Abstract:
|
When choosing a sample size for a confidence interval, we describe how to exactly account for: I.) using a variance estimate from a previous study; and ii.) the truncation that results if the new study is planned only because the previous study achieved significance (or if the new study is planned only because the previous study did not achieve significance). We extend the methodology of Jiroutek, Muller, Kupper, and Stewart (2002, in review), who derived new methods for choosing a sample size to compute well-behaved confidence intervals for scalar parameters in the General Linear Multivariate Model with Gaussian errors. Their approach centers on the simultaneous consideration of hypothesis testing and confidence interval properties, which leads to the improved alignment of sample-size calculations with study objectives. Simple examples are presented to illustrate the substantial bias that may result by failing to account for either issue. The results parallel previous work for power calculations (Taylor and Muller, 1995; Muller and Pasour, 1997).
|