Re: st: analyzing clustered data

[Steven Samuels <samplerx@earthlink.net>]

On Dec 15, 2006, at 5:17 PM, Cathy L. Antonakos wrote:

> I have survey data from medical workers in ICUs. There are 462  
> surveys altogether, from 25 ICU's, in 8 hospitals. There is no  
> explicit survey sampling plan, so I believe I have to choose  
> between using multilevel modeling, and using clustered robust  
> standard errors (ref: UCLA Stata Library "Analyzing Correlated  
> (Clustered) Data"). I am planning to use clustered robust standard  
> errors. (The intra-class correlation is .24.)
>
> Although most of the variables in the regression model are at the  
> level of medical worker, there are 3 that are not. Two are ICU- 
> level variables (unit type, average patient risk score), and one is  
> hospital-level (hospital size).
>
> I have little experience with this, so I am wondering if the  
> following is correct.
> (1) Get robust standard errors by identifying "ICU" as the main  
> clustering variable. (2) Include in the regression model the  
> variables "unit-type," "average patient risk score" and "hospital  
> size" as control/independent variables.
Cathy,  I strongly suggest that you seek guidance at the Michigan  
Survey Research Center or Center for Statistical Consultation or the  
Department of Biostatistics.   I will ask you some of the questions  
that an expert consultant might ask.

Questions: What is the purpose of the study?  Do the eight hospitals  
constitute the "target population"? Do they represent a larger  
population of hospitals?  Were they a sample of a larger population?   
(The last two questions are NOT equivalent!)  What time-periods do  
the data represent?  Are the periods sampled or fixed?  Are time or  
calendar trends possible? Are some outcomes related to the length of  
stay in the ICU, and what is the relation of this to risk score?  Do  
you expect differences by shift? Were the surveys from a sample of  
medical workers? What kind of sample? What were the response rates?   
What characteristics were related to nonresponse.  Would weighting be  
helpful to better represent the population of workers?

The answer to these questions are important.  for example, if you are  
concerned only about the eight study hospitals, then "hospital" will  
be a stratum variable. If, however, the study hospitals represent  
others, then they may a first-level PSU, and ICU will be the second- 
level PSU. The proper approach might be a full multilevel model with  
fixed and random coefficients for  ICU and ICU-level variables and,  
possibly, for  hospital and hospital-level variables. You have to  
decide if you will use STATA's survey commands  or others.


I think that you need faculty-level consultation to help you to  
clarify your purpose, decide on your approach, and carry it out.


Steven



*
*   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/