ABSTRACT
Sociologists often argue that social context matters. Features of the social context, not just the characteristics of individuals, help produce aggregate outcomes such as the distribution of economic rewards, or paths of development. Multilevel designs where individuals are nested within social contexts provide a strong design for observing both contextual effects and the aggregate outcomes those effects might produce. We present an analysis of migration in rural Thailand, in which survey respondents are
nested within villages, providing annual reports on migration for the 1980s and 1990s. Rural-urban migration has propelled economic development as rural migrants remit their earnings back to their villages and return with news of economic opportunities for friends and family members. Though our data describe thousands of individual migration decisions, our interest focuses on aggregate differences across villages. The rural northeast of Thailand varies tremendously in the degree to which villages are integrated into the urban economies further south. The evolution of inequality in migration across villages is thus important for our understanding patterns of poverty and development in the rural areas of countries experiencing rapid growth. Hierarchicalmodels provide avaluable tool for studyingmultilevel sociological data such
as the Thai migration surveys (Mason et al., 1983; Western, 1999). In sociology and demography, panel surveys of individuals and households, survey data frommany countries, and pooled time series data from US states and cities have all been analyzed with hierarchical models (DiPrete and Forristal, 1994). Sometimes sociological applications have studied the heterogeneity of parameters across units, thoughmore commonly hierarchicalmodels offer away to account for clustering in inferences about fixed parameters. In these cases, random effects are a nuisance, integrated out for correct inference. Hierarchical models are common in sociology, but applied research often neglects two
important topics. First, sociological analysis of hierarchical models rarely provides a detailed examination of model fit. In our analysis of the Thai migration data we study the fit of several alternative models by comparing the deviance information criterion (DIC) and posterior predictive statistics. Model fit is an important applied topic because sociological theory is often indifferent to alternative specifications of random effects. The structure of random effects may also have important implications for substantive conclusions. In
like inequality in a response across units or response variable quantiles-may be sensitive to the specification of random effects. A second limitation of applied sociological research with hierarchical models is that these aggregate implications of model estimates typically go unexamined. Our analysis of rural-urban migration in Thailand examines several hierarchical models. In our analysis, Markov chain Monte Carlo (MCMC) computation for hierarchical models provides a convenient framework for studying aggregate patterns of variation by simulating migration given different hypothetical distributions of covariates.
The Thai migration data are based on the Nang Rong Survey∗ of men and women aged 13-41 from 22 villages in the Nang Rong district of northeastern Thailand (Curran et al., 2005). We combine data from twowaves (1994 and 2000) of the life history survey. The 1994 wave begins withmen andwomen aged 13-35 in 1994, and asks about respondents’ migration experiences since the age of 13. This design is replicated in 2000: men andwomen aged 18-41 are asked about their migration behavior starting at the age of 13. Some respondents were living away from the village at the time of the survey, and they were followed up and interviewed.† We merge these data with household censuses conducted in 1984, 1994, and 2000 to obtain household and village characteristics. The resulting data contain information onmigration of 6768 respondents nestedwithin 22 villages over a 16-year time period from 1984 to 2000 (N = 93,914). Our interest focuses on how the level of migration in a village might subsequently pro-
mote more migration among individuals. Figure 24.1 shows the distribution of village migration rates, y¯jt = ∑i yijt/njt, from 1984 to 2000. The survey data are retrospective, and the age distributions vary over time. The figure displays the migration rates for men and women aged 18-25, the age group that we observe every year. Migration rates generally increase until 1996. In 1984, around a quarter of young residents in Nang Rong left their district for at least two months. By 1996, the migration rate for the region had increased to about 50%. In 1996, the Asian financial crisis precipitated recession in Thailand. Migration rates declined over the next four years. In some villages, migration declines were particularly steep, with migration rates falling to around 10%. Trends for a high-migration and low-migration village are also shown in the plot. These trends share some common features, such as the increase in migration in the first decade and the decline from 1996. Part of our substantive interest focuses on how the accumulation of migration experi-
ences within villages is associated with an individual’s likelihood of migration. Migration for an individual may become more likely if they live in a village in which many others have migrated. This phenomenon, called the cumulative causation of migration, occurs because prior migration generates resources or influence that make individuals more likely
to migrate (Massey, 1990). Extensive empirical evidence documents how past migration becomes a primary engine for future migration flows, eventually diminishing the importance of alternative explanations (Garip, 2008; Massey and Espinosa, 1997; Massey and Zenteno, 1999). We study the effect of social context by a constructing a “village trips” variable that
records the number of trips taken in a village in the years preceding the current year. A scatterplot of village trips and annual village migration rates for the 1984-2000 period is shown in Figure 24.2. In any given year, villages with the highest migration rates have a history of high levels of migration. This pattern is not surprising, but it remains an open empirical questionwhetheravillage’shistoryofmigration is associatedwithan individual’s likelihood of migration, after accounting for their own history of migration, their family’s migration history, and other covariates. To study the effect of village trips for these multilevel data we write several hierarchical
logistic regression models. For respondent i (i = 1, . . . , ntj) in village j ( j = 1, . . . , 22) in year t (t = 1984, . . . , 2000), yijt denotes the binary migration outcome, taking the value 1 if the respondent travels away from the village for more than two months in the year, and 0 otherwise. Individual-and village-level covariates are collected in vectors, xijt and zjt. In each of the following logistic regressions, yijt, conditional on fixed and random effects collected in the vector θ, is assumed to be Bernoulli, P( y|θ) = py(1− p)1−y, with expectation E( y) = p and likelihood L(θ; y) = ∏P( yijt|θ). If we consider only the panel aspect of the data design, we can fit a respondent-level
random effect, αi, to allow for the correlation of observations for the same respondent,
yielding the logistic regression:
