Performance of bias-correction methods for exposure measurement error using repeated measurements with and without...

Area of Science:

• Biostatistics
• Epidemiology
• Occupational Health

Background:

• Measurement error in exposure assessment can introduce bias into effect estimates.
• Independent replicates can correct bias, but are often expensive.
• Two-stage (2S) studies offer a cost-effective alternative to single-stage (1S) studies by collecting replicates on a subsample.

Purpose of the Study:

• To compare the performance of bias-correction methods in single-stage (1S) and two-stage (2S) study designs.
• To evaluate methods including instrumental variable (EVROS IV), regression calibration, and simulation extrapolation.
• To assess performance under varying levels of exposure measurement error.

Main Methods:

• A simulation study was conducted to compare bias-correction methods.
• Methods evaluated: EVROS IV, regression calibration, and simulation extrapolation.
• Two-stage study data handling included ignoring missing data or using multiple imputations.

Main Results:

• Regression calibration demonstrated the best performance (lowest root mean square error) for small to moderate measurement error in both 1S and 2S designs.
• The EVROS IV method outperformed regression calibration when measurement error was severe and grouping was adequate.
• Simulation extrapolation performed poorly across both designs with moderate to large measurement error.

Conclusions:

• Regression calibration is a robust method for correcting bias due to measurement error, particularly in 1S and 2S designs with small to moderate error.
• The implementation of regression calibration in statistical software (Stata) requires careful consideration and may benefit from multiple imputation.
• For severe exposure mismeasurement, the EVROS IV method is a viable alternative, provided effective grouping strategies are employed.