Agent Based Modeling of Land Use Change: The Case of Shade Coffee in Mexico.
- Author(s): Marcos-Martinez, Raymundo
- Advisor(s): Baerenklau, Kenneth A.
- et al.
This research focuses on addressing methodological issues that impact the performance of spatially explicit discrete choice agent-based land use models that are estimated with remotely sensed data. The empirical setting considers land use transitions between agroforests, perennial crops, grass and corn, and fallow lands during the period 1984 - 2006 in a Mexican coffee growing region in which relatively high deforestation rates were observed. As a starting point, a Mixed Conditional - Multinomial Logit model is implemented to highlight assumptions and limitations associated with this standard modeling approach. The results indicate that this model produces theoretically inconsistent parameter estimates for the revenue variable associated with three out of four land uses considered in the analysis. To investigate whether those counterintuitive marginal effects are generated from misclassified land use data, a Latent Multinomial Logit (LMNL) model is implemented. This approach allows the identification of land use observations that have a high likelihood of being wrongly classified. A reconfiguration of the dataset based on the LMNL model increased the magnitudes of the marginal effects of the analyzed land use drivers in the theoretically expected directions. Next, because static land use models require limiting assumptions that potentially oversimplify the behavioral process followed by landowners, a structural dynamic discrete choice model of land use decisions is implemented under the assumption that land managers are forward-looking and act to maximize their discounted flow of current and future expected utility within a stochastic environment. A comparison between static and dynamic models shows that the directions of the marginal effects corresponding to time-invariant parcel-specific variables generally have the expected directions independent of the selected modeling approach. More importantly, the marginal effect estimates for the revenue variables of the agroforestry and perennial crops categories have the expected direction in the dynamic model. By contrast the myopic modeling approaches generate counter-intuitive results for the revenue variable that corresponds to perennial crops production, which affects the validity of those results for policy design. Finally, a policy simulation exercise shows the sensitivity of welfare estimates to the discount factor selected as representative of the true value used by decision makers.