Correct analysis of crossover trials

Hi everybody

I was wondering if anyone here had any expertise regarding analysis of crossover trials.

I'm writing my thesis at the moment and was reading through an article that is considered central to my field. It's a cross-over clinical trial with one placebo and one active treatment. There is suffcient wash-out time in the trial.

I think there is a problem in the stats of the study though and was interested in the possible effects such an error would produce as I would like to touch on this in the discussion section of my thesis.

The data in the study have been analysed using students t-test or wilcoxon rank sum test if not normally distributed.

What they've done is to analyse everything as pre(placebo)-/post-(active compound) treatment in effect disregarding what sequence the treatments were given in.

As I understand the reason you have a placebo control in a parallel group design is to control for effects of time.

I found an article about analysing crossover designs (DOI: 10.3238/arztebl.2012.0276 or ) and as far as I can tell the way you control for effects of time in crossover design is by using a stat method that takes into account what sequence treatments are given in.


1) Is my understanding of the need for analysing data in a crossover trial regarding controlling for effect of time correct?

2) What would the effect be when analysing data in the way described? Increased type I/II error etc. Does anyone know??


Omega Contributor
This is all given I don't know the context of the study!

I have not run a cross-over before, but I am familiar with the logic. Well, if treatments are randomized everything seems golden. Though, as you point out an investigator should control for time. This is because there could be residual affects of treatment, broken stable treatment variable assumption or perhaps disproportion non-adherence in the second group, as well as changes in the disease course. Though, this last one gets addressed with randomization most likely.

Without seeing the paper, and not wanting to read another paper, I would say that they should have investigated time as a potential confounding covariate. Though, if this was a funded, well designed study beyond this possible concern, there may be a chance the investigators did examined this and found no association or concerns. So they just neglected to report this in the publication. Moreover, I see no issue and just forget or opt not to even address it.

Well, how may it bias the results, if there truly was a time effect. This one is up to you, since we don't know the content. All depends on how efficacious the treatment was and what you suspect could have happened. Feel free to make a list of suspected effects of time and we can brainstorm its impact!