National Supported Work Demonstration, Ex-Offenders

The National Supported Work Demonstration was a randomized evaluation, begun in 1975, of a subsidized work program that targeted four groups of unemployed individuals: adult women on AFDC, ex-addicts, ex-offenders, and youth who had dropped out of school. The program operated at 15 sites nationwide by local nonprofits. Participants were provided with temporary employment (for no longer than 18 months) with gradually increasing expectations for productivity and attendance and/or gradually decreased supervision. Because the precise parameters of interventions included in the evaluation varied from site to site, Hendren and Sprung-Keyser (2020) provide separate MVPFs for each of the four groups targeted by the program. For ex-addicts and ex-offenders, Hendren and Sprung-Keyser (2020) rely on the results in Hollister et al. (1984), who examine impacts up to 27 months following random assignment. This description walks through the MVPF calculation for ex-offenders.

MVPF SUMMARY

_________

Hollister et al. (1984) find a post-training earnings gain of $304 for the ex-offender target group. Hollister et al. (1984) report a program cost of $4,153 including $2,489 of payments to participants. At a tax rate of 12.9% (derived from CBO estimates), the sum of earnings changes and payments to participants correspond to an increase in tax payments of $369. Finally, Hollister et al. (1984) report additional savings to the government via reduced welfare utilization of $219 and reduced administrative costs from non-NSW training programs of $168, as well as reduced payments to participants from these substitute programs of $32.

Hendren and Sprung-Keyser (2020) in their baseline specification assume that participants value the program as the sum of its impact on post-tax income. This implies an MVPF of \frac{304+2,489-369-219-32}{4,153-369-219-168}=0.64.

If participants instead value the program at its cost instead of its impact on post-tax income, the MVPF is \frac{4,153}{4,153-369-219-168}=1.22.

However, in neither of these cases do Hendren and Sprung-Keyser (2020) observe information on the sampling uncertainty of the estimates from Hollister et al. (1984), so we are unable to report confidence intervals on these estimates.