Before turning to the evidence on whether work-based programs are likely to have long-term impact on wages, it is important to understand exactly the question being asked and the conditions under which different types of evidence answer this question. The question is much more difficult to answer than is commonly assumed.
The claim that taking a low-wage job will improve future wages is a claim about an actual event — the person's actual outcome — and a hypothetical outcome — what would have happened if the person had taken the job. Clearly, the counterfactual is not observable. The inability to see people in both states (having taken the job and not having taken the job) is the fundamental issue in evaluation studies. If we could only rewind the tape and have the person follow the alternative path, we would have the answer. This is, however, not possible, so the question can only be answered by making assumptions that allow the researcher to compare other people's outcomes to infer what would have happened to the person had he/she followed the alternative path. These "identifying assumptions" are the key to answering the evaluation question.
Randomized experiments require the weakest assumptions. Under this approach, people are randomly assigned to a "treatment" group and a "control" group. A treatment (e.g., job placement) is randomly assigned to a subset of the participants. In a classical experiment, the effect of the treatment is measured by the difference in outcomes between the treated and the control groups. The only assumption necessary is that the treated group would have had the same outcomes as the control group had they not received the treatment. This assumption is justified by random assignment. If people are truly assigned randomly to the treatment and control groups, then the only systematic difference between the two groups is the treatment.(1) Therefore, differences in outcomes between the two groups reflect the pure effect of the treatment, not unobserved differences in outcomes that would have taken place even without the treatment.(2)
The second method of measuring the impact of early labor market experience on future outcomes is to use nonexperimental data. For example, entry-level jobs may be more plentiful in some areas than others. The resulting variation in early work experience, which may affect future wages, potentially takes the place of random assignment. It is, however, now necessary to assume that this nonexperimental assignment is random. If people with skills live in areas with more plentiful jobs or higher wage growth, then it is not possible to separate the effects of having obtained a job from the effects of having more skills.(3) Therefore, if one were to tabulate the adult wages of persons classified by their initial employment status, or to estimate regressions with adult wages as the dependent variable and initial employment status as one of the independent variables, one would implicitly be assuming that initial employment status was random. However, more motivated people are more likely to find jobs initially and are more likely to receive high wages later in life. Comparing their wages with the wages of less motivated people who did not have the early labor market experience tells us little about the wages the latter group would have received had they landed a similar job.(4)
While nonexperimental studies have these drawbacks, they may still be useful if they can be used to obtain bounds on the possible future effects of work on the target population. For example, it might be reasonable to assume that nonworkers would have gained less from early labor market experience than those who did work. In this case, nonexperimental evidence would give us an upper bound on what the nonworkers would have gained from work had they had early market experience. For example, knowing the impact of work experience on future wages for the welfare mothers who did work may give us an upper bound on the potential gains from work for those who did not work.