|
Abstract:
|
When merging multiple longitudinal data, missing patterns are likely to be heterogenous between sources and homogenous within source. Existing work for handling missing data from multiple sources focuses on cross-sectional study. In this paper, we proposed a method to select variables when combining datasets for analysis based on multiple imputation, which impute missing covariates based on all covariates from complete data source and partial covariates from source observed including complete data source. We constructed estimating equations with imputed data for each source and then aggregate the information across sources. The objective function consisted of function of estimating equations (generalized method of moments) and SCAD penalty. We showed the consistency, sparsity, and normality of the proposed estimator with fixed number of covariates. Also, with tuning parameter selected by BIC criteria, the true model can be consistently identified. We use numerical experiments to evaluate the performance of the proposed methods and also apply the proposed method to a real data application.
|
