While random-assignment experiments have been considerable conceptual appeal, the validity and reliability of results depends crucially on a number of design and implementation issues. This paper reviews the major experiment in class size reduction – Tennessee’s Project STAR – and puts the results in the context of existing nonexperimental evidence about class size. The nonexperimental evidence uniformly indicates no consistent improvements in achievement with class size reductions. This evidence comes from very different sources and methodologies, making the consistence of results quite striking. The experimental evidence from the STAR experiment is typically cited as providing strong support of current policy proposals to reduce class size. Detailed review of the evidence, however, uncovers a number of important designs and implementation issues that suggest considerable uncertainty about the magnitude of any treatment effects. Moreover, there is reason to believe that the commonly cites results are biased upwards. Ignoring consideration of the uncertainties and possible biases in the experiment, the results show effects that are limited to very large (and expensive) reduction in kindergarten or possible first grade class sizes. No support for smaller reductions in class size i.e., reductions resulting in class sized greater than 13-17 students) or for reductions in larger grades is found in the STAR results.
