Control of Selection Effects in the Evaluation of Social Problems

Evaluations involving nonrandom assignment to treatment or control groups are vulnerable to an accidental or intentional confounding of a selection effect with the treatment effect. The resulting selection bias is compensated with two techniques, discriminant analysis and base expectancy analysis, which model the selection process that generates the treatment and control groups. These models permit separate estimation of the selection and treatment effects in the final results. These techniques are applied to the evaluation of a college program in a maximum-security prison.