Module 5
Lesson 6
Welcome back!
I will now return to the issue that was introduced in the previous lesson.
When we perform disproportionality analysis we need some way to handle the issue of random
variability or else we run a clear risk of detecting a lot of false associations.
To be able to tackle this issue we often create decision rules for our disproportionality measures,
that tell us when we should consider a drug-event combination disproportionately reported, and
these rules all have built-in protection against this issue of random variability.
For the IC we often write like this: IC 025 > 0. But what does that mean?
Well, the IC 025 is nothing but the lower endpoint of a 95% uncertainty interval for the IC, as we
explained in the previous lesson. We need to recall a specific example we were looking at there. We
had two drug event combinations, one with two reports and one with 20 reports. They both started
out at the same value for the relative reporting ratio and then we computed IC values with the
shrinkage and with the uncertainty intervals. And as you see here, the IC 025 is actually below 0 for
the combination with two reports whereas it's above zero for the combination with 20 reports.
Please note that if this value is indeed above zero this just means that we're highly confident that we
have observed more reports than can be reasonably expected in our database. It does not mean that
there has to be a causal link between the drug and the event but there are other measures of
disproportionality than the IC that are used by other organizations.
Specifically we have introduced the PRR -the proportional reporting ratio and the ROR -the reporting
odds ratio, in earlier lessons and neither of these measures are typically used with shrinkage. That
means we have to use other tools to handle the issue of random variability. Decision rules for PRR or
ROR tend to become a little bit messy because they often contain several conditions that have to be
fulfilled at the same time. One condition that is typically always there is that we put a requirement
on the number of reports for the drug-event combination, and that requirement is often set at 3 or 5
reports. This condition kind of makes sense because as we have seen earlier, the issue of random
variability is most prominent at low numbers of reports.
In addition, these rules often contain one or two more conditions and those conditions can be
requirements either on the value of the PRR or the ROR itself, or they could apply to uncertainty
intervals around those values, or they could apply to some kind of statistical test.
Now there are too many variations of these decision rules for PRR and ROR to cover them all here,
but I will present one specific example to make it hopefully a bit clearer.
And so, this specific decision rule that we're looking at here is for the PRR. It has two components:
first, there is a requirement on the so-called PRR 025 value, which is the lower endpoint of a 95%
uncertainty interval for the PRR, and that value has to be above 1. Or if we look at it on a log2 scale it
has to be above 0. The second condition in this decision rule is that we require at least three reports
on the drug-event combination. So, if we then go back again to our example with the two drug-event
combinations, what does this specific decision rule look like? Well, the PRR values themselves are
basically the same as the relative reporting ratio values that we computed earlier so they are just
above 200 or about 8 on a log2 scale. As I said earlier with the PRR there is no shrinkage, but we can
add 95% uncertainty intervals and as you see, the lower end points of those intervals are clearly
above 0 on the log2 scale on both of these drug-event combinations. So both of them passed the first
condition of the decision rule. However, trivially the drug-event combination that has two reports
does not pass the condition of having at least three reports whereas the combination with 20
Module 5
Lesson 6
reports obviously does fulfill that requirement. So in this specific case with this specific PRR decision
rule we get the same result in the end as we did earlier with the IC, but of course this would not
always be the case.