This paper presents propensity score matching methods for non-experimental causal studies. The authors propose using these methods to estimate treatment effects in observational studies where the treatment group differs significantly from the control group. They use data from the National Supported Work Demonstration (NSW) experiment to demonstrate the effectiveness of their methods. The NSW experiment involved randomly assigning participants to treatment (on-the-job training) or control groups. The authors compare their non-experimental estimates to the experimental benchmark results from the NSW study. They show that their methods successfully identify a subset of control units comparable to the treatment units, thereby reducing bias due to systematic differences between the groups. The paper discusses the theory behind propensity score matching, including the role of randomization in causal inference and the use of the propensity score to reduce the dimensionality of the matching problem. The authors propose algorithms for propensity score matching, including exact matching on covariates and methods for handling incomplete matching. They also discuss the use of replacement in matching, which allows for more accurate estimates when there are few comparable control units. The authors analyze data from the NSW experiment, using two non-experimental control groups from the Current Population Survey (CPS) and the Panel Study of Income Dynamics (PSID). They compare the treatment effects estimated using their methods to the experimental benchmark. They find that their methods produce accurate estimates of the treatment effect, particularly when there are few comparable control units. The paper concludes that propensity score matching is a valuable tool for causal inference in non-experimental studies, especially when there are few comparable control units. The authors also highlight the importance of checking the comparability of treatment and control units in terms of pre-treatment characteristics.This paper presents propensity score matching methods for non-experimental causal studies. The authors propose using these methods to estimate treatment effects in observational studies where the treatment group differs significantly from the control group. They use data from the National Supported Work Demonstration (NSW) experiment to demonstrate the effectiveness of their methods. The NSW experiment involved randomly assigning participants to treatment (on-the-job training) or control groups. The authors compare their non-experimental estimates to the experimental benchmark results from the NSW study. They show that their methods successfully identify a subset of control units comparable to the treatment units, thereby reducing bias due to systematic differences between the groups. The paper discusses the theory behind propensity score matching, including the role of randomization in causal inference and the use of the propensity score to reduce the dimensionality of the matching problem. The authors propose algorithms for propensity score matching, including exact matching on covariates and methods for handling incomplete matching. They also discuss the use of replacement in matching, which allows for more accurate estimates when there are few comparable control units. The authors analyze data from the NSW experiment, using two non-experimental control groups from the Current Population Survey (CPS) and the Panel Study of Income Dynamics (PSID). They compare the treatment effects estimated using their methods to the experimental benchmark. They find that their methods produce accurate estimates of the treatment effect, particularly when there are few comparable control units. The paper concludes that propensity score matching is a valuable tool for causal inference in non-experimental studies, especially when there are few comparable control units. The authors also highlight the importance of checking the comparability of treatment and control units in terms of pre-treatment characteristics.