Laura,
If your variables are x1, x2, x3, you could transform them to two log ratios. That is,
create new variables, z1= log(x2/x1) and z2=log(x3/x1) using natural logarithms. The
choice of which variable serves as the common denominator is arbitrary, so if one of your
variables has no missingness, choose that one. The transformed variables are now
independent, unconstrained, and contain all the information of the original variables. If
you model them as distributed by a Normal (the working assumption in the base Amelia model
with no additional options), then the original x1, x2, x3 are distributed additive
logistic normal.
Then include the z's (and omit the x's) from the imputation model. When you
have the z's imputed, transform them back into the x's (if the x's are
needed for your model purposes). The reverse transformations are:
x1=1/w; x2=exp(z1)/w; x3=exp(z2)/w; where w=1 + exp(z1) + exp(z2)
Key cites on this transformation, if you want more information, are:
Aitchison, J. 1986. ``The Statistical Analysis of Compositional Data.'' London:
Chapman and Hall
Katz, Jonathan and Gary King. 1999. ``A Statistical Model for Multiparty Electoral
Data.'' American Political Science Review 93(1)(March): 15-32.
Some papers with imputation using this model are:
Honaker, James, Jonathan Katz and Gary King. 2002. “A Fast, Easy, and Efficient Esti-
mator for Multiparty Electoral Data.” Political Analysis 10(1): 84-100.
Frisina, Laurin, Michael Herron, James Honaker and Jeffrey B. Lewis. 2008. “Ballot
Formats, Touchscreens, and Undervotes: A Study of the 2006 Midterm Elec- tions in
Florida.” Election Law Journal 7(1):25-47.
But these are both cases where the imputation is of a counter factual "what would
have happened if some variable in the composition that was forced to be zero, was not
forced to be zero." Your application might be much simpler and more straightforward
if you don't need to use the additional information of bounds to add constraints from
a factual observation to a counter factual estimation.
The hardest problem in this model, is if some of the observed x's contain zeros, thus
the logarithms become log(0) or log(infinity). If this is the case in your data, let me
know, and I'll send you some review notes on the half dozen popular ways people deal
this this. Aitchison includes the simplest of these, which is to just change the 0's
to a small value.
Best,
James.
--
James Honaker, Senior Research Scientist
//// Institute for Quantitative Social Science, Harvard University
________________________________________
From: amelia-bounces(a)lists.gking.harvard.edu [amelia-bounces(a)lists.gking.harvard.edu] On
Behalf Of laura mayoral [lm139(a)nyu.edu]
Sent: Thursday, September 27, 2012 7:34 AM
Cc: amelia(a)lists.gking.harvard.edu
Subject: Re: [amelia] Relationships to other variables
Dear all
I have to impute three variables that have to verify that are all between 0 and 1 and that
the sum of the three adds up to 1. Do you have an idea of how I could introduce the latter
restriction in AmeliaView?
Thanks a million in advance!
Laura Mayoral
On Fri, Sep 7, 2012 at 4:46 PM, Gary King
<king@harvard.edu<mailto:king@harvard.edu>> wrote:
you can do a multivariate transformation. e.g., if you have X, Y as your variables, you
could do a=X+Y and b=X-Y. then if you know b > 0, you could enter into amelia say
B=ln(b) and then no matter what value Amelia imputes for B, you'd be ok
Gary
--
Gary King - Albert J. Weatherhead III University Professor - Director, IQSS
<http://iq.harvard.edu/> - Harvard University
GKing.Harvard.edu<http://gking.harvard.edu/> -
King@Harvard.edu<mailto:King@Harvard.edu> -
@kinggary<http://twitter.com/kinggary> - 617-500-7570<tel:617-500-7570> - Asst
495-9271 - Fax 812-8581
On Thu, Sep 6, 2012 at 6:27 PM, Patrick Lam
<plam@fas.harvard.edu<mailto:plam@fas.harvard.edu>> wrote:
Is there a way to bound variables via relationships with other variables in the dataset
when multiply imputing with Amelia? For example, if we have a household income variable
and a household savings variable (both with some missingness) in our dataset, is there a
way to specify that savings must be less than income for each observation in the imputed
datasets? Priors and logical bounds currently in Amelia don't seem like they are set
up to do this exactly.
Thanks!
--
Patrick Lam
Department of Government and Institute for Quantitative Social Science, Harvard
University
http://www.people.fas.harvard.edu/~plam
--
Amelia mailing list served by HUIT
[Un]Subscribe/View Archive:
http://lists.gking.harvard.edu/?info=amelia
More info about Amelia:
http://gking.harvard.edu/amelia
Amelia mailing list
Amelia@lists.gking.harvard.edu<mailto:Amelia@lists.gking.harvard.edu>
https://lists.gking.harvard.edu/mailman/listinfo/amelia
--
Amelia mailing list served by HUIT
[Un]Subscribe/View Archive:
http://lists.gking.harvard.edu/?info=amelia
More info about Amelia:
http://gking.harvard.edu/amelia
Amelia mailing list
Amelia@lists.gking.harvard.edu<mailto:Amelia@lists.gking.harvard.edu>
https://lists.gking.harvard.edu/mailman/listinfo/amelia