I'd be very hesitant to use indicator variables to model missing
variables. The group of subjects with the missing values will be a
mixture of subjects from the other categories of the variable.
Therefore, the relative risk (or the odds ratio) in this group would be
biased (even if missing values occur at random). When combining the
stratum specific relative risk to obtain an adjusted relative risk, we
will be summarizing biased and unbiased estimated and the adjusted
relative risk would be biased. See "Vach W, Blettner M. Biased
estimation of the odds ratio in case-control studies due to the use of
ad hoc methods of correcting for missing values for confounding
variables. Am J Epidemiol 1991;134:895-907" for a good discussion of
these issues.
Leonelo Bautista
-----Original Message-----
From: [email protected]
[mailto:[email protected]] On Behalf Of Dupont,
William
Sent: Friday, April 16, 2004 7:54 AM
To: bill magee; [email protected]
Subject: st: RE: RE: RE: Imputing values for categorical data
Bill
What you propose sounds reasonable to me. However, I recently submitted
a paper that did this and was trashed by a referee, in part because of
how I was handling missing values. Using an indicator variable for
missing values has the advantage as it gives you some idea as to whether
you are, in fact, dealing with nonignorable missing data. However, this
approach does not appear to be in fashion at this time.
My own approach to data analysis is to attempt to use methods that
1. I think are reasonable,
2. are widely accepted within the biostatistical community, and
3. avoid ignoring or attacking sacred cows that are dear to likely
referees.
My own sense is that, at least in medical statistics, multiple
imputation is becoming a very popular way of dealing with missing data.
I also feel it is a sensible approach, particularly if it is only used
for confounding variables or if your study design gives you reason to
believe that the missing data is missing at random.
I would be interested to know what other Statalisters think about using
indicator variables to model missing values.
Bill
-----Original Message-----
From: bill magee [mailto:[email protected]]
Sent: Friday, April 16, 2004 5:11 AM
To: Dupont, William
Subject: RE: RE: RE: Imputing values for categorical data
Hi Bill --
Rather than imputing a missing categorical control or confounding
variable, such as gender, wouldn't it usually be better to just include
a category for missing (e.g. a dummy for female, a dummy for missing,
with male as the excluded contrast group)?
bill magee
*
* 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/
*
* 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/
