Should regression calibration or multiple imputation be used when calibrating different devices in a longitudinal...

Area of Science:

• Biostatistics
• Epidemiology
• Longitudinal Data Analysis

Background:

• Measurement devices in longitudinal studies can change between visits.
• Calibration studies assess between-device differences but introduce missing data.
• Statistical adjustment is crucial for accurate longitudinal data analysis.

Purpose of the Study:

• To compare regression calibration and multiple imputation for adjusting between-device differences in longitudinal studies.
• To evaluate methods for handling missing data in calibration studies.
• To identify the most reliable statistical method for longitudinal pulse wave velocity studies with device changes.

Main Methods:

• A simulation study was conducted using linear regression.
• Scenarios mimicked real-world longitudinal studies, specifically for pulse wave velocity.
• Regression calibration and multiple imputation (fully stochastic and predicted mean matching) were compared.

Main Results:

• Both regression calibration and multiple imputation were largely unbiased.
• Estimating standard errors presented challenges for both methods.
• Multiple imputation with predicted mean matching closely matched empirical standard errors.
• Fully stochastic multiple imputation underestimated standard errors by up to 50%.
• Regression calibration with bootstrapped standard errors showed moderate improvement over fully stochastic imputation.
• Regression calibration was more efficient than multiple imputation methods.

Conclusions:

• Multiple imputation with predicted mean matching is recommended for longitudinal studies with potential device-related errors.
• Fully stochastic imputation and regression calibration may lead to inaccurate standard error estimates.
• Accurate adjustment for measurement errors and missing data is vital in longitudinal research.