Re: st: Using the cluster command or GLS random effects?

[Buzz Burhans <wsb2@cornell.edu>]

Sandra,

The "two approaches do the same thing" in the sense that they both account 
for the lack of independence of pupils within schools, however, they 
provide different estimates for error.  Remember two issues with this type 
of data...the pupils are not independent within school, i.e. pupils form 
the same school are likely to be more similar that pupils form different 
schools for all kinds of putative reasons such as socio-economic factors or 
school educational management factors , and this lack of independence 
within school engenders the need for a cluster approach whether by xtreg or 
the cluster option.

 Secondly,  the "error" or noise associated with a fitted value for a 
pupil contains at least two components, one for the unique noise associated 
with the pupil, and another for the noise due to variance between schools. 
Regress with the (cluster) option relaxes the assumption of independence, 
and therefore, compared with regress without the cluster option, increases 
the error term to accommodate the violation of the assumption that the 
errors are independent, but leaves both the noise associated with 
differences between pupils and noise associated with differences between 
schools in the error term.  Xtreg is different. In the "random effect" 
model, xtreg fits an additional parameter, the Ui term, or random school 
term, which accounts for the differences between schools, thus the residual 
error term contains the "within school" variance between pupils, but the 
between school portion (which remained in the regress model) is now removed 
and accounted for by the weighted "between" estimator, and thus the error 
is reduced.

Buzz Burhans
wsb2@cornell.edu

At 01:16 PM 7/17/03 +0100, you wrote:

> Dear all,
>
> I am using a repeated cross-section of pupil-level data to regress exam
> attainment on various characteristics. Since pupils are clustered in 
> particular
> schools, I need to correct the standard errors for clustering at school-level.
>
> I could adopt one of the following approaches:
>
> regress Y X, cluster(school)
> xtreg Y X, re (i=school)
>
> So the first approach corrects standard errors by using the cluster command.
> The second approach uses a random effects GLS approach.
>
> I thought that the two approaches do the same thing and should give the
> same results. However, I find that the standard errors are alot smaller
> using the second approach.
>
> Does anyone know how the two approaches differ from one another?
>
> Thanks,
>
> Sandra
>
>
> *
> *   For searches and help try:
> *   http://www.stata.com/support/faqs/res/findit.html
> *   http://www.stata.com/support/statalist/faq
> *   http://www.ats.ucla.edu/stat/stata/

*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/