Estimating the Effect of Training Programs with Non-Random Selection

This study reports on efforts to use the Continuous Longitudinal Manpower Survey to estimate the effect that manpower training programs have had on participants' earnings. Estimation techniques are developed to control for nonrandom selection into the program based on individual unobservables which are either constant and/or changing over time, as well as non-random selection because of "creaming" by program administrators. The study finds that, in general, fixed effects estimators are sufficient to eliminate the bias created by non-random selection. While women appear to benefit substantially from manpower training programs, no significant earnings effects were found for men participants. G OVERNMENTAL involvement in postschooling employment and training programs for disadvantaged and hard to employ workers began with the passage of the Manpower Development and Training Act of 1962. Although funding levels and objectives have changed frequently, training programs of one type or another have been an increasingly significant factor in the operation of the U.S. economy since that time. For almost a decade the Comprehensive Employment and Training Act of 1973 (CETA) had major responsibility for provision of training, serving over four million participants during 1979.1 The purpose of training programs, as defined by CETA, was "to provide training and employment opportunities for economically disadvantaged, unemployed, or underemployed persons which will result in an increase in their earned income."2 Despite nearly a decade of massive expenditures on training under CETA, surprisingly little is known about the program's effectiveness in achieving its goal of human resources development. This study reports on efforts to use the recently available Continuous Longitudinal Manpower Survey (CLMS) to estimate the effect that CETA has had on participants' post-training earnings.3 The plan of the paper is as follows. Section I develops a range of estimation techniques, given alternative assumptions about the structure of the error term. A series of nested hypotheses tests is suggested as a method for choosing which estimation technique(s) fulfills the necessary assumptions for unbiased estimation of training effects. Section II describes the data that are available from the CLMS. The framework developed in section I is then used to test which of four available comparison groups are appropriate for unbiased estimation and presents the empirical results and their limitations. Section III summarizes the findings of