Power and sample size for cluster randomized and stepped wedge trials: Comparing estimates obtained by applying design effects or by direct estimation in GLMM

TL;DR

This paper compares two methods for calculating power and sample size in clustered and stepped wedge trials: using design effects with standard software versus direct estimation with GLMM, highlighting their similarities and differences.

Contribution

It provides a detailed comparison of two approaches for power estimation in clustered designs, illustrating their application and differences through examples.

Findings

Both methods generally produce similar estimates, validating each other.

Small differences in power estimates highlight the importance of specifying hypotheses carefully.

Power estimation is sensitive to assumptions about correlation structures.

Abstract

When observations are independent, formulae and software are readily available to plan and design studies of appropriate size and power to detect important associations. When observations are correlated or clustered, results obtained from the standard software require adjustment. This tutorial compares two approaches, using examples that illustrate various designs for both independent and clustered data. One approach obtains initial estimates using software that assume independence among observations, then adjusts these estimates using a design effect (DE), also called a variance inflation factor (VIF). A second approach generates estimates using generalized linear mixed models (GLMM) that account directly for patterns of clustering and correlation. The two approaches generally produce similar estimates and so validate one another. For certain clustered designs, small differences in power estimates emphasize the importance of specifying an alternative hypothesis in terms of means but also in terms of expected variances and covariances. Both approaches to power estimation are sensitive to assumptions concerning the structure or pattern of independence or correlation among clustered outcomes.