Subsemble estimation has attractive properties for statistical modeling: it has been shown to perform as well as the oracle model asymptotically at minimizing the difference in estimation risk, and dramatically reduces the computational complexity of analyzing a single spatially correlated random variable.
This research explores application and modifications of subsemble estimation to effectively estimate and predict in multivariate spatial models. We consider weighting schemes for combination algorithms based on the spatial subsample quality as well as modeling techniques for accuracy and gains in computational speed. The goal of this work is to develop intelligent methods for partitioning spatial data while preserving spatial structure, determine whether spatial structure and balance are useful weighting measures for combining subsemble estimators, and investigate whether gains in computational efficiency are balanced by potential losses in prediction accuracy. These techniques are evaluated using simulation studies and applied to examples from epidemiological data.