The page referred states that:
"If there are only a few out of bounds predictions, a reasonable strategy is to round
the imputations below zero up to zero and likewise those above one down to one."
Is there a practical threshold guideline, perhaps a percentage of observations, that
constitutes "a few"?
And does this advice hold for any out of bounds predictions, irrespective of the magnitude
of distance from the bound? If not, are there practical guidelines for determining when
even a few observations are so far out of bounds that the corrective measures are
required.
Thanks in advance.
----- Original Message -----
From: Gary King <king(a)harvard.edu>
Date: Monday, June 7, 2004 5:00 pm
Subject: Re: [amelia] impossible values
Usually if you transform the data before Amelia, impute, and then
transform back, you can avoid this problem. See
http://gking.harvard.edu/amelia/node22.html for more details about
thisand other procedures in more complicated situations.
Best of luck,
Gary King
: 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 (617) 812-8581 :
On Mon, 7 Jun 2004, Leonelo Bautista wrote:
I'm imputing values for missing blood lipids
(roughly 20% out of
3000> observations) in a cross-sectional data set. Some of
the
imputed values are
not possible (for example, negative triglycerides
levels). I'm
not sure how
to handle these imputed values. Should I just
ignore then and go
ahead with
the analysis? Is it possible to impose
restrictions on the model
(using> Amelia) to make all triglyceride values
positive?
Any suggestions will be appreciated.
Leonelo E. Bautista, MD, DrPH
Assistant Professor
University of Wisconsin Medical School
Population Health Sciences
610 Walnut Street, 703 WARF
Madison, WI 53726-2397
Phone: (608)265-6176
Fax: (608)263-2820