‘Can You Fix It?’ Using Variance-Based Sensitivity Analysis to Reduce the Input Space of an Agent-Based Model of Land Use Change

Abstract

The growing body of knowledge on agent-based modeling (ABM) points to statistical and systematic uncertainty as the most challenging obstacles to developing parsimonious models. To decrease ABM dimensionality, a comprehensive uncertainty analysis (UA) and sensitivity analysis (SA) are employed, where input uncertainty is propagated through the model using Monte Carlo simulations, resulting in probability distributions (PDs) of outputs. The PDs are further summarized using variance—a simple yet succinct measure of result variability. The variance is apportioned to model inputs using decomposition, in order to quantify which of them and to what extent affect the variability of ABM results. The calculated sensitivity indices represent fractional contributions of each input to output variance. To simplify the model, inputs with low sensitivity values can be set to constants (e.g. mean), effectively decreasing model input space.

In this chapter, I apply UA and SA to simplify an ABM of agricultural land conservation, in which farmer agents make decisions on participation in a land conservation program. A positive decision leads to the conversion of land use from row crop/pasture to fallow. The results of the ABM are maps of land use change, which are summarized using a number of metrics, from total fallow land acreage, through various measures of compactness and contiguity, to cost of land retirement. The probability distribution of each output variable is used separately in variance-based SA, suggesting different ABM simplifications depending on the type of output variable. The chapter concludes with a list of practical guidelines on choosing the right path to building a transparent land use ABM using the presented UA and SA methodology.