it is not entirely clear, I haven't tested it, and the problem will be hard in an absolute sense for all algorithms, but I suspect that Amelia will be relatively easier to converge, because more of it depends on em and deterministic rather than stochastic convergence, but no better at getting the uncertainty right.
its probably worth a try.
Gary King
-----Original Message-----
From: "Paul von Hippel" <von-hippel.1(a)sociology.osu.edu>
Date: Tuesday, Jun 21, 2005 9:29 pm
Subject: Re: [amelia] collinearity and convergence
Thanks for your comments. It's complicated to explain why, but I actually have an unusual situation where my analysis can make use of the collinear variables. So my question is whether Amelia will handle collinearity
better than competing algorithms. King, Honaker, Joseph, and Scheve claim that Amelia's algorithm has fewer convergence problems than leading
alternatives, and I'm wondering if those advantages hold up when some variables are nearly collinear.
Gary King said:
> I don't know the details of what SAS/MI does, but collinear variables
don't help predict much and will slow down convergence. You might as well
create an index from the collinear variables. Unless you're mainly
interested in the conditional effects of one given the other (in which
case your data won't have much info about it anyway), you won't lose much
in terms of bias, and the algorithm will go faster too.
Gary
> ---
Gary King
David Florence Professor of Government,
Director, 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, 21 Jun 2005, Paul von Hippel wrote:
>> I have been using the MI procedure in SAS, which converges slowly if at
> all when the variables are nearly collinear. I am considering trying
> Amelia -- will Amelia have an easier time with collinearity?
>
>
> --
> Paul von Hippel
> Department of Sociology / Initiative in Population Research
> Ohio State University
>
> -
> 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.
-
Amelia mailing list served by Harvard-MIT Data Center
[Un]Subscribe/View Archive: http://lists.hmdc.harvard.edu/?info=amelia
In response to a query from Carl Klarner regarding the compatibility of
GAUSS 6.0 and Amelia:
I use Amelia for GAUSS (2.1) with GAUSS 6.0.54. The key thing I needed
to do to get it running was in include the "pgraph" library. Hence, I
have the following statement immediately following my "new" command:
library amelia, pgraph;
--Kent Currie
-
Amelia mailing list served by Harvard-MIT Data Center
[Un]Subscribe/View Archive: http://lists.hmdc.harvard.edu/?info=amelia
I have been using the MI procedure in SAS, which converges slowly if at
all when the variables are nearly collinear. I am considering trying
Amelia -- will Amelia have an easier time with collinearity?
--
Paul von Hippel
Department of Sociology / Initiative in Population Research
Ohio State University
-
Amelia mailing list served by Harvard-MIT Data Center
[Un]Subscribe/View Archive: http://lists.hmdc.harvard.edu/?info=amelia
Looking through the archives, I saw that a user ran into memory
limitations when using Amelia under Gauss. Could I run into the same
limitations using Amelia for Windows? I have about 5,000 cases, and I may
want to impute as many as 50-100 variables.
Thanks!
--
Paul von Hippel
Department of Sociology / Initiative in Population Research
Ohio State University
-
Amelia mailing list served by Harvard-MIT Data Center
[Un]Subscribe/View Archive: http://lists.hmdc.harvard.edu/?info=amelia
Thanks for your note.
the user's group is the Amelia listserv. There's info on how to subscribe
at our web page. I'm CCing the group to see if anyone has had any
problems with Gauss 6.
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 Fri, 17 Jun 2005, Carl Klarner wrote:
> Dear Professor Gary King:
> My name is Carl Klarner and I am a professor at Indiana State
> University. Thank you for the help you've given me in the past and for
> putting up so much free software on your Web site.
> I am having a problem using Amelia. My dataset has 69 variables, and is
> 769 cases long. The good news is that 67 variables are fully observed,
> which cuts down on the amount of memory that is needed considerably.
> The stand alone version of Amelia can only define about a dozen
> variables as "fully observed." For this reason, I purchased Gauss4.0.
> It worked fine, I was able to use Amelia and define many variables as
> "fully observed." Recently I purchased Gauss6.0, and now Amelia
> won't run with it. Do you know of any compatibility issues between
> Gauss6.0 and Amelia? Do you know of any user's groups that deal with
> Amelia? I contacted Aptech systems, and they helped me as much as
> possible. They told me that the next step should be to contact you.
> Thank you very much for your time, I'm sure you're extremely busy.
> Sincerely,
> Carl Klarner
>
>
-
Amelia mailing list served by Harvard-MIT Data Center
[Un]Subscribe/View Archive: http://lists.hmdc.harvard.edu/?info=amelia