Cluster Randomization

At this month's journal club, we reviewed a study which used "cluster randomization."  This is a study strategy which does basically what it says on the tin: instead of being randomized individually, test subjects are randomized in clusters.  In the case of the study we talked about at journal club, the unit of "clustering" was the individual ICU.

We thought it was worth going into cluster-randomization in a little detail, apropos of a very interesting piece in this month's first issue of Annals of Internal Medicine.  You may recall from Dr. Bhuket's Grand Rounds on cirrhosis earlier this year that the American Association for the Study of Liver Diseases' (AASLD) guidelines on screening for hepatocellular carcinoma in cirrhotics currently recommend ultrasound examination every six months, without measuring alpha-fetoprotein.

The authors point out that this recommendation (which was graded as a level I) is based on a single trial cluster-randomized trial.  In the analysis of their data, however, the authors calculated the 95% confidence intervals of the hazard ratio they used to report their findings as though the patients had been individually randomized.  This is a basic statistical mistake.

This case is instructive for all students of the medical literature, which is so complex and difficult to navigate that even an august body such as the AASLD can be misled, yet so democratic that anyone who understands basic principles of medical statistics can understand where the mistake lies, and potentially be the first to recognize it.