Sample size adjustment to maintain power in cluster randomized trials

Abstract

Adequately powered sample size calculation for cluster randomized trials primarily depends on the primary
outcome variable's distribution, effect size, average clulster size an-d intra-cluster correlation (ICC) estimates.
Furthermore, the ICC estimate depends on the outcome variable's distribution, cluster size, and number of clusters.
Researchers ofteni cdesign cluster trials basecd on ICC estimates fromn previous trials or from-l a simulation. At
the end of a trial, the distribution of the outcomne variable changes due to thie intervention in the intervention
grotup. Furthermore, since the ICC estimnate depends on the outcome variable's distribution, the ICC estimate will
often. change at the end of a trial. This change in the IC_C estimate effects the samiiple size requiirement and power
for the trial. If the ICJC estimate at the end of the trial is less than the ICC used for samiiple Si7e calculation then
the trial will be over powered. On the other hand, if the ICC estimilate at the end of the trial is greateif than the
ICC usect for sample size calculation then- the inferences will be based orn an urnder powered trial. Therefore, we
need to adjust for this predicted ICC change during the desigin phase when calculatin-g the requiredl sarnple size.
We conducted a sitnulationi trial by varying the outcome variable proportions, cluster sizes, and number of available
clusters to determine the chanige in ICC estimates and contidence intervals for ICC estimates for different
scenarios. Simulation results suggest adjustinig the sample size during the design phase for potential changes to
the ICC estimate at the end of the study in order to preserve the appropriate power- tand in some situations save
trial cost. Two differenit trial examples were usect to demIolnstrate differenit tvpes of samnple size adtjustments during
thie desiginl phiase.