Estimation of Causal Effects from Observational Study of Job Training Program

In the social sciences, researchers are often interested in measuring the effect of a treatment or intervention. For example, many labor economists are interested in measuring the effect of job training programs on increasing salaries or reducing unemployment. Measuring the ”causal effect” of an intervention is relatively straight forward when the treatment is given to randomly selected individuals. In this case the difference in the average outcome of the treatment and control groups can be directly compared. Unfortunately in many policy relevant situations economists do not have data from a randomized controlled experiment. Instead they have data on the characteristics and outcomes of a set of individuals who received the treatment, but no directly comparable control group. We tried two approaches to estimating the effect of a job training treatment on wages. The first approach predicts counterfactual wages for the treated individuals based on models trained on a large sample from the general population. The second approach matches the treated individuals to people from the general population based on propensity scores. We find that the second approach yields more plausible estimates of the effect of job training.