Jonas Ranstam MedicReS World Congress 2013

MedicReS Good Clinical Research CME June 7-8 2013 | Istanbul Turkey

Practical and statistical aspects of randomization and blinding in clinical trials

Jonas Ranstam PhD

Department of Clinical Sciences

Lund University

Sweden

Outline Why randomize?

- Statistical principles and phenomena

- Historical examples

How can randomization and blinding be achieved in practice?

- Simple randomization

- Block randomization

- Blocking, or stratified randomization

- Cluster randomization

- Dynamic allocation

- Blinding

Statistical issues with randomization

- Common misunderstandings

- How to do it

Suggestions

- Regulatory guidelines

- Reporting guidelines


Why randomize?


Why not test a drug on a single patient?

Problem

A treatment effect estimate from a single patient does not provide any

information on estimation uncertainty.


Why not test a drug on a single patient?

Problem

A treatment effect estimate from a single patient does not provide any

information on estimation uncertainty.

Solution

Use a random sample from the infinite population of current and

future patients, and calculate the uncertainty of the findings using

statistical inference.


Why not use one group of patients?

Problem: Temporal variability within patients

The response to a treatment usually varies over time. A consequence

is the regression-to-the-mean effect, which often is misunderstood as

a “placebo effect”.


----------------------------------------------------------------------- Variable Obs Mean SD Min Max Change (95% Ci) p-value ----------------------------------------------------------------------- All observations Baseline 100 150 25 92 206 Follow up 100 150 25 91 209 0 (-7, 7) ≈1

Baseline 150+ mmHG Baseline 49 171 14 150 206 Follow up 49 150 25 91 209 20 (11, 28) <0.001 -----------------------------------------------------------------------

Regression-to-the-mean

Simulated blood pressure data using Stata

n: 100

Distribution: Gaussian

Baseline: mean 150 (SD 25) mmHg

Follow up: mean 150 (SD 25) mmHg

Correlation baseline-follow up: 0

Treatment effect: None

Establishing a control group

Problem: Temporal variability within patients

Solution: Use a control group

An untreated group from the same source of subjects and with

identical inclusion/exclusion criteria will have the same regression-to-

the-mean.



Problem: Confounded treatment effects

Treatment response may depend on known and unknown factors (like

age, gender, co-morbidity, etc.), which may be unevenly distributed

between the two groups.



Amberson JB Jr, McMahon BT, Pinner M. A clinical trial of sanocrysin in pulmonary tuberculosis.

Am Rev Tuberc 1931;24:401–435.

The 24 (tuberculosis) patients were then divided into

two approximately comparable groups of 12 each.

The cases were individually matched, one with another,

in making this division. ... Then by a flip of the coin, one

group became identified as group I (treated group) and

the other as group II (control).


Problem: Confounded treatment effects

Solution: Alternation

This methods was used before 1948.


Medical Research Council. Clinical trial of patulin in the common cold. Lancet 1944; ii: 373-5.

The first properly controlled multicentre trial

Patulin and the common cold

Patulin (Penicillium patulum) was considered for cancer treatment during

1943. The director of the Imperial Cancer Research Fund's laboratories,

WE Gye, had a severe cold and sprayed his nasal passages with patulin.

His blocked nose cleared within an hour and the following morning his

cold had gone. A controlled trial was initiated.

Clarke M. The 1944 patulin trial of the British Medical Research Council. J R Soc Med. 2006;99: 478–480.


Eligible patients, suffering from common cold, were alternately (every other) allocated to patulin or a placebo group.

------------------------------------------------

Group Patient number

------------------------------------------------

Patulin: 1, 3, 5, 7, 9, 11, 13, 15, 17, 19...

Placebo: 2, 4, 6, 8, 10, 12, 14, 16, 18, 20...

------------------------------------------------



Conclusion

“No evidence was found that patulin is effective in the treatment of the

common cold”


Why is alternation not used today?

Problem: Allocation outcome is predictable

An investigator can interfere with the allocation of patients by selecting

patients to participate in the trial, creating selection bias.


Why is alternation not used today?

Problem: Allocation outcome is predictable

Solution: Randomize and conceal the outcome

Keep the allocation outcome concealed until the patient is entered into

the trial.


Theobald GW. Effect of calcium and vitamin A and D on incidence of pregnancy toxaemia.

Lancet 1937;2:1397-1399.

The first randomized trial

Reporting and observer bias

Problem: Subjectivity

Randomized groups are comparable at baseline, but measurements

made during the trial can be biased by patients' and or investigators'

subjective belief.


Reporting and observer bias

Problem: Subjectivity

Solution: Blinding

Keep the allocation outcome concealed, not only at randomization but

during the entire trial: “blinding”.


Medical Research Council. Streptomycin treatment of pulmonary tuberculosis: a Medical Research Council

investigation. BMJ 1948;2:769-782.

The first randomized and blinded trial

How can randomization and blinding be achieved in

practice?


Alternation

Two groups, TRT and PBO, 20 patients

TRT PBO TRT PBO TRT PBO TRT PBO TRT PBO

TRT PBO TRT PBO TRT PBO TRT PBO TRT PBO


Randomization

Simple randomization

Two groups A and B, 20 patients

A B A B A A B A A A A B A B B B B B B A

Whether A or B stands for placebo is kept secret, at least in double

blinded trials.


Randomization

Stata program

set obs 20

gen grp="A"

replace grp="B" if _n>10

gen rnd=uniform()

sort rnd


Randomization

Concealment of outcome

Opaque envelopes, telephone service, Internet service

Problem

It may be necessary to break the code (unblind) in case of a medical

emergency. This must be clearly document, and the code should be

kept unbroken for other patients.


Randomization

Simple randomization

Two groups, A and B, 20 patients

A B A B A A B A A A A B A B B B B B B A

Problem: may create unnecessary imbalance

Especially in the case of performing an interim analysis.


Randomization

Block randomization

Two groups A and B, 20 patients, block size 4

[A B A B] [A A B B] [A B A B] [A B B A] [B A B A]

Problem

Does not balance known prognostic factors.


Randomization

Stratified (block) randomization

Two groups A and B, 20 patients, block size 4,

stratified (blocking) on gender

Males: [A B A B] [A A B B] [A B A B] [A B B A]

Females: [B A A B] [B A B A] [B B A A] [A B B A]


Randomization

Cluster randomization

A cluster randomized trial is a trial in which individuals are randomized

in groups.

Examples

General practices, hospitals, families, school classrooms, football

teams.



An antihistamine and seasickness trial was performed using two troop

transport ships in the north Atlantic during 1916. On one of the ships

the troops received treatment. The troops on other ship were controls.

Randomization

Dynamic allocation

The number of previously enrolled patients having the same

prognostic factor as the new patient is counted.

The new patient is then allocated to the treatment arm for which the

sum of the previously enrolled patient prognostic factor counts is

smallest.


Randomization

Dynamic allocation


Randomization

Dynamic allocation

Next patient is a man aged 23


Randomization

Dynamic allocation

Next patient is a man aged 23


Randomization

Dynamic allocation

Several methods exist (e.g. Pocock & Simon, Perry, Heritier, etc.)

The allocation must include a chance element. It must not be

deterministic.

Statistical analyses must be conditional on all factors included in the

algorithm.


Blinding

Open label

Sometimes the circumstances of a trial make it impossible to

conceal the identity of the treatment patients are receiving.

Blinding is particularly difficult in surgical trials, although

standardized wound dressings, sham operations, or independent

outcome evaluation may be useful.


Blinding

Single blind

The patient does not know if treatment or placebo has been

administered.

Double blind

Neither the patient, nor the physician, knows if treatment or placebo

has been administered.


Blinding

Double blind is better then single blind

More complex logistics.

Not always possible to achieve.


Blinding

Triple blind

When the code list has been lost.

This is not a good thing.


Blinding

Blinding

Test drug vs. placebo

Blinding

Identical shape, weight and color


Blinding

Identical shape, weight and color


Blinding

Double dummy technique


Statistical issues


Statistics is often misunderstood

The null hypotheses tested in hypothesis tests are about the

parameters of a hypothetical population, not about the observed

sample.

The p-value, SEM, and confidence interval describe the uncertainty

in a generalization.


Statistics is often misunderstood

Randomization has consequences for the statistical analysis.

We will discuss some common misconceptions about the

consequences of randomization.


Misconception:

The aim of randomization is balance

Randomization protects against systematic imbalance (selection bias)

but not against random.

Randomization makes sure that the null hypothesis is true, and that

assumptions underlying statistical methods (like homogeneous

variance) are fulfilled.


Misconception:

Blocking is not necessary

Yes it is, blocking for known predictors reduces estimation errors.

The design and analysis of experiments should take into account all

of the major sources of variability that are expected to influence the

responses.

“Block what you can, randomize what you cannot block”.


Misconception:

Only compliant patients should be analysed

Excluding patients from the analysis is a potential risk for selection

bias.

The ITT population consists of all randomized subjects, without

exclusion because of non-compliance, protocol deviations,

withdrawal, or anything else that happens after randomization.

The prognostic balance generated by the randomization is

maintained in the ITT population.


Misconception:

Only compliant patients should be analysed

The ITT approach is often combined with an analysis of the per-

protocol (PP) population. This is defined as the subset of the ITT

population that completed the study without major protocol

violations.


Misconception: It can be tested if the

randomization has been successful

Many published trial reports include a Table 1. with hypothesis tests of

imbalance at baseline, after randomization.


Frobell RB, Roos EM, Roos HP, Ranstam J, Lohmander LS. A Randomized Trial of Treatment for Acute Anterior

Cruciate Ligament Tears. N Engl J Med 2010; 363:331-342.

It can be tested if the randomization has been successful

Misconception:

It can be tested if the randomization has been

successful

With randomization we know the null hypothesis is true.

The risk of a false positive test outcome will depend on the specified

false positive rate (the significance level).

With 5% significance level and 20 tested null hypothesis at least one

false positive test should be expected.


Misconception:


successful

If imbalance is detected it cannot be adjusted away using

epidemiological methods.

With flawed randomization the results of the trial will also be flawed.


Misconception:


successful

Imbalance at baseline is a property of the sample, not the population.

This common misconception indicates confusion of sample and

population.


Misconception:

Blocking factors can be ignored in the

analysis

Factors used for stratifying the randomization are not always

included in the analysis.



Misconception:

Blocking factors can be ignored in the

analysis

This leads to an incorrect variance estimate.

Include randomization stratification factors as covariates in the

statistical analysis.


Misconception:

Cluster randomized trials can be analyzed as

any other trial



An antihistamine and seasickness trial was performed using two troop

transport ships in the north Atlantic during 1916. On one of the ships

the troops received treatment. The troops on other ship were controls.

Misconception:


any other trial

Having a unit of analysis other than the unit of randomization can

create dependency problems.


Misconception:


any other trial

Randomizing clusters and analyzing subjects leads to an

underestimation of variance, the extent of which is known as the

variance inflation factor, IF.

IF = 1 + (m − 1) ρ

ρ = intraclass correlation and

m = average cluster size

Misconception:


any other trial

A standard patient number calculation leads to an underpowered

trial.

A standard statistical analysis exaggerates the statistical

significance.


Final suggestions


ICH-E9 Statistical Principles

http://www.ema.europa.eu/pdfs/human/ich/036396en.pdf


CONSORT Statement

http://www.consort-statement.org/


Thank you for your attention!


Date post:	20-Aug-2015
Category:	Health & Medicine
Upload:	medicres
View:	395 times
Download:	1 times

Jonas Ranstam MedicReS World Congress 2013

Health & Medicine