This paper by Rajeev H. Dehejia and Sadek Wahba, published as NBER Working Paper 6829, discusses the use of propensity score matching methods in non-experimental causal studies. The authors address the challenge of estimating treatment effects in settings where the comparison group is not well-matched to the treatment group, and where selecting a subset of similar comparison units is difficult due to a high-dimensional set of pre-treatment characteristics. They propose and implement these methods using data from the National Supported Work (NSW) experiment, comparing their estimates to benchmark results from the experimental treatment group. The paper makes three main contributions: (1) it discusses and extends propensity score matching methods, which are new to the economics literature; (2) it highlights the importance of comparability between the treatment and control groups in terms of observable characteristics; and (3) it demonstrates that these methods can produce accurate estimates of treatment impact even when there are very few comparable comparison units. The authors use the NSW experiment, where participants were randomized into treatment (on-the-job training) and control groups, to set a benchmark estimate of the treatment effect. They then pair experimental treated units with non-experimental comparison units from the Consumer Protection Survey (CPS) and the Panel Study of Income Dynamics (PSID). By comparing the estimates obtained using their methods to the experimental benchmark, they show that their approach successfully focuses attention on the small subset of comparison units that are comparable to the treated units, thereby reducing bias due to systematic differences between the groups. The paper also explores the limitations of standard matching approaches and proposes algorithms to address issues of incomplete matching. It discusses the use of the propensity score to reduce the dimensionality of the matching problem and the importance of choosing an appropriate distance metric for matching. The authors conclude that their method is valuable for constructing suitable control groups in non-experimental settings, as it effectively highlights the need to adjust for observable differences between the treatment and control groups.This paper by Rajeev H. Dehejia and Sadek Wahba, published as NBER Working Paper 6829, discusses the use of propensity score matching methods in non-experimental causal studies. The authors address the challenge of estimating treatment effects in settings where the comparison group is not well-matched to the treatment group, and where selecting a subset of similar comparison units is difficult due to a high-dimensional set of pre-treatment characteristics. They propose and implement these methods using data from the National Supported Work (NSW) experiment, comparing their estimates to benchmark results from the experimental treatment group. The paper makes three main contributions: (1) it discusses and extends propensity score matching methods, which are new to the economics literature; (2) it highlights the importance of comparability between the treatment and control groups in terms of observable characteristics; and (3) it demonstrates that these methods can produce accurate estimates of treatment impact even when there are very few comparable comparison units. The authors use the NSW experiment, where participants were randomized into treatment (on-the-job training) and control groups, to set a benchmark estimate of the treatment effect. They then pair experimental treated units with non-experimental comparison units from the Consumer Protection Survey (CPS) and the Panel Study of Income Dynamics (PSID). By comparing the estimates obtained using their methods to the experimental benchmark, they show that their approach successfully focuses attention on the small subset of comparison units that are comparable to the treated units, thereby reducing bias due to systematic differences between the groups. The paper also explores the limitations of standard matching approaches and proposes algorithms to address issues of incomplete matching. It discusses the use of the propensity score to reduce the dimensionality of the matching problem and the importance of choosing an appropriate distance metric for matching. The authors conclude that their method is valuable for constructing suitable control groups in non-experimental settings, as it effectively highlights the need to adjust for observable differences between the treatment and control groups.