Bayesian Melding of a Forest Ecosystem Model with Correlated Inputs

Abstract

Abstract Bayesian melding, a method for assessing uncertainties in deterministic simulation models, was augmented to make use of prior knowledge about correlations between model inputs. The augmentation involved the use of a nonparametric correlation induction algorithm. The modified Bayesian melding technique was applied to the process-based forest ecosystem computer model PnET-II. The Bayesian posterior distribution for this analysis did not reflect prior knowledge of input correlations for five input pairs tested unless the correlations were explicitly accounted for in the Bayesian prior distribution. For other input pairs not known to be correlated prior to the analysis, numerous significant posterior correlations were identified. For one such pair of model inputs, a moderate posterior correlation was substantiated by empirical evidence that had not previously been taken into consideration. We conclude that, when possible, efforts should be made to account for prior knowledge of correlated inputs; however, Bayesian melding may elucidate input correlations in its posterior sample, even when no prior knowledge of such correlations exists. FOR. SCI. 48(4):701–711.