Hi,
I'm currently using Stata 9.2 to analyze a survey
sample - the Nationwide inpatient sample that uses
three different weights (pweight = discwt), stratum
weight (nis_stratum) and psu weight (hospid)
However, I'm running into a problem when I try
specific subgroup analyses.
When I try to run logistic regression of the outcome
(say, death from pneumonia) on specific subgroups (say
age group 65-85) or in the presence of a specific
complication (say, acute renal failure), I get the
odds ratio estimates but no p-value or confidence
intervals, and I get this message "Missing standard
error due to stratum with single sampling unit". I
don't get this message for any of the univariate
analysis or for some other subgroups (especially those
with larger numbers, though that isn't necessarily the
case always).
I did look for strata with single PSU using the svydes
command. It was suggested to me to try to combine the
different strata with single PSUs in order to avoid
getting this error. However, for certain specific
subanalyses, nearly all the strata in the dataset are
listed as having a single PSU. I'm not sure that if I
try to combine almost all strata into one or two, then
the entire design of the stratified survey will be
lost and my estimates will not be accurate.
Have others who are using stata for survey analysis
run into similar problems? Is there any other way
around this problem? or is using SUDAAN the only
answer?
Thanks,
Ashwin
