MODELING STAKEHOLDER PREFERENCES WITH PROBABILISTIC INVERSION

摘要

A panel of 64 experts ranked 30 scenarios of human activities according to their impacts on coastal ecosystems. Experts were asked to rank the five scenarios posing the greatest threats and the five scenarios posing the least threats. The goal of this study was to find weights for criteria that adequately model these stakeholders' preferences and can be used to predict the scores of other scenarios. Probabilistic inversion (PI) techniques were used to quantify a model of ecosystem vulnerability based on five criteria. Distinctive features of this approach are: 1. A model of the stakeholder population as a joint distribution over the criteria weights is obtained. This distribution is found by minimizing relative information with respect to a noninformative starting distribution, but makes no further assumptions about the interactions between the weights for different criteria. Criteria distributions with dependence emerge from the fitting procedure. 2. The multicriteria preference model can be empirically validated with expert preferences not used in fitting the model.