Sensitivity analysis for random measurement error using regression calibration and simulation-extrapolation.

Area of Science:

• Biostatistics
• Epidemiology
• Statistical Modeling

Background:

• Sensitivity analysis for random measurement error is crucial in statistical modeling.
• Regression calibration and simulation-extrapolation are methods for this analysis when validation data are absent.
• A direct comparison of these methods for sensitivity analysis has been lacking.

Purpose of the Study:

• To compare the performance of regression calibration and simulation-extrapolation for sensitivity analysis of random measurement error.
• To evaluate bias, mean squared error (MSE), and confidence interval coverage of both methods.
• To provide guidance on method selection in the absence of validation data.

Main Methods:

• A simulation study was conducted to compare regression calibration and simulation-extrapolation.
• The study evaluated linear and logistic regression models.
• Performance metrics included bias, MSE, and confidence interval coverage across various reliability, sample size, replicates, and R-squared values.

Main Results:

• Regression calibration yielded unbiased results with median bias of 0.8%.
• Simulation-extrapolation showed significant bias (median -19.0%) and lower confidence interval coverage (median 85%).
• Simulation-extrapolation offered a slight efficiency gain (lower median MSE), but at the cost of accuracy and reliability.

Conclusions:

• Regression calibration is the preferred method for sensitivity analysis of random measurement error due to its unbiasedness and reliable confidence interval coverage.
• Simulation-extrapolation, while offering marginal efficiency gains, is less suitable due to introduced bias and poor coverage.
• The findings support the routine use of regression calibration in relevant statistical analyses.