Title: Computationally efficient sequential regression imputation for multilevel datasets
Authors: Murat Recai Yucel - State University of New York (United States) [presenting]
Tugba Akkaya-Hocagil - State Univervisty of New York at Albany (United States)
Abstract: An alternative approach based on variable-by-variable imputation is considered when either the joint model is nonsensical due to non-applicability or due to high dimensionality. Joint models can be nonsensical in situations when survey item can only be applicable to certain sub-groups. We improve computational aspects of the previous methods relying on traditional sampling techniques such as Gibbs sampler under linear mixed-effects models. In particular, we eliminate the adverse effect of increased inter-dependency on random-effects and its impact on convergence. The improved algorithm greatly reduces computation times in the context of variable-by-variable imputation in surveys or other data problems with large number of variables subject to missingness. We present a comprehensive simulation study as well as real data example from nationally-conducted survey.