st: Regional dummies in cross section treatment effects

[Henrik Wiig <henrik.wiig@nibr.no>]

Hi!

I got a cross section dataset of 1200 households were we assume households in a specific type of community was imposed a treatment T more or less as a natural experiement.

The survey was collected with the idea of comparing the outcome variable Y (% share between 0 and 1) from households in such treatment communities with households in controll communities.

1) If the assumption of natural experiment holds, then a ttest comparing means between HH in treatment and control should be enough to indicate whether treatment has had a significant impact or not on Y, correct? What other types of test could I use. The dependent variable Y is a share truncated on both sides, i.e. 0 and 1.

2) Sceptics would argue treatment might be correlated with variable X1 and X2. I hence run a Propensity Score Matching model in the following way

pscore T  X1 X2, pscore(ps1) comsup
attk Y T, pscore(ps1) logit comsup boot reps(500) detail

I then get the following messeage:

The balancing property is not satisfied
Try a different specification of the propensity score

but the numbers keep pouring out anyway...

Her comes the main questions: What does the balancing requierement really do? My original understanding is as follows:

Compare Y in household with treatment to Y in household in control group if the following conditions are satisfied:

1) pscore_T is more or less equal to pscore_C
2) X1_T is more or less equal to X1_C
3) X2_T is more or less equal to X2_C

However, now someone tells me that common support implies that more extreme HH are thrown out until the mean of X1_T is more or less equal to the mean of X1_C, and similar for X2_T and X2_C.

To my understanding this implies that we do not compare households with extreme values of X1 and X2 in T and C, even though they are comparable between eachother while not the rest of the households in the sample. Is this not throwing away useful information?

If I want to keep such extreme pairs, what do I do?

3) The communities are taken from 10 different districts, which might have different unobservable cultural features that might affect the outcome variable Y. The original idea was hence just to compare households T with C within each district, not between districts using PSM. Then, I supposed that all comparable pairs was aggregated into one file, giving one global result of difference in means of Y between treated and control group.

How do I do this?

Hopefully someone out there can help so that I canstart writing my paper....

Henrik





t




*
*   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/