0:00
My name is William F. Rosenberger.
I'm an University Professor
and Chairman of the Department
of Statistics
at George Mason University.
I also have written two books
on the subject of randomization.
This is part two of "An introduction
to randomization
for clinical trials"
recorded for Henry Stewart Talks.
I recommend that you watch
the first part of "An introduction
to randomization for clinical trials"
before proceeding to part two.
0:28
Now, I want to discuss another criterion
which is randomization
as a basis for inference.
0:36
So the third great property
of randomization
is that it provides
a basis for inference,
that's assumption-free
and relies only on the way
in which the subjects were randomized.
The early clinical trialists
were aware of the importance
of randomization-based inference,
but as I mentioned earlier,
they had limited computer resources
to implement it.
Nowadays, we can run
a randomization test,
or as the FDA calls it,
a "re-randomization" test in seconds,
just by modifying the program
that we used to generate
the initial sequence.
Unfortunately, many students
are not taught
randomization tests anymore,
or even told
that the usual population model
does not apply in clinical trials.
Randomization tests
are particularly useful
for small clinical trials,
where standard large sample theory tests
may not apply.
1:29
So if we think about
the usual population model,
based on random sampling
from a population,
we can think of two populations,
populations A and B.
And then we do random sampling.
And then presumably,
the sample that we get
will be an "i.i.d" sample
with the same population model
as what we sampled from
and parameters
based on those populations.
We call it θA and θB.
Clinical trials
do not follow this model,
so often what's done
is we invoke a population
or a random sampling model
to describe a clinical trial
in order to conduct inference.
In this case, we have basically
an unspecified patient population
because they really aren't
populations of patients
taking on experimental therapy
or even taking a placebo.
Then there's some undefined
sampling procedure from this population
that produces "n" patients
and then randomization is conducted.
And we get "nA" patients on treatment A,
"nB" patients on treatment B.
And we assume this "i.i.d" model
with parameters θA and θB.
