Skip to main content
Search form
  • 00:00

    [MUSICAL GESTURE][Introduction to Pairwise Randomization]

  • 00:14

    DAVID TORGERSON: My name is David Torgerson.I am a professor at University of Yorkin the department of health sciences,and I'm also director of the York Trials Unit.My research interests are developing randomized trials,how the methods and randomized trials are used,and how we can improve the methodologyconduit to randomized controlled trials.

  • 00:36

    DAVID TORGERSON [continued]: This tutorial is about how pairwise randomization.This is a method of randomizationthat was described by Daniels back in 2003,and it's not been widely used ever since.I think this is a shame, because it has the potentialto be more widely used and addressessimilar problems that I'll talk about in a minute.

  • 00:57

    DAVID TORGERSON [continued]: And the current talk is based on papersthat he published in the Research Methodsin Medicine and Health Sciences journal--a new journal produced by SAGE.[Why should researchers randomize?]We randomize for four main reasons.

  • 01:20

    DAVID TORGERSON [continued]: First of all, it deals with temporal problems--so anything that occurs during the conduct of the trialwill affect all groups at the same time.It deals with a mathematical phenomenon known as regressionto the mean.This is when you select participantsbased on some variables, such as pain score,

  • 01:41

    DAVID TORGERSON [continued]: to take part in a trial.And then there is a natural tendencyfor that variable to change towards getting better.And that will affect both groups equally.We also do it to eliminate selection bias.Selection bias a real problem in non-randomized trials,and selection bias means patients

  • 02:02

    DAVID TORGERSON [continued]: have been selected into treatment groupsin a way that is related to outcome.So they might be selected because for treatmentbecause they have to be more ill than usual,and so if you compare them with patientswho have not received treatment, you'renot comparing like with like.So randomization ensures that patientshave a similar morbidity when they'rerandomized into two groups.

  • 02:23

    DAVID TORGERSON [continued]: And finally, the fourth reason is statistical influence.Most of our statistical tests are based on the notionthat you're comparing random samples.And one of the few areas where holds trueis in a randomized trial, because you take a large sampleand you randomize it into two or more groups.And that's essentially the same as randomly select from thatlarger sample.

  • 02:45

    DAVID TORGERSON [continued]: [How does randomization help solve these problems?]We can eliminate selection bias through randomization,and this also deals with temporal effects.Now although randomization, in principle, is very simple--it's like tossing a coin--

  • 03:06

    DAVID TORGERSON [continued]: in practice, it becomes somewhat more different.So people often use forms of restricted randomization, whichis no longer tossing a coin.So tossing a coin is simple randomization.There's a perceived problem with simple randomizationin that just as you can toss a coin and get heads 10 timesin a row, so that may occur in a randomized controlled trial.

  • 03:29

    DAVID TORGERSON [continued]: So this has certain problems.So in small sample sizes, you mayget very big numerical imbalance in the group.That is not a problem when you have a large sample size.And by large, I mean 1 to 200 patients.So it's not that large before the play of chance,

  • 03:49

    DAVID TORGERSON [continued]: in short, varies equal groups.The other problem is logistics problems.So if you are a surgeon, for example,and you are randomizing people to surgery,you will want to set up numbers of slots ready for patientsto randomized.You want to have some how to predictin the next month how many treatments youcan deliver to patients.

  • 04:09

    DAVID TORGERSON [continued]: Also, there are statistical reasonsthat, if you have many centers in the trial,we know the center affects the outcome.So if you've got a surgical trial againwhere surgeons will be of different caliberor have different techniques, then the outcome of the patientmay be related to the center.

  • 04:31

    DAVID TORGERSON [continued]: Or that center has patients from, say,higher or lower socioeconomic groups from different centers,which may affect outcome.So to try and deal with that, most randomized trialsuse some form and restricted randomization.Restricted randomization is when you have the randomization

  • 04:51

    DAVID TORGERSON [continued]: algorithm such that the total numbers of participantsare equal between the two groups over a set period of time.So a common way of doing this is by blocking.So block randomization is when you take a blockof allocations-- and a common block size is four--

  • 05:11

    DAVID TORGERSON [continued]: so every four patients that you randomize,you can guarantee that two will get interventionand two will get the control.There's a problem with block randomizationin that we can introduce subversion,and subversion is unfortunately more widespreadthan we like to believe.

  • 05:31

    DAVID TORGERSON [continued]: And this is when researchers or clinicians try and subvertrandomization so that a named patient getsone of the treatments-- usually the intervention treatment.So in the past, this has happened, typically,when people will open randomization envelopesin advance to see what the treatment is in the envelope

  • 05:53

    DAVID TORGERSON [continued]: rather than opening the envelope whenthey've recruited the patient.And one the first trials I ever undertook was a surgical trialI was involved in as a junior researcher.And of the five sites we were involved in, three of the siteswere opening the envelopes in advanceto ensure that older patients gotthe conventional treatment and the younger patients

  • 06:14

    DAVID TORGERSON [continued]: got the new treatment so that theywould show that the new treatment would work.And so that trial had to be abandoned and restarted.Now block randomization opens the door a little bitto subversion.So if you're using the block of four allocationsand you keep a record of the allocations,you can always predict the fourth allocation.So for example, if you know the first in the first three

  • 06:37

    DAVID TORGERSON [continued]: allocations to x the interventionand the third one that control youknow the fourth will get its control.So if you could hold patients backto be scheduled for that randomization,you can subvert the randomization.So as I said, this is not uncommon.Now, if we move to, how does the pairwise randomization deal

  • 06:59

    DAVID TORGERSON [continued]: with that problem?Well, Daniels knew that center stratification,which is very common, is really dangerous for subversion.So if you are center stratified--you know and you've worked out that the block size is four--you can then ensure that patientsget what you want them to get and not randomly.

  • 07:23

    DAVID TORGERSON [continued]: So Daniels came up with this ideathat instead of randomizing patients one at a time,we could recruit patients in pairs.So you wait till you get two patients--they don't have to be matched-- so it's notmatched randomization, which is different--when you try to match patients on some characteristic.Here we're trying to just pair them.

  • 07:43

    DAVID TORGERSON [continued]: It doesn't matter whether it's a man and a woman or just a pair.You don't have to wait for two women or two men.You just get two patients.Then you put the pair into the computer programthat does the randomization.The computer program then randomizesone pair to the intervention and one member of the pair

  • 08:04

    DAVID TORGERSON [continued]: to the control.So it's impossible, then, to predict who gets what.So this prevents subversion at the center level.Now, as I've said before, it's been underutilized.And a problem in the trial that we had

  • 08:25

    DAVID TORGERSON [continued]: was a surgical trial where the treatment is surgeryversus conservative care.And surgery doesn't need to be done rapidly.It was elective surgery.Now, the problem is what we wanted to find outwas what was the impact of surgery

  • 08:45

    DAVID TORGERSON [continued]: 12 weeks after surgery took place.Now, unfortunately, we don't knowwhen the surgery takes place.So we would randomize patients to surgery,and then they'd go on a waiting listfor surgery, which may vary.So if they got a cancellation, it might be next week.But more typically, it's two to three months after theyget on the waiting list.So we don't know exactly when the patient gets the surgery.

  • 09:09

    DAVID TORGERSON [continued]: So at this point, I thought pairwise randomizationmay be really useful.Because instead of randomizing them one at a time,if we randomize them in pairs, thenwhen the surgical member of the got surgery,it means we could start the clock ticking for when wewere going to follow them up.So when the surgical pair got the surgery,

  • 09:33

    DAVID TORGERSON [continued]: we would then start counting to 12 weeks.So 12 weeks later, both members of the pairwould get their outcome assessmentto see how well they were doing post-surgical treatment,whereas obviously the intervention patient didn'tget the treatment.And the surgical patient did.And so at 12 weeks, we'd follow them up.This may be 14, 16, or 20 weeks after randomization,

  • 09:56

    DAVID TORGERSON [continued]: but it was 12 weeks after treatment.More recently, a trial has come into our trials in it thatwants to evaluate another treatment,and this is introducing showers to patients who find

  • 10:23

    DAVID TORGERSON [continued]: it difficult to take a bath.So these patients are elderly.They find it difficult to get in and out of the bath,whereas they can walk into a shower,but they live in houses that just have a bath.So the intervention is at the momentthe patients go onto the waiting list for the local counselto introduce a shower--so they're on this waiting list for some variable length

  • 10:46

    DAVID TORGERSON [continued]: of time, and what we wanted to find out was whether--or we want to find out--if we can accelerate the waiting list so that patients getthe treatment-- the shower--straightaway in, let's say, the next monthrather than waiting 6 or 7 or even 12 months on the waitinglist.Well, this improves their quality of life.

  • 11:10

    DAVID TORGERSON [continued]: So again, just like the surgical example,we know when the intervention patient gets a shower--and that's quite quick--but we don't know when a control patient will get their shower.And what we want to do is measure outcomesjust before the control patient gets their showerin place in the control group.

  • 11:32

    DAVID TORGERSON [continued]: So for this we though pairwise randomization wouldbe a solution, and so we build that into the trial proposal--to enable us to follow two people at the same time whenthe control patient--just before control patient gets their shower.So in this instance, we will have a variable follow up.

  • 11:54

    DAVID TORGERSON [continued]: So some patients-- it will be like 6 months--some patients it might be 12 months--till they get a shower.And that's when they both get followed up.And so in some ways, that's quite useful,for us to get a lot more information.We ought to say we'll get a variation in the followup patients.And so we can look to see how the difference in outcomes

  • 12:17

    DAVID TORGERSON [continued]: are for the patients at say, 6 months or 12 months--which, if we hadn't used pairwise randomization,would not be possible.[What are the disadvantages of using pairwise randomization?]I hope I've convinced you that pairwise randomization isa really useful technique to consider

  • 12:38

    DAVID TORGERSON [continued]: for your randomized controlled trial,but there is a good reason that ithasn't been more widely used.And this is it can only be used whenyou don't have to randomize quicklyand you've got enough patients thatare coming through in a regular fashionthat you can have pairs available.So for an example I gave-- of elective surgery--

  • 12:60

    DAVID TORGERSON [continued]: that's fine.Because patients don't need surgery in a rush.However, we did another trial wherewe have to have surgery vs conservative care--but it's emergency surgery.So somebody's coming in with a fractured bone,and so the treatments there are to have surgery almost

  • 13:21

    DAVID TORGERSON [continued]: straightaway or to conservative care-- putthe fractured limb in a cast.Now, we can't pairwise randomization,because we can't ask the patients to wait until they geta second patient to be a pair.So in circumstances where you need to randomize quickly,pairwise randomization isn't an option.

  • 13:43

    DAVID TORGERSON [continued]: So it can't be used.So we need to think of other randomization methodsthat in those circumstances.Key disadvantages-- you need at least two participantsavailable at any one time before you can randomize,and if you have multiple arms-- obviously,if we had three arms, you'd need three participants.So this is a disadvantage where you

  • 14:03

    DAVID TORGERSON [continued]: it may be a thing of urgency to randomize,and it can't be done unless you have the pairs or tripletsor how many arms you've got.[What is the difference between matched and pairwiserandomization approaches?]Well, matched randomization is when you take a subject

  • 14:25

    DAVID TORGERSON [continued]: and match them on, say, their gender or age--and, say, you might take someone who's in their 40s--find somebody else who's in their 40s who'sa male as well--and then they match on age and gender.Then at that point, you randomize them.Now, the difference between pairwise and matchedis we're not interested in matching them

  • 14:46

    DAVID TORGERSON [continued]: on a characteristic.So we just take the first two peoplewho are available to be randomized,and they form a pattern that we randomize.And this has advantages over matched randomization,because matched randomization requiresspecialist statistical techniquesto cope to deal with the matching,

  • 15:08

    DAVID TORGERSON [continued]: whereas pairwise randomization-- we're notmatched on characteristics.So we don't need to use that.We can just use standard statistical approachesto trial analysis.[MUSICAL GESTURE]


David Torgerson, PhD, Professor at University of York, discusses pairwise randomization in research, including reasons for using, suitable types of studies, and disadvantages.

Looks like you do not have access to this content.

Introduction to Pairwise Randomization

David Torgerson, PhD, Professor at University of York, discusses pairwise randomization in research, including reasons for using, suitable types of studies, and disadvantages.

Copy and paste the following HTML into your website