Statistical Problems in Monitoring Adverse Drug Reactions

Abstract

This paper proposes a probabilistic basis for monitoring adverse reactions to drugs, with allowance for the heterogeneities and irregularities inevitable among spontaneously submitted reports. The model partitions the population of reports into subpopulations. Each subpopulation includes clusters, a cluster being defined as a homogeneous group of reports on one drug set from one subpopulation in unit time. Signals from clusters can be more informative than those obtainable from data aggregated over the whole population.Reports in one cluster on a particular event (drug-reaction association) are assumed to conform to a Bernoulli distribution, with a parameter varying between clusters of the same subpopulation according to a two-parameter beta distribution and a cluster size that follows a negative binomial distribution. The four parameters of the compound model are estimated by maximum likelihood. The one-sided numerical cumulative sum test (NCST) is proposed as a signalling procedure; its performance characteristics are studied by Monte Carlo sampling. Data from the UK Committee on Safety of Medicines are used in an example.