Table 2 Summary of the characteristics of the various randomised, comparative trial designs (it is assumed for all designs that the control group is a placebo)

Parallel groups

(comparison between groups)[24]

Patients are assigned to a treatment group for the duration of the trial.

Randomisation to one of two or more treatment groups, with a pre-specified randomisation ratio.

Design simple to understand and to implement.

Larger sample size often required, compared with other designs.

Treatment groups can have different numbers of patients.

Difficulties with recruitment possible, if placebo-controlled.

Analysis and interpretation of results is simple.

Cannot estimate the contribution of inter- and intra-patient variability to the overall variability.

Factorial

Can answer two or more questions with one trial

Patients are randomised twice, once for treatment A or placebo and then for treatment B or placebo

Time-saving for the trial sponsor

Need to be sure that there is no interaction between treatments A and B

Requires fewer patients to obtain the answer to two or more questions

Cross-over

Patient receives both of two treatments, A and B, in a pre-specified sequence. Patients act as their own control.

Randomisation to a pre-specified treatment sequence.

Smaller sample size than parallel groups.

Stable chronic diseases (assumes patient’s state is comparable at the start of both periods of treatment).

Endpoint must not be sensitive to learning processes.

Requires a wash-out period between treatment periods. Follow-up is at least twice as long compared with corresponding parallel group trials. The analysis must confirm the absence of treatment - period interaction

Results depending only on within-patient variability.

Often used in healthy volunteers (for phase 1 clinical trials)

Latin square

More than two treatments to compare

Randomisation to a pre-specified treatment sequence.

Same as for cross-over design.

Same as for cross-over design, except carry-over is controlled (similar properties as those for Latin square design).

Balanced design, i.e. every treatment (or dose) appears only once in each sequence and each treatment period.

N-of-1[14, 19, 23]

Only one patient and design aims to assess effects of several treatments in one individual

The order of treatment(s) and placebo periods are randomly assigned for the patient

Provides an estimate of individual effectiveness (personalized medicine)

Same as for cross-over design. Needs a stable, chronic disease

Patients are more likely to have better adherence to treatment, and understand their disease and treatment better

Delayed start[10, 12, 16]

Two phases: initial placebo controlled phase (patients randomised to treatment or placebo) followed by active control phase (all patients receive treatment) – those in the initial placebo group have a delayed start

In first phase, patients randomised to early start group (treatment) or delayed start group (placebo)

Allows more patients to receive active treatment

At the start of the second phase, the patients are not comparable. No real blinding for the second period; carry over effect possible.

Can distinguish effects on symptoms and effects on the disease evolution

Randomized placebo-phase[13]

All patients receive the tested treatment in the end – but have varying lengths of time on placebo.

Randomisation of time from enrolment to starting tested treatment

Can be used for disease-modifying therapies, in diseases with a rapid, unfavourable evolution. All patients receive active treatment

Variable length of placebo period reduces statistical power

Assumes that a response will occur sometime after an effective treatment is given, so that patients who start the treatment earlier should, on average, respond sooner

Low and intermediate potency therapies show large variability for response

Limited ability to estimate size of treatment effect

Stepped wedge[11]

All patients receive tested treatment in the end. Intervention allocated sequentially to participants (either as individuals or clusters of individuals)

For a 5-step wedge design, all patients start with control then for the following five time periods individual or clusters randomised to treatment to finish in the last period with all patients receiving tested treatment

Useful when there is a prior belief that treatment will do more good than harm

There might be a risk of contamination between intervention participants, and a need for blind assessment of outcome

Also, when an innovation cannot be delivered concurrently to all units

Randomised withdrawal

Used to assess treatment continuation in patients who are responding to the treatment.

Randomisation of responders to continue treatment or switch to placebo

Reduces the time on placebo since only responders are randomised to placebo.

For use in chronic diseases, Not suitable for unpredictable diseases (e.g. spontaneous remission) or those with slow evolution. The treatment effect is overestimated since only responders are included (and compared to placebo)

All patients initially receive the tested treatment; responders are randomised to continue treatment or to receive placebo

Can assess if treatment needs to be continued or can be stopped

Possible carry-over effect for adverse effects.

Early escape

Patients withdrawn if they satisfy a priori failure criterion

Randomisation to active treatment or placebo

Reduces the time on placebo or in treatment failure.

Difficult to define a binary failure/success outcome.

Analyse failure rate, so minimises exposure to ineffective treatment

Only short-term efficacy evaluated.

Loss of power if significant number of patients ‘escape’

Three-stage[15]

Initial randomised placebo-control phase, a randomised withdrawal stage for responders, and a third randomised phase for placebo non-responders who subsequently respond to treatment

Randomisation to treatment or placebo and randomised withdrawal for responders

Three separate (independent) assessments of efficacy which are then combined (Fisher’s method) to derive a single overall p-value.

Applicable only to chronic conditions where both response to therapy and withdrawal of therapy can be assessed.

Care should be taken to allow the withdrawal phase to be sufficiently long so that the drug can be completely washed out and the clinical effects of therapy reversed.

Subjects may barely meet criteria for being a responder and would consequently forgo active treatment even though they may have benefited from it.

Since fewer patients may be available in the initial stage of the trial, the ability to precisely determine initial response rates may be less than with a traditional randomized trial design.

May be less suited for controlled assessment of safety

Fewer patients required compared with parallel group design.

Reduces the time on placebo or non-efficacious treatment.

May evaluate the efficacy of a therapeutic agent in a particular patient subpopulation when efficacy in the general patient population has already been established.

Adaptive randomization designs -play the winner[17, 20]

An adaptive randomization design.

The probability of being randomised to one group is modified according to the results obtained with previous patients. It favours the group with favourable results (play the winner), or penalise the group with unfavourable results (drop the looser); it can be generalised to multi-treatment clinical trials, and delayed responses (Generalized drop the looser)

Reduces the number of patients receiving a less effective treatment.

Unequal sample size reduces power.

Need to have binary outcome, (success/failure)

In some situations, the number of patients who have actually received one of the treatments is very low.

Could improve patient recruitment due to better satisfaction

-drop the losers[18]

-generalised drop the loser[21, 22]