Multilevel spatial and statistical analyses to examine the relationship between population and environment: A case study of the Ecuadorian Amazon

Multilevel spatial and statistical analyses are conducted to examine population-environment relationships in the Ecuadorian Amazon using a representative sample of migrant colonists and communities in 1990 and 1999. An ipsative response vector representing the proportion of land allocated to five uses (forest, pasture, perennial crops, annual crops, and swamp or fallow) is specified for each farm and the number of children born to women between 1990 and 1999 are the outcomes of interest. Statistically, the goals are to: (1) examine the application of constrained multilevel models; (2) compare and contrast implicit and explicit methods for controlling spatially correlated observations; and (3) decompose variance to control for nested data in models of land use and fertility. Contextually, the goals are to: (1) identify the most relevant spatial neighborhood(s) for sampled households; (2) fit multilevel models to explore community-level factors on land use and fertility; and (3) evaluate the determinants of fertility in a rural agricultural frontier using spatial exploratory tools and statistical models. These studies demonstrate the importance in controlling for hierarchical and spatial effects in population-environment research and in properly defining spatial neighborhoods for rural areas. A series of spatial neighborhoods are examined for land use and fertility models, which create the hierarchical data structure. Convergence problems for the ipsative land use data using restricted maximum likelihood (REML) led to the implementation of iterative generalized least squares (IGLS), although diagnostics revealed deviation from model assumptions. Spatial autocorrelation is controlled implicitly in the multilevel model by decomposing variance into place and neighborhood effects, and explicitly in the spatial seemingly unrelated regression (spSUR) using spatial weight matrices. The spSUR is favored to estimate the multidimensional feedback occurring between autocorrelated observations, but since farms are located in clusters that are typically separated by long distances, the multilevel model is preferred since autocorrelation is minimal. If the clusters were contiguous or located in closer proximity, the same conclusion may not have been reached. Finally, spatial exploratory analysis and multilevel Poisson models demonstrate the severe lack of family planning resources available for women to control their fertility in these frontier environments.