The importance of multimodel projections to assess uncertainty in projections from simulation models

Abstract

Simulation models are increasingly used to gain insights regarding the longterm effect of both direct and indirect anthropogenic impacts on natural resources and to devise and evaluate policies that aim to minimize these effects. If the uncertainty from simulation model projections is not adequately quantified and reported, modeling results might be misleading, with potentially serious implications. A method is described, based on a nested simulation design associated with multimodel projections, that allows the partitioning of the overall uncertainty in model projections into a number of different sources of uncertainty: model stochasticity, starting conditions, parameter uncertainty, and uncertainty that originates from the use of key model assumptions. These sources of uncertainty are likely to be present in most simulation models. Using the forest dynamics model SYMFOR as a case study, it is shown that the uncertainty originated from the use of alternate modeling assumptions, a source of uncertainty seldom reported, can be the greatest source of uncertainty, accounting for 66–97% of the overall variance of the mean after 100 years of stand dynamics simulation. This implicitly reveals the great importance of these multimodel projections even when multiple models from independent research groups are not available. Finally, it is suggested that a weighted multimodel average (in which the weights are estimated from the data) might be substantially more precise than a simple multimodel average (equivalent to equal weights for all models) as models that strongly conflict with the data are given greatly reduced or even zero weights. The method of partitioning modeling uncertainty is likely to be useful for other simulation models, allowing for a better estimate of the uncertainty of model projections and allowing researchers to identify which data need to be collected to reduce this uncertainty.