st: Population averaged vs random effects logit

[Jack Needleman <needlema@ucla.edu>]

We have a data set of hospital discharges with 280,000 discharges from a 
3-year pooled sample of hospitals.  There are 1181 hospitals in the 
sample, and 1574 hospital-years of data.  841 hospitals are in the 
sample for only 1 year, 287 for two and, 53 for three years.

We are analyzing a mixed model with a substantial block of patient 
variables, 10 hospital-level variables that are the focus of the 
research, and year dummies.  For most of the sample, as noted above, 
there is only a single set of hospital-level variables.

I would have assumed that the correct model was a random effects logit, 
but in reviewing the literature came across two articles that argue that 
if you only have a single set of observations at the cluster level, the 
population averaged model is more appropriate:

Neuhaus JM. Statistical methods for longitudinal and clustered designs 
with binary responses. Stat Methods Med Res. 1992;1(3):249-73. Review.

Young ML, Preisser JS, Qaqish BF, Wolfson M. Comparison of 
subject-specific and population averaged models for count data from 
cluster-unit intervention trials. Stat Methods Med Res. 2007 
Apr;16(2):167-84.

The results from the RE model are substantially closer to zero.  The 
purpose of the research is to explore the association of one of the 
hospital level variables (staffing) on patient outcomes.  I would 
welcome any thoughts from the list members on which model should be 
preferred.  Many thanks.

--
Jack Needleman, PhD
Department of Health Services
UCLA School of Public Health


*
*   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/