It is difficult to obtain precise and accurate estimates of the true mycotoxin concentration of a bulk lot when using a mycotoxin-sampling plan that measures the concentration in only a small portion of the bulk lot. A mycotoxin-sampling plan is defined by a mycotoxin test procedure and a defined accept/reject limit. A mycotoxin test procedure is a complicated process and generally consists of several steps: (1) a sample of a given size is taken from the lot, (2) the sample is ground (comminuted) in a mill to reduce its particle size, (3) a subsample is removed from the comminuted sample, and (4) the mycotoxin is extracted from the comminuted subsample and quantified. Even when using accepted test procedures, there is uncertainty associated with each step of the mycotoxin test procedure. Because of this variability, the true mycotoxin concentration in the lot cannot be determined with 100% certainty by measuring the mycotoxin concentration in a sample taken from the lot. The variability for each step of the mycotoxin test procedure, as measured by the variance statistic, is shown to increase with mycotoxin concentration. Sampling is usually the largest source of variability associated with the mycotoxin test procedure. Sampling variability is large because a small percentage of kernels are contaminated and the level of contamination on a single seed can be very large. Methods to reduce sampling, sample preparation and analytical variability are discussed.

In this study, a number of probability distributions that have been used to model the occurrence of aflatoxin in peanuts are compared. Two distributions, the compound gamma and the negative binomial, are shown to have special appeal in that both can be justified by reasoning from the fundamental biological and stochastic processes that generate the aflatoxin. Since method of moments and maximum likelihood give consistent estimates of parameters in both models, practical considerations suggest using the former. One hundred twenty data sets, each consisting of fifty observations, were not sufficient to provide goodness-of-fit tests to establish either as superior to the other as a model. Both models fit the data well, appreciably better than other models examined. An attractive aspect of the compound gamma and the negative binomial distributions is that, as a consequence of their theoretical underpinnings, both involve parameters that have meaningful interpretations. In the compound gamma, the alpha parameter reflects the shape of the kernel-to-kernel aflatoxin content distribution, the lambda parameter reflects the number (or frequency) of contaminated kernels in the sample, and the beta parameter is a scale parameter. In the negative binomial, the two parameters can be used as measures of mean or location and shape.

The 1987 United States aflatoxin testing plan for shelled peanuts was designed with a final accept level of 25 parts per billion (ppb) total aflatoxin. Some of the importers of U.S. peanuts use aflatoxin testing plans with accept levels lower than 25 ppb. For example, the accept level of a testing plan used in The Netherlands is 5 ppb Bi or 10 ppb total aflatoxin. Whenever export lots are re-tested for aflatoxin by an importing country, some lots accepted in the United States will be rejected by the importing country's aflatoxin testing plan. Computer models were developed to determine the effects of decreasing the final accept level of the U.S. testing plan on the number of lots accepted and rejected in the United States and the number of exported lots accepted and rejected by The Netherlands testing plan. Decreasing the final accept level of the U.S. testing plan from 25 to 5 ppb increased the number of lots rejected in the United States by 371% while reducing the number of exported lots rejected by 51%. For every additional 8.3 lots rejected in the United States, one less export lot will be rejected.