Drug development and manufacture is a highly regulated industry. Most product tablet batches must pass the potency content uniformity test given in article <905> of the United States Pharmacopoeia (USP). The USP<905> test is a complicated two stage test with limits on 10 or 30 individual tablet potency results and their RSD (coefficient of variation). Consistent need for a second stage of testing or high batch failure rates result in higher analytical costs, failed batch disposal and replacement costs, loss of valuable active substance, supply shortage, and safety/efficacy concerns. When a new drug product is ready for launch, or when an existing product is modified, it is of interest to the FDA, potential patients, and the manufacturer to predict how often future batches of the product will fail the requirements of USP<905>.
The traditional approach to prediction, based on a one-way variance component model and inexact confidence bounds, will be described. Alternative approaches based on parametric bootstrap or Bayesian methods will be proposed. Some theoretical basis and implementation details will be given. Application of the traditional and alternative approaches to actual manufacturing data will be illustrated. Comparison of frequentist performance will be made using Monte-Carlo simulations. The results presented will indicate that the traditional approach should be used with caution and that modern computing capabilities and efficient sampling algorithms offer promising solutions to such risk assessment challenges.