|
Abstract:
|
Mixed-effects models are popularly applied in statistical practice with longitudinal data for which a challenge arises from the missing values on either outcome or predictor variables. For the purposes of model selection, we propose two imputation-based strategies: ITS (impute, then select) and SIAS (simultaneously impute and select). ITS applies the method of multiple imputation directly to first generate several complete data sets, then conduct Bayesian variable selection on each imputed data set, and finally combine the selection results. SIAS embeds the data imputation and variable selection steps within one integrated Gibbs sampler. By applying the two strategies with various Bayesian variable selection algorithms to both practical and simulated data sets, we evaluate their performance and compare them with traditional methods (step-wise, AIC, BIC, etc.).
|
