How to Address Measurement Noise in Bayesian Model Averaging

When confronted with the challenge of selecting one out of several competing conceptual models for a specific modeling task, Bayesian model averaging is a rigorous choice. It ranks the plausibility of models based on Bayes’ theorem, which yields an optimal trade-off between performance and complexity. With the resulting posterior model probabilities, their individual predictions are combined into a robust weighted average and the overall predictive uncertainty (including conceptual uncertainty) can be quantified. This rigorous framework does, however, not yet explicitly consider statistical significance of measurement noise in the calibration data set. This is a major drawback, because model weights might be instable due to the uncertainty in noisy data, which may compromise the reliability of model ranking.

We present a new extension to the Bayesian model averaging framework that explicitly accounts for measurement noise as a source of uncertainty for the weights. This enables modelers to assess the reliability of model ranking for a specific application and a given calibration data set. Also, the impact of measurement noise on the overall prediction uncertainty can be determined.

Technically, our extension is built within a Monte Carlo framework. We repeatedly perturb the observed data with random realizations of measurement error. Then, we determine the robustness of the resulting model weights against measurement noise. We quantify the variability of posterior model weights as weighting variance. We add this new variance term to the overall prediction uncertainty analysis within the Bayesian model averaging framework to make uncertainty quantification more realistic and “complete”.

We illustrate the importance of our suggested extension with an application to soil-plant model selection, based on studies by Wöhling et al. (2013, 2014). Results confirm that noise in leaf area index or evaporation rate observations produces a significant amount of weighting uncertainty and compromises the reliability of model ranking. Without our suggested extension, this additional contribution to prediction uncertainty could not be detected and model ranking results would be misinterpreted. We therefore advise modelers to include our suggested upgrade in the Bayesian model averaging routine.