Uncertainty in site classification and its sensitivity to sample size and indicator quality – Bayesian misclassification rate in ecological risk assessment

The aim of this study was to quantify uncertainty when assigning field investigation sites according to their species community composition to either undisturbed or disturbed reference sites by use of ecological indicators. In ecological risk assessment this problem arises when selecting control investigation sites or defining reference species communities. Uncertainty is quantified using a Type II error or misclassification rate. A probabilistic Bayesian model is used to integrate a priori domain knowledge, assess the error rate and come to recommendations about an adequate sample size. Application is demonstrated using data from a case study investigating off-crop arthropod communities in German grassy field margins and consequences for impact assessment of pesticides on terrestrial ecosystems. The model allows calculating statistical power when using such a classification system. By means of stochastic simulations, recommendations about experimental design and indicator size are derived. The study shows that to develop a classification system to typify newly observed sites a well-balanced ratio of undisturbed and disturbed sites as well as a high relevance of reference sites are needed. For the given data set, a much larger number of reference sites as well as increased relevance of selected reference sites would be needed to achieve a good classification result. An optimal number of indicators is calculated allowing for a compromise between sampling error and indicator quality. Uncertainty for correct assignment of an investigation site is compared using indicators for disturbance and reference conditions. Finally, misclassification rate is proposed as a new measure for indicator quality.