Rules for assessing compliance with percentile standards commonly limit the
number of exceedances permitted in a batch of samples taken over a defined
assessment period. Such rules are commonly developed using classical stati
stical methods. Results from alternative Bayesian methods are presented (us
ing beta-distributed prior information and a binomial likelihood), resultin
g in "confidence of compliance" graphs. These allow simple reading of the c
onsumer's risk and the supplier's risks for any proposed rule. The influenc
e of the prior assumptions required by the Bayesian technique on the confid
ence results is demonstrated, using two reference priors (uniform and Jeffr
eys') and also using optimistic and pessimistic user-defined priors. All fo
ur give less pessimistic results than does the classical technique, because
interpreting classical results as "confidence of compliance" actually invo
kes a Bayesian approach with an extreme prior distribution. Jeffreys' prior
is shown to be the most generally appropriate choice of prior distribution
. Cost savings can be expected using rules based on this approach. (C) 2001
Elsevier Science Ltd. All rights reserved.