Content area
Full text
Introduction
The “Guideline for the Content of Statistical Analysis Plans in Clinical Trials,” published in 2017 and referred to here as the 2017 SAP guidelines, outlines a minimum set of 32 items that should be included in the statistical analysis plans (SAPs) for clinical trials [1]. To accommodate early phase (phase I and non-randomized phase II) clinical trials, the guideline was extended in 2022, modifying 25 of the initial items and adding a further 11 items [2]. Although it is not explicitly stated, these guidelines are generally focused on clinical trials in which the units of randomization are individual patients.
The use of cluster randomization, in which the unit of randomization is a cluster rather than an individual, has been steadily increasing over the past decades [3]. Common reasons for cluster randomization include the evaluation of complex interventions when it is infeasible to randomize individuals, the need to simplify trial processes, the need to avoid within-cluster contamination, or when the objectives pertain to the cluster level [4, 5, 6–7]. Cluster randomized trials (CRTs) are known to have many additional complexities in their design, execution, and analysis, compared to individually randomized designs [5, 6, 8, 9]. They also have unique reporting requirements: the CONSORT guidelines for reporting parallel arm randomized trials were extended to cluster randomized designs in 2004 and updated in 2012 [10]. Additional extensions were later developed to accommodate novel cluster-randomized designs: specifically, the stepped wedge cluster randomized trial (SW-CRT) in 2018 [11, 12] and the cluster randomized cross-over design in 2024 [13].
Some of the key analytical considerations for cluster randomized trials are outlined in Table 1.. Most importantly, clustering should always be allowed for in the statistical analysis using one of several available methods [14]. Additionally, when the number of clusters is small—less than about 40—a “small sample correction” is usually needed to avoid biased estimation of the standard errors [14]. Furthermore, cluster randomized trials often need to recruit participants post-randomization, which increases the risk of imbalances between the arms at baseline [15, 16, 17–18]. This may then necessitate statistical adjustment in the primary analysis, for example, using direct covariate adjustment or a propensity score approach [19, 20]. There are numerous other complexities, including, according to the CONSORT extension for CRTs, the...