How to conduct a regression discontinuity on stata

[Danula Gamage]

Dear All,

I am examing the impact of a year-long review period on a job outcome variable.
Background context - There are 15 different private providers of employment services and for each of them I observe the job outcomes gained over time (monthly observations). After the first 10 months, there is a year-long review period in which performance of job outcome are reviewed and base on the performance they are rewarded.

I want to examine whether there is a jump in the job outcomes when the review period starts. The time period of the non-review period is June-2011- March-2012 and review period is Apr-2012 - March 2013.

I far as I am aware you can conduct this by regressing the job outcome variable with the independent variable including time dummies (note that all my independent variables are dummies). Extract the coefficients of time dummies (ie. June-2012 - March-2013), and introduce another dummy variable for the review period. Then regress the coefficient of the time dummies on a polynomial (with degree 2) of the month dummies and the dummy variable. The coefficient of the dummy variable will show whether there is a jump or not.

My question is whether the mentioned method is correct. If so how could I do this on stata. (I am fairly new to stata)
Hope I am clear enough.

Many Thanks
Danula

[Weiwen Ng]

Danula,

I've got no specific experience with regression discontinuity in Stata. However, it doesn't sound like this is a regression discontinuity problem to me. In RD, you have a situation where you assigned some treatment based on a hard cutoff on a score. In your case, there could be an RD setup if people were rewarded for their job performance based on a cutoff in some performance score. And so, you're doing a regression on the people who are just above the cutoff (treated) vs. just below (untreated). In this setup, however, it sounds like there's endogeneity that you have to consider.

What you say you want to do sounds more like a difference-in-differences regression for the effect of the review period.

Unless I'm missing something, of course.

[Richard Williams]

Maybe see

https://www.statalist.org/forums/for...sults-in-stata


https://www.stata.com/meeting/columb...5_cattaneo.pdf

[Danula Gamage]

Thank you, both for the replies.

[Neil Farmer]

Danula,

From your description, it sounds like the running(threshold) variable for your RD design is time. While it is not unheard of to perform an RD with this type of running variable, another option is to perform an event study. Like Weiwen stated, it sounds like there is an endogeneity problem here that might be visibly seen in an event study.