Re: st: question about cluster() option in regression

[Steven Samuels <sjhsamuels@earthlink.net>]

--

Thus there is nothing "improper" about the cluster option. The  
difference in estimated coefficients is due to a difference in how  
observations are weighted. In -reg- with or without a cluster option,  
all observations get the same weight, unless you specify otherwise.  
Multi-level models, on the other hand, estimate variances of the  
random effects and use the information to give observations different  
weights.

Contrary to your impression, this phenomenon is not "rare" in  
statistical software. In SAS, for example, you will find the same  
difference between estimates from: 1) the GENMOD and SURVEYREG  
programs, which have equivalent cluster options, and 2) the multi- 
level MIXED program.

The main difference between the two approaches, cluster or  
multilevel, is that the cluster() option provides model-free standard  
errors. The multi-level programs require a correct model for the  
variance structure, for example that standard deviations are constant  
at each level. If the model for the variance structure is correct,  
estimates from multi-level programs will be more efficient, and  
standard errors will be more precisely estimated. In Stata and SAS,  
you can combine the advantages of both: fit a multi-level model, but  
get cluster-robust standard errors. In -gllamm-, you can go further,  
fit a multi-level model but also account for clusters that are not  
part of the model.

-Steve

On Feb 12, 2009, at 10:11 AM, ronggui wrote:

> Hi all,
>
> When the data is clustered rather then independent, there are many
> possible ways to handle it such as dummy variable, panel data models
> (fixed or random effects), GEE and multilevel model (e.g.
> http://www.gseis.ucla.edu/courses/ed230bc1/notes3/cluster.html). Of
> course, cluster() option of reg is one quick choice too. Yet, I have
> noticed that the result (especially coef) from reg differs that from
> multilevel model. Furthermore, this kind of adjustment is rare in
> other statistical software. All of this makes me cast doubt on the
> practices of regression with cluster() option in the Stata way. I
> wonder if there is paper on this, especially if  this adjustment is
> proper or improper? Besides, can we predict the change of level of
> significance by using cluster()? Thanks in advance.
*
*   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/