Handling missing patient-reported outcomes in longitudinal clinical trials: a simulation study.

Area of Science:

• Biostatistics
• Clinical Research Methodology
• Health Outcomes Research

Background:

• Missing data in longitudinal Patient-Reported Outcomes (PROs) poses a significant challenge to data analysis.
• Incomplete PRO data can lead to biased results and erroneous conclusions in clinical trials.
• Understanding the performance of different methods for handling missing data is crucial for reliable research.

Purpose of the Study:

• To compare the effectiveness of multiple imputation, machine learning techniques, and complete case analysis in handling missing PRO data.
• To evaluate these methods under various missing data mechanisms (MAR, MNAR) and proportions.
• To identify optimal strategies for analyzing longitudinal PRO data with missing values.

Main Methods:

• A comprehensive Monte Carlo simulation study involving 36,000 datasets.
• Varied sample sizes (100-300), missingness proportions (20%-60%), and missing data mechanisms (MAR, MNAR).
• Compared Multivariate Imputation by Chained Equations (MICE), Random Forest, k-Nearest Neighbors (KNN), and Complete Case Analysis (CCA).

Main Results:

• Multivariate Imputation by Chained Equations (MICE) and Random Forest demonstrated superior performance, yielding the least biased and most precise estimates.
• Complete Case Analysis (CCA) exhibited substantial and clinically significant bias under high missingness proportions (60%) with Missing Not at Random (MNAR) mechanisms.
• k-Nearest Neighbors (KNN) was found to be the least accurate imputation method across evaluated conditions.

Conclusions:

• Complete Case Analysis (CCA) is not recommended for longitudinal PRO data due to its high risk of bias, particularly under MNAR conditions.
• Employing robust methods such as MICE and Random Forest within a sensitivity analysis framework is essential for ensuring the validity of PRO research findings.
• Advanced imputation techniques offer more reliable results than simple methods when dealing with missing PRO data, especially in scenarios where patient health influences data completion.