Regression Analysis of Factorial Designs with Sequential Replication

摘要

A heuristic sequential procedure for determining the number of replications per factor-level combination, such that the variances of the average responses become approximately constant is examined. The resulting least squares estimator B remains unbiased. Confidence intervals around B (two-sided intervals) are not exact, even if the pilot sample size is as high as 25. These intervals are too tight if the standard normal variable is used. The intervals can become conservative, if the Student statistic is used.