14 Jul
2003
14 Jul
'03
4:05 p.m.
It will compute a probability density of the missing data given the data.
so the procedure i'm suggesting will be computing part of the density that
you don't need. so you can ignore it.
Gary
: Gary King, King(a)Harvard.Edu http://GKing.Harvard.Edu :
: Center for Basic Research Direct (617) 495-2027 :
: in the Social Sciences Assistant (617) 495-9271 :
: 34 Kirkland Street, Rm. 2 HU-MIT DC (617) 495-4734 :
: Harvard U, Cambridge, MA 02138 eFax (928) 832-7022 :
On Mon, 14 Jul 2003, Feinstein, Zachary wrote:
Thanks for the reply. Indeed you are correct about
the "are you pregnant"
analogy. There is no sensible answer to those questions for those people.
Your solution to just impute everything, then know that particular cells is
intriguing. On one hand, it is quite simple. On the other hand, I was
under the impression that an iterative process went on behind the scenes in
estimating the missing values. Your response leads me to believe that it is
iterative in the sense that it only uses information from data that exists
and does not substitute, then go through another iteration (think of a
cluster analysis procedure where people are first assigned, then can move
around a bit, depdenting upon what procedure you are using).
Once again, many thanks for getting back with me so quickly.
Zachary
-----Original Message-----
From: Gary King [mailto:king@harvard.edu]
Sent: Friday, July 11, 2003 9:14 PM
To: Feinstein, Zachary
Cc: Amelia Listserv
Subject: Re: Way to Do Missing Value Imputation But Also Have
Structurally Mis sing Data Unhampered?
a structural missing element would be the answer to the question "are you
pregnant?" and the respondent is male. you don't ask the question because
there is no sensible random variable that this answer would produce. if
you havethis kind of problem, thenyou could merely impute this along
with the rest but ignore the imputations for the structural zeros.
however, maybe what you have instead is just questions you decided not to
ask some respondents for lack of time. this is often called matrix
sampling, and there's no problme both imputing and using theimputations
from this setup. in fact, since you determine the missingness pattern,
you know that the asumptions of amelia apply (assuming you choose the
respondents in a sensible way of course).
Gary
On Fri, 11 Jul 2003, Feinstein, Zachary wrote:
I have a dataset of let's say 200
respondents by 10 variables. This is
for
market research. Occasionally people say
Don't Know or state they do not
wish to respond to certain questions. For the analyses I will be using
the
results for, I need to estimate the missing data
here. A poor-person's
way
of doing it would be mean-substitution.
Note that certain groups of people were not asked certain questions though
based on skip patterns. I cannot remove their column or row from the 200
X
10 matrix and I still wish to figure what to do?
Note that the scales are ordinal/continuous from 1 to 10. I made it so
that
if they said Don't Know, the data becomes
missing. I used zeroes to
define
the cells where it is structurally missing and
quite Kosher. Perhaps I
should use a character to make it explicit that it is not part of the
scale
in the future.
Any help on how to use Amelia with this kind of situation is greatly
valued
and appreciated. Thank you.
--
Zachary S. Feinstein
Methodologist
Harris Interactive
zfeinstein(a)harrisinteractive.com
http://www.harrisinteractive.com
phone: (952) 541-7161
435 Ford Road, Suite 250
Minneapolis, MN 55426