Use of a urinary sugars biomarker to assess measurement error in self-reported sugars intake in the Nutrition and Physical Activity Assessment Study (NPAAS)

Abstract

Background

Measurement error (ME) in self-reported sugars intake may be obscuring the association between sugars and cancer risk in nutritional epidemiologic studies.

Methods

We used 24-hour urinary sucrose and fructose as a predictive biomarker for total sugars, to assess ME in self-reported sugars intake. The Nutrition and Physical Activity Assessment Study (NPAAS) is a biomarker study within the Women’s Health Initiative (WHI) Observational Study, that includes 450 post-menopausal women aged 60–91. Food Frequency Questionnaires (FFQ), 4-day food records (4DFR) and three 24-h dietary recalls (24HRs) were collected along with sugars and energy dietary biomarkers.

Results

Using the biomarker, we found self-reported sugars to be substantially and roughly equally misreported across the FFQ, 4DFR and 24HR. All instruments were associated with considerable intake- and person-specific bias. Three 24HRs would provide the least attenuated risk estimate for sugars (attenuation factor, AF=0.57), followed by FFQ (AF=0.48), and 4DFR (AF=0.32), in studies of energy-adjusted sugars and disease risk. In calibration models, self-reports explained little variation in true intake (5–6% for absolute sugars; 7–18% for sugars density). Adding participants’ characteristics somewhat improved the percentage variation explained (16–18% for absolute sugars; 29–40% for sugars density).

Conclusions

None of the self-report instruments provided a good estimate of sugars intake, although overall 24HRs seemed to perform the best.

Impact

Assuming the calibrated sugars biomarker is unbiased, this analysis suggests that, measuring the biomarker in a subsample of the study population for calibration purposes may be necessary for obtaining unbiased risk estimates in cancer association studies.

INTRODUCTION

The associations between sugars consumption and cancer have long been studied, yet, are difficult to demonstrate in an epidemiologic context (1–3). One of the major limitations in nutritional epidemiologic analyses is unreliability of self-reported dietary data, which can lead to severely distorted disease association estimates and reduced statistical power to detect an association (4).

Food frequency questionnaires (FFQs) are the most widely used dietary assessment instruments in population studies, despite known large measurement errors (ME) both random and systematic (5, 6). Other self-reports such as 24-h dietary recalls (24HRs) (7) and food records (FRs) (8), although generally considered to have better ME properties than FFQs, they too are affected by random and systematic ME (7, 9, 10). These findings have prompted researchers to incorporate reference dietary biomarkers, as objective measures of intake, in sub-samples of their cohorts, to assess and correct for ME in self-reported intake. So far, four classes of biomarkers have been described (11). Recovery biomarkers are based on a known recovered proportion of intake over certain period of time. Following transformation, recovery biomarkers adhere to a classical ME model and generate unbiased estimates of intakes (12), which can be used to assess ME in self-report instruments (7, 10) or to develop calibration equations for calibrating (i.e. correcting) self-reported intake to be applied in diet-disease risk models of association studies (13). Concentration biomarkers provide correlate, rather than a direct measure of intake (14), yet when combined with self-reported intake were shown to improve reliability of risk estimates and to increase the statistical power to detect an association (15). Replacement biomarkers replace estimates of intake for nutrients or compounds difficult to measure or with no food composition data available, and depending on their innate characteristics may be used as recovery (e.g. 24-hour urinary sodium) or more commonly as concentration biomarkers (e.g. serum phytoestrogens). Predictive biomarkers, the most recently described class of biomarkers, predict intake after being calibrated to account for certain level of bias, estimable from a feeding study and assumed to be stable across populations (16). Following calibration, similar to recovery biomarkers, they, too, can be used as reference instruments.

Recently, based on findings from two controlled feeding studies, 24-hour urinary sucrose plus fructose was suggested as a predictive biomarker for total sugars intake (17). Although predictive biomarkers exhibit more complex relationship with true intake than recovery biomarkers, this relationship is assumed stable, thereby distinguishing predictive biomarkers from less specific concentration biomarkers (16). Based on a novel ME model for predictive biomarkers (16), the sugars biomarker was calibrated for use as a reference instrument in biomarker studies of dietary ME or disease-association cohort studies having available 24-h urine collections in a sub-sample of participants.

The Nutrition and Physical Activity Assessment Study (NPAAS) is a biomarker study involving a subset of 450 participants in the Women’s Health Initiative (WHI) Observational Study. Prentice et al (10) compared FFQ-, 24HRs- and 4DFR-energy and protein intake in NPAAS against respective recovery biomarkers, doubly-labeled water (DLW) and 24-h urinary nitrogen. We now compare estimates of total sugars intake from these three self-report instruments against the 24-h urinary sugars biomarker. After calibrating the biomarker, we use it as a reference measure to evaluate the ME structure of self-reported sugars, and to estimate attenuation factors and correlations with true intake, two important parameters that determine how well each instrument will be able to detect and estimate disease risks associated with total sugars intake. We also develop calibration equations that predict total sugars intake, based on the objective predictive sugars biomarker given self-report and other covariates, that could be applied in future WHI association studies for more reliable disease risk estimation.

MATERIALS AND METHODS

Participants

The NPAAS is an ancillary study to the WHI Observational Study, a prospective study of 93,676 postmenopausal women aged 50–79, enrolled during 1994–1998 at 40 U.S. clinical centers (18, 19). Between 2007 and 2009, 450 participants aged 60–91 at the time of the NPAAS, were recruited from nine WHI centers. The NPAAS design has been previously described in detail (10). To ensure adequate statistical power to assess the effect of racial/ethnic groups, BMI and age, recruitment was conducted to oversample African-American and Hispanic women, women with BMI <18.5 kg/m² and ≥30 kg/m², and those who were 50–59 years at WHI enrollment, respectively. Among women who were screened, 20.6% were eligible and willing to participate. The study was approved by the institutional review boards of all participating institutions.

Study design

Participants visited local WHI centers twice. At Visit 1, they provided informed consent, had their body weight and height measured, completed the FFQ and the WHI Personal Habit Questionnaire (PHQ) assessing participants’ recreational physical activity (20). Information on participants’ demographic characteristics was available from the WHI database. During Visit 1, participants received training on how to keep a 4DFR and started the two-week DLW protocol for measurement of energy expenditure, which was concluded at Visit 2 (10). In the two weeks between Visit 1 and 2, participants completed the 4DFR and collected 24-h urine on the day prior to Visit 2. Three 24HRs were administered following Visit 2. Approximately 20% of the NPAAS participants (N=88) completed the “reliability study”, which involved repeating the whole study protocol six months after the baseline study.

Dietary Assessment

Participants’ diet over the previous three months was assessed at NPAAS baseline using the WHI self-administered semi-quantitative FFQ (21), designed to assess diet in a multiethnic and geographically diverse population, and inquired about 122 foods or food groups, including 19 adjustment questions and four summary questions. The FFQs were reviewed by study staff, checking for missing responses in the presence of the participants. As part of the 4DFR training, participants viewed 25-minute instructional video and received instructions along with a serving size booklet containing photographs and measuring devices. They completed the 4DFR between Visit 1 and 2 over four alternate days, including one weekend day. Upon return, the 4DFR was reviewed in the presence of participants. For the three 24HRs, the first was administered 1–3 weeks after Visit 2 and the other two approximately monthly thereafter by a certified interviewer over the phone (70% weekday and 30% weekend) with computer-assisted data entry using the NDSR software (Nutrition Data System for Research; Nutrition Coordinating Center, University of Minnesota, Minneapolis) based on the USDA multi-pass method (22). The NDSR version 2005 was used to calculate total sugars intake (g/day), a sum of sucrose, fructose, glucose, lactose, galactose and maltose, and total sugars density (g/1000 kcal) for the three instruments. Based on information provided on the 4DFR, we calculated percent meals eaten at home.

Urinary sugars biomarkers

The day before Visit 2, participants collected a 24-hour urine sample. They were asked to record any missed voids or spillage during the collection. To assess urine completeness, participants took three 80 mg tablets of para-aminobenzoic acid (PABA) on the day of urine collection, one with each meal (23). In this paper, we analyzed 384 participants from the primary and 78 from the reliability study, who provided urine samples, had <2 missed voids or <240 ml of missed/spilled urine, and had sufficient specimen available for laboratory analysis, without excluding urine samples based on PABA (see Supplementary Material for results with PABA exclusions).

Sucrose and fructose in urine were quantified by a colorimetric enzymatic assay (Sucrose/D-Glucose/D-Fructose; Boehringer Mannheim, R-Biopharm, Roche, modified to be run on a Microplate reader in 96-well plates. Samples were run in duplicate, and glucose, sucrose and fructose standards (5, 10, 20, 30, 50 and 100 mg/L) and quality control urine samples were included in each run. When %CV for duplicate samples was >15%, samples were re-analyzed. When applicable, urine samples were diluted and re-analyzed to obtain values within the range of the calibration curve. The level of detection (LOD) was 1 mg/L. The intra- and inter-assay CV for the method were <18% and <25% for concentrations <5 mg/l and <10% and <20% for concentrations ≥5 mg/l, respectively, for both sucrose and fructose. Urinary excretion of sucrose and fructose (mg/day) was calculated based on 24-h urine volume (ml/day). One hundred and thirty nine sucrose and seven fructose values were below LOQ, in which case a value of 0.5 mg/L was imputed.

Biomarker for energy intake

We used total energy expenditure, assessed by DLW as previously described (10). It has been shown that in weight-stable individuals, two-weeks of total energy consumption can be objectively estimated by this approach (24, 25). Approximately 6.5% of the samples were excluded from the analysis; half due to low tracer enrichments or lack of equilibration, while the others due to dilution space or other external reproducibility issues. DLW-based estimates of energy expenditure (kcal/day) were used to express the biomarker-based total sugars density (g/1000 kcal).

Statistical methods

Calibration of the urinary sugars biomarker

We used the calibration equation for the total sugars biomarker developed by Tasevska et al (16), based on data from a UK-based 30-day feeding study with 13 subjects aged 23–66 years, consuming their usual diets and collecting 24-h urines for 30 days under highly controlled conditions (17),

$$M_{ij}^{\ast }=M_{ij}-1.67-0.02\times S_i+0.71\times A_i,$$ (A)

where (M_ij) is log-transformed sugars biomarker, S_i is an indicator variable that equals 0 for men and 1 for women, and A_i is log-transformed age. The calibrated biomarker $M_{ij}^{\ast }$ satisfies the following ME model,

$$M_{ij}^{\ast }=T_i+u_{Mi}+{\epsilon }_{\mathit{Mij}},$$ (B)

where T_i is log-transformed true usual intake of total sugars, u_Mi is a person-specific bias (random effect) having mean zero and variance ${\sigma }_{u_M}^2$, and ε_Mij is within-person random error. In order to use the calibrated biomarker as a reference measure in the ME model for self-reported sugars intake, we would need to know the variance ${\sigma }_{u_M}^2$, or some related value such as the ratio of the variance of u_Mi to the variance of T_i + u_Mi, calculated as,

$$\kappa =\frac{{\sigma }_{u_M}^2}{{\sigma }_T^2+{\sigma }_{u_M}^2}.$$ (C)

In our analysis, we used the value κ = 0.218, estimated from the feeding study (16).

ME model for self-reported sugars intake

The ME model for self-reported sugars intake has been described in detail (16). Briefly, for individual i, let Q_ij, F_ij and R_ij denote log-transformed reported intake on the j^th application of the FFQ, 4DFR and 24HR, respectively. Let X_Qi, X_Fi and X_Ri be vectors of covariates that affect the relationship between reported and true usual intake T_i. The ME model is:

$$\begin{array}{l}{Q}_{ij}={\beta }_{Q0}+{\beta }_{QT}{T}_i+{\beta }_{QX}{X}_{Qi}+u_{Qi}+{\epsilon }_{\mathit{Qij}},\\ F_{ij}={\beta }_{F0}+{\beta }_{FT}{T}_i+{\beta }_{FX}{X}_{Fi}+u_{Fi}+{\epsilon }_{\mathit{Fij}},\\ R_{ij}={\beta }_{R0}+{\beta }_{RT}{T}_i+{\beta }_{RX}{X}_{Ri}+u_{Ri}+{\epsilon }_{\mathit{Rij}},\\ M_{ij}^{\ast }=T_i+u_{Mi}+{\epsilon }_{\mathit{Mij}},\end{array}$$ (D)

where $M_{ij}^{\ast }$ is the calibrated biomarker described in the previous section; u_Qi, u_Fi, u_Ri, and u_Mi, are person-specific biases with zero means and variances ${\sigma }_{u_Q}^2,{\sigma }_{u_F}^2,{\sigma }_{u_R}^2$ and ${\sigma }_{u_M}^2$, respectively; and e_Qij, e_Fij, e_Rij and e_Mij, are within-person random errors with zero means and variances ${\sigma }_{{\epsilon }_Q}^2,{\sigma }_{{\epsilon }_F}^2,{\sigma }_{{\epsilon }_R}^2$ and ${\sigma }_{{\epsilon }_M}^2$, respectively.

The main assumption is that the person-specific biases of the FFQ, 4DFR and 24HR (u_Qi, u_Fi, u_Ri), though possibly correlated with each other, are independent of the person-specific bias of the biomarker (u_Mi). The parameters in model (D) were estimated using maximum likelihood. Log age and log BMI were included as components of the covariate vector X_Qi = X_Fi = X_Ri,, as it is believed these variables may affect the relationship between self-report and true intake (26).

The estimated parameters in the ME model can be used to estimate two characteristics that are important for evaluating the effect of ME in self-reported intake on the estimation of diet-disease associations; the attenuation factor (AF) is the multiplicative bias in the estimated association due to using self-reported rather than true intake, while the Pearson correlation coefficient (CC) between self-reported and true intake is related to the loss of power to detect the association due to ME. Formulas for calculating AF and CC are given in Tasevska et al (16).

We also developed calibration equations for predicting true intake by regressing $M_{ij}^{\ast }$ on Q_ij, F_ij or R_ij, and other covariates that may be related to self-reported sugars and/or true intake. Potential covariates included age, BMI, physical activity [metabolic equivalents per week], smoking status, dietary supplement use, race, education, annual income and percentage of meals eaten at home. The latter covariate was considered because low percentage of meals eaten at home was associated with energy under-reporting in NPAAS (27). Metabolic equivalents values for physical activity reported on the PHQ were generated as previously described (28). The calibration model included log age and log BMI. Backward selection (P=0.1) was used to select other covariates that were significant predictors of log true intake. We reported R² adjusted for within-subject error in the biomarker, using repeated biomarker measurements from the reliability study (N=88). We also reported unadjusted R², because of the possibility of correlated ME for the biomarker collected six months apart.

We performed sensitivity analyses to examine the effects of excluding incomplete 24-h urine collections based on PABA (Supplementary Material Section A), and of calibrating the biomarker using a ME model that did not include age (Supplementary Material Section B).

RESULTS

Table 1 reports baseline characteristics for the participants. Geometric means and 95% confidence intervals for self-reported and biomarker-based total sugars intake (g/d) and total sugars density (g/1000 kcal) are presented in Table 2. Compared to biomarker-based estimates, self-reported total sugars appeared considerably misreported on all three instruments (Figure 1–3), with geometric means approximately half those of the biomarker. The under-reporting of total sugars density was less striking, with the ratio of self-reported to biomarker-based means ranging from 0.6 to 0.7. Geometric means of daily sucrose and fructose urinary excretion were 10.7 mg/d and 20.0 mg/d in the primary and 12.5 mg/d and 22.3 mg/d in the reliability study, respectively. Sugars excretion, particularly sucrose excretion, was highly variable between participants, with almost a third of participants having urinary sucrose concentration below LOD.

Table 1.

Demographics and Lifestyle Characteristics Among NPAAS Participants (Primary Study, N = 450; Reliability Study N = 88) Collected at WHI (1994–1998) or NPAAS (2007–2009) Baseline

Characteristic	N	%
Age (years)a
60–64	79	17.5
65–69	165	36.7
70–74	110	24.4
75–79	59	13.1
≥80	37	8.2
Body mass indexa
<18.5	9	2
18–5–24.9	143	31.8
25.0 – 29.9	125	27.8
≥30	173	38.4
Race/ethnicity
Non-Hispanic White	288	64
African American	84	18.6
Hispanic	64	14.2
Asian/Pacific Islander	8	1.8
Other	6	1.3
Annual income
< $20,000	43	9.9
$20,000–$49,999	176	40.6
≥$50,000	215	49.5
Education level
None to some high school	16	3.6
High School graduate	48	10.7
Some college	157	35.1
College graduate or higher	226	50.6
Smoking status
Current	21	4.7
Former or never	424	95.3
Recreational physical activity (metabolic equivalents/week)
<5	111	25.0
5 – 19.9	190	42.8
≥20.0	143	32.2
Meals eaten at homeb (%)
< 75.0	103	22.9
75.0–94.9	264	58.7
≥95.0	83	18.4

Abbreviations: NPAAS, Nutrition and Physical Activity Assessment Study; WHI, Women’s Health Initiative;

^aAge and body mass index were collected at NPAAS baseline, whereas all other variables were obtained at WHI baseline.

^bMeasured by 4DFR.

Table 2.

Geometric Means With 95% CI and Coefficients of Variation (CV) for Total Sugars Intake and Total Sugars Density Assessed by FFQ, 24HR, 4DFR and Urinary Sugars Biomarker Among NPAAS (2007–2009) Participants (Primary Study, N = 450; Reliability Study N = 88)

Instrument	Total sugars intake (g/day)				Total sugars density (g/1000 kcal)
	N	Geometric Mean	95% CI	CV	N	Geometric Mean	95% CI	CV
FFQa	450	82.4	(78.7,86.4)	0.54	450	56.6	(55.0,58.3)	0.32
FFQb	88	82.8	(75.1,91.3)	0.50	88	55.1	(51.4,59.1)	0.34
4DFRa	450	86.5	(83.3,89.8)	0.42	450	53.5	(51.9,55.1)	0.33
4DFRb	88	83.8	(77.4,90.7)	0.39	88	52.0	(49.0,55.2)	0.29
1st 24HRa	398	80.1	(75.8,84.6)	0.60	398	52.3	(50.1,54.6)	0.46
2nd 24HRa	425	80.2	(76.0,84.6)	0.61	425	54.0	(51.8,56.3)	0.46
3rd 24HRa	421	82.0	(77.9,86.3)	0.57	421	53.5	(51.4,55.6)	0.43
24HRa (3-Day Avg.)	440	85.1	(81.8,88.5)	0.44	440	54.8	(53.2,56.4)	0.32
1st 24HRb	86	78.7	(69.9,88.7)	0.61	86	52.7	(48.2,57.8)	0.45
2nd 24HRb	83	81.5	(71.2,93.4)	0.70	83	51.4	(45.8,57.8)	0.58
3rd 24HRb	82	72.3	(62.7,83.4)	0.74	82	48.8	(43.4,55.0)	0.59
24HRb (3-Day Avg.)	87	82.4	(75.1,90.5)	0.47	87	53.1	(49.4,57.2)	0.36
Biomarkera,c	384	158.8	(145.6,173.1)	1.06	351	78.3	(71.6,85.6)	1.03
Biomarkerb,c	78	173.9	(142.9,211.6)	1.09	73	82.9	(67.3,102.2)	1.14

Abbreviations: CV, Coefficients of Variation; CI, Confidence Intervals; FFQ, Food Frequency Questionnaires; NPAAS, Nutrition and Physical Activity Assessment Study; 4DFR, 4-day food records; 24HR, 24-h dietary recall;

^aCollected in the primary study.

^bCollected in the reliability study.

^cCalibrated using the measurement error parameters generated from the feeding study (17).

^dEnergy intake estimated by DLW measurement of total energy expenditure.

Figure 1.

Association Between Log Biomarker-Based and Log FFQ-based Total Sugars Intake Among NPAAS Participants

Figure 3.

Association Between Log Biomarker-Based and Log 3d average 24HR-based Total Sugars Intake Among NPAAS Participants

ME model parameters for the FFQ, 4DFR and 24HR are reported in Table 3. For total sugars (g/d), the slope in the regression of reported on true intake (β_Q1, β_F1 or β_R1) was approximately 0.2 for all instruments. This slope measures the extent to which bias in self- report is related to true intake: β_Q1=1 indicates no intake-related bias, while β_Q1<1 indicates the extent to which participants tend to under-report high and over-report low intakes (a “flattened slope phenomenon”). A value of 0.2 denotes large intake-related bias. Energy adjustment using total sugars density did not improve the intake-related bias for any of the instruments.

Table 3.

Measurement Error Structure for log Total Sugars Intake and log Total Sugars Density Assessed by FFQ, 24HR and 4DFR Among NPAAS (2007–2009) Participants (Primary Study, N = 450; Reliability Study N = 88)

	Variance of true Intake ( ${\sigma }_T^2$)	Instrument	Slope in regression of reported on true intake (βQ1 or βR1)	Variance of person-specific bias ( ${\sigma }_{u_Q}^2$ or ${\sigma }_{u_R}^2$)	Correlation of person- specific bias with FFQ person-specific bias (ρuQuR)	Variance of within-person error ( ${\sigma }_{{\epsilon }_Q}^2$ or ${\sigma }_{{\epsilon }_R}^2$)
Total sugars intake (g/day)	0.26 (0.07)	FFQ	0.21 (0.09)a	0.15 (0.02)		0.06 (0.01)
		4DFR	0.20 (0.08)	0.07 (0.01)	0.62 (0.07)	0.06 (0.01)
		24HR	0.20 (0.08)	0.07 (0.01)	0.63 (0.07)	0.16 (0.01)
Total sugars density (g/1000 kcal)	0.23 (0.07)	FFQ	0.21 (0.08)	0.06 (0.01)		0.03 (0.01)
		4DFR	0.14 (0.07)	0.05 (0.01)	0.84 (0.07)	0.04 (0.01)
		24HR	0.22 (0.08)	0.04 (0.01)	0.70 (0.08)	0.11 (0.01)

Abbreviations: FFQ, Food Frequency Questionnaires; NPAAS, Nutrition and Physical Activity Assessment Study; ME, measurement error; 4DFR, 4-day food records; 24HR, 24-h dietary recall;

^aStandard error (all values in parenthesis).

Note: All the parameters were estimated using FFQ-, 24HR- and 4DFR-ME models adjusted for BMI and age, and biomarker ME model adjusted for age.

The variance of the person-specific bias in total sugars intake was much larger for FFQ (0.15) than for 4DFR (0.07) or 24HR (0.07). Energy adjustment led to smaller person-specific biases for all instruments (0.04–0.06). Person-specific biases in both 4DFR and 24HR were highly correlated with person-specific bias in FFQ [≈0.6 for total sugars, and 0.70 (24HR) and 0.84 (4DFR) for density]. The variance of within-person error in 24HR was three times as large as that in FFQ or 4DFR, and remained the greatest after energy adjustment.

The AF measures the level of attenuation (under-estimation) of disease risk due to ME in self-reported intake; AF=1 indicates no attenuation, while AF<1 indicates the extent of attenuation. Total sugars intake measured by one 24HR had the smallest AF (0.20), due to the large within-person errors in 24HR (Table 4); increasing the number of 24HRs increased the AF (two 24HR=0.29; three 24HR=0.34). The AF for 4DFR (0.33) was similar to the average of three 24HR, while the AF for FFQ (0.22) was similar to one 24HR. Energy adjustment improved AFs for FFQ (0.48) and three 24HR (0.57), but not 4DFR (0.32). Results for CCs were qualitatively similar to those observed for AFs (Table 4).

Table 4.

Attenuation Factors for Self-Reported Intake and Correlation Coefficients Between Self-Reported and True Intakes for log Total Sugars and log Total Sugars Density Among NPAAS (2007–2009) Participants (Primary Study, N = 450; Reliability Study N = 88)

Instrument	Total sugars intake (g/day)		Total sugars density (g/1000 kcal)
	Attenuation Factor	Correlation with true intake	Attenuation Factor	Correlation with true intake
FFQ	0.22 (0.08)a	0.22 (0.08)	0.48 (0.13)	0.32 (0.09)
4DFR	0.33 (0.10)	0.26 (0.08)	0.32 (0.13)	0.21 (0.09)
Single 24HR	0.20 (0.06)	0.20 (0.06)	0.32 (0.07)	0.26 (0.07)
Avg. of two 24HR	0.29 (0.09)	0.24 (0.08)	0.48 (0.11)	0.32 (0.08)
Avg. of three 24HR	0.34 (0.10)	0.26 (0.08)	0.57 (0.13)	0.36 (0.09)

Abbreviations: FFQ, Food Frequency Questionnaires; NPAAS, Nutrition and Physical Activity Assessment Study; ME, measurement error; 4DFR, 4-day food records; 24HR, 24-h dietary recall;

^aStandard error (all values in parenthesis).

Note: All the parameters were estimated using FFQ-, 24HR- and 4DFR-ME models adjusted for BMI and age, and biomarker ME model adjusted for age.

Calibration equations to predict biomarker-based “true” sugars intake given self-reported intake and other covariates are presented in Table 5. For total sugars (g/d), self-reported intake alone explained little of the variation in “true” intake (Adjusted R²=5–6%), and none of the other covariates was a strong predictor; among investigated covariates, only smoking status, education and physical activity were retained (Adjusted R²=16–18%). Using total sugars density modestly increased the percentage of variation explained by 24HR (18% vs. 6%) and FFQ (10% vs. 5%), but not 4DFR (7% vs. 6%). The final sugars density model explained 29–40% of variation in “true” intake, leaving a considerable portion unexplained. R² unadjusted for within-subject error in the biomarker had markedly lower values compared to adjusted R² across all instruments.

Table 5.

Regression Calibration Equations From Regression of log Calibrated Biomarker^a on log Total Sugars Intake and log Total Sugars Density Assessed by FFQ, 4DFR and Three 24HRs Among NPAAS (2007–2009) Participants (Primary Study, N = 384; Reliability Study N = 78)

Variable	Total sugars intake (g/day)			Total sugars density (g/1000 kcal)
	FFQ	4DFR	Avg. of three 24HR	FFQ	4DFR	Avg. of three 24HR
Intercept	3.51 (2.50) c	2.72 (2.50)	4.63 (2.45)	0.49 (2.53)	0.49 (2.51)	0.64 (2.49)
Log self-reported intake	0.23 (0.08)	0.30 (0.10)	0.26 (0.10)	0.44 (0.13)	0.36 (0.13)	0.52 (0.14)
Log Age	0.17 (0.52)	0.25 (0.53)	−0.02 (0.52)	0.77 (0.53)	0.85 (0.54)	0.73 (0.53)
Log BMI	−0.11 (0.20)	−0.07 (0.20)	−0.19 (0.19)	−0.37 (0.20)	−0.38 (0.20)	−0.46 (0.20)
Smoking status (current)	−0.64 (0.24)	−0.63 (0.24)	−0.59 (0.24)	−0.56 (0.24)	−0.62 (0.25)	−0.57 (0.24)
Education:
< HS graduate	0.55 (0.23)	0.53 (0.23)	0.51 (0.22)	0.64 (0.24)	0.65 (0.24)	0.41 (0.24)
HS graduate	0.30 (0.14)	0.29 (0.14)	0.23 (0.14)	0.43 (0.15)	0.44 (0.15)	0.36 (0.14)
Some college	0.16 (0.09)	0.16 (0.09)	0.13 (0.09)	0.18 (0.09)	0.19 (0.09)	0.16 (0.09)
College graduate	0	0	0	0	0	0
Physical activity
Square-Root METs	0.04 (0.02)	0.04 (0.02)	-
For Self-reported intake only
R2	0.01	0.02	0.02	0.03	0.02	0.04
Adjustedd R2	0.05	0.06	0.06	0.10	0.07	0.18
Multiple
R2	0.06	0.06	0.04	0.09	0.09	0.10
Adjustedd R2	0.18	0.17	0.16	0.34	0.29	0.40

Abbreviations: FFQ, Food Frequency Questionnaires; NPAAS, Nutrition and Physical Activity Assessment Study; METs, metabolic equivalents; 4DFR, 4-day food records; 24HR, 24-h dietary recall;

^aCalibrated using the measurement error parameters generated from the feeding study (17).

^bRegression calibration models automatically included self-reported intake, log age and log BMI as covariates: additional covariates were chosen by backward selection (P = 0.1) from the following covariates: square root physical activity (METs/week), smoking status (current, former/never), supplement use (yes, no), race (black, white, Hispanic, other), education (< HS graduate, HS graduate, some college, college graduate), annual income (< $20,000, $20,000-$49,999, ≥$50,000), and percentage of meals eaten at home (< 75, 75–94, > 94).

^cStandard error (all values in parenthesis).

^dR² and multiple R² are adjusted for within-subject error in the biomarker.

Excluding incomplete urines based on PABA had only modest effect on the biomarker-predicted intake estimates, and ME parameters, AFs, CCs and regression coefficients for self-reported intakes (Supplementary Tables 1–4). Omitting age from the ME model for the biomarker had a dramatic effect on the geometric means of the biomarker-predicted intake (making them considerably smaller and closer to the self-reported means) (Supplementary Table 5), but little effect on other findings (Supplementary Tables 6–8). Yet, geometric means measure intake on a group level only, and give no indication of an instrument’s ability to measure individual’s intake, which is of interest in a cohort study.

DISCUSSION

Using the urinary sugars biomarker to predict “true” intake of total sugars in these post-menopausal women, we found self-reported sugars to be substantially and roughly equally misreported across the FFQ, 4DFR and 24HR. This is the first study to assess ME properties and ME-associated disease risk attenuation for a FR with regard to sugars using a biomarker. Investigation of instruments’ ME structure revealed considerable level of intake-related and person-specific biases, as well as within-subject error for all three instruments. Person-specific bias in 24HR and 4DFR was strongly correlated with bias in the FFQ. We found that sugars density measured by three 24HRs would result in the least attenuation of the association of sugars with disease risk, followed by FFQ and lastly by 4DFR. In this analysis, self-reported sugars and participants’ personal characteristics recovered only some of the variation in “true” intake.

In this population we observed large intake-related bias, comparable in magnitude across the three instruments. Interestingly, the bias did not improve with energy adjustment (even becoming greater) for the 4DFR, implying that other macronutrients may be larger contributors to intake-related bias in energy relative to sugars. The FFQ had almost double the person-specific bias compared to 4DFR and 24HR, which improved with energy adjustment. Another study of similar design using the sugars biomarker to investigate misreporting on FFQ and 24HR (16), reported similar levels of person-specific bias in females. Furthermore, similar to the OPEN study, we found the person-specific bias in 24HR strongly correlated with the bias in FFQ. The same was apparent for 4DFR, suggesting that the use of 24HR and 4DFR as reference instruments for FFQ-based total sugars may provide a very incomplete ME correction procedure. Whereas the intake-related bias, which was considerable in all three instruments, may ultimately cause inflation of the risk estimate through creating the flattened slope phenomenon (29), the person-specific bias and within-person random error have an opposing effect, causing risk attenuation (30). The fact that the AFs<1 across all self-reports for both absolute and energy-adjusted sugars intake indicates that the person-specific bias and within-person random error overrode the effect of intake-related bias, causing underestimation of true effect in a disease model with self-reported sugars. The AF for three 24HRs was most favorable and much improved with energy adjustment (AF=0.57), followed by the FFQ (AF=0.48). In contrast, AF for 4DFR became less favorable with energy adjustment (AF=0.32), suggesting that errors in 4DFR-based sugars were independent from errors in the energy estimate. Based on the AFs, in an association study using self-reports, a true RR=2 for a given change in energy-adjusted intake will be observed as 2^0.57=1.5 for three 24HR, 2^0.48=1.4 for an FFQ and 2^0.32=1.2 for 4DFR. In comparison to NPAAS women, OPEN women had lower AFs (0.33 for FFQ and 0.35 for two 24HRs for sugars density).

Using calibration equations, self-report instruments explained very little variation in “true” sugars intake. BMI, age and ethnicity have been identified as sources of systematic bias in dietary self-reporting, and their incorporation in equations may provide certain ME adjustment and “strengthen the signal” (9, 10). None of the covariates we investigated was a particularly important predictor of “true” sugars intake. Our final models explained a total of 16–18% of “true” absolute sugars and 29–40% of “true” sugars density variability, lower than observed for energy, protein and protein density in an earlier analysis of NPAAS (10). Being a smoker was associated with lower sugars consumption, whereas lower education with higher consumption, consistent with the latest reports on added sugars intake in the U.S. (31, 32). Although BMI would be expected to be a significant determinant of sugars intake, and associated with sugars misreporting (33), we did not find BMI to improve predictability of the calibration equation. In the original (17) and a later feeding study (34), BMI showed no effect on the performance of the biomarker, i.e. BMI was not a predictor (17) or an effect modifier (34) in the association between the sugars biomarker and known intake. In contrast, a recent feeding study found gender and %body fat to be significant predictors in the regression of the sugars biomarker to known dietary sugars intake, in a study design originally developed to investigate the effect of high vs. low glycemic load (GL) diets (35). Participants were fed two constant isocaloric diets, which were identical in macronutrient composition (%energy) and only differed by GL and fiber content. Given all participants received the same diets adjusted to fit each participant’s energy requirement, it may have been that gender and %body fat, being major determinants of energy requirement, were significant predictors of urinary sugars in this study (35).

Almost a third of our participants had urinary sucrose concentration below the LOD, similarly to what was found in the OPEN study (16). Sucrose is not a dominant sweetener in the U.S. diet, and it can be expected that urinary sucrose excretion will be low in this population.

Given the high burden of the study and the stratified recruitment algorithm which provided the intended enrollment for age, race and BMI ranges, the response rate of 20.6% was judged to be adequate. Moreover, the intake estimates for energy, protein and protein density, calculated by calibration equations generated in the Nutrient Biomarker Study (NBS), an earlier WHI biomarker study with a response rate almost double the rate in NPAAS, were highly correlated (r=0.95–0.96), indicating good transferability of the calibration equations (9, 10). A potential limitation of this analysis is that our women were ≥60 years of age, whereas the study used to derive the sugars biomarker included subjects 23–66 years (17). Yet, excluding age from the calibrated biomarker did not appreciably change any of the findings. We acknowledge that the calibrated sugars biomarker was derived from a feeding study of a limited sample size (17). The biomarker contains certain level of bias, possibly arising from person-specific differences in absorption, gastrointestinal mucosal integrity, hepatic metabolism, or renal excretion of these two nutrients, determined by genetic factors, physiologic or medical conditions, and those may differ between different populations. Hence, more well-designed feeding studies collecting multiple biomarker measurements per participants from different populations are needed to further explore the biomarker’s characteristics and ‘behavior’, and to confirm that the calibration parameters quantified from the original feeding study are stable and indeed applicable in different populations.

In summary, none of the self-report instruments provided a good estimate of sugars intake, although overall 24HRs seemed to perform the best. While 4DFR was comparable to multiple 24HRs in measuring absolute sugars, its performance did not improve with energy adjustment. Lower performance of the 4DFR compared to the 24HR and FFQ with regard to sugars density is surprising, given its administration most closely corresponded with the time-frame of biomarker collection (i.e. within 2 weeks), whereas the 24HRs were administered a month apart starting 1–3 weeks after the biomarker assessment, and the FFQ, inquiring about usual diet over the preceding three months, was completed two weeks before biomarker collection.

Assuming that the calibrated sugars biomarker provides unbiased estimate of total sugars, this analysis suggests that measuring the biomarker in a subsample of the study population for calibration purposes may be necessary for obtaining unbiased risk estimates and reliable 95% confidence intervals in disease association studies. Accordingly, the regression calibration equation for the FFQ developed here, can be used in future WHI analyses to calibrate FFQ sugars intake and examine its associations with cancer and other disease outcomes.

Figure 2.

Association Between Log Biomarker-Based and Log 4DFR-based Total Sugars Intake Among NPAAS Participants