Seemingly unrelated measurement error models, with application to nutritional epidemiology

Area of Science:

• Nutritional Epidemiology
• Biostatistics
• Measurement Error Models

Background:

• Accurate nutrient intake assessment is crucial in nutritional epidemiology.
• Food Frequency Questionnaires (FFQ) are common but have measurement errors.
• Biomarkers offer more accurate intake data but are often costly or invasive.

Purpose of the Study:

• To develop and evaluate Seemingly Unrelated Measurement Error Models (SUMEMs) for nutritional epidemiology.
• To assess the measurement error properties of FFQs for protein and energy intake.
• To improve statistical efficiency in estimating nutrient-disease relationships.

Main Methods:

• Development of marginal linear mixed measurement error models for individual nutrients.
• Combination of marginal models into a multivariate measurement error model.
• Comparison of a "full" model versus a scientifically motivated "reduced" model.
• Utilizing the Akaike Information Criterion (AIC) for model selection.

Main Results:

• The "full" SUMEM provided no efficiency gains over separate models.
• The "reduced" SUMEM demonstrated considerable statistical efficiency gains (e.g., 40% decrease in standard errors).
• AIC effectively distinguished between the models, favoring the efficient "reduced" model.
• Theoretical and practical issues were identified with the Bayesian Information Criterion (BIC) in this context.

Conclusions:

• SUMEMs, particularly the "reduced" model, offer significant statistical efficiency improvements in nutritional epidemiology.
• The AIC is a suitable criterion for selecting the appropriate SUMEM.
• Careful model specification is essential for realizing efficiency gains in measurement error modeling.