Bayesian Case Series — Bayesian Case Series Analysis

The standard SCCS uses classical (frequentist) conditional Poisson regression and produces a maximum-likelihood estimate and confidence interval. A Bayesian case series places a prior distribution on the incidence rate ratio and produces a full posterior distribution. The Bayesian version formally incorporates prior evidence, supports sequential updating as new data arrive, and yields a credible interval that can be directly interpreted as a probability statement about the parameter.

A weakly informative log-normal prior on the log incidence rate ratio — centred at 0 (corresponding to IRR = 1) with moderate variance — is a common default that allows the data to dominate unless they are very sparse. If historical data from similar compounds or pre-clinical studies exist, a more informative prior can be justified, but sensitivity analyses under alternative priors should always be reported so readers can assess how much the conclusions depend on prior assumptions.

Yes, but with important caveats. Spontaneous reporting systems suffer from under-reporting, reporting biases (notoriety bias, Weber effect), and absence of denominator data. Bayesian case series applied to spontaneous reports can estimate disproportionality but not true incidence rate ratios. Methods like Bayesian Information Component (BIC) or the Gamma Poisson Shrinker were specifically developed for spontaneous reporting databases and are more appropriate for that data source.

There is no universal minimum, but extremely sparse data — fewer than five events in total across all cases — will yield highly uncertain posteriors regardless of the prior. In practice, Bayesian case series is most useful when at least 10–30 events are available. The Bayesian framework handles small samples better than frequentist SCCS because the prior stabilises estimates, but a genuinely underpowered study cannot be rescued by the statistical framework alone.

Standard case series methods assume either a single non-recurrent event or independent recurrent events. If events are recurrent and the occurrence of one event affects the probability of a subsequent event (e.g., seizures), the independence assumption is violated. Extended SCCS models for recurrent dependent events have been developed by Farrington and colleagues and should be used instead of the standard formulation.