Recommendation	Rationale
Longitudinal observational data should be further explored as a complement to signal detection using individual case reports but cannot currently replace individual case reports for this purpose	Individual case reports of suspected harm from medicines have a proven value for safety signal detection. However, they are not optimal for detecting increased rates of multifactorial adverse drug reactions with high background incidence. Longitudinal observational data provide the basis for epidemiological evaluation of such associations and should in principle enable their initial identification. However, we lack evidence to suggest that signal detection in longitudinal observational data can match the performance of signal detection in individual case reports for all drugs and medical events. In our evaluation of historical safety signals from the EMA, none of the positive controls could be detected in the THIN database at an early stage, whereas this was possible in VigiBase for some of the signals, even when we considered only the subset of the UK individual case reports [72]
Safety signal detection in longitudinal observational data should include clinical, pharmacological and epidemiological review of identified temporal associations	Clinical review of statistical signals is fundamental in evaluating signals arising from spontaneous report databases. In our study of structured assessment for prospective identification on safety signals in electronic health records, three out of four temporal associations identified in the initial screen could be dismissed from further evaluation after initial review. Without review, the majority of the highlighted associations would have been false positives [72]
To the extent possible, temporal associations detected in longitudinal observational data should be further explored with statistical graphical methods	In our prospective identification study, in-depth review of the chronograph temporal patterns proved a valuable component of the expert review. Univariate measures of temporal association may over-simplify or obscure the underlying patterns in such rich, complex and often long records [72]
Safety signal detection in longitudinal observational data should account for limitations of the underlying data and take measures to ensure appropriate interpretation. In selecting the data set for analysis, one should account for both its size and scope (which drugs and diagnoses it captures) and for the fact that effective review of identified temporal associations requires expert knowledge of the underlying data, which is particularly relevant for large heterogeneous data sets	Our retrospective evaluation against historical safety signals for European centrally authorised products showed that none of them could be detected in THIN with the method we used, prior to the initial signal at the EMA. In many cases (to be further specified once we have the data), this was because of the drug not being available on the UK market at the time, or the drug or medical event not being reliably captured in primary care
Future research should explore the relative merits of performing safety signal detection in longitudinal observational data for groups of medicinal products and medical events, instead of or in parallel with that of individual products and events	In our comparison to published epidemiological studies, a common discrepancy was that they performed analysis for all drugs in a class together and/or for a number of related medical events together, which improves power. However, our detailed review often found substantial and important differences among different drugs in the same class or among different medical events in the same category.