------------ 2016-05-23
Thanks for your kind note after the webinar on Friday.  I am glad that you found it helpful.  It is hard for me to have any sense of how people are reacting when all I can see is the slides on my monitor.

The multi-period design you describe sounds like a stepped wedge design.  Is that right?  I had a few slides on stepped-wedge designs, 82-84, and provided some recent references.  In this design, all groups start out in the control condition, and at regular intervals, one or more groups is switched to the intervention condition, usually in a random order.  By the end of the study, all groups have provided both control observations and intervention observations.  Such designs often require fewer groups, though not dramatically fewer.  We discussed those issues in Rhoda et al., 2011 in the reference list.

The issue of informed consent has received considerable attention in recent years.  The Collaboratory published a paper recently on that issue (Anderson et al., 2015 in the reference list).  In all health-related GRTs, the local IRB must decide whether individual participants are required to give consent, and if so, whether they are required to give written consent.  In most of my school-based health promotion studies, our intervention was adopted as part of the school’s curriculum and so was not subject to consent requirements; however, we were required to obtain consent for measurement activities, though usually passive parental consent.  In studies conducted in health care systems, local IRBs have to judge whether the intervention or data collection activities require more than minimal risk, and is not, they may waive consent requirements.  That has happened in several of the Collaboratory trials, for example.  In other cases, individual consent may be required.

On slide 43, I noted that mixed-model ANOVA/ANCOVA in which more than 2 time points were explicitly modeled are not appropriate.  These “repeated measures analyses” assume that the group-specific trends are homogenous within a condition, and that may not be true.  If it is not true, but you employ that analysis anyway, you will have an inflated type 1 error rate.  We document this issue in a 1996 Stat in Med paper (Murray et al., 1996, in the reference list).  The random coefficients analysis allows each group to have its own trend and intercept, so that heterogeneity may hurt power, but it does not affect the type 1 error rate.

I hope these comments are helpful.

David Murray

***************************************************
David M. Murray, Ph.D.
Associate Director for Prevention
Director, Office of Disease Prevention
david.murray2@nih.gov


---------------------- 2016-05-29

I have not worked on a trial with that kind of cross-over pattern where the unit of assignment is a group or cluster.  In my work, it is rarely possible to remove an intervention once initiated, then reintroduce it later.  So I have no experience with that particular variation.  Because the groups are crossed with study conditions (each would be A at times and then B at times), the impact of the intraclass correlation is greatly diminished.  Scott & Holt (1982) provide the general formula in slide 14 that shows that the DEFF is in part a function of the product of two ICCs.  ICCy is the usual ICC for the dependent variable.  ICCx is the ICC for the exposure or independent variable.  In a GRT, ICCx=1 because all participants in the same group get the same treatment, and only that treatment.  In your case, ICCx would be <1, because the group or cluster is exposed to both treatments.  So the impact of the clustering would be reduced.

If your treatment involves bathing ICU patients with antibiotics, you might want to chat with Susan Huang, PI on one of the Collaboratory Projects.  Here is a recent paper…

1.
Huang SS, Septimus E, Hayden MK, Kleinman K, Sturtevant J, Avery TR, Moody J, Hickok J, Lankiewicz J, Gombosev A, Kaganov RE, Haffenreffer K, Jernigan JA, Perlin JB, Platt R, Weinstein RA, Agency for Healthcare R, Quality DN, Healthcare-Associated Infections P, the CDCPEP. Effect of body surface decolonisation on bacteriuria and candiduria in intensive care units: an analysis of a cluster-randomised trial. Lancet Infect Dis. 2016;16(1):70-9.

They have already dealt with the consent issues and might be of help to you.

On the >2 time point issues, the problem is that we cannot know, in advance, if individuals or groups (or clusters) are on the same trajectory or not.  The more familiar repeated measures models are fine if they are.  But if the groups within a condition are on heterogeneous trajectories, then the random coefficients model should be used.  Details in the Murray et al. 1998 paper.

David