Main Menu Browse by Day or Program Author Keywords

411 All crossed up? Using cluster-randomized crossover clinical trials in healthcare epidemiology

Sunday, April 3, 2011
Trinity Ballroom (Hilton Anatole)
Nicholas G. Reich, PhD , Johns Hopkins Medical Institutions, Baltimore, MD
Aaron M. Milstone, MD, MHS , Johns Hopkins Medical Institutions, Baltimore, MD
Trish M. Perl, MD, MSc , The Johns Hopkins Hospital, Johns Hopkins University Bloomberg School of Public Health, Baltimore, MD
Background: Gathering evidence about effective clinical interventions in healthcare settings poses scientific and logistical challenges. Increasingly, medical researchers rely on complex study designs to learn about and compare interventions. Cluster-randomized crossover (CRXO) clinical trials are one example of such a design. Multi-level models are often used for analyzing data from CRXO studies, although little work has been done to verify that these models produce unbiased estimates of the intervention effect, especially for small sample or varying cluster sizes. Little theoretical work has been done to evaluate how and when the CRXO study design is an appropriate design to answer a complex clinical question.

Objective: To summarize the pros and cons of a CRXO design in healthcare epidemiology settings. To review statistical methods used for analyzing CRXO trials.

Methods: We conducted a systematic review of CRXO studies in the published medical literature, paying particular attention to the types of outcome measures and the statistical methods used. Characteristics of the studies are presented in tabular form. The observed methods are written down in a standardized, generalizable format and compared with recommended methods (when such recommendations exist).

Results: Both individual-level and cluster-level analyses have been used for analyzing data from CRXO studies. There is not a standard method for analyzing binary or count outcome data from CRXO studies.  Methods used for this type of data have not been validated by formal statistical studies to evaluate their effectiveness. We propose a standardized method for analyzing binary or count outcome data from CRXO trials.

Conclusions: The CRXO design takes advantage of within-cluster comparisons to improve statistical efficiency.  However, crossover studies take longer to complete than a simple cluster-randomized study, and they introduce other variables including a study-period effect. Our proposed methods for analyzing binary or count data from CRXO trials fill in an important methodological gap in healthcare epidemiological research methods.