Correlated measurement error--implications for nutritional epidemiology.

Area of Science:

• Nutritional epidemiology
• Statistical modeling in health research

Background:

• Dietary assessment instruments often yield correlated nutrient intakes and correlated measurement errors.
• The impact of correlated errors on multivariate regression in observational studies is under-explored.

Purpose of the Study:

• To investigate the effect of correlated measurement errors between dietary variables on regression coefficients.
• To assess how varying levels of variable and error correlation influence bivariate regression estimates.

Main Methods:

• Utilized a multivariate error structure model to analyze bivariate linear regression coefficients.
• Examined effects of measurement precision, variable correlation, and error correlation.
• Applied the model to plasma vitamin C prediction using dietary data from the EPIC-Norfolk study (FFQ and 7DD).

Main Results:

• Zero error correlation yields estimates reflecting measurement precision and confounding; bias towards null observed.
• Non-zero error correlation below 0.7 has minor effects; above 0.8, effects become substantial and precision-dependent.
• High error correlation (e.g., fat-energy > 0.9 with FFQ) significantly impacts regression coefficients.
• Energy adjustment effects differ markedly between nutrients (e.g., vitamin C vs. fat).

Conclusions:

• Substantial error correlation can greatly alter bivariate regression estimates, with variable-dependent effects.
• Energy adjustment impacts vary widely; it can mitigate regression dilution for some nutrients (e.g., fat) but reduce true variance.
• Understanding error correlation is crucial for accurate interpretation of dietary epidemiology findings.