Thanks for sending along this additional information, Paul. A central
part of my read on the additional articles focuses on the relationship
between the proportion of missing data and the relative bias. Although
the proportion of missing data is relatively small in dataset with which
I'm working (more than half of the variables with missing data have a
missing rate of less than 1%; the largest percentage of missing data is
12-13% among a small number of variables), and thus, there could be
options for using the normal model of estimation and rounding for binary
variables that have the lowest rate of missing data if necessary, I've
been persuaded...Some of the dichotomous variables I'm using were formed
from continuous variables. I've decided to use the continuous variables
in the imputation model, and, if there remains a need for dichotomous
values in the analysis models (as I believe there will), I will
dichotomize the continuous variables post-imputation. I think this
route will help resolve my quandary and reduce the potential for bias in
resulting estimates.
Thanks, again, for all of the information and input.
Malitta Engstrom, Ph.D., L.C.S.W.
Assistant Professor
University of Chicago
-----Original Message-----
From: Paul von Hippel [mailto:von-hippel.1@osu.edu]
Sent: Wednesday, August 03, 2005 3:02 PM
To: Engstrom, Malitta; Gary King
Cc: amelia(a)latte.harvard.edu
Subject: RE: [amelia] Commands for Imputing Dichotomous Variables in
AMELIA
My reading of their article is that it's best to avoid rounding imputed
values if possible.
Here are two more pertinent articles:
http://www2.sas.com/proceedings/sugi30/112-30.pdf
http://www2.sas.com/proceedings/sugi30/113-30.pdf
At 03:34 PM 8/3/2005, Engstrom, Malitta wrote:
Many, many thanks for this information. It sounds like
experiments
conducted specifically with AMELIA indicate that using the normal model
underlying AMELIA to impute dichotomous variables (which will be
rounded
to 0 or 1) will not create biased estimates. The
article by Horton,
Lipsitz, and Parzen suggests that such action should be taken with
caution as there may be a risk of biased estimates in certain
situations
when the imputed values are dichotomized. From my read
of their
article, it looks like: (1) higher rates of missing data may result in
greater bias; (2) their demonstration of "a simpler model" may not
apply
to all situations; and (3) when given an option, it may
be better to
select a categorical variable which is not dichotomous for imputation
using the normal model and rounding the variable to reduce possible
bias.
As you might have guessed, what prompted my question was the imputation
of a dataset with more dichotomous variables than could be accommodated
using the "AMORDS" command in AMELIA; thus it requires the use of the
"ANOMS" command for some of the dichotomous variables. Given the
experiments conducted specifically with AMELIA, it seems that using
this
second configuration should be okay. If there is any
further
information
I should consider, have missed something in my
synthesis of the
information, or if there are any specific tests I should conduct to
check for potential bias, I'd be grateful for the input.
Thank you, again, for your thoughtful (and helpful!) responses.
Malitta Engstrom, Ph.D., L.C.S.W.
Assistant Professor
University of Chicago
-----Original Message-----
From: Paul von Hippel [mailto:von-hippel.1@osu.edu]
Sent: Tuesday, August 02, 2005 2:09 PM
To: Gary King; Engstrom, Malitta
Cc: amelia(a)latte.harvard.edu
Subject: Re: [amelia] Commands for Imputing Dichotomous Variables in
AMELIA
Here's a pertinent article by Horton et al:
http://www.biostat.harvard.edu/~horton/tasround.pdf
It shows that, when a normal model is used to impute a dichotomous
variable, it's best not to dichotomize the imputed values.
Best --
Paul
At 03:04 PM 8/2/2005, Gary King wrote:
Experiments seem to indicate that for imputation,
you can use the
normal
model underlying Amelia, even though you
wouldn't want to use it for
the
analysis model. if you choose options
appropriately, Amelia will
dichotomize the resulting simulations for convenience of subsequent
analysis. But if that's not necessary for your analysis model, then
using
the continuously imputed variable should work
fine.
Gary King
---
Gary King Institute for Quantitative Social Science
Harvard University, 34 Kirkland St, Cambridge, MA 02138
http://GKing.Harvard.Edu, King(a)Harvard.Edu
Direct 617-495-2027, Assistant 495-9271, eFax 812-8581
On Tue, 2 Aug 2005, Engstrom, Malitta wrote:
>Dear Dr. King,
>
>I'm hoping you can provide some information, or direct me to a
resource
>for gathering information, regarding the
imputation of dichotomous
>variables with AMELIA. Specifically, I'm wondering about how
>problematic it is to use the "ANOMS" command with the imputation of
>dichotomous variables. In statistical consultations, I've been
informed
>that it should not be a problem to use this
command; however, since
the
>>program directs one to use the "AMORDS" command with the imputation
of
>>dichotomous variables, I wanted to gather
additional information
about
>>it.
>>
>>I would be most grateful for any information you can share on this
>>topic.
>>
>>Thank you,
>>Malitta Engstrom
>>Assistant Professor
>>University of Chicago
>>-
>>Amelia mailing list served by Harvard-MIT Data Center
>>[Un]Subscribe/View Archive:
http://lists.hmdc.harvard.edu/?info=amelia
>-
>Amelia mailing list served by Harvard-MIT Data Center
>[Un]Subscribe/View Archive:
http://lists.hmdc.harvard.edu/?info=amelia
Paul von Hippel
Department of Sociology / Initiative in Population Research
Ohio State University
300 Bricker Hall
190 N. Oval Mall
Columbus OH 43210
614 688-3768
Office hours TThF 3-5pm
I read email every weekday at 3.
Paul von Hippel
Department of Sociology / Initiative in Population Research
Ohio State University
300 Bricker Hall
190 N. Oval Mall
Columbus OH 43210
614 688-3768
Office hours TThF 3-5pm
I read email every weekday at 3.
-
Amelia mailing list served by Harvard-MIT Data Center
[Un]Subscribe/View Archive:
http://lists.hmdc.harvard.edu/?info=amelia