Using cluster analysis in program evaluation

The conventional way to measure program impacts is to compute the average treatment effect; that is, the difference between a treatment group that received some intervention and a control group that did not. Recently, scholars have recognized that looking only at the average treatment effect may obscure impacts that accrue to subgroups. In an effort to inform subgroup analysis research, this article explains the challenge of treatment group heterogeneity. It then proposes using cluster analysis to identify otherwise difficult-to-identify subgroups within evaluation data. The approach maintains the integrity of the experimental evaluation design, thereby producing unbiased estimates of program impacts by subgroup. This method is applied to data from the evaluation of New York State's Child Assistance Program, a reform that intended to increase work and earnings among welfare recipients. The article interprets the substantive findings and then addresses the advantages and disadvantages of the proposed method.