[George Ford]

You can't use DID if all units are treated by the law. You need a control group in the treated period.

As for pre-trends, say you are studying employment of some sort and you have limited data on a certain type of employment. You might look look at a broader but similar type of employment covering a long time for the areas in your sample to see if those trends are the same.

[Matthew Williams]

You can't use DID if all units are treated by the law. You need a control group in the treated period.

As for pre-trends, say you are studying employment of some sort and you have limited data on a certain type of employment. You might look look at a broader but similar type of employment covering a long time for the areas in your sample to see if those trends are the same.

Dear Prof. George,

Thank you from your insightful comments. Do you think that it makes sense if I use propensity score matching in this setup. For example, the year 2006 is used as the control and the years 2010 and 2012 as the treatment, controlling for other individuals and household characteristics? Thanks!

[George Ford]

Are the regressors unbalanced between the treated and control groups? (search covbal)

[Matthew Williams]

Are the regressors unbalanced between the treated and control groups? (search covbal)

Dear Prof. George,

Thank you for the useful suggestion about -covbal- command. However, since this command does not show p-values or any statistics to determine whether the test is statistically significant difference or not, I am not sure how to interpret the results appropriately. Would you mind sharing your insights on the following example?

covbal policy educ_1 if sampl, wt(wmweight)


             |             Treated             |             Control             |        Balance       
             |      Mean   Variance   Skewness |      Mean   Variance   Skewness |  Std-diff  Var-ratio 
-------------+---------------------------------+---------------------------------+----------------------
      educ_1 |  .0463901   .0442412   4.313346 |  .1193525   .1051252   2.348206 | -.2669854   .4208429 
--------------------------------------------------------------------------------------------------------

where policy is an indicator taking the value of one if in post-law and zero otherwise, educ_1 is individuals' education, sampl is the restricted sample, and wmweight is the sampling weight. Thank you.

[George Ford]

0.25 is considered "big" (some go as low 0.10)

you don't want t-test, since that depends on sample size.

[George Ford]

Imbens & Wooldridge 2009 paper in the JEL has a good discussion of covariate balance.

https://www.aeaweb.org/articles?id=10.1257/jel.47.1.5

[Matthew Williams]

Imbens & Wooldridge 2009 paper in the JEL has a good discussion of covariate balance.

https://www.aeaweb.org/articles?id=10.1257/jel.47.1.5

Thank you so much for the useful reference.

[Aisha Khan]

treatment0 = demonstration
treatment1 = harvester use (since all harvester use has a demo, this is fine)


treatment0*post gives the effect of the demo
treatment1*post gives you the effect of the harvester

should be no effect of demo without harvester use.

To get the effect of the demo, you need harvester use without a demo, which you don't have. You are limited to testing the effect of the harvester.

George Ford thank you for the help.
Reading on the literature, I find that some papers use the difference-in difference-in differences approach.
I am wondering if that is applicable to the research problem of demonstrations and adoption
where the demonstration is the treatment and adoption is the factor that can vary the impact of the demonstration within the treatment group.

[George Ford]

DiD is what you should be using. I think this approach would work. You can have demo with harvester, and but not harvester without demo. Thus, the question you can ask is "what is effect of demo+harvester" and what is the effect of "demo no harvester". I can't imagine a demonstration of a harvester would improve productivity not using the harvester (like a placebo effect), but I guess that's an empirical question. Should be zero. If not, then there may be a selection issue to investigate. Are more (or less) productive farmers given the demo, or is it random?

[Aisha Khan]

George Ford
the halqa and tehsils (higher geographical unit) were purposively selected. The selection criteria for tehsils were soil quality and and availability of services providers. From each tehsil, two halqa were selected based on the access to paved roads. From each halqa villages were selected based on the access to roads. Very large villages or those with extreme socio-economic diversity were not included.
The treatment halqa were randomly selected following a cluster randomized controlled trial methodology.

while the demonstrations were on harvester, they also included 'post harvest management techniques' and motivations for farmers. So is it possible that "demonstration" is the treatment but those who attended demonstrations and adopted the benefit the most.

[George Ford]

Does this look right?

1. Can't get demo
a. no harvester
b. yes harvester

2. Can get demo (very detailed selection criterion that may be correlated with outcomes)
a. no demo
i. no harvester
b. yes demo
i. no harvester
ii. yes harvester

I'd be concerned mostly about parallel trends between 1 and 2, and secondly covariate balance.

I might start with group 2 to see what you get.

If this outline is correct, I might start a new post with the outline and ask for ideas on how best to specify a model.