Mixed Up: Modeling for Context

Hendricks Brown, PhD
Northwestern University Feinberg School of Medicine

Daniel Almirall, PhD
University of Michigan

Robert Gibbons, PhD
University of Chicago

Don Hedeker, PhD
University of Chicago, Public Health Sciences

Carlos Gallo, PhD
Northwestern University Feinberg School of Medicine

Naihua Duan, PhD
Columbia University

ABSTRACT:
This presentation provides a background into design and analysis of interventions or implementation strategies that are initially randomized, then afterwards are conducted in group or network settings where the units randomized can no longer be treated as independent. Such designs include individually randomized group assigned trials, where the group context is an active ingredient in delivering one arm of the trial. Also included are implementation trials that involve formal learning collaboratives where the sites interact with one another. A wide variation of such designs occur, including trials with rolling entrances and exits to groups, network based interventions, and so-called rollout trials. It is important to take into account such non-independence in analysis, because otherwise the critical values ordinarily used in test statistics are too small and therefore erroneously finding significance more often than they should. Examples are given in multiple contexts, and appropriate statistical procedures are given. To increase appropriate statistical testing, we provide tools to conduct such analyses across different statistical platforms. A shiny R program that accounts for some of these procedures is demonstrated.