In a parallel group design, patients are randomly allocated either to receive an experimental treatment or to a control group. The randomised allocation is performed to ensure that the characteristics of the patients in each arm of the study are similar. If there are systematic differences between the study arms, then any differences in outcome may not be wholly attributable to the intervention. Alternatively, such confounding due to group imbalances may mask or attenuate a treatment effect, causing promising interventions to be overlooked. The balancing of study arms in a parallel design is therefore crucial in order to preclude spurious estimates of the effect of the treatment. However, simple random allocation does not guarantee the balance of characteristics between study arms, particularly for small trials. Investigators may therefore wish to implement a stratified randomisation procedure, whereby patients are categorised according to one or more factors that are believed to be associated with the study outcome variables and are then randomised to receive either treatment or control in blocks. This prevents large imbalances between the groups in relation to the stratification factor/s; at most the imbalance will be half the size of the block. As an example, investigators may wish to achieve similar numbers of smokers in each arm of a trial for Raynaud’s phenomenon, due to the effects of smoking on the vasculature. This could be achieved by stratifying by smoking status. Prior to randomisation, a patient’s smoking status is established, and the patient is randomly allocated to the treatment or control arm using a randomisation list for smokers or non-smokers. At the analysis stage, adjustment should be made for the stratification variable using regression methods. Such an approach may be preferable to the exclusion of smokers from the study, as the latter strategy has an impact on the generalisability of the findings.

In addition to details of the randomisation procedure employed, trial reports should present separate summaries of the baseline characteristics for each arm of the study cohort so that a reader may judge to what extent the randomisation has been successful in the balancing of potential confounders [1]. Tests of statistical significance of differences in baseline characteristics between groups are to be avoided, as they are inappropriate for two reasons. Firstly, any differences between randomised groups must be due to chance, making the test redundant. Secondly, there may be a substantive confounding effect even if no significant difference between groups is found, making the test uninformative [6]. Whichever method of randomisation is employed, it is imperative that the randomisation list is concealed from the persons entering patients into the study, or else the allocation procedure will not be truly random. The steps taken to ensure allocation concealment should also be described in the trial report.

Patients in the control arm may receive either a placebo or an active treatment. In a parallel-group trial of losartan for Raynaud’s phenomenon (including both patients with primary Raynaud’s phenomenon and those with systemic sclerosis-related Raynaud’s phenomenon), Dziadzio et al. [7] used the calcium channel blocker nifedipine as a comparator drug. The use of an active control brings several potential advantages: it allows for a direct comparison to be made between an experimental and an established treatment; it may help to maintain the blinding of participants and investigators to the treatment allocation; it may be considered more acceptable from an ethical perspective to treat patients with an active drug rather than mere placebo, and this may be more acceptable to patients considering whether or not to participate. A benefit of the increased acceptability of using an active comparator is that investigators may then be more inclined to include patients with more severe symptoms, knowing that they will not be given placebo and thereby withheld treatment. These are precisely those patients for whom there is greatest need to identify effective treatments, so it is vital that they are included in trials. Problems with interpretation may arise however when no evidence of superiority is found from a direct comparison of experimental treatment versus active comparator. In such circumstances, it cannot be said that either treatment has been shown to have an effect. A common but misguided approach in this situation is to declare an effect to have been demonstrated in one or both groups on the basis of within-group changes. This reasoning cannot be justified, as a fundamental principle motivating the status of the RCT as a gold standard in research is that effectiveness of a treatment can only be declared on the basis of a direct comparison between randomised groups [8]. Investigators may therefore wish to include a placebo group in a trial, either as the sole comparator arm or alternatively as a third arm in addition to the experimental treatment and active control groups. The inclusion of a placebo arm may be recommended for several reasons [9]. The placebo arm acts as an internal standard, and a comparison between treatment and placebo arms serves to provide a direct assessment of the effect of the treatment over and above non-specific placebo effects. The inclusion of a placebo arm also enables investigators to distinguish adverse events attributable to the experimental treatment from those spontaneous events related to the disease.

In an “AB/BA” crossover design, all patients recruited to the study receive both the experimental treatment “A” and the control “B”, which may be active or placebo (e.g. [5]). The order in which these are received is randomly allocated, so that a participant may undergo a period of A followed by a period of B, or else this sequence will be delivered in reverse. The two periods are separated by a washout phase, which must be of sufficient length to ensure that there are no persisting effects of the intervention delivered in the first period upon commencement of the second. Failure to include a sufficient washout phase in the design of a crossover study will compromise the investigators’ ability to make a valid inference relating to the effect of the experimental treatment, as carryover effects from the first period will obfuscate and potentially interfere with any effects in the second. These effects cannot be isolated at the analysis stage, so it is imperative that they are precluded through thoughtful design [10].

Senn notes how crossover trials may further be subject to period effects, where some secular trend in the experiment means that outcomes in the second period would be higher or lower than those in the first even if no treatment was administered. In the case of Raynaud’s phenomenon, this could plausibly manifest if all of the assessments in period 1 took place within a short timeframe, all of the assessments in period 2 were similarly grouped, and temperatures between the two periods of assessment were different enough to impact upon patient outcomes. Under such circumstances, it may be prudent to use one of the two methods described by Senn to adjust for period effects at the analysis stage. In practice, patients will usually be recruited into the study and assessed at different times, so that some participants will have completed period 2 before others have begun period 1. This means that a period effect may be apparent for some patients (e.g. those for whom temperatures were substantially different between treatment periods) but not others. Senn notes that although such interaction effects might add to the variability in the results, they should not prevent a valid assessment of the treatment effect. Advantages of the crossover design arise from the fact that each patient acts as their own control. Consequently, treatment effects are evaluated on the basis of within-patient comparisons, removing the effects of confounding and reducing the required sample size. The removal of sources of between-patient variability is particularly useful in a condition as heterogeneous as Raynaud’s phenomenon.