A longitudinal transition imputation model for categorical data applied to a large registry dataset

Area of Science:

• Statistics
• Biostatistics
• Longitudinal Data Analysis

Background:

• Imputing longitudinal categorical covariates with multiple waves and predictors presents challenges including implausible transitions, collinearity, and overfitting.
• Existing methods struggle with the complexity of time-dependent categorical variables in large datasets.

Purpose of the Study:

• To develop and evaluate a flexible 3-stage imputation methodology for longitudinal categorical covariates.
• To address issues of implausible transitions, collinearity, and overfitting in complex datasets.
• To improve the accuracy and reliability of statistical analyses involving time-dependent categorical data.

Main Methods:

• A simulation study using Belgian general practitioner morbidity registry data with smoking as the covariate of interest.
• A 3-stage approach involving multiple imputation (MI) using fully conditional specification (FCS) or MI for predictive variables in wide format, followed by a joint Markov transition model for imputation.
• Comparison of the proposed method against complete case analysis and listwise deletion using bias and root mean square error (RMSE).

Main Results:

• The proposed methodology effectively imputes time-dependent categorical covariates while preserving transition plausibility.
• It successfully resolves issues of overfitting and collinearity, allowing for the utilization of confounders.
• Performance evaluation via simulation demonstrated the method's advantages over traditional approaches.

Conclusions:

• The developed 3-stage imputation framework offers a robust solution for handling longitudinal categorical covariates.
• This method enhances the validity of statistical inferences by maintaining data integrity and addressing common imputation challenges.
• A companion R package is available to facilitate the application and replication of this methodology.