[Simulation study on missing data imputation methods for longitudinal data in cohort studies]

Area of Science:

• Biostatistics
• Epidemiology
• Data Science

Background:

• Missing data is a common challenge in longitudinal cohort studies.
• Effective imputation methods are crucial for reliable data analysis and subsequent statistical modeling.

Purpose of the Study:

• To compare the performance of eight common missing data imputation techniques.
• To evaluate their impact on longitudinal data and subsequent multivariate analyses.
• To provide guidance for selecting appropriate imputation methods in cohort studies.

Main Methods:

• A simulation study was conducted using R language software.
• Longitudinal data with missing values were generated using the Monte Carlo method.
• Imputation methods were evaluated based on average absolute deviation, average relative deviation, and Type I error in regression analysis.

Main Results:

• Mean imputation, k-nearest neighbor (KNN), regression imputation, and random forest demonstrated stable and comparable imputation effects.
• Hot deck imputation performed less effectively than the aforementioned methods.
• K-means clustering and expectation maximization (EM) algorithm showed the poorest and most unstable results.
• Mean imputation, EM algorithm, random forest, KNN, and regression imputation effectively controlled Type I error, while multiple imputations, hot deck, and K-means clustering did not.

Conclusions:

• Mean imputation, KNN, regression imputation, and random forest are recommended for missing data in longitudinal studies under the missing at random mechanism.
• Multiple imputations and hot deck can be suitable when the missing data ratio is low.
• K-means clustering and EM algorithm are not advised due to their instability and poor performance.