Semiparametric imputation using conditional Gaussian mixture models under item nonresponse.

Area of Science:

• Statistics
• Data Science
• Econometrics

Background:

• Item nonresponse is a common challenge in surveys, requiring robust imputation techniques.
• Parametric imputation lacks robustness, while nonparametric methods struggle with high-dimensional data (curse of dimensionality).
• Semiparametric imputation offers a balance of robustness and flexibility.

Purpose of the Study:

• To propose a novel semiparametric imputation method that is more flexible than existing Gaussian mixture models.
• To address the challenges of high-dimensional covariates in imputation.
• To enhance the accuracy and applicability of imputation techniques in complex surveys.

Main Methods:

• Developed a new semiparametric mixture model with a conditional Gaussian assumption for the study variable.
• The marginal distribution of auxiliary variables is not restricted to Gaussian.
• Incorporated a penalty function into the conditional log-likelihood for high-dimensional covariate applicability.

Main Results:

• The proposed mixture model demonstrates greater flexibility and better approximation compared to Gaussian mixture models.
• The method effectively handles high-dimensional covariate problems.
• Successful application to the 2017 Korean Household Income and Expenditure Survey.

Conclusions:

• The proposed semiparametric imputation method offers a robust and flexible alternative for handling item nonresponse.
• It provides a valuable tool for analyzing complex, high-dimensional survey data.
• The method enhances the reliability of statistical inferences from incomplete datasets.