Nonresponse Issues in Public Policy Experiments, with Emphasis on the Health Insurance Study

Missing data problems, including nonresponse problems, are troublesome because the sampling probabilities are unknown, and cannot be determined even with large samples. Judgment must be used in the analysis and should be reflected in expressions of uncertainty. During the design of a study, many choices must be made between collecting more observations or devoting resources to reducing nonresponse or estimating its effects. Mean square error appears to be a more useful concept in making these tradeoffs than unbiasedness. Dual data collection systems are helpful for measuring nonresponse.