Citation

Atkinson G. Sportscience 2007; 11(online): 12-15.

Affiliation

Research Institute for Sport and Exercise Sciences, Liverpool John Moores University, Liverpool L3 2ET, UK

Copyright

(Copyright © 2007, Internet Society of Sports Science, Physiology Department, Otago Medical School)

DOI

unavailable

PMID

unavailable

Abstract

In an applied field like sport and exercise science, inferences based on estimation of true effect sizes are usually more important than inferences about statistical significance. Inferences about estimation are conventionally made using confidence intervals, which are associated with several critical judgments. The most important decision concerns the smallest effect size that is practically or clinically important. A recently published new approach to sample size estimation also raises issues of judging the appropriate coverage probability of a confidence interval (e.g. 90 or 95%) as well as the degree of overlap between confidence limits and the smallest worthwhile effect. It is these a priori rationalized decisions that underpin the mathematics of confidence intervals, the probabilistic inferences made from them and associated issues like sample size estimation and claims that a statistical approach is too conservative or liberal. First, I discuss that the “null” in the null hypothesis testing process does not always need to be set at zero. If the smallest worthwhile effect itself is selected as the null value, then this process not so isolated from practical significance. Second, I contrast ideas on boundaries of overlap between confidence limits and the smallest worthwhile effect with other published guidelines on using confidence intervals to interpret study results. It is these differences in delimited probability coverage that govern the apparently lower sample sizes required for the new approach. Third, I illustrate how critical the decision on smallest worthwhile effect size can be for accuracy of study conclusions, and question whether uncertainty in this decision process might, in some instances, compromise the accuracy of the inferential statements that are made following statistical analysis. KEYWORDS: confidence intervals, null hypothesis, Type I and II statistical errors, smallest worthwhile effect.