Sample size calculations for trials in health services research.

The current orthodox way of estimating sample size for a trial is through a power calculation based on a significance test. It therefore carries the assumption that this test should be the centerpiece of the statistical analysis. However, it is increasingly the case that confidence intervals are preferred to significance tests in summarising the results of trials, particularly in health services research. We believe that the way sample size is estimated should reflect this change and focus on the width of the confidence interval rather than on the outcome of a significance test. Such a method of estimation is described here and shown to have additional advantages of simplicity and transparency, enabling a more informed debate about the proposed size of trials.