Why%20to%20Randomize%20a%20Randomized%20Controlled%20Trial?%20(and%20how%20to%20do%20it)

1
Why to Randomize a Randomized Controlled Trial?
(and how to do it)

• John MatthewsUniversity of Newcastle upon Tyne

2
Schema of a simple trial
Rx group 1
Randomize
Eligible patients
Rx group 2
3
Outline of talk

• Many aspects to a trial this talk focuses on
  just two
• Why you should randomize
• benefits of doing so
• dangers of failing to do so
• How to randomize
• often glossed over unspecified

4
Why Randomize?

• Compare groups at the end of the trial
• Difference is because of the Rx
• For this you need comparable groups
• Purpose of randomization is to make the treatment
  groups comparable
• Ensures that only difference in groups is due to
  trial treatments

5
How does it do it?

• Each group is a random sample of eligible
  patients, so both are representative of that same
  population
• In this sense they are comparable
• same proportions of males, stage IV tumours,
  ambulant cases, elderly patients etc.
• Anything which subsequently changes the groups
  will destroy this balance.

6
Why Randomize?

• Other benefits are
• Randomization is largely unpredictable
• Why this is a good thing and why it might not
  obtain will emerge in the talk
• Randomization provides a valid basis for
  statistical inference
• This is important but is not addressed at all in
  this talk

7
What is wrong with non-randomized studies?

• Two main types of study, those with and those
  without concurrent control groups

8
Non-randomized studies II

• Without concurrent controls
• Uncontrolled
• cannot really make much of such studies if there
  is any variation in outcomes.
• Historical controls
• type of patient may change, due to eligibility
  criteria
• environment changes, due to trial
• data quality often quite different between groups

9
Non-randomized studies III

• Non-randomized concurrent controls
• Alternation
• Odd/Even hospital no. or date of birth
• First letter of surname
• Difficult to argue that one group is different
  from another but allocation is predictable, so
  bias can arise from selection of patients see
  Keirse (1988)
• so randomization must be unpredictable

10
Features of a RCT

• Provide reliable evidence of Rx efficacy
• Essentially simple
• Much attendant methodology
• ensure reliability of evidence
• give credibility to results
• CONSORT statements www.consort-statement.orgindic
  ate good practice in trial reporting

11
How to Randomize

• Toss a coin
• Essentially the right thing to do
• Try not to do it in front of the patient
• More sophisticated implementations possible

12
Is coin tossing OK?

• OK for big trials
• For small trials, such simple randomization can
  lead to imbalance in group sizes

13
Example trial with 30 patients

• If 30 patients are in a trial randomized using
  coin tossing there is a 14 chance of 1515 split
• For 1614 chance is 27
• Worse than 2010 is 10
• Why worse?
• Because imbalance leads to loss of power

14
Alternatives

• Could use a restricted randomization scheme
• legitimate, intended to protect power
• but often not mentioned in trial report see
  Altman Doré, 1990 Schulz et al., 1994
• Needs to be done properly
• Only ensures similar numbers in groups
• Combine with stratification to ensure
  comparability for prognostic factors

15
Random Permuted Blocks

• An allocation sequence is, e.g.,A,B,A,A,A,A,B,B,B
  ,Ai.e. 6 As, 4 Bs
• This sequence built up by using a computer to
  toss a coin
• Random Permuted Blocks (RPBs) is an alternative
  method which ensures imbalance can never be
  substantial

16
RPBs II

• All sequences of length 4 comprising 2 As and 2
  Bs are1. AABB 2. ABAB 3. ABBA4. BBAA 5.
  BABA 6. BAAB
• Generate random sequence of numbers 1 to 6, say
  6,5,2,6, and substitute from above to give
  allocation sequence ofBAAB BABA ABAB BAAB

17
RPBs III

• Such sequences cannot be more than two out of
  balance
• Must be in exact balance after 4, 8, 12, etc.
  patients have been recruited
• So RPBs are, to some extent, predictable
• To avoid this, vary block length at random use
  blocks of length six (3t) as well as 4 (2t)

18
Is it enough to equalise numbers?

• No, can still have imbalance in important
  prognostic factors
• E.g. two groups of size 15 one comprises 14
  young children and the other comprises 14
  adolescents in a trial for diabetes
• Stratify recruitment with respect to age
• i.e. use separate allocation sequence within each
  stratum

19
Stratification

• RPBs can be used without stratification
• Stratification without using RPB (or an
  equivalent device) is nonsensical
• Separate allocation sequence in each stratum can
  become cumbersome with many prognostic factors
• e.g. ambulant/not, over/under 55, M/F gives 8
  allocation sequences

20
Minimisation

• More complicated, in principle
• ensures balance on each factor separately, not
  for all combinations
• keeps track of patients already in trial,
  computes an imbalance score and allocates to
  minimise this
• can include a random element
• Less cumbersome, in practice
• largely because you need a computer
• Good if there are many prognostic factors

21
How to serve it all up

• Methods for delivering randomisation sequences to
  the clinic are important.
• They hold the key to ensuring adequate
  concealment of the allocation until the patient
  has been randomized.

22
Implementation methods

• Need to separate the person who generates
  allocation from those who assess eligibility
• Third party schemes
• Telephone randomization service
• Pharmacy randomization
• Web-based service?
• Envelopes
• Serially numbered, sealed and opaque

23
Then what?

• You will have two groups that are comparable and
  free from bias
• Well, sort of
• You have the best start, certainly
• Drop-outs, protocol violations etc. etc. disturb
  the comparability
• Might not have been comparable to start with!
• Need to allow for baseline imbalance and
  stratifying variables

24
Conclusion

• Randomization is needed in all clinical trials
• As with most aspects of trial design, the details
  of how you randomize are important
• The analysis needs to respect the design (esp.
  stratification) and make sensible adjustment for
  baselines
• All looking more awkward if there isnt a
  statistician involved.
• Some details given atwww.mas.ncl.ac.uk/njnsm/tal
  ks/titles.htm