Imputation methods for high-dimensional mixed-type datasets by nearest neighbors.

Area of Science:

• Biostatistics
• Data Science
• Biomedical Informatics

Background:

• Modern biomedical research frequently encounters datasets with numerous variables of mixed data types (continuous, multi-categorical, binary).
• Missing observations within these variables pose a significant challenge for downstream analyses requiring complete data matrices.
• Existing imputation methods often rely on specific distributional assumptions, limiting their applicability.

Purpose of the Study:

• To propose an advancement over the non-parametric nearest neighbor imputation method.
• To develop an imputation technique that effectively leverages the associations among variables without requiring specific distributional assumptions.
• To enhance the accuracy and performance of imputation for mixed-type data.

Main Methods:

• A weighted version of the Lq distance metric was developed for mixed-type data.
• The method utilizes information from a subset of important variables to inform the imputation process.
• Performance was evaluated using diverse simulated and real-world datasets from various application areas.

Main Results:

• The proposed weighted nearest neighbor imputation method demonstrated smaller imputation errors compared to existing approaches.
• The method achieved superior performance in imputation tasks across different datasets.
• Effective imputation was achieved even in scenarios with fewer samples than variables.

Conclusions:

• The novel imputation method offers a robust and effective solution for handling missing mixed-type data in biomedical research.
• Its non-parametric nature and efficient use of variable associations make it a valuable tool for data preprocessing.
• The method's performance, particularly in high-dimensional or low-sample settings, highlights its practical utility.