What Statistical Models Can Better Detect Land-Change Mechanisms?

Statistical models are widely used to explore relationships between aggregate variables in a land-use system by inductively fitting empirical data, for example with regression models. Seldom do researchers know whether, and if so how much, the relationships thus obtained reflect the underlying mechanisms. This is partially due to the uncertainties in the system of interest, especially when the system is a complex adaptive system. This research addresses this issue by integrating a spatial agent-based model (ABM) and several regression models. Driven by a set of hypothetical residential location rules represented by associated parameters, the ABM generates a series of land-use patterns on a 2-dimentional virtual landscape over a number of time steps. For land pixels randomly sampled from the landscape, we recorded their emergent land-change dynamics, and calculated each pixel's characteristics (variables used in the ABM) of interest at selected time steps. Using these data, we applied survival analysis, logistic regression, and multivariate regression to discover the relationships between the land-change dynamics and the variables of interests. The findings suggest that these statistical models can capture the relationships encoded in the ABM in terms of some metrics (e.g., signs and significance status of regression coefficients) to varying degrees, and that survival analysis outperforms the other models. Our approach has substantial implications in calibrating and validating agent-based models, testing plausible scientific hypothesis, conducting sensitivity analysis, and many other model-based scientific endeavors.