Design effects for highly clustered count data with varying length of follow-up--applying to the analysis of ICD shocks.

Highly clustered count data are commonly seen in medical device clinical studies such as cardiac rhythm management. For instance, anti-arrhythmia shocks delivered from an implantable cardioverter-defibrillator (ICD) often occur as "storms", i.e., multiple shocks within a short period of time. There are unique challenges in designing and analyzing clinical studies with such response data. In general, there is less amount of data information available when events occur as "storms" as compared to independent events, and the efficiency loss may be substantial in practice. In addition, event incidences of subjects may be highly unbalanced and length of follow-up may vary from subject to subject. In this paper, the design effect for such highly clustered and unbalanced count data is studied. Varying length of subject follow-up is taken into account by the use of a random variable that may depend on external covariates. Sample size adjustment for event correlation and imbalance is discussed. The practical application of the strategy is illustrated through a clinical study of ICD shock therapy.