Communicating expected uncertainty in a geostatistical survey to support co-design with users of information

Much research has examined communication about the uncertainty in spatial information to users of that information, but an equally challenging task is enabling those users to understand a priori measures of uncertainty for surveys of different intensity (and cost) at the planning stage. While statisticians can relate sampling density to measures of uncertainty, such as prediction error variance, these do not necessarily help information users (e.g. agronomists, soil scientists, policymakers and health experts) to make rational decisions about how much of the budget should be assigned to field sampling to produce information of adequate quality. In this exploratory study, we considered four ways to communicate uncertainty associated with predictions made based on data from a geostatistical survey, to determine an appropriate sampling density to meet an information user's expectations. The first method, offset correlation, is a measure of the consistency of kriging predictions made from data on sample grids with the same spacing but different origins. The second and third methods are based on the conditional prediction distribution: the second is the width of the prediction interval, whereas the third is the overall probability that, at a site where the true value of the variable indicates the need for an intervention, the contrary is indicated by the prediction. Fourth, the implicit loss function is a method that allows the user to reflect on the valuation of losses from decisions based on uncertain information implicit in selecting some arbitrary sampling density. All of these methods depend on the model of spatial dependence for the target variable, but they interrogate it in different ways and do not provide the same information. The evaluation of the four communication methods was carried out using a questionnaire that gathered the opinions of experienced participants (with experience in survey planning) about the effectiveness of the method and the comprehensibility of the uncertainty measure and its trade-off with the sampling effort. Our results show significant differences in how participants responded to the methods: the conditional probability and implicit loss function approaches were not well understood, whereas the offset correlation was the most understood. During feedback sessions, the information users highlighted that they were more familiar with the concept of correlation, with a closed interval, in this instance of [0, 1], which is likely to account for the more consistent responses regarding this method. Offset correlation will likely be more useful to information users with little or no statistical background and to those who are unable to express their requirements with respect to information quality based on other measures of uncertainty. However, the results should not be generalized due to the small sample size, and there is the need for a more in-depth study with a larger sample size to explore this further.