An agent-based model of household dynamics and land use change

Abstract

Climate change will alter patterns of land use, but its effects may also change social organization at multiple levels. This article proposes a spatially explicit agent-based model of migration land use to explore the dynamics of response to floods and droughts in Nang Rong, Thailand to explore this possibility.Within an environmental setting of marginality, villages and households vary in their vulnerability. The model integrates multiple inter-related units: individuals, land parcels, households, subfamilies, social networks, and communities. The inclusion of explicit social networks as resources to be mobilized is an especially innovative element. This article uses the language of mathematics and statistics to describe the model under construction and facilitate comparisons with other spatially explicit agent-based models.

Keywords:

• agent-based model
• land use change
• migration
• social networks
• Thailand

1. Introduction

Across multiple social and spatial scales, marginal populations are particularly likely to be affected by extreme weather events and climate change, partly because of their attachment to marginal environments. At the global level, projected impacts of climate warming show countries that are already at significant disadvantage globally to be the most likely to suffer from flooding associated with future sea-level rise, e.g. Bangladesh. At a regional level, when Hurricane Katrina struck the southern coast of the US, already marginalized populations were disproportionately affected. In New Orleans, the lower ninth ward was devastated and recovery has been extremely slow. The contrast with Bourbon Street is stark.

In the examples just given, the joint distribution of the population with respect to environmental vulnerability and economic and social disadvantage is an important part of the story. The poor are relegated to more vulnerable locations, which expose them to greater risks, which in turn make it difficult to accumulate the resources to improve economically. Disasters large and small have the potential to exacerbate inequalities at multiple levels. Low resilience to endogenous and exogenous dynamics affects vulnerability through feedbacks and places the poor and disadvantaged at the edge of chaos.

Processes involved in the perpetuation of environmental vulnerability and economic and social disadvantage operate globally and regionally, but also locally. Our study focuses at the local level, examining the dynamics of responses to extreme weather events currently to develop some intuition about local response in the context of future global climate change.

The focus of our study is the dynamics of response to flood and drought in Northeast Thailand. We are building a spatially explicit agent-based model that relates household dynamics to land use change and that incorporates social networks and community characteristics as endogenous elements. The model draws on analyses of out-migration, remittance behavior, return migration, marriage, and subsequent residence decisions of household members, of land use at the parcel and household level, and of the consequences of migration, remittances, and land use for household assets. Social ties between households, which are crucial to the sharing of resources and the exchange of help in difficult times are critical to the processes we examine in networks. The centrality of households within village-based kin networks and the cohesiveness of these networks within villages are hypothesized to affect migration, remittance, return migration, marriage, and subsequent residence decisions as well as land use. In turn, households become more and less central, and villages more and less cohesive, in response to demographic change.

This paper has two purposes. The first is to structurally describe an agent-based model which is under construction that incorporates rich social and biophysical data in a spatially explicit manner. The inclusion of explicit social networks as resources to be mobilized is an especially innovative element. The second is to use the language of mathematics and statistics to describe our model and to facilitate comparisons with other spatially explicit agent-based models.

2. Background

To anticipate the potential impact of increased intra- and inter-annual variability in rainfall as a consequence of global climate change requires us to consider a wide range of potential human responses (cf. Davis 1963), and feedbacks involving these responses. The latter are likely to be especially significant at the local level. In response to frequent flooding, or prolonged drought, individuals might move away in search of an alternative or supplementary livelihood. Those who have already moved may be more likely to send resources back in response to need, but probably less likely to return, at least in the short run, disrupting patterns of circular migration. Marriages may be postponed, or accelerated, depending. Land currently used may be used in a different way, or abandoned; land not currently in use may acquire new commercial value (e.g. Bilsborrow 1987). Households might get poorer, but some are likely to be better off. Social networks condition the response, whether it be a change in migration, marriage, land use, or some combination. Social networks are also changed by these behaviors, becoming more or less cohesive, more or less permeable. Social inequality may widen, or narrow, with consequent implications for parcel fragmentation or consolidation.

Typically, responses to social and environmental change are studied in isolation, e.g. a change in land use (Caldas et al. 2007) or increased out-migration, even though these responses are likely to be inter-related: land use and its productivity affect household resources, which affect and are affected by migration patterns. Further, because land use and migration are both cause and consequence of the other, using standard regression approaches to examine hypotheses about the effect of one on the other requires heroic assumptions about the errors (specifically, the assumption of no covariance between these errors and the included independent variables). Without statistical ‘fixes’, results will be biased and incorrect. More importantly, considering elements of a larger more complex reality one at a time provides a myopic and largely static view of the dynamics of human behavior. Our holistic approach uses agent-based models to explicate dynamic responses to exogenous local-scale shocks, including potential tipping points where the entire system changes.

As an approach and methodology, agent-based models are at the forefront in many fields. Our agent-based model builds on and integrates these disciplinary traditions and perspectives. In geography and land use science, spatially explicit agent-based models have been developed to describe land use change and landscape dynamics in a variety of settings (see Parker et al., this volume). Our model includes these elements but also includes feedbacks involving migration, changes in household composition, and impacts of agent interaction. We are able to pursue feedbacks from land use to and through migration because we have an appropriate data set and because we will explicitly build them into the model. Agent-based models of land use change have less demographic detail at the household level (e.g. Rouchier, Bousquet, Barreteau, Le Page, and Bonnefoy 2001; Bharwani et al. 2005; Brown, Page, Riolo, Zellner, and Rand 2005) and have not incorporated important feedbacks. Demographic change is treated as an external ‘driver’. Demographic change is included endogenously in agent-based models of mobility, residential segregation, and tipping points (e.g. Bruch and Mare 2006; Macy and van de Rijt 2006; also see Schelling 1971, 1972). To date, however, these models have not been made spatially explicit (Macy and Willer 2002) or linked to land use. Explicit and endogenous social networks have not yet been joined with community effects in any of these models, quite possibly because the data needed to do this have not been available heretofore (Entwisle 2007).

We also incorporate a systems perspective. In contrast to marginality, which may characterize land, households, or communities, resilience is a characteristic of a system. Study of system dynamics is significant because it can uncover relations between drivers and system changes that are specific and explanatory for a system while simultaneously revealing relations of pattern and process that are shared among systems more generally. Patterns and drivers can be analyzed together in multidimensional state space where areas of particular sensitivity can be observed and analyzed (Wilson 1978); newer tools for quantification allow us to move beyond visualization of these system dynamics (Malanson, Zeng, and Walsh 2006). While complexity analyses in some fields are able to use many alternative realizations of a given system through simulations, we too can do this for each village, and, because each village can be analyzed as a separate realization of the underlying dynamics, we can tie our simulations back to the real Nang Rong much more powerfully than in other land use simulations because we have so many villages.

3. Setting, site, and situation

Nang Rong District, Thailand, our study site, serves as a laboratory to explore interactions among people, place, and environment through theories, data, and methods emanating from the social, natural, and spatial sciences (Entwisle, Walsh, Rindfuss, and Chamratrithirong 1998; Walsh, Rindfuss, Prasartkul, Entwisle, and Chamratrithirong 2005). Our multi-dimensional data have been explicitly gathered and processed to examine the complex interplay between people and the environment. The data come from multiple sources – demographic surveys, administrative records, maps, satellite images, aerial photographs, and field observations. The data are multilevel. On the social side, they cover individuals (including migrants), households, and villages for the period from 1984 to 2000/01. On the spatial side, they cover pixels, plots, village territories, and the district, 1954 to the present. The data are linked, over time, across scales, and across social, biophysical, and spatial domains (e.g. Walsh et al. 2005). Our data are unique in having complete, spatially referenced social networks for 51 villages (Rindfuss et al. 2004a; Entwisle et al. 2007), making it possible to incorporate social networks as an endogenous element in a way that is a real departure from the past (Entwisle 2007).

The district occupies approximately 1300 km² in Northeast Thailand. See Map 1. The environmental setting is one of marginality. Soil fertility is relatively low, drainage is poor, and generally speaking, there is a limited natural resource base. The climate is monsoonal, with rains arriving late spring–early summer. The rice harvest occurs in December, followed by a long dry season with few agricultural activities. Precipitation, however, is quite unpredictable. Figure 1 plots rainfall patterns in Nang Rong over the past 50 years. The trend is clearly downward; and the variance has increased. During the decade of the 1990s, there were both years of significant wetness (1993, 1999) and years of significant drought (1992, 1997).

Map 1. Nang Rong, Thailand.

Figure 1. Nang Rong 0.05 Degree Grid Cell: Annual Precipitation, 1950–1999.

The rhythm of life in Nang Rong is governed by the monsoon and seasonality of agriculture and also by major ups and downs in the Thai economy. Through the 1980s and most of the 1990s, the Thai economy grew at a remarkable rate. This growth was concentrated in Bangkok and the Eastern Seaboard, primarily in the manufacturing and service sectors (Singhanetra-Renard and Prabhudhanitisarn 1992; Phongpaichit 1993; Pejarandonda, Santipaporn, and Guest 1995; Mills 1999), and was fueled by migration from rural areas, including Nang Rong. Growth was sharply curtailed by the devaluation of the Thai currency in 1997, although there was some recovery by 1999 (ESCAP 1999; World Bank 1999). Migration, both permanent and temporary (including seasonal), is common (Fuller, Lightfoot, and Kamnuansilpa 1985; Fuller, Kamnuansilpa, and Lightfoot 1990; Guest 1998; Korinek, Entwisle, and Jampaklay 2005). Nang Rong farmers must deal with shocks originating in the national and international economy as well as those environmental in nature.

Both villages and households vary in their vulnerability. In Nang Rong, villages consist of clusters of dwelling units surrounded by agricultural lands. The ideal site for a village is a small hill or rise proximate to lowlying lands suitable for paddy rice cultivation (Riethmuller, Scholz, Sirisambhand, and Spaeth 1984; Phongphit and Hewison 2001). Figure 2 shows the location of villages in Nang Rong relative to inundation patterns revealed in a rare July 2000 satellite image from the rainy season (typically, clouds obscure the imagery during the rainy season). Except for villages located in the southwest, which was settled most recently in response to an increase in world demand for cassava, an upland crop, villages line up along the edges of the inundated area. It is worth mentioning that 2000 was a relatively wet year. Indeed, a few villages appear to be so close to the inundated zone that flooding might have been a problem.

Figure 2. Nang Rong Villages and Inundated Areas.

Figure 3 shows the location of the plots farmed by residents of a few specific villages in our data set. In many (although not all) instances, farmers living in villages just beyond areas of inundation are farming plots within the inundated zones. A significant advantage of the Nang Rong data is the ability to link households and plots and to inter-relate the experiences of household members, household wealth, plot characteristics, and land use outcomes. It is difficult to see in Figure 3, but in fact, the plots farmed by village residents vary significantly in their relative elevation, slope, and accessibility to water, the main biophysical characteristics that determine suitability for rice cultivation in this context (Phongphit and Hewison 2001). In years of flood or drought, only some of the plots would produce successful harvests – which ones would depend on the particular event, its severity, and its geographic reach.

Figure 3. Inundated Areas and Parcel Linkages: Villages Center-to-Parcel Links.

More and more, farmers in Nang Rong are engaged in market agriculture. Although most farmers grow rice, both a subsistence and cash crop in this setting, some also grow upland cash crops such as cassava and sugar cane. In recent years, more farmers have been growing tree crops such as eucalyptus and para-rubber, as well as fruit trees. The agent-based model will include all three – rice, upland crops, tree crops – as land use choices. What farmers choose to grow will depend on their resources and plot characteristics, including their title to the land. Titles can range from the equivalent of squatters' rights to a clear title permitting the land to be sold, mortgaged or passed on to the next generation. The model assumes that there is no market for land, which is currently the case as shown by qualitative data collected in 2004. The model does, however, allow land to be divided as new households are formed and the older generation dies off. We do expect a land market to emerge eventually, and this will be a topic of interest to explore in future extensions (see paper by Parker et al., this volume).

4. Agent-based approach

Agent-based simulations are microsimulations in which agents – in our case, households – interact dynamically with one another and with their environment to produce change at multiple levels. In these simulations, agents are autonomous decision-making entities that have unique characteristics, are represented by algorithms, and interact with each other according to rules that define the relationships between agents and their environment. These types of models allow us to develop candidate explanations for specific landscape patterns, spatially simulate landscape patterns over space and time, assess pattern–process relations, and examine likely future scenarios of change, including an increase in the frequency of extreme weather events accompanying global climate change.

Feedbacks are fundamental to the dynamism of human and ecological systems. These feedbacks are of two types. One involves endogenous relations among key variables: for example, the risk of migration depends on household assets, which in turn depend on prior productivity of specific plots of land; use of specific plots depends on household assets, which in turn depend on prior migration and remittance behavior. Interconnections among migration, land use, and household assets operate across individual household members, households as collective units, kin networks and villages on the social side and the geographic site and situation of specific land plots and village territories on the spatial side. Our modeling approach thus solves a general problem in the population-environment field, i.e. lack of fit between spatially discrete social data and spatially continuous environmental data (Rindfuss, Walsh, Turner, Fox, and Mishra 2004b; Entwisle and Stern 2005). Such problems are exacerbated in settings such as Nang Rong, where households are clustered in villages, away from the land they farm, but are present in all settings.

The other type of feedback involves interaction among agents, i.e. the influences that neighbors have upon each other. Migration is cumulative (Massey et al. 1993; Massey 1999), such that the behavior of one potential migrant helps to define the context of possibilities for subsequent potential migrants. In empirical research, this feature of migration is typically captured through measures of the percent of potential migrants in a locale who have migrated (e.g. Massey and Espinosa 1997). Social network ties are also important to migration (Massey, Alarcón, González, and Durand 1987; Grasmuck and Pessar 1991; Stark 1991; Hondagneu-Sotelo 1994; Pessar 1999; Curran and Saguy 2001; Palloni, Massey, Ceallos, Espinosa, and Spittel 2001; Brown 2002), and are endogenous in a similar way. Normative expectations and social influences are thought to operate with respect to remittance behavior as well (e.g. Vanwey 2004). The model incorporates these influences by including changing measures of social network ties at the household and village level, as well as incorporating dynamic measures of migration, return migration, and remittance behavior into the evolving village network structure. The influence of neighbors on land use is incorporated similarly. Feedbacks involving neighborhood composition have been incorporated in agent-based models of mobility and residential segregation (Bruch and Mare 2006; Macy and van de Rijt 2006), but we know of no simulations involving migration and remittance behavior.

5. Overview of the computational model

We now provide a heuristic overview of the spatial simulation that is at the heart of the model we are constructing; we provide a more formal, mathematical explication below.

Our agent-based simulation iterates over multiple types of inter-related units: individuals, land parcels, households, social networks, and communities. Households are a point of integration for the model: individuals come together to form households, which are embedded in social networks and communities; land parcels are owned, managed or used by households. Villages are composed of households, and social networks consist of ties among those households. The model assumes that households have decision-making authority for the management of land parcels, but could later be modified to include communities, businesses or other types of organizations. The model makes no assumptions about the nature of household decision-making (e.g. autocratic, democratic, consensus).

Each element of the model has attributes and can experience demographic, social, economic, and/or biophysical processes. Individuals are characterized by age, sex, marital status and years of education. They experience the possibility of mortality, out-migration, return migration, marriage, establishing a new residence locally, and, for women, giving birth. When not residing in the community they can remit to the origin household, influencing that household's assets. The agent-based model computes, that is, simulates, each of these processes on an annual basis, updating and taking into account attributes of the individual as well as the individual's household, social network ties to other individuals and households, and community.

Land parcels also have attributes, e.g. size, distance from the village, potential for flooding and topographic settings, land use type, and soil suitability for various agricultural uses. Depending on these attributes, household resources, and environmental factors such as the timing and amount of rainfall, the household makes a choice about how to use the land parcel, e.g. rice cultivation, and experiences some level of productivity. Since rice is the dominant crop it receives focused attention, but parcels may be used for other crops as well.

Individuals and parcels are non-nested units, but households influence and are influenced by both, serving to link the two. The basic demographic processes of fertility, mortality, migration, and marriage determine the composition of the household. The extent to which individual migrants remit affects wealth. Household composition and wealth, in turn, influence how households use land parcels, and the use of land parcels can subsequently affect the household's wealth as well as indirectly affect its composition.

Households are nested in villages. Village characteristics affect households, the decisions they make about land use, and also the behavior of individual members. As households change, the aggregated characteristics of villages change.

Social networks are the social ties among individuals and, through them, households in a village. We focus first on kin ties, but other kinds of ties can be considered. When offspring marry and create new households, new ties between households are created. In other words, the behavior of individuals has consequences for the entire web of inter-relationships among residents of a village. When individuals leave the village, or die, ties between households weaken or disappear. Households vary with respect to their position within these networks, e.g. central vs. isolated. Villages vary with respect to the overall structure of the network, e.g. more or less cohesive. Social networks appear in the model as characteristics of households and of villages, updated after each iteration.

It should be evident that causality runs directly from individuals to the household and indirectly to land parcels, as well as in the reverse direction. It also runs upward from individual to household to social network to community, and from plot to household to landscape, and back again. The model does not attempt to partition the causality as one might do in an instrumental variables statistical approach. Instead, the proposed agent-based simulation allows for feedbacks among the model elements as it relates patterns to processes, within a complex adaptive system context. Thus, the simulation is focused on the dynamics of the system.

The processes just described are estimated within the agent-based model annually using a set of stochastic operations or ‘rules’. These operations or rules follow one of four approaches, depending on the process: classic demographic projection approach; multiple regression approach; arithmetic approach; and scenario testing approach. We now briefly describe each.

The projection approach is straightforward. Consider mortality. For each individual, we will use the q_i column (which is the probability of dying at age i) from a single year life table to calculate which individuals die in year t. We will use period life tables available for Thailand as a whole, adapting them to mortality conditions in Nang Rong. This projection approach will be used for mortality, fertility, and education.

The regression approach involves analyzing our Nang Rong data, building on previous analyses as well as undertaking new analyses, while drawing on the relevant substantive and theoretical literature. Consider out-migration. We will estimate a regression model that predicts out-migration based on characteristics of individuals, their households, social networks, and communities. The regression coefficients will then be used in the agent-based simulation to determine who migrates out in a given year. The regression approach will be used for individual out-migration, return migration, the sending of remittances, marriage, and postnuptial residence; for land parcel use and productivity; and for household wealth and household splits. The possibility of biased coefficients coming from the regression analyses (see above arguments) is handled through sensitivity testing of the model's parameters.

The projection and regression approaches will be used in conjunction with an arithmetic approach to derive household characteristics (e.g. number of members 13–55) and village characteristics (e.g. percent of migrants remitting) on an annual basis. For example, measures of household age structure will depend on projection approaches for mortality and fertility, and regression approaches for out-migration, return migration, marriage, postnuptial residence, and household formation. Arithmetically, births will be added, deaths subtracted and so forth. Everyone ages a year at a time. The model keeps track of those currently in the household and those who are not.

The operation of social networks in the agent-based model is realized through changes in measures of household ties such as network centrality and measures of network structure at the village level such as cohesiveness. Centrality is measured as the number of other households linked by close kin ties to the reference household. Cohesiveness is measured as the average number of households mutually reachable by close kin ties within a village. These and other measures are calculated from a sociomatrix (see below) that incorporates the results from each iteration of the model and then is used to update the social network measures. At each subsequent iteration, their effects are included in the same way as other household and village variables in the regression approach.

Finally, the scenario testing approach will allow us to run ‘experiments’ altering components in the model or in any of the three prior approaches. Our main interest is the impact of exogenous deleterious events, e.g. droughts, floods, and economic crisis.

6. Social networks

An innovation in our modeling is the incorporation of endogenous social networks. Social ties are the conduits along which information and resources flow (e.g. Cain 1978, 1983; Coleman 1988; Friedkin 1982), and according to which social influence is exerted (e.g. Portes and Sensenbrenner 1993). Social networks are structured collections of such ties. Our previous research, and also the general literature, shows that the position of individuals and households within village-based networks and network structure at the village level may affect decisions about migration, remittance, marriage, and land use (Korinek et al. 2005; Jampaklay 2006; Rindfuss, Kaneda, Chattopadhyay, and Sethaput 2006; Entwisle, Faust, Rindfuss, and Kaneda 2007; Tong and Piotrowski, in review). What has not been addressed is the impact of these outcomes on the network position of individuals and households or the structure of networks at the village level. Village-based social ties are weakened by migration, and severed by death. New ties are created through marriage, household formation, and fertility. The omission of feedbacks from demographic outcomes to social network structure and position is a fundamental shortcoming of the literature on social network effects and has the potential to bias results. Our project is the first to incorporate these feedbacks.

Kin networks are crucial to the sharing of resources and the exchange of help. Kinship connections are based on the location of the individual's spouse, parents, and siblings. In the data, we have information about each person's mother and father; for married residents, their spouse; and for people aged 18–35 years, their siblings. If these relatives were not in the individual's household but were in the village, we determined the household in which they resided. If they were outside the village, we determined the village (if in Nang Rong District), district, other province in Thailand, or the country, if abroad (Rindfuss et al. 2006). We use this information to initialize conditions for each village. Marginal households strongly linked to other households in the village may be better able to weather bad years, in the long as well as the short run. Living in a highly cohesive village may likewise be a benefit. The position of individuals and households within networks as well as their overall structure is both cause and consequence of decisions to move, to remit, to return, to marry, to form a new household, and to have children.

Operationally, both for deriving measures from the data and as part of the dynamic simulation, we begin with a sociomatrix for each type of kinship tie (Wasserman and Faust 1994). In the sociomatrix rows and columns index individuals within a village and the entry in a cell records the presence or absence of the specific kinship relationship between the row individual and the column individual. For each of the 51 villages, there are two basic sociomatrices of first-degree kinship ties: parents and spouses. Second-degree kinship ties are then constructed from the first degree ties using matrix multiplication. For example, if the matrix P records ties from child to parent and S records ties to spouse, then the matrix product SP gives ties to a spouse's parents. Taking gender of parent into account differentiates mother-in-law from father-in-law. Combining parent ties, P, and child ties, C, the matrix product PC gives sibling ties. From these networks we construct individual level measures of ties to kin of different degrees. For each person in a village we determine not only whether they have parents or a spouse in the village, but we also count the number of first- and second-degree kin in the village and subsequently aggregate these up to the household level. We break the counts down by the gender of the relative to yield separate counts of the numbers of female and male kin. We also determine whether there is a first- or second-degree kin link to a household in the top 10% of the asset distribution in the village (Edmeades, Entwisle, Rindfuss, and Thongthai, 2005). We then collect together the ties of household members to create a sociomatrix at the household level. Measures of village cohesion summarize at the village level the structure of ties between households within a village (e.g. Entwisle et al. 2007).

7. Model description

An overview of the proposed agent-based simulation was given above. Here we provide a more formal description, using the language of mathematics. There are three components: the individual component, the land use component, and the wealth component. We will discuss them in this order. It is important to remember that the model allows for feedbacks from household assets to both individual behavior and land use, from individual behavior to assets, and from land use to assets. It also includes feedbacks from all of these to the community or village level.

All the equations below are for the agent-based simulation, and hence there are no error terms. The β’s, γ ’s, and so forth represent coefficients from the regression results. Variables and coefficients are both subscripted. Subscripts track specific variables (K variables at the individual level, L variables at the household level, Q variables at the parcel level, and S variables at the village level) and units of observation (individual i, in household j, in village v, or parcel p used by household j in village v, and time t). The subscripting of coefficients according to unit of observation indicates that a small amount of individual stochasticity is added to each to maintain heterogeneity in the system. For example, the coefficient indicating effects of individual characteristics on nuptiality ζ _kijv in equation (1)(1) below include subscripts for the individual as well as that person's household and village. This individual subscript indicates that there is stochasticity associated with the behavior of that individual (e.g. the individual might be unusually risk-averse). The standard approach is to lump all stochasticity (i.e. that due to errors in the equations, to random coefficients associated with the characteristics of the individual agents, and to measurement error) in the selection of a choice or outcome of the equation by applying a pseudorandom number test to the probabilities for each choice (or to use coarser scale selection of uses among multiple parcels or within watersheds; e.g. Hoffman, Kelley, and Evans 2002; Sengupta et al. 2005). We believe, however, that it will be more informative to include this stochasticity more explicitly on the right-hand side of the equations for each decision and to allow those unmeasured characteristics of, say, individuals to operate in the same way over time. Following up on the previous example, we expect individuals who are unusually risk-averse at one point in time to also be unusually risk-averse at other points in time as well. We do not show equations for the projection or arithmetic operations in the agent-based model, nor for the regression equations. We list the variables that will be used, such as age, education or plot size, but to keep this paper a manageable length we do not describe their operationalization.

7.1. Behavior of household members

Each household consists of a set of I members, each of whom is characterized by his or her age, sex, marital status, and years of education. Our first simulations initialize the household roster with data from the 1994 household survey.

Household members and their characteristics are incremented annually by subjecting each one to a sequence of conditional risks. The projection approach is used for mortality, fertility and education, using standard techniques (Shryock and Siegel 1976; Preston, Heuveline and Gulliot 2001). We use national data (e.g. survival probabilities and fertility rates), except for education, for which we use probabilities based on our Nang Rong data. All risks are age- and sex-specific, and all involve a small stochastic element to introduce individual heterogeneity.

We begin with mortality. Each member at time t–1 is subjected to a risk of dying. If they die, they are removed from the household roster for time t and all subsequent times. If they survive, age in time t is incremented by one. If a female household member is married and aged less than 45, we use the rural Thai age-specific, parity-specific fertility schedule to subject them at time t–1 to the risk of having a birth. If they have a birth, that child is added to the household. Since non-marital fertility is very unusual in rural Thailand, unmarried women are not subjected to the parental risk. If a surviving member is of school age, they are subjected to a risk of continuing their education. If they do, then years of education is incremented by one for time t; if not, then this characteristic does not change. To keep things simple, we model youngsters as in school until the end of compulsory education. In subsequent work, we will model obtaining education past the compulsory level.

The first component of the agent-based model that incorporates regression results is nuptiality. Surviving household members aged 15 + and never married are at risk of marrying. The nuptiality equation in the agent-based simulation is:

(1)

where N_ijvt is the nuptiality probability for individual i of household j of village v at time t. There are multiple variables at the individual, household and village level. To keep equation (1)(1) and subsequent equations manageable, the subscripts k, l, and s on the ζ’s and X’s indicate multiple individual, household, and village variables, respectively. The use of subscripts i, j, and v for the ζ’s represents heterogeneity in the effect of the individual, household, and village variables. This heterogeneity is assumed constant over time. The subscripting of the ζ’s also indicates that these are the equations used in the simulation, not those used to estimate the coefficients.

Since marriage is close to universal in rural Thailand, equation (1)(1) primarily affects the timing of marriage. The variables included in equation (1)(1) are based on the rich literature on age at marriage in Thailand (Chamratrithirong, Morgan, and Rindfuss 1988; Guest and Tan 1994; Jampaklay 2006; Limanonda 1979, 1983, 1992). Important individual-level variables include age, gender, education and prior migration history. Household wealth along with other household variables is included. At the village level, we will compute the mean and variance of age at first marriage for those resident in the village in 1996 and who first between 1996 and 2000. Social networks are also important. Ties outside the village expand the marriage market and hasten marriage. These are measured at both the household and village level. These measures capture the extent to which there are different marriage cultures across villages. We assume stable marriages. Divorce is not very common in rural Thailand.

In Northeast Thailand, residence with the wife's parents is the preferred arrangement in the first few years of marriage, but there is flexibility (Chamratrithirong et al. 1988). Some live in the husband's parents' house, others establish their own nuclear household, and some migrate. A major factor determining post-nuptial residence is the wealth of the bride's and groom's families, as well as local job opportunities (Chamratrithirong et al. 1988). So in a year a person marries they need to be assigned one of the following post-nuptial residence outcomes: (a) remained in origin household, (b) established a nuclear household in origin village, and (c) left the origin village:

(2)

where PNR_ijvt refers to the three post-nuptial residence categories and the φ coefficients are obtained from a multinomial logistic regression analysis. If the married person stays in the origin household, then we add the spouse to the household using classic census hot-deck procedures to impute the characteristics of the spouse (Shryock and Siegel 1976; Weisberg 2005). If a new household is established or the couple does not reside in the village, then the person who married is removed from the household.

Unlike such East Asian countries as China or Japan, in rural Thailand it is common for a newly married couple to reside in the household of the bride's or groom's parents for a while, and then move out, setting up their own household. Sometimes the marriage of a younger sibling or the birth of a child can be the triggering event. Sometimes it is simply having accumulated sufficient wealth to permit the establishment of a new household. Indeed, as the general level of wealth in Nang Rong has increased, average household size has decreased (5.5 persons per household in 1984 and 3.8 in 2000). Within the agent-based model, subfamilies, that is recently married couples and their children, are subjected to the risk of leaving the household and setting up their own household as follows:

(3)

where HS_fjvt is the probability of subfamily f splitting from household j in village v in time period t. The ’s come from regression analyses that build upon the work by Chamratrithirong et al. (1988), Piotrowski (2006), and Rindfuss, Piotrowski, Thongthai, and Prasartkul (2007) that shows the importance of wealth, and anything that might contribute to the accumulation of wealth in a setting transitioning from a subsistence to a monetized economy, as well as demographic events to household members (marriage, migration, fertility, and mortality). We are evaluating two different assumptions about the division of land: (1) assign half of the land used by the origin household to the new household; (2) assign a proportionate amount that depends on the number of siblings still living in the origin household. Based on qualitative fieldwork done in 2004, there is not much of a market for land in Nang Rong. Inheritance is the main mechanism of transfer.

The allocation of labor between nonagricultural activities outside the village and household-based farm activity in the origin village is one of the most important decision points in the model. All household members between the ages of 15 and 55 will be subjected to the possibility of out-migrating for a job in a factory, on a construction site, in the service sector, or as an agricultural laborer in some location outside Nang Rong. This is based on the age, sex, marital status, and education of each member (see above), the size and age structure of the household (as calculated from above), its wealth (see below), ties to other households (see above), and village characteristics, including the Gini coefficient for inequality of wealth distribution in the village, a measure of village network cohesion, and the percent of 1984 residents who are no longer resident in 1994. The regression model is based on work using the Nang Rong data (Adamo and Rindfuss 2005; Korinek et al. 2005; Vanwey 2005) as well as the general literature on internal migration from rural areas in developing countries (e.g. Harris and Todaro 1970; Lucas and Stark 1985; Stark and Taylor 1989, 1991; Massey et al. 1993). The out-migration equation in the proposed agent-based simulation is:

(4)

where M_ijvt is the migration probability for individual i of household j of village v at time t.

If the household member is a migrant, either as a result of a decision to move in year t or some earlier year, the risk of remittance is calculated as:

(5)

where the γ ’s come from a regression equation with a small random element added to introduce heterogeneity. The individual variables are age, sex, marital status, years of education, and years since migration began. Household variables are amount of land owned, ownership of various assets (such as a motorcycle or sewing machine), whether there is a nonagricultural worker in the household, number of children aged 12 and under, number of persons aged 55 and under, and number of first- and second-degree kin ties to other households in the village (see below). Village variables include a measure of social network cohesion and the proportion of migrants from the village who are remitting. The variables included in equation (5)(5) and the underlying regression analysis are based on prior analyses of the Nang Rong data (Vanwey 2004; Korinek and Entwisle, 2006) and the general literature on remittances (e.g. Dinerman 1978; Stark and Lucas 1988; Massey et al. 1993; De Jong, Richter, and Isarabhakdi 1996; Osaki 2003).

Household members who are migrants are also subject to a risk of return to the origin household:

(6)

where the α’s come from a regression equation with a small random element added to introduce heterogeneity. The individual variables are the same as those for remittances, except that the village variable is the proportion of migrants from the village who return. Again, equation (6)(6) and the empirical analysis of return migration on which it is based, is informed by prior work (Korinek, Entwisle, and Jampaklay 2004, 2005; Tong and Piotrowski, under review) and the more general literature on return migration (e.g. Ramos 1992; Lindstrom and Massey 1994).

At this point all household members, including migrants, have been updated to t for mortality, out- and return migration, fertility, age, education, marriage, postnuptial residence, and remittances. The arithmetic component of the agent-based simulation will update the household roster, creating counts of the number of migrants and the number of migrants remitting as well as a set of aggregated household characteristics.

7.2. Land use

There are three steps to this part of the model. In the first step, we determine whether the household is engaged in agriculture. In the second step, conditional on household involvement in agriculture, we consider specific plots farmed by that household. For each plot, we determine whether rice or some other crop is grown on that particular plot. In the third step, for those plots where rice is grown, we determine how much rice is produced. These steps translate into three equations.

Whether or not the household is engaged in agriculture is hypothesized to depend on its demographic structure, its resources, and village characteristics:

(7)

where A is the probability of engaging in agriculture for household j of village v at time t. The household variables include the aggregated characteristics of the members such as the number of members aged 13–55 and the number over 55, the amount of land owned, and the ownership of various assets. It also includes first- and second-degree kin ties to other households which, for households headed by older individuals, may provide for the household's needs. The village variables are location relative to Nang Rong town and the presence of a rice bank in the village. The coefficients for the agriculture decision equation are based on a binomial logistic regression using our data.

If a household is engaged in agriculture, its choice of land use for each plot used depends on the characteristics of the plot, the household, and the village:

(8)

where U is the probability of one of three land uses (i.e. rice, other, and fallow), for plot p of household j of village v at time t. The plot variables are its crop suitability, plot size, and plot distance to the household's dwelling unit. The household variables are number of plots farmed by the household, agricultural assets such as ownership of a thresher, and the number of members aged 13–55. The village variables are existence of a rice bank, location relative to Nang Rong town, and the percent of plots in rice and in other crops. The coefficients for the land use equation are based on a multinomial logistic regression using our data.

If the plot is used for rice, its yield is hypothesized to be a consequence of parcel characteristics, household labor and resources, and village characteristics:

(9)

where Y is the agricultural yield in m³/ha, for plot p of household j of village v at time t. The plot variables are its crop suitability, location, size, and use of chemical fertilizer, pesticides, and/or herbicides. The household variables are the number of plots farmed by the household, agricultural assets, and the number of members aged 13–55. The village variables are the percent of households using chemical fertilizers, pesticides, and herbicides and location relative to Nang Rong town. The coefficients for the land use equation are based on a regression model using our data from 2000, which includes plot level information on rice harvested.

7.3. Wealth

Asset ownership, and its determinants and consequences, is the final component of the agent-based model. In the year-to-year simulations, the wealth of the household is an important feedback to household dynamics because assets affect the likelihood of out-migration and potentially increase as a consequence of remittances; they affect the likelihood that a household grows rice, and the productivity of rice if they do, as well as increasing or decreasing as a result of the harvest. We measure household wealth in terms of ownership of productive agricultural assets, consumer assets, and mixed assets, measured in terms of their value in Thai currency (baht). Because each type of asset responds somewhat differently to migration and remittances (Entwisle and Tong 2005), and in addition, likely affects land use decisions somewhat differently, we consider each separately in the simulations, but for purposes of exposition just show one equation here.

Assets are modeled as a function of household characteristics, the number of migrants, the number of remitters, the total amount of rice produced on the parcels farmed by the household, the area of plots in some other agricultural use, and village characteristics.

(10)

The analyses in equation (10)(10) and the underlying regression analyses are based on prior analyses of the Nang Rong data (Entwisle and Tong 2005) as well as the general literature on the impact of remittances (especially Rozelle, Taylor, and Debrauw (1999) and Taylor, Rozelle, and De Brauw (2003); but also, e.g. Durand, Parrado, and Massey (1996), Lindstrom (1996), and Massey and Parrado (1998)).

8. Implementation

The agent-based simulation is under construction as a spatially explicit, object-oriented program. We are using the REPAST (Recursive Porous Agent Simulation Toolkit) ABS platform, because it provides a tool kit that is advanced, flexible, and allows us to develop the experiments that we need (Tobias and Hofmann 2004; Robertson 2005; Railsback, Lytinen, and Jackson 2006). REPAST is a free opensource tool kit initially created at the University of Chicago and now maintained by the Argonne National Laboratory, and managed by the non-profit group ROAD (Repast Organization for Architecture and Development). Through links to ESRI ArcGIS software, Agent Analyst (i.e. agent-based modeling extensions for ArcGIS), and Python database environment, REPAST is supported through spatially explicit databases, visual interface, agent templates, an automated Monte Carlo simulation framework, dynamic access and modification of agent properties, agent behavioral equations, libraries of genetic algorithms, social network modeling tools, and an integrated GIS.

In the simulation, each household has individuals and parcels of land attached to it, and as shown above, the equations will be applied to the members and the parcels as well as the household itself. Time steps will be annual, corresponding to the agricultural cycle and patterns of seasonal off-farm employment. The t–1 to t step begins with the monsoon in spring and runs through the long dry season. At each iteration, we will apply the projection, regression-based, and arithmetic equations for each household independently. Interaction among households occurs when a household at time t evaluates the land use of all other households in the village at t–1, including use of fertilizers, pesticides, herbicides as well as the particular crops grown. It occurs when household members evaluate the migration behavior of individuals living in other households and when migrants evaluate the remittance behavior of others from their village. And it occurs when household members are influenced by kin ties to others in the village, migrants by the centrality of their origin household within these networks, and both by the cohesiveness of these networks at the village level. All variables are updated as necessary for use in the next iteration. We will simulate a single village at a time, and for its initial conditions each village will match one of the 51 survey villages.

Compared to previous agent-based models of land use, our model involves a much more detailed treatment of household dynamics (including nuptiality, migration, return migration, remittances, and household fission as endogenous elements) as well as the inclusion of feedbacks involving social ties between households and social networks within villages. Doing this complicates the modeling – what is gained? To answer this question, we will model outcomes with and without key feedbacks, in essence ‘breaking’ parts of the model to see what difference it makes. For example, we will examine land use and household assets dynamically in the simulation, but treat migration and remittances as exogenously (as others have treated them), rather than endogenously. To do this, we would remove the migration feedback equations ((2)(2) and (3)(3)) from the agent-based model. We will also examine migration, remittances, and household assets dynamically in the simulation, but treat involvement in agriculture, land use, and the productivity of plots exogenously rather than endogenously (removing 4–6 from the model). Comparing these model outcomes to those based on a full complement of feedbacks will speak to the ‘value-added’ of the agent-based simulation approach, and more specifically, the particular elements that we have chosen to highlight.

9. Conclusion

In many settings, intra- and inter-annual variability in precipitation is a dominant factor in agricultural production. Certainly this is true in Nang Rong, where the success of the rice crop depends on the timing and amount of the annual monsoon. Risk in agricultural production is due in large part to unpredictable precipitation patterns. Unpredictability, i.e. variability in precipitation patterns, is likely to increase with climate change. Floods and droughts are likely to become more frequent, more severe. To study the short- and long-run effects of these changes, and response to them, we are building a spatially explicit agent-based model, linking people to their environment in important new ways. Migration, marriage, household formation, remittance behavior, land use, and wealth accumulation are all endogenous, constituting potentially important feedbacks involving human behavioral response. Significantly, we incorporate dynamic social networks. The position of individuals and households within networks as well as their overall structure is both cause and consequence of decisions to move, to remit, to return, to marry, to form a new household, and to have children. Our goal is to understand system response.

Agent-based simulations make it possible to study patterns at the village and landscape level as an outcome of the migration behavior of individuals and land use behavior of households. Feedbacks and nonlinearities in the system make it unlikely that these patterns are a simple aggregation of results obtained in regression analysis. The complex nature of systems emerges from nonlinearities due to interactions involving feedbacks occurring at one or more lower levels within the system (Cilliers, 1998; Malanson, 1999; Crawford, Messina, Manson, and O'Sullivan 2005). Complex systems not only evolve through time, their past is co-responsible for its present behavior. In the Nang Rong setting, households interact through endogenous and exogenous processes to create a dynamic system that is space and time dependent, where feedbacks between human activities, social change, land use change, and ecological dynamics have the potential to produce nonlinearity. Social inequality, i.e. the marginalization of some households and villages more than others, is an emergent phenomenon within this approach (Krugman 1995).

Climate change will alter patterns of land use, but ultimately, its effects may reach far beyond, potentially changing social organization at multiple levels. Marginal households are likely to farm marginal land to begin with, so are likely to be particularly vulnerable to increases in risk. Social inequalities may increase. We expect that the degree to which this occurs will depend on the patterning of social ties, however. Marginal households strongly linked to other households in the village may be better able to weather bad years, in the long as well as the short run. Living in a highly cohesive village may likewise be a benefit. We cannot predict now exactly how things will turn out because the patterning of social ties changes over time and these changes will be in part due to how households respond to increased risk. To see whether feedbacks contribute significantly to our understanding of environmental marginality and socioeconomic marginality and to resilience as a system characteristic, in future work we will selectively break these likes and assess the consequences for model performance.

Acknowledgments

We would like to thank Philip McDaniel for preparing the maps and Bridget Riordan for assistance with manuscript preparation. The National Institute for Child Health and Human Development supported the development of the model and the collection of data on which it is based (R21 HD051766, R01 HD25482); the National Science Foundation (BCS-0728822) supported its application to questions of social and environmental marginality.