Thank you very much for the thoughtful replies.
My aim in presenting the Hosmer Lemeshow statistics for each bivariate
combination was simply to demonstrate that the variable of interest remained
a robust predictor (i.e. not only is the odds ratio unchanged, but model fit
remains non-rejected) in the presence of other strongly strongly co-linear
predictors. I realize that this statistic cannot be used to compare one
bivariate model against the next (and indeed, there are some missing
observations that would make such comnparison problematic with any method).
I am in the process of digesting your paper and tracking down the cited
refernces. Thanks for the leads.
-----Original Message-----
From: [email protected]
[mailto:[email protected]] On Behalf Of Maarten buis
Sent: Sunday, January 29, 2006 4:42 PM
To: [email protected]
Subject: Re: st: How can I access goodness-of-fit p-value in a program?
(re-send with subject line)
> > The bivariate logistic models consist of one continuous dependent
variable
> > of interest which is included in every model (var1, below) together with
a
> > different (mostly dichotomous) variable (var2) for each model, i.e.:
> >
> > .logistic death var1 var2
> >
> > The point is to assess the change in the odds ratio associated with var1
> > with inclusion of var2.
>
>
> It looks like you are doing some sensitivity analysis or extreme bounds
analysis like
> (Sala-i-Martin 1997), is that right?
>
> Sala-i-Martin, Xavier, 1997. "I Just Ran Two Million Regressions,"
American Economic Review,
> American Economic Association, vol. 87(2), pages 178-83.
>
I have two additional comments on this enterprise:
1. The ommited variable bias in logistic regression works a bit different in
logistic regression
than in OLS regression: Omitted variables will effect estimated effects of
the variable of
interest even if the omited variable is uncorrelated with the variable of
interest. I have a
working paper (Unobserved heterogeneity in logistic regression) on my
website
(http://home.fsw.vu.nl/m.buis/) which shows this with some simple examples.
2. Have you taken care of missing values (i.e. have you checked that the
sample size is the same
in each model)? If not, than you may make some strange comparisons. You
could either choose to
only use data that are completely observed on all variables that are used in
at least ones model,
or do some multiple impution (e.g. by using -ice-, see -findit ice-)
HTH,
Maarten
-----------------------------------------
Maarten L. Buis
Department of Social Research Methodology
Vrije Universiteit Amsterdam
Boelelaan 1081
1081 HV Amsterdam
The Netherlands
visiting adress:
Buitenveldertselaan 3 (Metropolitan), room Z214
+31 20 5986715
http://home.fsw.vu.nl/m.buis/
-----------------------------------------
___________________________________________________________
Yahoo! Messenger - NEW crystal clear PC to PC calling worldwide with
voicemail http://uk.messenger.yahoo.com
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
