Announcement

Collapse
No announcement yet.
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • How to use OLS to estimate a rough ATE (Average Treatment Effect)?

    Hi,

    I am using traditional NSW(National Supported Work Demonstration) dataset to do couple of matching estimations. I am trying to get the ATE of the work training in various approaches.The dependent variable is “re78(income in 1978)", the treatment variable is “treatment” where “treatment=1” represents group who get the training, the control variables are black, hispanic, married. One of the approach is to use OLS directly to get the ATE (I know this approach would have a bad estimation result but I still want to get an idea of this approach).

    Here is my thought of doing it:

    reg re78 black hispanic married
    predict unre78, xb
    The unre78 represents the estimated unobservable treatment effect. For example, as for group of "treatment=1”, the unre78 contains rough data of assumed income in 1978 if these people do not get the work training. Then I could get the individual treatment effect, and a rough ATT by “re78-unre78” if treatment=1. But for the control group, I do not think unre78 contains data of assumed income in 1978 if these people actually get the work training. Thus I consider myself thinking in a wrong way.

    So could someone please helping me with this ATE estimation question? I also intend to figure out how to get the Individual Treatment Effect (since I do not think doing a “re78-unre78” would generate the ITE)?

  • #2
    The unre78 represents the estimated unobservable treatment effect. For example, as for group of "treatment=1”, the unre78 contains rough data of assumed income in 1978 if these people do not get the work training.
    Maybe I don't understand you properly, but I don't think that's right. I think unre78 will just contain an estimate of income irrespective of any treatment effect.

    I think that if you want rough estimates of what average income would be with and without treatment, you need something like
    Code:
    reg re78 i.treatment i.black i.hispanic i.married
    margins treatment
    margins, dydx(treatment)
    Or have I missed the point?

    Comment


    • #3
      Originally posted by Clyde Schechter View Post
      Maybe I don't understand you properly, but I don't think that's right. I think unre78 will just contain an estimate of income irrespective of any treatment effect.

      I think that if you want rough estimates of what average income would be with and without treatment, you need something like
      Code:
      reg re78 i.treatment i.black i.hispanic i.married
      margins treatment
      margins, dydx(treatment)
      Or have I missed the point?
      Hi, Clyde,

      To be clearer about my question, I quote the original question that I need to solve below:

      Estimate an OLS regression of re78 on the following set of explanatory variables(black, hispanic, married) in the experimental sample(which it the combination of people being treated and people being untreated).Use this estimates to calculate hypothetical earning in the unobserved treatment state. Construct individual treatment effects and an estimate of the ATE.
      So is this original description better to understand? Sorry for my bad question description. Really appreciate if you could help me again with this!

      Comment


      • #4
        So maybe they want you to do this:

        Code:
        regress re78 i.black i.hispanic i.married
        predict re78_estimate_tx_unknown, xb
        That gives you individual level predictions of re78 when treatment is unknown.

        How you are supposed to know get an estimate of the ATE from that eludes me. The ATE is the expected difference in outcome between treatment and control groups. The predicted value when re78 is unknown is some indeterminate blend of those group averages. I honestly don't say how you can proceed from there.

        Evidently I'm missing something.

        That said, the code I showed in #2 will get you an actual estimate of the ATE.

        Comment

        Working...
        X