Hello everyone. My name is Lieven Kennes;
I'm a Professor of Statistics at the University of Applied Sciences Stralsund.
Today, I will be talking about the detection of and adjusting
for selection bias in randomized controlled clinical trials.
After a short introduction into the topic and after having a look at the negative impacts,
selection bias can still happen even in randomized controlled clinical trials.
I'm going to introduce a mathematical model to address this problem and
show how this solution performs in comparison to the traditional approach.
In clinical trials, we often want to compare two or more treatments.
For the theory presented in this talk, however,
we will restrict our considerations to the comparison of just two treatments.
Now, looking at the two treatments to be compared,
one of these treatments is typically the new treatment under
investigation and the other treatment is a control treatment,
which depending on the stage of development,
is either an active control treatment or a placebo.
To conduct our clinical trial,
the subjects we enroll have to be distributed among these two treatment groups.
Now, it is apparent that these two treatment groups have to be homogeneous otherwise,
I cannot assign the observed difference to an actual treatment difference,
but I have to hold the group heterogeneity responsible for this observed difference.
If the patients, or the investigators, the clinicians,
are allowed to choose the treatment,
it is likely that groups will not be homogeneous.
Those groups will probably be different due to
certain observable or even unobservable characteristics.
The bias that, then, might result from this selection is referred to as selection bias.
To prevent selection bias and, of course, also other types of bias,
randomization is usually performed.
Randomization is the act of randomly assigning subjects to treatment groups.
So, in our case,
randomly assigning subjects to these two treatment groups under investigation.
Now, it's often believed that selection bias is completely eliminated by randomization.
However, as we will see,
this is not entirely true.