Adjustments for Center in Multicenter Studies: An Overview

The inattention of published studies to departures from the assumption behind most statistical methods, the independence of observations, is overlooked, and investigators need to identify centers, incorporate the concept of centers into their designs, estimate the design effect, and adjust confidence intervals and P values appropriately.

Abstract

Increasingly, investigators rely on multicenter or multigroup studies to demonstrate effectiveness and generalizability. Authors too often overlook the analytic challenges in these study designs: the correlation of outcomes and exposures among patients within centers, confounding of associations by center, and effect modification of treatment or exposure across center. Correlation or clustering, resulting from the similarity of outcomes among patients within a center, requires an adjustment to confidence intervals and P values, especially in observational studies and in randomized multicenter studies in which treatment is allocated by center rather than by individual patient. Multicenter designs also warrant testing and adjustment for the potential bias of confounding by center, and for the presence of effect modification or interaction by center. This paper uses examples from the recent biomedical literature to highlight the issues and analytic options.