Previous considerations of power analysis for common designs have involved separate approaches for each design and test. More recent work has extended the usual power analysis framework to include the t-tests for continuous symmetric responses, logistic regression models for binary, nominal and ordinal responses, Poisson regression for count responses, as well as other models for response with high degrees of skewness, into a single class of analyses based on likelihood ratio tests. Knowledge of the available power estimates for these designs will allow greater freedom in choosing appropriate responses for clinical studies. Much work has also been done on developing a power framework for multivariate and clustered designs, such as repeated measures and multilevel designs, and to describe sample size approaches for Bayesian statistical methods.

In this presentation, we will describe these developments in the context of nursing research studies. We will briefly demonstrate how sample size is calculated for a simple two-sample t-test, and extend this via the generalized linear model framework to comparisons of groups on count responses. Next, we will extend these univariate comparisons to repeated measures designs, and to a complex study where repeated measures are nested within counseling groups. We will compare the Bayesian and classical approaches for this complex design. These demonstrations will describe how the power analyses should be described in a grant protocol, and how to investigate the sensitivity of these sample size estimates to changes in the assumptions and the estimates on which they are based.