Re: st: test or ftest command

[Joerg Luedicke <joerg.luedicke@gmail.com>]

On Wed, Jun 8, 2011 at 1:28 PM, Lisa Marie Yarnell
<lisayarnell@yahoo.com> wrote:
> Hello Stata group!  I am trying to do an ftest (or other test possibly using the "test" command) comparing two models: one regression model in which clustering of persons within schools is modeled, and one regression with clustering of persons within schools *not* modeled.
>
> I am concerned that these are not really nested models, however.  Can I still use the test or ftest command?  Is there another command that I should be using?  Our basic question is: "Is it important to model the clustering of persons wihtin schools, based on comparisons of model fit?"  Here are the models:
>
> (1) regress w3bmi_ms bio_sex w3_yrs w1faminc black hispanic mothed selfed sup_dens ffchdens nat_dens con_dens veg_dens act_dens med_hval, cluster (scid)
>
> (2) regress w3bmi_ms bio_sex w3_yrs w1faminc black hispanic mothed selfed sup_dens ffchdens nat_dens con_dens veg_dens act_dens med_hval
>
> Can I use the test command for this?  When I tried, I got the error message below; Stata treated "cluster(scid)" as a variable in the model:
>
> regress w3bmi_ms bio_sex w3_yrs w1faminc black hispanic mothed selfed sup_dens ffchdens nat_dens con_dens veg_dens act_dens med_hval, cluster (scid)
>
> test cluster(scid)
> cluster not found
>

The above two models only differ in the way the standard errors are
calculated. The models itself are the same and, thus, there is no
difference in model fit. If I understand your problem correctly ("Is
it important to model the clustering of persons within schools"), this
boils down to the question to what extent you observe any intra-class
correlation. With students nested in schools this is usually the case
(I guess) and the standard way to handle these type of data in
educational research is a multilevel model. However, to start, I would
suggest looking at possible intra-school correlation. For example you
could try:

 xtreg w3bmi_ms bio_sex w3_yrs w1faminc black hispanic mothed selfed
sup_dens ffchdens nat_dens con_dens veg_dens act_dens med_hval,
i(scid)

"rho" can then be interpreted as a measure of intra-class correlation.

You could also use a maximum-likelihood estimator in which case Stata
returns a likelihood ratio test where the model with the random effect
is compared to the model without random effect:

xtreg w3bmi_ms bio_sex w3_yrs w1faminc black hispanic mothed selfed
sup_dens ffchdens nat_dens con_dens veg_dens act_dens med_hval,
i(scid) mle

This is the same as a mixed model using xtmixed with varying
intercepts across schools and maximum-likelihood estimation:

xtmixed w3bmi_ms bio_sex w3_yrs w1faminc black hispanic mothed selfed
sup_dens ffchdens nat_dens con_dens veg_dens act_dens med_hval ||
scid:, mle


HTH,

J.

*
*   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/