re: Re: st: propensity score matching question

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

Hi Nyasha,

Is there a reason why you could not match the treated individuals in each
period to controls in those respective periods? In other words, those
treated in the first period (July 2009 - Dec 2009) would be matched to
available untreated individuals during that period. 

Note that matching is intended to replicate an RCT (by matching on observed
characteristics and assuming that there is no meaningful residual
confounding). Thus, you would be matching on pre-intervention
characteristics only in each cohort (without looking at the outcomes).

If the untreated individuals do not temporally align with the treatment
periods, you can still match, since what is of primary importance here is
that the treated and matched controls are balanced on baseline observed
characteristics. Of course, if "time" is an important covariate to consider,
you will have to ensure that the controls are matched temporally.

As for your last question about differing observation periods: that could be
a major problem if you believe there is a dose-response here. You should
rely on content expertise to answer that question, but you should also test
that in the data. If there is a dose-response, you will have to ensure that
you are comparing treated to controls over the same duration. If you
arbitrarily "standardize" the time-frame under study (such is the case when
people "annualize" data), you may be introducing measurement error.

I hope this helps

Ariel


Date: Tue, 7 Aug 2012 13:27:50 +0200
From: Nyasha Tirivayi <ntirivayi@gmail.com>
Subject: Re: st: propensity score matching question

Dear Peter

Thanks for your response. The study is evaluating the impact of food
aid on adherence to medication. The hypotheses is that food aid
improves adherence to medication. I have three groups of the treated
i.e.  first group was on food aid from july 2009 to dec 2009, second
group from may 2009 to Jan 2010  and third group from december 2009 to
July 2010. The first group has about 36 people, the second around 65
and the last one abour over a 100.

The comparison group has over 1000 people. Since I do not have the
same baseline for everyone, is propensity score matching still ideal
for my evaluation? How would I account for the varying times in
program participation by the three treated group?   If not, what other
method can I use in Stata.

Kindly advise

Regards

N. Tirivayi
Maastricht
Netherlands


*
*   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/