Thanks for the reply. In the example that I gave I think the number
of parameters was reduced by 50%, which seems pretty substantial, and
I think this should generalize to more realistic scenarios. Or do I
have that wrong? Any other thoughts are welcome.
On Fri, Sep 18, 2009 at 5:25 PM, Gary King <king(a)harvard.edu> wrote:
Ameila I (but not II) has a feature like this which
allows you to impute
conditional on the observed variables. It does save parameters, but not as
many as you'd think.
Gary
---
http://gking.harvard.edu
On 09/18/2009 11:23 AM, Cyrus Samii wrote:
I often work with survey datasets for which n<p(p+3)/2 but where the
number of variables with missing data is a lot less than n. My
intuition tells me that in these cases, there is no reason to impute
from a model with p(p+3)/2 parameters. It seems excessive and
requires that we fiddle with dropping variables. I know we could
ridge the covariance matrix, but I'd like to bracket that for now to
consider the question of whether we really need those p(p+3)/2
parameters in the first place.
As an example, suppose we have data with 4 variables, A,B,C, and D.
We have complete data for A and B. C and D exhibit arbitrary
missingness (maybe monotone, maybe not). In this case, a full joint
normal approach such as Amelia's algorithm would estimate parameters
for a 4-dimensional normal distribution---that is, 4 means and 10
covariances, or 14 parameters. Suppose as an alternative, we impute
using a bivariate normal model, where C and D are modeled as bivariate
normal, with means that are regressions onto A and B, and conditional
on A and B, C and D covary according to Cov(C,D) not zero. This
distribution is thus characterized by 4 regression coefficients,
Var(C), Var(D), and Cov(C,D)---that is, 7 parameters (i.e. half as
many as in the full joint model) . Is there any reason that the
second, bivariate, approach wouldn't be preferred? (If so, the
obvious follow-up is, why are we trying to build more model than we
need?)
Thanks for any illumination!
Cyrus
--
Cyrus Samii
Political Science
Columbia University
cds81(a)columbia.edu
Burundi Survey:
-
Amelia mailing list served by Harvard-MIT Data Center
[Un]Subscribe/View Archive: