|Rosner Preis, Sarah - Harvard School Of Public Health|
|Spiegelman, Donna - Harvard School Of Public Health|
|Zhao, Barbara - Harvard School Of Public Health|
|Willett, Walter - Harvard School Of Public Health|
Submitted to: American Journal of Epidemiology
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 10/24/2010
Publication Date: 2/22/2011
Citation: Rosner Preis, S., Spiegelman, D., Zhao, B., Moshfegh, A.J., Baer, D.J., Willett, W.C. 2011. Random and correlated errors in gold standards used in nutritional epidemiology: implications for validation studies. American Journal of Epidemiology. 173:683-94.
Interpretive Summary: Measuring food intake food in free-living populations is important in epidemiologic studies designed to assess the association between diet and health. These types of studies provide insight into the role of diet in preventing diseases that may take many years to develop, such as cancer. Food frequency questionnaires (FFQ) are a tool that has been used to determine food intake. However, these questionnaires are prone to errors, including measurement errors. To develop better ways to correct for measurement error, data were used from two validation studies that used recovery biomarkers, the Observing Protein and Energy Nutrition (OPEN) and the Automated Multiple-Pass Method Validation (AMPM) Study. In the OPEN study, it was not possible to adequately account for within-person variation in protein and energy intake estimated by recovery biomarkers since these measures were collected too closely together. The result of this timing is that there could be an overestimation of the error correlation between the food frequency questionnaire and 24-hour recall, and an under-estimation of the FFQ’s validity. In the AMPM Study, repeated doubly labeled water and urinary nitrogen samples were collected one year apart. This timing allowed for estimation of the relationship (correlation) between energy intake, as well as the protein density of the diet. Correlations between true protein density intake and the FFQ were higher in the AMPM study compared to the OPEN study (0.42 vs 0.36, respectively). When assuming no biomarker scale bias, the de-attenuation factor was 0.33 in OPEN and 0.52 in AMPM, similar to the regression slope estimates for dietary recall of the two studies. Analyses that adjust for measurement error using standard methods would often be substantially less biased than analyses which ignore measurement error, even when recovery biomarkers do not exist or are otherwise infeasible to include in validation studies.
Technical Abstract: The measurement error correction de-attenuation factor was estimated from two studies using recovery biomarkers. One study, the Observing Protein and Energy Nutrition (OPEN), was unable to adequately account for within-person variation in protein and energy intake estimated by recovery biomarkers, which could have resulted in overestimation of the error correlation between the food frequency questionnaire (FFQ) and 24-hour recall (24HR), and an under-estimation of the FFQ’s validity. In the other study, the Automated Multiple-Pass Method Validation (AMPM) Study, repeated doubly labeled water (DLW)and urinary nitrogen (UN) were collected one year apart, yielding intra-class correlation coefficients of 0.43 for energy (DLW) and 0.54 for protein density (UN/DLW), indicating important within-person variation in these biomarkers. Correlation between true protein density intake and the FFQ was 0.36 in OPEN and 0.42 in AMPM, and the de-attenuation factor was 0.15 in OPEN and 0.30 in AMPM when assuming no 24HR scale bias. When assuming no biomarker scale bias, the de-attenuation factor was 0.33 in OPEN and 0.52 in AMPM, similar to regression slope estimates for dietary records data on the FFQ of 0.49 in OPEN and 0.32 in AMPM. Standard measurement error methods may adequately adjust for bias due to measurement error in energy-adjusted nutrients.