model, but these kinds of models tend to converge much better when you
center your variables. When that does not work, I would look at the
sub-sample used for estimating your -clogit- model. Since -clogit- is
a fixed effects model it only uses the variation within groups and
drops groups that do not show variation. Within that restricted sample
there may be a case of perfect prediction that would not show up in
pooled or random effects models. Alternatively, I could imagine
trouble arising from variables that show really small within group
variation.

Hope this helps,
Maarten

--------------------------
Maarten L. Buis
Institut fuer Soziologie
Universitaet Tuebingen
Wilhelmstrasse 36
72074 Tuebingen
Germany


http://www.maartenbuis.nl
--------------------------

*
*   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/