The microcomputer has increased interest in performance testing, which samples what a person can do rather than what he or she knows. Conventional psychometric theory is based on knowledge tests, but in performance testing the unit of analysis is a trial, and it is unreasonable to assume that mean performance and interim correlations are independent of order of administration. For example, performance typically improves with practice. Both reliability and temporal stability frequently encounter optima as a performance test is lengthened. Scoring all trials administered may not yield the best predictive validity; rather, scoring a subset of trials frequently yields higher predictive validities. Subset analysis serves the same ends in performance-test theory as does item analysis in conventional psychometrics. Serial averaging and its applications (reliability and stability optima, optimal scoring for predictive validity, and subset analysis) are explained and illustrated. Results from the Project-A computer-administered tests, served as the database. Ten performance tests were used with two samples of college undergraduates (102 in each) and samples of Army enlisted people ranging from 8,892 to 9,269. Twenty-three references are included, and 12 tables and 11 graphs provide illustrative data. (SLD)