Sensitivity Analysis in Agent-Based Models of Socio-Ecological Systems: An Example in Agricultural Land Conservation for Lake Water Quality Improvement

Socio-ecological systems are dynamic and nonlinear. To account for this complexity, we employ agent-based models (ABMs) to study macro-scale phenomena resulting from micro-scale interactions among system components. Because ABMs typically have many parameters, it is challenging to identify which parameters contribute to the emerging macro-scale patterns. In this paper, we address the following question: What is the extent of participation in agricultural land conservation programs given heterogeneous landscape, economic, social, and individual decision making criteria in complex lakesheds? To answer this question, we: [1] built an ABM for our model system; [2] simulated land use change resulting from agent decision making, [3] estimated the uncertainty of the model output, decomposed it and apportioned it to each of the parameters in the model. Our model system is a freshwater socio-ecological system - that of farmland and lake water quality within a region containing a large number of lakes and high proportions of agricultural lands. Our study focuses on examining how agricultural land conversion from active to fallow reduces freshwater nutrient loading and improves water quality. Consequently, our ABM is composed of farmer agents who make decisions related to participation in a government-sponsored Conservation Reserve Program (CRP) managed by the Farm Service Agency (FSA). We also include an FSA agent, who selects enrollment offers made by farmers and announces the signup results leading to land use change. The model is executed in a Monte Carlo simulation framework to generate a distribution of maps of fallow lands that are used for calculating nutrient loading to lakes. What follows is a variance-based sensitivity analysis of the results. We compute sensitivity indices for individual parameters and their combinations, allowing for identification of the most influential as well as the insignificant inputs. In the case study, we observe that farmland conservation is first and foremost driven by the FSA signup choices. Environmental criteria used in FSA offer selection play a secondary role in farmland-to-fallow-land conversion. Farmer decision making is mainly influenced by the willingness to reduce the potential annual rental payments. As the case study demonstrates, our approach leads to ABM simplification without the loss of outcome variability. It also shows how to represent the magnitude of ABM complexity and isolate the effects of the interconnected explanatory variables on the simulated emergent phenomena. More importantly, the results of our research indicate that some of the parameters exert influence on model outcomes only if analyzed in combination with other parameters. Without evaluating the interaction effects among inputs, we risk losing important functional relationships among ABM components and, consequently, we potentially reduce its explanatory power.