When are two events worrisome?

Steve Simon

2006-12-04

I was attending a workshop on developing a risk management plan for a new drug. The presenter, Nawab Qizilbash from Oxon Clinical Epidemiology, gave an interesting example. Suppose you are running a clinical trial with 50 total patients. In the treatment group, you notice two adverse events and in the placebo group you notice zero adverse events. Should you stop the trial?

The group discussion started with questions about statistical significance. What is the confidence interval? one participant asked. The speaker showed some statistics. With two events versus zero in a sample of 25 in each group, Fisher’s Exact Test is not statistically significant. The Chi-square test is not statistically significant.

Discussion then moved to the type of disease being treated and the severity of the side effect. If the effect was very severe, such as death, then you might want to stop the trial. Dr. Qizilbash then raised some specific scenarios.

Would you stop the trial if the adverse event was a rash? Would you stop the trial if the event was neutropenia? Would you stop the trial if the event was aplastic anemia? Again the discussion focused on the severity of the events.

Dr. Qizilbash then pointed out some background data for the last event. Data shows that the rate of aplastic anemia is approximately 1 in 1,000 in the general population. The relative risk of the treatment group to the background rate is .04 / .001 = 40, which is indeed statistically significant.

When you are assessing safety events, especially safety events that occur rarely, you need to compare not just with the concurrent control group, but with a historical control background rate. There are indeed problems with a comparison to a historical control group, but you do have the advantage that the historical control group is large enough that you can define precise estimates even for very rare events. The historical comparison should not replace the comparison to the concurrent control group, but should supplement it.

This is a very important message. The audience (including me, I’m afraid to admit) was focusing on the severity of the event. That is indeed important, but of greater importance is that certain severe events are also (thankfully) rare events. If an event is rare enough in the background population, then that rarity may make two events by themselves to be a very worrisome problem.

You can find an earlier version of this page on my old website.