Was this in your Statistics Textbook? III. Design and Analysis

Abstract

SummaryThat experimental design is fundamentally more important than the formal analysis of variance of results is emphasized. Attention is drawn to the many weaknesses of split plot designs, and the case for adopting other types of confounding whenever possible is stated. Split plots are sometimes necessary but rarely desirable. The need for randomization is stressed, with a warning against abandoning this principle before fully considering the dangers. The error in analysing repeated observations as though they are from split plots is discussed at length. After brief comment on variance heterogeneity, a final section looks at how measures of the precision of treatment means should be presented. The two main points made are that standard errors of treatment means should always be included in tables of results, and that existing multiple range significance test procedures have little or no practical utility.