Prediction and Statistical Overkill Revisited

作者：Douglas E. Miller, Joseph T. Kunce · 发表于：Measurement and Evaluation in Guidance · 年份：1973 · DOI：10.1080/00256307.1973.12022590 · 被引用次数：130 · 研究领域：Efficiency Analysis Using DEA、Advanced Causal Inference Techniques

An empirical study of statistical overkill investigated the generalizability of multiple regression equations as a function of the subject/variable ratio. Data from small, medium, and large samples of clients selected from one state's 1968 rehabilitation program were used to develop the equations. Data from three large samples of rehabilitation clients in an adjacent state's program were used to evaluate the generalizability of the original equations. Twelve client background variables were used to predict a criterion of client's salary. Findings showed that equations developed on samples with less than a 10 to 1 ratio failed to generalize. Inferences of the effects of a low subject/variable ratio on results of other statistical procedures were made. In addition, comparable predictive efficiency was found between the use of powerful and simplified methods.