When Capacity Is Scarce: Targeting, Experts, and Fairness

Once a provider accepts that Mentoring and Challenges fit different people, a second problem appears: there may not be enough places for everyone. Dare IT’s oversubscribed programs turn matching into an allocation question. This chapter explains why the person most likely to get a technology job is not necessarily the person who benefits most from a program, how the study’s 86% and 23% targeting results should be interpreted, and why a data-assisted rule must still declare its objective and limits.

Allocation sounds technical, but the underlying distinction is ordinary. Imagine two applicants. One is likely to find a job without help; the other is less likely to succeed alone but could change substantially if given the right intervention. If a provider selects only the first applicant, the program may report strong outcomes among participants while creating little additional employment. If it selects the second, the measured improvement may be larger even though the participant’s final employment probability remains lower. A useful allocation rule must ask not only who will succeed, but who will be changed by access.

The counterfactual behind “benefit”

The relevant quantity is a treatment response: the difference between what would happen to an applicant with a program and what would happen to that same applicant without it. The second outcome cannot be observed for the same person at the same time, so researchers estimate it from randomized comparisons and models. That unseen alternative is the counterfactual.

This is different from predicting the participant’s final outcome after treatment. An applicant may have a high predicted chance of employment with Mentoring because they already have strong credentials, a dense network, and good local opportunities. But if they would probably find a job without Mentoring, the program’s incremental effect may be modest. Another applicant may have a lower final probability but a much larger gap between the untreated and treated scenarios because the program addresses a severe network barrier.

The Dare IT study exposes this distinction through mentor selection. Mentors selected applicants with higher baseline employment prospects than non-selected applicants. That is useful evidence of judgment: mentors were not choosing randomly. But mentor selection was not identical to selecting people with the largest treatment response. A selection process can be good at spotting promising candidates while still leaving room to identify people for whom the intervention adds the most.

The study therefore compares several objectives. One can select the people most likely to succeed without treatment, the people least likely to succeed without treatment, the people predicted to have the best outcome after treatment, or the people predicted to gain the most from treatment. These objectives answer different policy questions. The study finds that the treatment-effect objective performs better for maximizing measured program impact than simply choosing on baseline or post-treatment predictions.

That result should not be reduced to “always prioritize the highest predicted gain.” A provider may have an equity objective that gives additional weight to people facing the greatest disadvantage.