Re: st: propsensity scoring for 4 groups in observational design

["Ariel Linden, DrPH" <ariel.linden@gmail.com>]

Hi Stephen,

When the treatment variable is more than two levels, a different approach is
taken to create the propensity score.

While not the easiest of reads, I suggest you read: Robins JM, Hernán MA,
Brumback B. Marginal structural models and causal inference in epidemiology.
Epidemiol 2000;11:550?60. 

In particular, section 6 discusses multilevel treatment. In general, you
would use an ordinal or multilogit model to estimate the propensity score,
and then use the estimate corresponding to  true level of the treatment.

For a continuous treatment variable (e.g. a drug with increasing dosage),
modelling the propensity score is even more complex. Fortunately, there is a
user written stata program available called -doseresponse- , (but you really
would only need the sub-routine called -gpscore-)  

This program comes with an accompanying paper in the Stata Journal by
Michela Bia and Alessandra Mattei called "A STATA Package for the Estimation
of the Dose-Response Function through Adjustment for the Generalized
Propensity Score",  2008. Stata Journal Volume 8 Number 3.  

In general, this is complex stuff, but I hope this will guide you in the
right direction

Ariel



Date: Mon, 4 Oct 2010 14:39:59 -0700 (PDT)
From: oconnor7@u.washington.edu
Subject: st: propsensity scoring for 4 groups in observational design

Hello, 

I'm relatively new to Stata and very new to this listerv. I run Stata 11 for
Windows. I am wondering whether anyone has advice on obtaining propensity
scores for 4 groups in a longitundinal, observational study. We have a
sample of children who experienced a traumatic brain injury and have been
divided into 3 groups based upon severity of injury. We also have an arm
injury comparison group. It was recommended that I create a propensity score
for being in a group by using the logit command, followed by the grouping
variable and then covariates of interest. For instance: 

"logit inj_sev maxais gender age" 

After more research, it I decided to download pscore updates. However, I am
worried that the grouping variable has to be dichotomous (treatment/no
treatment) in order for this command to run correctly. When I enter the
following command: 

"pscore  new_inj_severity2  maxais_n iss_n, pscore(prop) logit"

I receive the following error message:

Estimation of the propensity score 

outcome does not vary; remember:
                                  0 = negative outcome,
        all other nonmissing values = positive outcome

I'm not quite sure where to go from here. I do have two main questions: 

1) Is there a better command for determining the propsensity score for four
groups?
2) Am I doing something else incorrect to receive teh error message above?


Any other tips for determining propensity scores is greatly appreciated.
Thank you for taking the time to read and consider responding to my query. 

Best, 
Stephen


Senior Fellow
Harborview Injury Prevention & Research Center
Department of Psychiatry & Behavioral Sciences
University of Washington at Harborview Medical Center



*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/