When measurement errors correlate with truth: surprising effects of nondifferential misclassification

Area of Science:

• Epidemiology
• Biostatistics

Background:

• Nondifferential misclassification and errors in variables are common in epidemiological research.
• Existing literature often assumes binary exposures or continuous variables under the classical error model, where errors are uncorrelated with true values.
• In classical models, imperfect exposure measurement typically attenuates associations, and measurement error in confounders can still allow for partial confounding control.

Purpose of the Study:

• To examine the impact of correlated errors in variables on epidemiological findings.
• To highlight the limitations of the classical error model in real-world epidemiological data, particularly for self-reported variables.
• To caution researchers about potential misinterpretations of results due to deviations from classical error assumptions.

Main Methods:

• The study reviews existing literature on measurement error in epidemiological studies.
• It contrasts the classical error model with scenarios involving correlated errors.
• The implications for exposure-outcome relationships and confounding are discussed.

Main Results:

• Deviations from the classical error model, where errors are correlated with true values, can lead to different outcomes than predicted by standard models.
• Imperfectly measured exposures with correlated errors may not consistently attenuate associations.
• Poor measurement, even when correlated with true values, can potentially explain positive findings, irrespective of disease status.

Conclusions:

• Epidemiologists must be cautious of measurement errors that deviate from the classical model, as these can significantly distort findings.
• Self-reported data, often prone to correlated errors, requires careful consideration.
• Understanding these nuances is crucial for accurate interpretation of epidemiological results and public health implications.