The randomized experiment was devised to test whether or not a job training program that began in 1976 would boost post-program earnings (real earnings in 1978). The program guaranteed a job for 9-18 months.

The randomized experiment was devised to test whether or not a job training program that began in 1976 would boost post-program earnings (real earnings in 1978). The program guaranteed a job for 9-18 months.

Real earnings in 1978 (re78)

The major design element that was good was that the randomized selection appears to have been properly implemented, because all key baseline variables show very similar distributions across treatment and control groups (see Table 1). However, design elements that were problematic were: (1) difficulty following up with both groups years afterward, and the resulting attrition; (2) agencies that operated the experiment at the local level provided treatment group members with different work experiences depending on the state and locality in which they worked, and also provided men and women with different work experiences that reflected conventional and prevailing gender norms.

Lalonde used this NSW study to establish an experimental (“gold standard”) benchmark for the impact of the program on 1978 earnings. He then attempted to recreate a realistic observational study (substituting administrative survey data for the NSW control group data) to explore which statistical approaches (if any) could reliably identify the experimentally-derived causal estimate.