All statistical analysis of clinical trials involves pooling results from patients who will differ in terms of demographic characteristics or prognostic factors. It seems natural to want to try to establish the extent to which a treatment effect varies by subgroup. However, the number of potential subgroups can be large and the number of patients in a trial is often such that only pooling all of them will make it possible to reach a useful conclusion. Hence many problems arise when looking at subgroup effects, especially if the investigation is not guided by pre-specified approaches. In this paper, some pitfalls in examining subgroups are discussed and some possible approaches to investigating interaction illustrated using two different examples: the BHAT and the ATAC studies.