Multimodel Bayesian analysis of data-worth applied to unsaturated fractured tuffs

To manage water resource and environmental systems effectively requires suitable data. The worth of collecting such data depends on their potential benefit and cost, including the expected cost (risk) of failing to take an appropriate decision. Evaluating this risk calls for a probabilistic approach to data-worth assessment. Recently we [39] developed a multimodel approach to optimum value-of-information or data-worth analysis based on model averaging within a maximum likelihood Bayesian framework. Adopting a two-dimensional synthetic example, we implemented our approach using Monte Carlo (MC) simulations with and without lead order approximations, finding that the former approach was almost equally accurate but computationally more efficient. Here we apply our methodology to pneumatic permeability data from vertical and inclined boreholes drilled into unsaturated fractured tuff near Superior, Arizona. In an attempt to improve computational efficiency, we introduce three new approximations that require less computational effort and compare results with those obtained by the original Monte Carlo method. The first approximation disregards uncertainty in model parameter estimates, the second does so for estimates of potential new data, and the third disregards both uncertainties. We find that only the first approximation yields reliable quantitative assessments of reductions in predictive uncertainty brought about by the collection of new data. We conclude that, whereas parameter uncertainty may sometimes be disregarded for purposes of analyzing data worth, the same does not generally apply to uncertainty in estimates of potential new data.

► Multimodel approach to optimum data-worth analysis within a Bayesian framework.
► Maximum likelihood approximations to reduce computational cost of MC simulations.
► Application to pneumatic permeability of unsaturated fractured tuff.
► Validation in selecting targets for additional data collection.
► Parameter uncertainty may be disregarded; data uncertainty must be considered.