Characterizing location preferences in an exurban population: implications for agent-based modeling

Powerful computational tools are becoming available to represent the behavior of complex systems. Agent-based modeling, in particular, facilitates an examination of the system-level outcomes of the heterogeneous actions of a set of heterogeneous agents: for example, patterns of land-use and land-cover change, such as urban sprawl as a result of residential location decisions. These new tools create new demands for data, and empirical studies of the selection behavior of residents. Using resident responses from the 2001 Detroit Area Study survey, we compared two alternative approaches to characterizing the heterogeneous preferences of agents; both based on a factor analysis of resident responses to questions about their reasons for moving to their current location. We used cluster analysis to identify how many and what types of residents there are, grouped by similar preferences. We also evaluated the relationships between socioeconomic and demographic characteristics and location preferences using regression trees, and evaluated the fit of the relationship to determine the degree to which socioeconomic characteristics predict preferences. The results showed that the preferences of resident exurbans of single-family homes in the Detroit metropolitan area were heterogeneous and that distinct preference groups do exist in resident populations, but are not well characterized on the basis of simple socioeconomic and demographic variables. We conclude that, given the heterogeneous nature of preferences and a relatively limited number of preference groupings observed in the survey respondents, agent-based models simulating resident behavior should reflect this diversity in the population and incorporate distinct agent classes of empirically derived preference distributions. DOI:10.1068/b3071 }Corresponding author: US Environmental Protection Agency, Office of International Affairs, 1200 Pennsylvania Ave, Mail Code: 2650R, Washington, DC 20460, USA; e-mail: fernandez. [email protected]. Recent developments in, and applications of, agent-based simulation in land-use modeling have created new opportunities for simulating microlevel (that is, individual and household) behaviors and evaluating their implications at aggregate levels, but also new requirements for data (Parker et al, 2003). The multiple useful roles of empirical data in informing these models remain to be completely determined. Empirical support for agent-based models can generally take either or both of two forms: (a) using observations about aggregate outcomes to evaluate or calibrate the model or (b) using data on individual or agent decisions to generate agent behavior. In the case of aggregate outcomes, it might be desirable simply to observe that some phenomenon results from a simple model of behaviors, as in the case of Page's (1999) model of city formation. Alternatively, model output can be compared with maps of the observed pattern using some spatial descriptor, such as the power-law distribution of cluster sizes (Batty and Longley, 1994; Rand et al, 2003) or other summaries of patch characteristics (Parker and Meretsky, 2004). At the level of agent behaviors, role-playing games have been used to collect information about decisionmaking behaviors for use in models that simulate these behaviors (Barreteau, 2003; Barreteau et al, 2001). Alternatively, and less expensively, agent behaviors are often determined through deductive reasoning (for example, Brown et al, 2004; Otter et al, 2001). Though this approach can be satisfying in the case of very simple models, and the sensitivity of models to assumptions can be evaluated, the importance of diversity in agent behaviors in complex systems (for example, Rand et al, 2002) suggests that it is worth some effort to characterize empirically the heterogeneity in an agent population. Indeed, representing diversity among agents is one of the important strengths of agent-based modeling, but it may be important to understand not just that diversity of behaviors exists, but something about its distribution and relation to agent characteristics. The research presented here supports a project that aims to develop agent-based models of land-use and land-cover change at the urban ^ rural fringe to help improve our understanding of human ^ environment interactions within this dynamic zone. The agent-based models developed in the project to date (Brown et al, 2002; Rand et al, 2002; 2003) model land-development patterns as a function of agent preferences for nearness to services and jobs, landscape aesthetic quality, and particular levels of density. The models have been populated with agents that have either identical preferences or preferences drawn from a single population, that is, characterized by a mean and standard deviation. The relative importance of preferences in determining the agent locations has been evaluated through parameter sweeping, that is, evaluating model behavior at different relative levels of preference. Alternative approaches might be to draw preferences from multiple populations, representing agent types, or to give each agent a unique set of characteristics and preferences. In the latter case, it might be possible to assign each agent a preference as a function of its socioeconomic and demographic characteristics (for example, age, marital status, race, and income). To judge the most realistic approach, we analyzed data collected through a survey of stated residential preferences conducted within the Detroit metropolitan area in 2001. The analysis involves comparing two alternative approaches to characterizing the heterogeneous preferences of agents; both are based on a factor analysis of resident responses to questions about their reasons for moving to their current location. First, we used cluster analysis to identify how many and what types of residents there are, grouped by similar preferences. Second, we evaluated the relationship between socioeconomic and demographic characteristics and location preferences by using regression trees, and evaluated the fit of the relationship to determine the degree to which socioeconomic characteristics predict preferences. The results of the two analyses were compared to answer the basic question: to what degree are preferences independent 800 L E Fernandez, D G Brown, R W Marans, J I Nassauer