propensity score matching with ordered outcome variable

[Foday Msheku]

Hi all,
I have a question on propensity score matching for the outcome variable that is not in a continuous form. For example, I collected the survey data as follows. to find out if energy access has effects on income.
Let's say I have outcome variable (income) likes this,
less than 10000
10001 and 20000
20001 and 40000
more than 100000

My treatment variable binary.

I have done the matching and here is my result below. How to interpret the result after matching?

[Jeff Wooldridge]

Using matching in this way doesn't make sense because your outcome variable has no quantitative meaning. It would be one thing if the response was something like a credit rating or health measure or happiness measure on a Likert scale, although the resulting estimates would not have cardinal meaning. Your situation is worse because the "ordered outcome" is a censored version of income and doesn't have meaning by itself. Sometimes standard methods should be used. I would use the -intreg- command because you have a variable censored in the bins. I'd include the treatment indicator (electricity access) and the control variables and interact the treatment with the mean-centered control variables. I'd probably transform the censoring points by taking the log so that the underlying model is this:

log(y) = b0 + b1*w + x*b2 + w*(x - xbar)*b3 + u

and then log(y) is censored. Then, b1^ will be interpreted as the average treatment effect of electricity access on log(y) and, after multiplication by 100, will have a percentage effect. You'll use log(10000), log(20000), and so on as the upper and lower censoring values.

It's been awhile since I've used intreg but it's not hard. The above is "regression adjustment" -- a perfectly appropriate method -- and allows handling the data censoring. Matching does not allow that.

[Foday Msheku]

Many thanks Jeff. This gives a very good insight to proceed with the work. All the time I thought it will be possible.