Measures of morbidity in clinical studies with recurrent skeletal complications.

Double-blind placebo-controlled clinical trials in metastatic bone disease considered counts of morbid events as the clinical outcome measure. These (recurrent) skeletal events were derived from a composite endpoint based on the necessity of medical interventions against bone pain or incident fractures. The trials were conducted for regulatory approval of therapy with bisphosphonates and were intended to demonstrate that an active treatment reduces the occurrence of skeletal events. The advanced morbidity of the patients led to a substantial amount of premature discontinuations, because of early death or because of progression of the basic disease. In some trials the discontinuations were unbalanced between treatment arms, where there were more dropouts under placebo. These effects contribute to a high portion of patients with no morbidity events. This background presents difficulties for the task of defining statistics that validly quantify skeletal morbidity in these trials while keeping the necessary simplicity so as to be accepted by health authorities and a clinical audience. Guided by this regulatory context, this paper compares the performance of five (simple) approaches to quantifying skeletal morbidity. Key criteria are bias and validity under different dropout scenarios, as well as the capability to demonstrate treatment effects. Based on theoretical considerations, two simulation studies, and real data results, a morbidity measure in the form of a smoothed rate was favored: this is a rate estimate in the individual patient of the form (y + c)(T + d)(-1), where y is the count of events, T is the patient's observation time, and c, d are constants to be appropriately chosen. This morbidity measure performed well in the simulation studies and showed a wanted insensitivity to dropout patterns. It outperforms the simple rate y/T due to the high portion of patients without events. Rates of this type were used before in the quantal bioassay and have a strong relationship to Bayesian approaches. Hence, these measures represent a good approach to handle the inherent difficulties of morbid event data in severely diseased patient populations, still offering sufficient simplicity.