Re: st: Scoring a Fixed Effects Model on another dataset

[Jeph Herrin <stata@spandrel.net>]

In this case, I think you can do it, but you want to have your two samples in the same dataset, with a flag to indicate 
which is which. Then once you get the fixed effects estimates for the test sample, carry them to the validation sample:

 u TEST_SAMPLE
 gen byte test=1
 append using VALIDATION_SAMPLE
 replace test=test==1

 xtset PROVIDER
 xtreg {independent variables} if test, fe
 predict score , xb
 predict feffects, u
 bys PROVIDER (feffects) : replace feffects=feffects[1]
 replace score=score+feffects

hth,
Jeph



On 6/12/2013 4:09 PM, William Sankey wrote:
> Right, I would only add that the "people" (here, Providers) in one
> sample are the same people in the second sample (aka, the same
> Providers). So hopefully this should be possible...I would agree that
> it probably would not be the case if the sample had different people.
> That would probably be more similar to a propensity score analysis?
>
> Thanks,
> Will
>
> On Wed, Jun 12, 2013 at 3:53 PM, Lucas <lucaselastic@gmail.com> wrote:
> > This is an interesting question.  I always thought that fixed effects
> > models give you very good causal effect estimates, but technically the
> > estimates do not apply beyond the sample.
> >
> > Think about it like this: the fixed, subject-specific terms (call them
> > b_0i) capture the effects of all unmeasured aspects of the person.  If
> > so, if you use the coefficients b_1 to b_k estimated using Sample 1 on
> > non-estimation Sample 2, you'll obtain predicted values based on those
> > coefficients, but your predicted values will not include the
> > subject-specific effects (i.e., no b_0i estimates) because those were
> > obtained only for each unique person in the estimation sample.
> >
> > Perhaps there is some way to do what you desire of which I am unaware
> > and, if so, I'd love to learn it.  But, my current knowledge indicates
> > what you want to do is not really appropriate for the fixed effects
> > model.
> >
> > Sam
> >
> > On Wed, Jun 12, 2013 at 12:31 PM, William Sankey <wsankey@gmail.com> wrote:
> > > Hi Statalist,
> > >
> > > I have two samples of the same data, my TEST_SAMPLE and
> > > VALIDATION_SAMPLE. I want to run a fixed effects model on my
> > > TEST_SAMPLE and collect the parameter estimates and apply them to the
> > > VALIDATION_SAMPLE. I believe I want to use the predict score command.
> > > After applying these coefficients to the VALIDATION_SAMPLE we want to
> > > compute the predicted value of my dependent variable as to compute the
> > > error.
> > >
> > > So far I believe I am on the right track with the following code:
> > >
> > > use TEST_SAMPLE
> > > xtset PROVIDER
> > > xtreg MARGIN {independent variables}, fe
> > >
> > > use VALIDATION_SAMPLE
> > > predict score
> > >
> > > At this point I'm not sure how to retrieve predicted values of the
> > > dependent variable to compute error. Also, if others have experience
> > > with this kind of procedure and used a different method they would
> > > recommend, I would appreciate it.
> > >
> > > Any insight would be greatly appreciated.
> > >
> > > Best,
> > > Will
> > > --
> > > William J. Sankey
> > > Johns Hopkins University
> > > MA Public Policy '12
> > > *
> > > *   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/