Counting recurrent events in cancer research.

In clinical research, outcomes of interest often recur in the same patient. Examples include asthma attacks, urinary-tract infections, migraines, injuries, seizures in epileptics, and admissions to a hospital. Examples in cancer research include recurrent mammary tumors (1), skin cancers (2), and fractures in patients with cancer metastatic to bone (3). However, appropriate statistical methods to compare the occurrence of recurrent events between treatment groups are not widely understood or are they broadly disseminated to the research community. In this issue of the Journal, Cook and Major (4) illustrate the errors that can occur with inappropriate analytic strategies and propose some principled alternatives. In the setting of recurrent events, research studies commonly count only the first occurrence of an outcome in a subject. For example, in the first publication of data from the pamidronate trial in women with breast cancer (5), used by Cook and Major to illustrate their methods, the primary analysis compared the time until a first important skeletal complication between the active and placebo groups. This analytic strategy is straightforward and unbiased, and it avoids the methodologic complications that can occur if a first event affects either risk of a subsequent event or compliance with ongoing randomized treatment in the trial (6). For example, occurrence of a first myocardial infarction markedly increases the risk of a subsequent event and can lead patients in the placebo arm of a randomized trial to initiate active therapy (7). However, consideration of only first events may lead to an inaccurate evaluation of the efficacy of a treatment. In particular, it can substantially underestimate potential benefits in terms of events prevented by a treatment. In a variety of settings, ranging from studies of health care utilization (8) to recurrent falls in high-risk persons (9), an appropriate analysis of recurrent events can yield clinical insights and an understanding of public health impact that would be missed in an analysis of only first events. Furthermore, studies that discard relevant information on subsequent outcomes can require far greater sample sizes than studies that appropriately use all available outcomes on study subjects. The major challenge to an appropriate analysis, including multiple events per person, arises because repeated events in the same person are almost never independent of one another. As noted by Cook and Major, the usual analysis of the “events per person-year” method assumes that all events occur randomly within each treatment group. More than 80 years ago, Green-