The Foundations of Randomized Clinical Trials

• Activity Number: 550
• Type: Invited
• Date/Time: Wednesday, August 1, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Memorial
• Author(s): Marvin Zelen*+
• Companies: Harvard School of Business

Abstract

Consider a multi-center randomized clinical trial. A local inference is defined as conclusions that only apply to the patients who entered the clinical trial; i.e. best treatment for the patients in the trial. A global inference is defined as conclusions that apply to the population with disease; i.e. best treatment for those with disease. Under what circumstances do the different inferences apply? There is a recruitment process for patients and centers in clinical trials. Patients and centers may be separately envisioned to be a random sample from a population of patients and centers or if they are not random samples they are each considered to be a collection of patients and centers. Consequently the recruitment process generates four models for making inferences; i.e. patients and centers may either be collections or random samples. In practice, patients and centers are each collections. As a result inferences in a trial, strictly speaking, are local inferences that may even differ for each center. Alternatively if the patients and centers are both random samples, the inference is a global one. The variance for comparing the average of two therapies is smallest for the local inference and largest for the global inference. Consequently the power for making a global inference will be less than making a local inference. A third inference associated with the model for collections of centers and random samples of patients, within each center, means that the therapies as given in the centers, refer to the treatment for disease, which could possibly be different for each center. Given these considerations if the aim of a trial is to make a global inference, the analysis should consider the appropriate variance associated with this inference, even though both patients and centers are in reality collections. In practice, analyses use variance terms that are appropriate for a local inference, but interpret the trial as a global inference.