shows the distribution of demographic and lifestyle characteristics in NPAAS, along with those for the remainder of the Observational Study cohort. The oversampling according to race/ethnicity, body mass index, and age at enrollment is evident. NPAAS women were somewhat more highly educated, more affluent, and more frequently engaged in recreational activities compared with other cohort members.
Baseline (1994–1998) Demographic and Lifestyle Characteristics of Participants in the NPAAS and Participants in the WHI Observational Study But Not the NPAAS
shows geometric means for biomarker and dietary assessments of energy, protein, and protein density, for assessments meeting quality control criteria. The geometric means of the self-report:biomarker assessment ratios are also shown. Each of the 3 self-report procedures underestimates energy substantially (20%–27%) and protein to a lesser extent (4%–10%), and each overestimates protein density compared with the biomarker (16%–25%).
Table 2. Geometric Means and 95% Confidence Intervals for Biomarker and Self-Report Assessments of Energy and Protein Consumption in the NPAAS (2007–2009), Along With Geometric Means and 95% Confidence Intervals for Self-Report:Biomarker Assessment Ratios (more ...)
shows some results from linear regression of log(self-report) − log(biomarker) on body mass index, age at NPAAS participation, and race/ethnicity. Each of the 3 self-report procedures shows evidence of systematic biases related to 1 or more of these factors, for both energy and protein. For 4-day food record and 24-hour dietary recall assessments, energy and protein underreporting was more severe among women with a high body mass index or a younger age, while black women tended to further modestly underestimate energy and to overestimate protein and protein density. Food frequency questionnaire systematic bias patterns included greater energy underestimation by younger women and substantially greater underestimation of both energy and protein by minority group women. Food frequency questionnaire bias for energy in relation to body mass index was greater (P < 0.05) in corresponding analyses that excluded the ethnicity variables from the regression model. Systematic biases were not evident for food frequency questionnaire protein density.
β Coefficients and Standard Errors From Regression of Log(Self-Report) − Log(Biomarker) on Body Mass Index, Age, and Ethnicity in the NPAAS (2007–2009) Among 450 Postmenopausal Women
Correlation coefficients between log-transformed biomarker and log-transformed food frequency questionnaire, 4-day food record, and 24-hour dietary recall assessments were, respectively, 0.196 (standard error (SE), 0.044), 0.297 (SE, 0.046), and 0.167 (SE, 0.051) for energy; 0.289 (SE, 0.042), 0.476 (SE, 0.043), and 0.403 (SE, 0.041) for protein; and 0.254 (SE, 0.041), 0.332 (SE, 0.049), and 0.264 (SE, 0.046) for protein density.
shows regression coefficients from linear regression of log(biomarker) on log(self-report), as well as body mass index, age, and ethnicity, thereby adjusting for the systematic biases noted in , while also allowing these study subject characteristics to help explain biomarker variation more generally. For energy, the resulting “calibration equations” that use food frequency questionnaire, 4-day food record, or 24-hour dietary recall assessments, respectively, explain 41.7%, 44.7%, and 42.1% of the biomarker variation. These percentages are much larger than those from analyses using the self-report data alone (3.8%, 7.8%, and 2.8%, respectively), with much of the added value deriving from body mass index and age. For protein, the food frequency questionnaire, 4-day food record, and 24-hour dietary recall-based calibration equations provide an explanation for 20.3%, 32.7%, and 28.4% of the biomarker variation. For protein density, the corresponding percentages are 8.7%, 14.4%, and 10.4%. Calibration equations are also shown using all 3 self-reports simultaneously with the other variables. The percentages of biomarker variation explained were 45.0%, 34.6%, and 15.5% for energy, protein, and protein density. The strongest self-report “signal” for each of the 3 nutritional variables arises from the 4-day food record, and the variation explained is not significantly greater than that from the calibration equation with only the 4-day food record self-report for energy (P = 0.67), protein (P = 0.10), or protein density (P = 0.23).
Table 4. Calibration Equation β Coefficients, Standard Errors, and Percentage of Biomarker Variation Explained as R2 From Regression of Log(Biomarker) on Log(Self-Report), Body Mass Index, Age, and Ethnicity in the NPAAS (2007–2009) Among 450 Postmenopausal (more ...)
The adjusted R2 values in suggest that the calibration equations recover a large fraction of the log-transformed consumption variation in the underlying dietary factor (e.g., 71%–77% for energy), using any of the self-report assessments, though less so for protein and protein density if the calibration procedure uses the food frequency questionnaire.
We also estimated measurement error correlations among pairs of assessment methods, under our measurement model and joint normality assumptions. For energy, the estimated measurement error correlation was 0.30 (SE, 0.05) for the food frequency questionnaire and 4-day food record, 0.30 (SE, 0.05) for the food frequency questionnaire and 24-hour dietary recall, and 0.50 (SE, 0.05) for the 4-day food record and 24-hour dietary recall. The corresponding numbers for protein were 0.35 (SE, 0.07), 0.33 (SE, 0.07), and 0.27 (SE, 0.18) and for protein density were 0.38 (SE, 0.14), 0.38 (SE, 0.12), and 0.40 (SE, 0.17).
compares food frequency questionnaire-based calibration equations between the 2 WHI biomarker studies. Dietary Modification trial women tended to be slightly younger and of higher body mass index compared with Observational Study women. A likelihood ratio test of equality of all coefficients is not significant for protein (P = 0.23) or for protein density (P = 0.95). This test is significant (P = 0.003) for energy, but the differences derive from coefficients for age and for Hispanic ethnicity, rather than from the food frequency questionnaire coefficient. The correlations between consumption estimates using the Nutrient Biomarker Study and NPAAS calibration equations are 0.95 for energy, 0.96 for protein, and 0.96 for protein density.
Table 5. Comparison of Calibration Equation β Coefficients and Standard Errors From Regression on Log(Biomarker) on Corresponding Log(Food Frequency Questionnaire), Body Mass Index, Age, and Ethnicity Between the NBS (2004–2006) and the NPAAS (2007–2009) (more ...)
provides scatterplots and correlation coefficients between NPAAS visit 1 and NPAAS visit 3, for women in the reliability subsample for log(biomarker) and each self-report. Food frequency questionnaire correlations are somewhat larger than those from the other self-reports, while the correlation for the protein density biomarker is low (r = 0.24).
Figure 2. Scatterplot of the Women’s Health Initiative Nutrition and Physical Activity Assessment Study (NPAAS; 2007–2009) primary versus reliability sample. Each plot provides the Pearson correlation for the log-measure. DLW, doubly labeled water; (more ...)
The WHI food frequency questionnaire aims to assess consumption over the preceding 3 months, whereas the 4-day food record and 24-hour dietary recalls target consumption over a few days or weeks, respectively, in proximity to biomarker assessment. Calibration equations of the type shown in were also carried out from reliability subsample data by averaging the visit 1 and visit 3 log(biomarker) assessments and using either the visit 3 log(food frequency questionnaire) or the average of visit 1 and visit 3 log(4-day food record) or log(24-hour dietary recall) assessments as predictor variables. These analyses led to somewhat higher percentages of biomarker variation explained, compared with . Specifically, for the food frequency questionnaire, 4-day food record, and 24-hour dietary recall, these percentages were, respectively, 52.3%, 58.1%, and 53.6% for energy; 24.8%, 42.6%, and 37.4% for protein; and 15.0%, 22.4%, and 20.0% for protein density. The percentages of variation explained by the food frequency questionnaire, 4-day food record, and 24-hour dietary recall data alone in these calibration equations were, respectively, 6.7%, 11.9%, and 4.3% for energy; 7.4%, 28.2%, and 18.1% for protein, and 4.9%, 12.3%, and 8.6% for protein density.
Calibration equations were also developed separately by race/ethnicity (white, black, Hispanic) and body mass index (<25.0, 25.0–29.9, ≥30.0). The “signals” from the self-report assessment were comparatively weaker for black women for each assessment procedure. Similarly, the signals for overweight and obese women were weaker than those for normal weight women for each assessment procedure. As shown in the Web Appendix, which is posted on the Journal
’s Web site (http://aje.oupjournals.org/
), the fraction of biomarker variation explained by these calibration equations was somewhat higher for Hispanic compared with black women, with white women intermediate; and somewhat higher for obese compared with normal weight women for energy, but higher for normal weight versus obese women for protein density, with overweight women intermediate.