University of Illinois Department of Economics
Spring 1998 Roger Koenker
As announced in the first class on Tuesday, January 27 we will
have a discussion about the group of "Treatment Effect" papers
which were available from Up-Close. At this point I have had
only three volunteers: do the rest of you know about email?
Here is the lineup: I will begin with some comments on AIR. Then we have:
Brian Weikel -- Imbens and Angrist (1994)
Tereko Hippo -- Rosenbaum's discussion piece
John Deke -- Heckman's Randomization as an IV
Gregory Kordas -- Heckman's discussion of AIR
I would like all of you to give some serious thought to the following questions as you prepare for class on Tuesday:
1.) Heckman's model of randomization in social experiments seems to presume that people volunteer to participate in the program and then are randomized into either the treatment or control group. The Rubin model presumes instead that people are randomly assigned to treatment and control groups in the first instance, but we are then able to observe whether they "comply" with the treatment protocol. The latter is definitely more in keeping with the clinical trial examples. How does this affect the notion of "what to estimate" in the two approaches.
2.) Consider a situation in which we have the Rubin setup characterized above, and we observe a continuous variable indicating degree of compliance, for example the proportion of the indicated drug dose actually taken by the patient, or the proportion of classes actually attended in a job retraining program. How would you estimated the treatment effect in this setup.
