st: impute or not

[Steve Pitts <pittssteve@gmail.com>]

I'm using Stata v11, and am in this fix:

I'm a timorous non-statistician using the public files from a national
multistage survey of emergency department visits. Length of stay (LOS)
is missing in about 8% of 254,000 observations (lots more in some
subgroups). I'd like to use LOS for other calculations (e.g. hourly
occupancy, for which I've created a loop incorporating arrival time
and LOS).

I think I've worked out how to use the mi programs for estimation
using LOS, e.g. "mi estimate: svy: mean lnlos, over(hour)". I'm
stumped on how to use imputed LOS in a program, though. Since I'm not
really planning to use occupancy in a model, just want the quantity
and its precision for each hour separately, do I need to do multiple
imputation? Listwise deletion reduces total occupancy with certainty,
never mind the potential for bias in various correlations, so I really
do want some form of imputation.

I know that ethnicity and other frequently missing variables are
imputed somehow in the public use data, and presented as if not
missing, for use in other analyses. How can I do that with LOS? I
checked out the Stata hotdeck program, and it too generates multiple
datasets, when I just want one, unless I'm not understanding
something. I thought that by using a random draw from other
observations with the same covariate pattern that are fully
nonmissing, the hot deck approach allows one to create a single
dataset without optimistically biased std errors one would get with
mean substitution. The Stata mi manual says there are two kinds of
persons: the imputer, and the data analyst. I want to be the analyst.

Thanks

Steve Pitts
Emory University Emergency Medicine
Atlanta, Georgia, USA

*
*   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/