We randomise to obtain groups which are comparable for everything except treatment, including known and unknown variables. The condition of the participant should not influence the choice of treatment.
Back to top.
How can we obtain comparable groups?
We might consider some kind of deterministic method, such as:
All these methods have been used. They are bad ideas because they involve open allocation — the person recruiting trial participants knows the next treatment and may be influenced in the recruitment. For example, they might accept a potential trial participant only because the next treatment is control and that is what they would get anyway, or reject because they think the participant needs the intervention treatment and decide to give it to them outside the trial. There is plenty of evidence that this has happened.
This is the reason Bradford Hill switched from recommending alternate to random allocation (Chalmers 1999).
To obtain comparable groups, we must use a non-deterministic method, so that we do not know what treatment the next research participant will receive before we recruit them. To do this, we allocate randomly and let chance decide.
We could use a physical method of randomisation, such as:
Such physical methods of random choice are widely used in other contexts, but they are poor things to use in clinical trials. These methods might be OK provided we stick to the allocation. However, people will subvert allocation in the perceived interest of the patient. This is known to have happened. For example, a coin can be flipped again, envelopes can be held up to a strong light so the contents can be read before opening, or opened and resealed if the desired treatment is not indicated.
I know that you would not do this (or do I?). Who do you trust? I trust myself, but would you trust me? The trial must be able to convince a sceptic that the allocation was blind. For this reason we need an audit trail, a numbered list of allocations to treatment determined apart from the recruiter.
It is much better to have allocation which is recorded independently of recruitment. A participant is recruited, agrees to be in the trial, and then someone else, not the recruiter, performs the allocation. Do avoid accusations of cheating, we do not usually use a physical method like flipping a coin or shuffling. Instead, we generate our randomisation using computer generated random numbers.
We could do this in advance and put the choice in numbered, sealed envelopes. Then after a participant has been recruited, the next envelope is opened and the treatment is indicated. This might be OK for a small trial under close personal control. Otherwise, envelopes may be tampered with.
In a placebo controlled drug trial, the drug/placebo may be made up by the trial pharmacy in randomly ordered numbered batches. This is conceptually the same as envelopes, but is much more tamper proof.
To avoid tampering, in non-drug trial we prefer randomisation by telephone or remote computer. This is essential if the trial is not under close personal control, e.g. a multicentre trial.
For telephone randomisation, we:
In a small, one site trial, the randomisation list could be locked in a secretary's filing cabinet, but everybody involved must be quite clear that it is at all times secret and confidential to that person.
For web-based computer randomisation, we:
Which we need depends on whether we need out of working hours access, what facilities we have available, etc.
Back to top.
Types of randomisation:
There are several variations on random allocation:
Despite randomisation, there will still be some small differences between the groups. Several methods are used in an attempt to making the groups more similar:
Back to top.
Cluster randomisation
Cluster randomisation is when, rather than allocate individual people to treatments, we allocate groups of individuals together, e.g. GP practices, hospital wards, schools. Problems arise because we may have a small number of clusters which are rather variable.
Trialists sometimes want to match clusters in pairs and allocate one to each treatment. We do not have the problem of lack of openness, because we usually have a list of clusters at the start anyway. But don’t match, it loses degrees of freedom, important with a small number of cluster, and it is difficult to get all the clusters into pairs of comparable similarity.
Stratification may be difficult because there are insufficient clusters in a stratum.
Minimisation is a good strategy. cluster randomised trials are very suitable for this.
Back to top.
Clinical Trials Units
Should you try do-it-yourself randomisation or should you use a professional clinical Trials Unit?
You could do it yourself if you have:
a small trial,
AND it is under personal control,
AND you have the skills.
You should use a Trials Unit if you have:
a large trial,
OR a multicentre trial,
OR more than one person recruiting participants,
OR you need experienced support.
Clinical Trials Centres are not free. Typical cost might be £10 per subject allocated. You should contact the trials centre before you apply for funding. They will help you. Trials centres also may provide many other trials services, including statistical analysis, economic evaluation, etc. They are strongly recommended for large trials.
People often ask "How do I find allocation software?" or "How do I find a trials centre?" I have produced and maintain an on-line directory of randomisation software and services. You can find this on my website: http://www-users.york.ac.uk/~mb55/ under "Statistics".
Back to top.
A question for you
And finally, what is wrong with this?
In a report of a study using 12 general practitioners' peer review groups. Welschen et al. (2004) reported that
"We allocated the 12 peer review groups who agreed to participate to groups A or B, to achieve comparability. We considered all possible compositions of groups A and B and chose the option of those groups resulting in comparability between group A and B in groups with a high or low volume of antibiotic prescribing, rural or urban working groups, and number of general practitioners per group. MMK, who was blinded to the composition of the groups, flipped a coin to determine whether group A became the intervention or control group."
Was that random allocation?
Chalmers I. (1999) Why transition from alternation to randomisation in clinical trials was made. British Medical Journal 319: 1372.
Welschen I, Kuyvenhoven MM, Hoes AW, Verheij TJM. (2004) Effectiveness of a multiple intervention to reduce antibiotic prescribing for respiratory tract symptoms in primary care: randomised controlled trial. British Medical Journal 329: 431-3.
To Martin Bland's M.Sc. index.
This page maintained by Martin Bland.
Last updated: 20 July, 2009.