Lately, I have been reading up about CRXO designs and they strike me as interesting because of the increased efficiency and clinical investigators like it because it’s logistically easier. The most common CRXO design is the one with 2 treatments/arms (A, B) and two periods AKA the AB/BA design. However, what interests me even more is the idea of a cross-sectional, multi-period single cluster design of the sort ABBAAB[…] with the cluster being randomized to a treatment at the beginning of each time period. The absence of a concurrent control group makes this design tricky, however intuitively it appears to me that the use of multiple (randomized) crossovers perhaps allow us to make a strong inference. Are there any useful references about analyzing/designing such a study, and what kind of inferences can be plausibly made from it? The most relevant paper I have found is: https://www.bmj.com/content/371/bmj.m3800.long but the authors say: “Here, we do not consider the special case of studies in one cluster.”
Since in a given period every subject receives the same treatment when there is only one cluster, the design is confounded with secular trends. I think you would need individual subject-level randomization (i.e., a traditional crossover study) to really pull this off. It may be worth looking at the N-of-1 literature to get a different perspective.
Sounds problematic allowing randomisation at each time-point, rather than a more typical structed design.
It would be very problematic if a cluster was allowed to be randomised to AAAAAAA!
Thanks! The pointer to n-of-1 literature was pretty helpful although secular trends are indeed a problem. A somewhat interesting suggestion here is to add a “control” outcome although that’s difficult in practice.