You are here:
INCREASING SCIENTIFIC POWER WITH STATISTICAL POWER
Citation:
Muller, K. AND V. Benignus. INCREASING SCIENTIFIC POWER WITH STATISTICAL POWER. U.S. Environmental Protection Agency, Washington, D.C., EPA/600/J-92/278 (NTIS PB92217199), 1992.
Description:
A brief survey of basic ideas in statistical power analysis demonstrates the advantages and ease of using power analysis throughout the design, analysis, and interpretation of research. he power of a statistical test is the probability of rejecting the null hypothesis of the test. he traditional approach to power involves computation of only a single power value. onsidering the more general power curve not only clarifies the concept, but also substantially increases the utility of power analysis. urther clarification and simplification arise from examination of the variables affecting power. or example, in ANOVA and regression (with Gaussian errors), power depends solely on sample size, population difference, and error variance. ence, one need only consider these three features in the most common analysis. he discussion leads to guidelines for reporting power analysis. raditionally power analysis has been used for study design in a prospective fashion. The greatest value in power analysis likely lies here. ower analysis can also aid in the evaluation of existing research. Retrospective power analysis may be preferred over re-analysis or meta-analysis. or example, power analysis can be used to help conclude that no scientifically important treatment difference exists. ny choice of the appropriate amount of power depends on: 1) opportunity costs, 2) ethical tradeoffs, 3) the size of effect considered important, 4) the uncertainty of parameter estimates, and 5) the analyst's preferences. lthough precise rules seem inappropriate, several guidelines are defensible. irst, the sensitivity of the power curve to particular characteristics of the study, such as the error variance, should be examined in any power analysis. econd, just as a small type I error rate should be demonstrated in order to declare a difference nonzero, a small type II error should be demonstrated in order to declare a difference zero. hird, when ethical and opportunity costs do not preclude it, power should be at least .84, and preferably greater than .90.
