Causal effects in non-experimental studies: Re-evaluation of training programs

Rajeev Dehejia and Sadek Wahba
Journal of the American Statistical Association, Volume 94, Number 448 (December 1999), pp. 1053-1062.

The need to use randomized experiments in the context of manpower training programs, and in analyzing causal effects more generally, has been a subject of much debate. Lalonde (1986)considers experimental data from the National Supported Work (NSW) Demonstration and nonexperimental comparison groups drawn from the CPS and PSID, and argues that econometricmethods fail to replicate the benchmark experimental treatment effect. This paper applies propensity score methods, which have been developed in the statistics literature, to Lalonde'sdataset. In contrast with Lalonde's findings, using propensity score methods, we are able closely to replicate the experimental training effect. The methods succeed because they are able flexibly to control for the wide range of observable differences between the (experimental) treatment group and the (non-experimental) comparison group.

Wagner Faculty