Comparison of Imputation Strategies for Incomplete Longitudinal Data in Life-Course Epidemiology

epidemiology
longitudinal data
missing data
multiple imputation
statistics
  • Topic: A study comparing the performance of three Multiple Imputation (MI) methods for handling incomplete longitudinal data in life-course epidemiology, focusing on the effect of longitudinal depressive symptoms on mortality.
  • Methods Compared: Normal Linear Regression (MICE), Predictive Mean Matching (PMM), and Variable-Tailored Specification were tested under nine scenarios varying missingness rate and mechanism (MCAR, MAR, MNAR).
  • Finding: All methods showed similar levels of bias in estimating the causal effect (Hazard Ratios), but Predictive Mean Matching (PMM) was identified as the most appealing strategy due to its consistently low root mean square error (RMSE) and competitive computation time.
Published

23 January 2026

PubMed: 37338987 DOI: 10.1093/aje/kwad139 Overview generated by: Gemini 2.5 Flash, 27/11/2025

Background and Objective

Incomplete longitudinal data—data with values missing at different time points for the same individuals—is a pervasive challenge in life-course epidemiology. If not handled correctly, this missingness can introduce bias and lead to incorrect statistical inference. Multiple imputation (MI) is the increasingly preferred method for addressing this, but real-world performance comparisons between different MI techniques are limited.

The objective of this study was to conduct a direct comparison of three Multiple Imputation (MI) methods using real-world longitudinal data to assess their performance, feasibility, and impact on causal effect estimates across various missing-data scenarios.

Methods: Benchmarking on HRS Data

The researchers compared three Multiple Imputation (MI) methods using data from the Health and Retirement Study (HRS):

  1. Normal Linear Regression (Multiple Imputation by Chained Equations - MICE)
  2. Predictive Mean Matching (PMM) (a non-parametric MICE method)
  3. Variable-Tailored Specification

The comparison was conducted under nine missing-data scenarios that represented combinations of: * Missingness Rate: 10%, 20%, and 30% missingness. * Missingness Mechanism: Missing Completely at Random (MCAR), Missing at Random (MAR), and Missing Not at Random (MNAR).

The study introduced record-level missingness to a complete sample of HRS participants and then used the imputed datasets to fit Cox proportional hazards models. The outcome of interest was the effect of four different operationalizations of longitudinal depressive symptoms on subsequent mortality.

Key Results and Conclusion

The comparison focused on three performance metrics: bias in hazard ratios, root mean square error (RMSE), and computation time.

  • Bias: Bias in the estimated hazard ratios was found to be similar across all three MI methods.
  • Consistency: The results were consistent across the four different ways the longitudinal exposure (depressive symptoms) was defined.
  • Performance: The authors suggest that Predictive Mean Matching (PMM) may be the most appealing strategy for imputing life-course exposure data. PMM consistently showed the lowest root mean square error (RMSE)—indicating greater precision—and maintained competitive computation times with few implementation challenges.

This study provides valuable, empirically-based guidance for researchers in life-course epidemiology regarding the choice of MI method for handling incomplete longitudinal exposure data.