the Robert Lalonde article presents four estimates of the annual difference in earnings ($s) for women in the NSW treatment group who received job training and women in the comparison group who did not receive the training. Importantly, the women were randomly assigned to the treatment and the comparison groups. Assume the experiment is ideal and the sense that the randomization was done successfully and there is no selective attrition. The year 1975 is prior to the training and 1979 is post-training. Estimates from Columns 2 and 4 come from bivariate regressions of earnings on an indicator for being in the treatment group, and estimates in Columns 3 and 5 come from multivariate regressions of earnings on an indicator for being in the treatment group and demographic controls.

a) Interpret the magnitude of the estimate in Column 4. What is the percentage increase associated with the training?

b) The estimates in columns 4 and 5 are different from one another. Is this difference evidence that one or both of these estimates are biased? If the answer is no, what accounts for the difference? Explain your answer.

Now drop the assumption that the experiment is ideal. Rather, assume that there is selective attrition from the control group. Specifically, those with the lowest expected earnings and weakest connection to the workforce are the most likely to attrit from the sample

c) Would you expect such attrition to bias the estimates? If so, in which direction? Explain your answer.

Estimated effects of training for AFDC recipients (women) Comparison Group Earnings Growth 1975-79 NSW Treatment Earnings Less Comparison Group Earnings Pre-Training Post-Training Year, 1975 Year, 1979 Unad Ad- Unad Ad- justed justed justed justed (2) (3) (5) Name of Comparison Group" Controls 2.942 (220) -17 (122) - 22 (122) 851 (307) 861 (306) Adjusted column estimates come from regressions that include controls