A new method of regression calibration - comparison with other methods of correcting covariate error

Oregon Convention Center 

Low-dose radiation risks must be extrapolated from groups exposed at much higher levels of dose. Recently, there has been much attention paid to methods of dealing with shared errors, which are common in many datasets, and particularly important in occupational and environmental settings.
We assess regression calibration (RC), Monte Carlo maximum likelihood (MCML), frequentist model averaging (FMA), 2D Monte Carlo+Bayesian model averaging (2DMC+BMA) methods with simulated datasets having mixed Berkson and classical errors. We test all these against a modification of RC, the extended regression calibration (ERC) method. 2DMC+BMA has poor coverage for the dose coefficients; coverage of RC, MCML, FMA is better, although uniformly too high for FMA and MCML, and best for ERC. Bias in predicted relative risk is generally smallest for ERC, and largest for quasi-2DMC+BMA and FMA methods, with RC and MCML exhibiting bias in predicted risk somewhat intermediate between the other two methods. In summary, the new ERC method performs well in comparison to previously proposed methods and is particularly suited to situations with low to moderate amounts of shared and unshared Berkson errors.

Covariate measurement error

Berkson error

Classical error

Regression calibration

Bayesian model averaging

Frequentist model averaging 

Section on Statistics and the Environment