Following up on Austin's comments: yes, it is perfectly reasonable to use
each monthly or yearly observation as covariates in the propensity score
model.
There are a couple alternative options available for this problem:
(1) roll up the periods into one aggregate value and propensity score match,
then check the balance on each period separately. Rosenbaum and Rubin have
suggested in more than one paper that the propensity score will tend to
achieve balance on these covariates, and I find this to be the case almost
100% of the time.
(2) there is a user-defined program in SAS called "traj"
(http://www.andrew.cmu.edu/user/bjones/index.htm) which takes multiple
observations per person and assigns them to a trajectory group (through a
clustering model). Once the subjects are assigned to the given trajectory
groups, you can further propensity score match them within the group. BTW,
the traj program is currently being written for Stata (yeah)!
A marvelous paper on this concept is:
Amelia Haviland, Paul R. Rosenbaum, Daniel S. Nagin. Combining Propensity
Score Matching and Group-Based Trajectory Analysis in an Observational
Study. Psychological Methods. 2007, Vol. 12, No. 3, 247-267
I hope this helps.
Ariel
Date: Tue, 8 Dec 2009 21:54:09 -0500
From: Austin Nichols <[email protected]>
Subject: Re: st: Time series and psmatch2
Joshua Hawley <[email protected]> :
Maybe we are not speaking the same language exactly, but no, this is
not "creating issues." If you want to model the propensity of
receiving some treatment in 1990 and you have the earnings history
from 1971 to 1989 for each control and treatment case, you can
condition on each year of earnings by including 19 variables for
earnings in your model (or a bunch of categorical variables measuring
e.g. membership in deciles in each year). There is no "repeated
measures" problem per se.
There is very good support for propensity scores in Stata; hard to get
much easier than -predict- after -logit- or -probit- (or -egen- for a
nonparametric model of propensity score).
On Tue, Dec 8, 2009 at 9:23 PM, Joshua Hawley <[email protected]> wrote:
> Thanks for the response.
>
> I've done that before, as is mentioned in Becker and Ichino. It works
fine.
>
> However, isn't this simply creating issues because of the repeated
measures
> over time?
>
> I guess I was looking for a time series version of the psmatch2. I'm
> wondering if there is one in R? I'm not sure of the mathematical problems
> that the repeated cross sectional strategy presents, but I don't like it
> from a conceptual basis. If you want to model functions over time, it
makes
> no sense to assume that the selection of the proper "control group" comes
> from repeated cross sectional measures.
>
> I wonder if you could do an approximation of a propensity score in HLM,
with
> the times nested in individuals?
>
> Anyone done anything about this in the past? Why don't we have better
> support for propensity scores in Stata?
>
> Josh
*
* 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/
