I also have had the problem that Frank is mentioning below and I do agree
that there are better ways to model "probability of being on treatment". I
have played with some non-parametric and semi-parametric models and have not
concluded that there is a best way. It really depends on the data. Sometimes
very simple parametric models work pretty well and other times it is better
to go non-parametric at the first stage and semiparametric at the second
stage.
Best,
Suna.
-----Original Message-----
From: Frank E Harrell Jr [mailto:fharrell@virginia.edu]
Sent: Thursday, April 24, 2003 3:03 PM
To: Barlas, Suna
Cc: s-news@lists.biostat.wustl.edu
Subject: Re: [S] Propensity score models
On Thu, 24 Apr 2003 14:17:05 -0400
"Barlas, Suna" <suna_barlas@merck.com> wrote:
> It is implemented in two stages. First, you basically fit a logistic (or
> probit) regression where dependent variable is probability of being on
> treatment and use every independent variable you have in possession. The
> problem is prediction and not estimation and therefore you would go with
the
> "kitchen sink" approach. After you obtain probability of being on
treatment
> for everybody in the data, you include that as an independent variable in
> the second stage model that may model treatment effect on some dependent
> variable. One of the most popular ways is to use quintiles to categorize
> propensity scores obtained from the first model and include that in the
> second model. The idea is similar to matching but in multidimensions. You
> try to match people on their propensity scores - this tries to get you
> "closer" to randomization.
>
> Best,
>
> Suna.
>
The main problem with that approach I've encountered is imbalances in the
outer quintiles, as these quintiles cover wider intervals of propensity. I
prefer to model the propensity parametrically using cubic splines of logit
of propensity. By the way, many analysts make the mistake of putting
propensity in the model as a covariate rather than logit propensity, even
when linearity is assumed.
Frank Harrell
>
> -----Original Message-----
> From: Peter Flom [mailto:flom@ndri.org]
> Sent: Tuesday, April 22, 2003 1:50 PM
> To: s-news@lists.biostat.wustl.edu
> Subject: [S] Propensity score models
>
>
> I recently attended a talk on Propensity Score Models by Rajeev
> Dehejia.
> It was very interesting. It's a method for dealing with sampling bias
> in nonexperimental studies.
>
> I was wondering if anyone had programmed this in S Plus?
>
> Thanks
>
> Peter
>
> Peter L. Flom, PhD
> Assistant Director, Statistics and Data Analysis Core
> Center for Drug Use and HIV Research
> National Development and Research Institutes
> 71 W. 23rd St
> New York, NY 10010
> (212) 845-4485 (voice)
> (917) 438-0894 (fax)
>
---
Frank E Harrell Jr Prof. of Biostatistics & Statistics
Div. of Biostatistics & Epidem. Dept. of Health Evaluation Sciences
U. Virginia School of Medicine http://hesweb1.med.virginia.edu/biostat
------------------------------------------------------------------------------
Notice: This e-mail message, together with any attachments, contains
information of Merck & Co., Inc. (Whitehouse Station, New Jersey, USA) that may
be confidential, proprietary copyrighted and/or legally privileged, and is
intended solely for the use of the individual or entity named in this message.
If you are not the intended recipient, and have received this message in error,
please immediately return this by e-mail and then delete it.
==============================================================================
|