The Reed Institute for Applied Statistics

Speaker Series

Prospective sample-size analysis is invaluable to research design, promoting wiser allocation of scientific resources and stronger ethics. Moreover, the process itself induces excellence and breadth in scientific planning by requiring the research team to delineate, critique, and tighten the research questions, study rationale, and many aspects of study design, including outcome measurements and analysis plans. Critically, the team must make reasonable conjectures about the "infinite datasets" representing the study populations. This is supported by ever-improving methods and software for computing power and/or required sample sizes under multiple plausible scenarios. In contrast, ritualistic sample-size and power computations, promulgated with little consideration of scientific context, are an empty exercise at best.

This tutorial will demonstrate the use of PROCs POWER and GLMPOWER (new in SAS 9), augmented with other SAS modules/macros (and a little Excel program) developed and distributed free by the presenters. In-depth examples will illustrate how to use this software, all within the context of the science at hand. This involves, in part, (1) positioning a study in a line of scientific investigation (early to middle to late in "The March of Science"); (2) sizing a study for precision of an statistical interval or power of a conventional hypothesis test; (3) considering positive and negative inference mistake rates (false discovery and false miss rates) rather than traditional Type I and II error rates; and (4) communicating power and sample size concepts and results to non-statistician investigators. Although the methods covered here will be frequentist, basic Bayesian concepts help to clarify and shape the planning process for both investigators and statisticians. Non-SAS users should have no difficulty applying these notions to other software systems.

Prerequisites: Familiarity with basic hypothesis testing and confidence intervals. Prior exposure to SAS will help, but is not necessary. This tutorial will be aimed at those with little exposure to modern sample-size analysis methods and strategies, a topic that is essential for many working statisticians, but is typically given short shrift in statistics education.