re: st: matching method

["Ariel Linden" <ariel.linden@gmail.com>]

Your overall pattern of missingness is obviously not missing at random,
which will pose problems for you regardless of method.

It is difficult to counsel regarding multiple imputation without reviewing
the data. You didn't provide any information regarding sample size, the
number of variables, distribution of those variables, etc.

Perhaps the simplest approach would be to generate a variable that would be
associated with the HRA variable in that it would be binary: 

. gen bin_hra = cond(hra !=.,1, 0)

You can then use bin_hra as a matching variable.

That said, this may be problematic if the HRA variable takes on many
categories. This approach does not address the distribution of the
missingness, just the presence of missingness itself.

I hope this helps

Ariel


Date: Tue, 26 Nov 2013 18:12:17 +0000
From: "Song, Lin" <Lin.Song@kingcounty.gov>
Subject: st: matching method

Hello Stata experts and statisticians,

I have a survey dataset with 20% missing in the HRA variable which is a
geographic locator.  Those with HRA missing are more likely to be younger
age, non-white race, lower income etc.

I would like to get a 1:1 match for each of the subjects with HRA missing
from the non-missing group.

Should I use propensity score matching (selecting y to be one of the many
indicator variables in the survey and t to be the missing/non-missing
group), hot deck matching or some other matching method?  With hot deck, I'm
concerned that some of the cells will become too thin even if I just use
three matching variables.

Thank you very much for your advice.

Lin



*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/