A comparison of sensitivity-specificity imputation, direct imputation and fully Bayesian analysis to adjust for exposure misclassification when validation data are unavailable.

PURPOSE

Measurement error is an important source of bias in epidemiological studies. We illustrate three approaches to sensitivity analysis for the effect of measurement error: imputation of the 'true' exposure based on specifying the sensitivity and specificity of the measured exposure (SS); direct imputation (DI) using a regression model for the predictive values; and adjustment based on a fully Bayesian analysis.

METHODS

We deliberately misclassify smoking status in data from a case-control study of lung cancer. We then implement the SS and DI methods using fixed-parameter (FBA) and probabilistic (PBA) bias analyses, and Bayesian analysis using the Markov-Chain Monte-Carlo program WinBUGS to show how well each recovers the original association.

RESULTS

The 'true' smoking-lung cancer odds ratio (OR), adjusted for sex in the original dataset, was OR = 8.18 [95% confidence limits (CL): 5.86, 11.43]; after misclassification, it decreased to OR = 3.08 (nominal 95% CL: 2.40, 3.96). The adjusted point estimates from all three approaches were always closer to the 'true' OR than the OR estimated from the unadjusted misclassified smoking data, and the adjusted interval estimates were always wider than the unadjusted interval estimate. When imputed misclassification parameters departed much from the actual misclassification, the 'true' OR was often omitted in the FBA intervals whereas it was always included in the PBA and Bayesian intervals.

CONCLUSIONS

These results illustrate how PBA and Bayesian analyses can be used to better account for uncertainty and bias due to measurement error.