Thursday, February 12, 2009

Day of an Assistant Professor (29)

CORRECTED on 22 Feb.

1. Attend a lunch-time seminar at Handels.

2. Continue working on the revision of the African democracy paper.

3. While waiting for regressions to be run, read Deaton (2009) again to summarize his points on randomized control trials (RCTs) in development economics:

(1) RCTs gives us an unbiased estimate of the average treatment effect due to the fact that the expectation is a linear operator, but it is not informative about other features of the distribution of the treatment effect such as median. (Pages 25-7)

(2) RCTs should be designed to learn not whether a particular project works but why it works, because "projects can rarely be replicated while the mechanism underlying success or failure will often be replicated and transportable" (pages 30-1). See Section 5 for more on this.

(3) Calculation of the standard error of the average treatment effect is not straightforward if the treatment effect is heterogenous, because the variance in the error term will then be different between the treated and the control group. (pages 31-3)

(4) Controlling for covariates to improve the precision of the treatment effect estimate can be hazardous with the presence of heterogeneous treatment effects. (Pages 34-6)

(5) If compliance to the treatment is incomplete, reduced-form evidence (ie. intention-to-treat effect) is more informative than the instrumental variables estimation because the latter yields the local average treatment effect which may not be of interest. (pages 14-20, 36-8)

No comments: