[MUSIC]

Hello, this is Janet Holbrook again.

And today we're going to be covering the
concept
of randomization, which is one of the real
core features
of randomized clinical trials and probably
the most familiar characteristic
to most people when we talk about
randomized clinical trials.
We have three sections, and for the first
section,
I will be going over the rationale for
randomization.
As I just alluded to, randomization is a
key feature
of the experimental design that a clinical
trial is based on.
So there are other features that are
indeed
as important as randomization and they
include things
like standardized treatment and having a
prospective plan
for data collection, adverse event
reporting, as well
as the regulatory requirements that go
along with clinical trials.
And these features tend to distinguish
clinical trials from other types
of observational studies and even from non
randomized studies in some cases.
So, it's important to recognize that the
concept of randomization, or the drawing
of lots, has been with us throughout
history and it even is referred
to in The Bible.
It, it is the idea of drawing a lot as a
way to ensure that the benefits and risk
of an activity or event are
equally shared or fairly shared across a
group of people is a very early concept.
One that's well illustrated with the
concept of a military draft.
Where the risk of war is shared across the
population
by drafting men mostly, into military
service by the mechanism of a lottery.
So this concept, which has been used in
many areas,
including how to divide some scarce
resource up, or the concept
of drawing the short straw.
We first see the idea of
randomization to be used in a medical
setting in the
17th century.
Van Helmont proposed that people be
randomly divided
or divided by lots to determine their

treatment because he wanted to evaluate

his form of care against more standard
forms of care
of the day, which included
bloodletting and other potentially harmful
interventions.
And so he proposed that a number
of patients be divided, and that at the
end the half that are treated by him,
according
to his procedures, be compared to the
other half
of the patients that were treated with
other procedures.
And then there'd be, there'll be a count
of how
many people had died in order to evaluate
bloodletting in healthcare.
So just by way of background, many times
we talk about or use the word random in,
in our usual speech, or in ordinary speech
and
such as, it's a common saying, random acts
of
kindness, or random acts of nature, random
acts of God.
You fill in who the random acts are of.
And what we mean by that is something
that's not
predictable and seems to be haphazard and
there's no causal relationship
that can be identified between things.
I sometimes think my neighbors must think
that I randomly mow my lawn.
It's very haphazard, my pattern.
But in the context of a clinical trial,
we have a much more rigorous definition of
random.
And what it means is that the process, a
random process is
one in which there is associated with
every legitimate outcome a probability.
So there's a legitimate probability
associated with whether
you'll be assigned to treatment A or
treatment B.
Such as with every coin toss, we assume if
it's
a fair coin there's a 50/50 chance of
heads or tails.
With dice throws there's a one in six
chance
of any particular side of the die coming
up.
And this is a much more rigorous
definition than
the lay definition.
The idea of randomization was first
introduced formally into experimental
design by Ronald Fisher.

Many of you are probably familiar with his

famous book, The Design of Experiments.
And the question that Fisher was trying to
address
was to how to allocate plots of soil in
agricultural
settings to different treatments.
There was no question about whether there
needed to be controls on the comparison of
the different treatments on plots.
It was just a matter of, of how the plots
were to be allocated because you can
imagine that even within a field there's
differences in sunlight and drainage.
And they wanted to have a method of fairly
distributing those effects among the
treatments.
And so what he proposed was to use random
allocation, in the allocation of lots.
Because he said by doing this it was the
most efficient and unbiased way
to estimate the measurement error, error
associated with a particular treatment.
And he proposed that any other way of
doing this would either over estimate or
under
estimate the kind of random error
associated with
how the yield was in a particular plot.
He also noted that there was some
safeguard against non-normality in
the outcome data, which was an added
benefit of using randomization.
So note that his arguments
about the use of randomization are all
kind of based on
its statistical properties for estimating
measurement error, and
that it's the most efficient way to do an
experiment.
And, there's not a lot of emphasis in his
work about the use of controls.
The person who really is responsible for
introducing randomization to medical
science, is Sir Austin Bradford Hill.
He was from an epidemiological
background and his concern was about
confounding
that was associated with the type of
people that got a particular treatment.
Because he noted that people actually were
quite variable
in how they may react or respond to a
treatment.
And there can be in some ways predicted,
and some of it's
unpredictable, was a problem in estimating
treatment effects.
And because of the idea of a selection
bias associated

with treatment assignment, that the

prognosis for the disease would be
somehow be related to what treatment a
patient was given.
Commonly referred to as confounding by
indication.
So that people are given treatments based
on their own unique characteristics.
And what Bradford Hill was suggesting, is
to get rid of that problem of
the selection bias, in which people get
which treatments, that random allocation
be used.
So the
critics of a clinical trial would be
unable to
say that the groups were not equal at the
start,
that they were somehow biased based on the
physician
and caretaker's opinion about what
treatment the patient should get.
Perhaps, you can think of it in some
senses that they might give the
sickest patients the new treatment because
they
hadn't responded to older treatments and
were
in more dire need of a new option.
And someone who was promoting a treatment
had a interest in the treatment
might do exactly the opposite and want the
new treatment to look good.
And so only give it to the best
candidates.
So Bradford Hill was very concerned about
these type of biases and
he proposed that randomization be
introduced
to clinical trials to eliminate this.
And indeed, he did design and was involved
in the first properly
randomized trial that is documented well
in the literature.
And led the way for randomization to be
used in other clinical trials.
And this was a trial of streptomycin for
tuberculosis in 1946.
Interestingly, the use of randomization in
this trial was based on
that it was a fair distribution of risk
and
benefits to patients in terms of getting
the new treatment.
But it was also because this was in the UK
right after World War II.
There were very limited supplies of
streptomycin.
Mounting a randomized clinical trial was
also argued to

be the only way to fairly allocate the

limited resources
to, to patients.
Which patients with TB were going to get
this new
medicine that had quite a bit of evidence
for its efficacy.
So as I've talked about, the rationale for
randomization is
to avoid selection bias and to avoid
confounding by indication.
That is, that there's some prognostic
factors that would be related
to treatment assignment, and thereby if
they aren't
somehow controlled, would bias our
assessment of treatment.
And it's important to realize that these
prognostic
factors, such as maybe age or severity of
disease or even clinical centers, can
influence outcomes
as strongly or more strongly as many
treatments.
Remember that clinical trials are really
designed to find small to moderate
treatment effects.
If there's very strong treatment effects,
many times we don't need a clinical trial.
So these prognostic factors can be quite
influential, and so
we use randomization to ensure their
balance across the treatment groups.
And that brings me to the next point, that
randomization
tends to produce comparable treatment
groups on know or unknown confounders.
But it's important
to note, it's not a guarantee.
It's still a probabilistic process, and
there
are times when you don't get perfect
balance.
That you may end up with what appears
to be an important imbalance in treatment
groups.
For example, you could end up with many
more women in one group than the other.
And going back to R A Fisher, the
randomization also assures
the validity of the statistical test that
we use generally to analyze trials.
They're based on the idea that the
treatment assignment, the primary variable
we're looking for and effect in was
randomly assigned across the population.
And perhaps a less important point but is
also a
nice feature of randomization is that it
gives us a

defined time-point for trial entry.

We know that we are going to measure
events and
attribute them to the appropriate
treatment group from randomization onward.
And so it's a very nice time zero that has
some
meaning in terms of how the patient was
treated as well.
I also want to briefly address the issue
of using
unequal allocation ratios for
randomization.
Pretty much up to now I've been talking or
referring
to randomization as sort of two groups one
to one allocation.
And that is certainly the most common
design for parallel treatment trials.
And one of the reasons the one to
one allocation is so commonly used, it's
the most
powerful design.
Other allocation ratios such as two to one
or four to one are typically less
powerful.
And I gave you an example here
of the power associated with different
allocation ratios.
But there may be reasons that you will
compromise on power or alternatively
accept maybe a larger sample size, because
you want to use a unequal
allocation ratio.
It may be important to acquire as much
experience as possible with the new
treatment including the side effects and
toxicity of that treatment.
So you may want to have as many people
as possible be exposed to the, the new
treatment.
That also may be an incentive for
recruitment.
If you're in the situation where patients
don't have
many options for how they'll be treated,
it may be
desirable to weight the randomization
towards the, the new treatment
as a tool for recruiting patients, so that
they know
that they have a greater chance of getting
the new treatment.
There may also be cost issues depending on
how much the treatments cost.
Very expensive treatments you, you may
have a different allocation
ratio in order to conserve resources.
Also, you can imagine that in some cases
that there may be

difference in the expected variance of the

outcome variable by treatment group.
So you may expect that one treatment may
have a more homogeneous effect than
another treatment.
And therefore you could argue for an
unequal
allocation ratio to optimize the power if
there's a
difference in the variance across the
groups that's expected.
I would say that that's probably an
unusual reason for having an unequal
allocation ratio.
And finally, I, I want to speak to the
definition
of selection bias.
Many of you who have taken other courses
with the epidemiology department have
heard people talk about selection bias as
a problem in the study group, the entire
group selected to be studied.
That somehow the association between the
intervention
or the risk factor and the outcome is
different in, in the group of people that
is studied than in the target population.
So the results from the study
really aren't generalizable to the target
population.
So that's sort of an epidemiological
definition of selection bias.
However, in the trial definition of
selection bias, we're really talking about
eliminating the bias associated with the
prognosis of the disease.
So again to eliminate the
confounding by indication.
And we want to, to break that link
between what the intervention actually is
and the patient's characteristics.
So it's more an issue of internal validity
when we're
talking about, in selection bias, in the
context of clinical trials.
So that wraps up this first section with
trying to give you some of the
underpinnings and
rationale for randomization.
In the next section, we're going to
be talking about different types of
randomization schemes.