Estimates of the production of adult salmon can be biased by assuming an incorrect form of random noise term in the descriptive model. Many authors of theoretical studies of stock–recruitment relations assume multiplicative, log-normal distributions of random noise, while other workers assume additive, normally distributed terms. Marine survival data for three species of salmon are used to determine which type of random variation term is most appropriate for salmon population models. Results indicate that a multiplicative, log-normal distribution is most consistent with the data. The random processes described by this type of noise term cause higher absolute variation in total adult returns as abundance of seaward migrants increases. This phenomenon creates difficulties in evaluating the success of artificial enhancement programs and it should be taken into account during planning of such programs and during subsequent evaluations of their success.Key words: marine survival, salmon, enhancement, log-normal, random, variability, parameter estimation, stock–recruitment, normal distributions