I am reminded about a study where thousands of people were shown to have improved outcomes by 20% when taking a cheap drug. It was an important study but the authors were asked to subdivide and they chose to use birth month to divide the total into twelve groups. This is a sequential type of sample within the full data set. They found that two of those sub groups had close to zero effect while a third subgroup had nearly double the overall average.

Sampling theory using, for example, binomial probability, would find that result to have near zero probability when the sub group is more than 1000 and the whole set is more than 12000 cases. Yet the authors successfully argued that this exercise demonstrated the need for large studies in epidemiology. Do others agree that actually it suggests that date of birth (at least in the mid lattitudes of the study group) has some effect on response to that drug?