Dear Imputers,
Can anyone help me with the following question. I have imputed a dataset and
now I want to analyse two regression models with it (model 1 is nested in model
2).
In order to conclude which model to prefer I want to have a pooled version of
the F-test.
My idea was to use the article of Meng about likelihood ratio tests with
deviances (Biometrika 1992).
I thought that the deviance in case of a normal model (ordinary regression) is
the residual sum of squares. Or do I have to divide with some estimate of the
variance? If this is the case, what estimate for the variance should I use then?
I am doubting because I did also the F-test for every imputed dataset, these
are all not significant (all p-values about 0.10). But the pooled version I
calculated is very significant (p=0.00001).
Thanks for any help or suggestions.
Karin Oudshoorn
---------------------------------
Yahoo! for Good
Click here to donate to the Hurricane Katrina relief effort.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
http://lists.utsouthwestern.edu/pipermail/impute/attachments/20050913/0028b2d3/attachment.htm
From arisilva <@t> ibge.gov.br Wed Sep 14 10:28:21 2005
From: arisilva <@t> ibge.gov.br ([email protected])
Date: Wed Sep 14 15:23:43 2005
Subject: [Impute] Seminar on Editing and Imputation
Message-ID: <[email protected]>
Skipped content of type multipart/alternative-------------- next part
--------------
A non-text attachment was scrubbed...
Name: SICI_Ingles.pdf
Type: application/octet-stream
Size: 148855 bytes
Desc: not available
Url :
http://lists.utsouthwestern.edu/pipermail/impute/attachments/20050914/14c0d8a5/SICI_Ingles-0001.obj
From rubin <@t> stat.harvard.edu Thu Sep 15 14:01:16 2005
From: rubin <@t> stat.harvard.edu (Donald Rubin)
Date: Thu Sep 15 14:01:40 2005
Subject: [Impute] Pooled version of Omnibus F-test
In-Reply-To: <[email protected]>
References: <[email protected]>
Message-ID: <[email protected]>
Hi Karin,
It sounds as if you must be doing something wrong. If it is simple for
you to do, also use the standard variance-covariance combining rule, as in
my MI book.
The result in the Meng & Rubin Biometrika article is asymptotically
equivalent, but should be simpler in your problem because it just uses the
ingredients of the likelihood ratio test, which, as you note, are the
residual sums of squares. It sounds to me that you're forgetting to
divide by degrees of freedom for a denominator somewhere when implementing
our result. There's a quite simple description of that result in a Stat
in Med review article by Rubin&Scherker (in the 1990s), and I also think
in the Little & Rubin (2002) book. I'm out of town so I don't have easy
access to page numbers of either. Sorry.
Best wishes, Don
On Tue, 13 Sep 2005, Karin wrote:
> Dear Imputers,
>
> Can anyone help me with the following question. I have imputed a dataset and
> now I want to analyse two regression models with it (model 1 is nested in
> model 2).
> In order to conclude which model to prefer I want to have a pooled version of
> the F-test.
> My idea was to use the article of Meng about likelihood ratio tests with
> deviances (Biometrika 1992).
> I thought that the deviance in case of a normal model (ordinary regression)
> is the residual sum of squares. Or do I have to divide with some estimate of
> the variance? If this is the case, what estimate for the variance should I
> use then?
>
> I am doubting because I did also the F-test for every imputed dataset, these
> are all not significant (all p-values about 0.10). But the pooled version I
> calculated is very significant (p=0.00001).
>
> Thanks for any help or suggestions.
>
> Karin Oudshoorn
>
>
> ---------------------------------
> Yahoo! for Good
> Click here to donate to the Hurricane Katrina relief effort.
--
Donald B. Rubin
John L. Loeb Professor of Statistics
Department of Statistics
1 Oxford Street
Harvard University
Cambridge MA 02138
Tel: 617-495-5498 Fax: 617-496-8057
