Essays in Social Experimentation: What Experiments are Needed for Manpower Policy?
指出，当项目效果因人而异时，随机实验只能估计平均效果而非边际效果，而后者对政策更重要，并提出了能同时估计两者的替代实验设计。
Experiments, where prospective participants are randomly assigned into experimental and control groups, are often regarded as the ideal approach to evaluation of manpower policies. This examination shows that such experimental designs can yield misleading and incomplete information about program impacts in cases where program effects vary among individuals. When individual heterogeneity is present, such designs at best provide unbiased estimates of average program effects, not of marginal effects. The latter are often more important for policy than the former. This paper proposes alternative experimental designs that permit estimation of the marginal as well as average program impacts.