1. Introduction
Nutritional epidemiological studies are dependent on the accuracy of the habitual dietary survey results they use. The purpose of a dietary assessment is usually one of the following: (1) to compare the mean or median food intake between groups, (2) to rank individuals belonging to the same group, or (3) to evaluate an individual’s habitual food intake [
1]. Epidemiological studies usually use food frequency questionnaires (FFQs), dietary records (DRs), or 24-hour dietary recalls (24 HRs) to assess a participant’s dietary habits. Of those, FFQs have been widely used in many epidemiological studies because they are easy to administer to more participants with a lower staff burden and cost than DRs or 24 HRs [
2]. However, actual dietary intake has wide variations, which are dependent on food culture and dietary habits of the target population. To limit the burden of the participants, FFQs can include only a limited number of foods to maintain high feasibility. Thus, FFQs need validation studies to accurately and precisely estimate the dietary intake of the target population for these purposes [
3].
In validation studies, energy and nutrient intake estimated by FFQs have been commonly compared with those estimated using DRs [
4,
5] or 24 HRs [
6], which are considered more robust methods than FFQs. However, reporting biases associated with variables such as age [
7,
8,
9], sex [
10,
11], body mass index (BMI) [
8,
9,
11,
12], smoking status [
8,
11], educational attainment [
8,
11], socioeconomic status [
8,
13], and social desirability [
13] cannot be entirely avoided when evaluating the dietary intake using self-reporting dietary assessment methods such as DRs, 24 HRs, and FFQs. Misreporting can easily happen intentionally or unintentionally in self-reporting dietary assessment methods, which leads to systematical or random errors. In fact, it is known that these self-reporting dietary assessment methods do not accurately estimate an individual’s usual dietary intake [
14,
15,
16,
17]. Therefore, calibration with objective biomarkers is necessary to reduce reporting biases to a bare minimum [
8,
9,
18].
For the energy intake (EI), when the body weight status is stable, the doubly labeled water (DLW) method is the preferred validation method when measuring the total energy expenditure (TEE) [
7,
8,
9,
10,
12,
15,
16,
19]. For example, a previous study has reported a comparison between the EI obtained by a FFQ and the TEE measured by the DLW method in Japanese middle-aged adults [
19]. However, to the best of our knowledge, the validation of the EI estimated by FFQ using the DLW method in the Japanese older adult populations has not been addressed [
20].
The population of people aged ≥65 years in Japan is rapidly increasing, being currently 35.6 million, a 28.1% increase as of 1 November 2018, and is even expected to increase by over 40% by 2055. Thus, epidemiological studies addressing health issues in the older population are critical to avoid a further increase in social burden and/or social security expenses. For that purpose, adequate tools assessing their food intake are required.
Our study group has launched an epidemiological study based on an older Japanese study, referred to as the Kyoto–Kameoka study. This study used a 47-item FFQ (in which one question is for alcohol) [
5,
21,
22,
23], which was originally developed and validated against DR for the middle-aged population [
4,
24]. This FFQ is currently being used in multi-site studies across the country, such as the Japan Multi-Institutional Collaborative Cohort (J-MICC) study [
25]. Recently, we have validated the FFQ against DR in our cohort [
5], but there are unresolved questions about the validity of FFQ against biomarkers like the DLW method, which does not rely on self-reporting tools.
In this study, we aimed to examine the validity of the EI estimated by the DR and 47-item FFQ against TEE measured by the DLW method and develop a calibration equation to address self-reporting bias for the EI estimated by FFQ.
3. Results
The characteristics of the participants in the present study are shown in
Table 1. The mean age was 72.2 (range: 65 to 84) and 73.5 (range: 66 to 88) years in women and men, respectively. The mean BMI of men was 23.0 (range: 16.8 to 31.1) and 22.7 (range: 14.3 to 30.0) kg/m
2 in women and men, respectively. The mean FFM was 32.8 and 43.9 kg in women and men, respectively. Notably, a low BMI (<18.5 kg/m
2) was found in 5 (10.0%) women and 5 (8.5%) men, while 14 (28.0%) women and 13 (22.0%) men were found to be overweight (≥25.0 kg/m
2). In addition, we found that 46 (81.6%) women and 57 (94.9%) men went out once a week and that 13 (26.0%) women and 22 (37.3%) men had high educational attainment (≥13 years).
Table 2 shows the comparison of TEE measured by DLW and EIs assessed with DR and FFQ. Among all participants, the mean TEE was 2175 (range: 1246 to 3435) kcal/day, and the mean EI estimated from the DR and 47-item FFQ were 1972 (range: 1306 to 2948) and 1774 (range: 736 to 3461) kcal/day, respectively. The ratios of EI assessed with the DR and 47-item FFQ to TEE measured by DLW were 0.91 (range: 0.57 to 1.52) and 0.82 (range: 0.20 to 1.61). The ratio of EI estimated by FFQ to TEE measured by DLW were significantly lower than ratio of EI estimated by DR to TEE (−0.09, 95% confidence interval: −0.13 to −0.05,
p <0.001). We observed similar results in the stratified groups, except in participants aged ≥75 years old.
The correlation coefficients of TEE measured by DLW and of EI assessed with the DR and 47-item FFQ are shown in
Table 3. Pearson’s and Spearman’s correlation coefficients of EI estimated by DR and this FFQ were significantly correlated with TEE measured by DLW. Moreover, there was no significant difference in the correlation coefficient between EI estimated by DR and this FFQ against TEE using Meng’s Z-test comparison
(p values are shown in the right column). We observed similar results in the stratified groups except in women.
Table 4 shows the results of the multivariate analysis of the linear model with TEE measured by DLW as the dependent variable and the EI estimated by this FFQ as the explanatory variable. The determinant coefficient (R
2) of the linear regression analysis was 0.36 for this FFQ. The age, sex, BMI, and EI estimated by this FFQ were included as significant independent variables in the model, while smoking, education, living, socioeconomic status, and physical activity (going out once a week) were not included. We developed an equation to calibrate the EI estimated by this FFQ using the multiple regression analysis. The models followed the equation:
This equation is modeled to calibrate the EI estimated by this FFQ, where ε is the calibrated mean EI and the intercept (
β0) is 1384.92 kcal in this FFQ. For binary variables, the coefficient of age (
β1) was −166.98 kcal for ≥75 years, and the coefficient for sex (
β2) was −354.72 kcal for women. For continuous variables, the coefficient for BMI (
β3) was 25.55 kcal/[kg/m
2], and the coefficient for EI (
β4) was 0.24 kcal with the 47-item FFQ. All regression coefficients are shown in
Table 4.
A review of previous and current studies on the developed calibration equation for energy intake from a self-reported dietary assessment and DLW is shown in
Table 5. Our results were similar to those of previous studies wherein the age, BMI, and EI estimated by FFQ were determined as significant independent variables in the calibration equation.
The comparison of the EI/TEE ratios according to the BMI group is shown in
Table 6. We calculated calibrated EI values using the equation (1) that we created based on our multivariate analysis (
Table 4). In addition, we determined EI/TEE from both uncalibrated and calibrated EI values divided by TEE measured by DLW. Using the Jonckheere-Terpstra trend test, we found no significant differences in the uncalibrated EI/TEE between the different BMI groups or between all types of dietary assessment. However, we observed that the higher BMI group tended to have lower EI/TEE ratios than the lower BMI group. Pearson’s correlation coefficient indicated that there was a significant correlation between BMIs and uncalibrated EI/TEEs (correlation coefficient; −0.19,
p = 0.048) for DR. However, there was no significant correlation between BMI and EI/TEE according to the two types of correlation analysis, and higher BMI tended to be associated with a negative EI/TEE in all uncalibrated dietary assessment methods. The calibrated EI/TEE ratios had lower correlation coefficients than the uncalibrated EI/TEE ratios in FFQ.
4. Discussion
We have recently reported the validation of this 47-item FFQ against the 7-day DR in older adults [
5]. However, even in DR, self-reporting methods have large potential error and bias. Therefore, we aimed to evaluate the precision and accuracy of EI values estimated by this FFQ against TEE measured by the DLW method in the current study. We have shown that EI estimated by FFQ correlates modestly with TEE measured by DLW. In addition, this study suggests that EI underreporting may be attenuated by our developed calibration equation based on our multiple regression analysis, which included variables affecting self-reporting biases in overweight individuals.
We have shown that EI/TEE ratios estimated using the FFQ are lower than those estimated by DR (
Table 2). Moreover, it has been reported that, in comparison with longer questionnaires and DR, short dietary survey questionnaires tend to underestimate the energy and nutrient intake more [
41]. However, a previous study has reported the development of several FFQs to assess the habitual Japanese dietary intake, where the median number of food items was 45 (ranging from 9 to 169) [
20]. The number of food items in our FFQ was similar to this median. In addition, our FFQ was developed based on foods that contribute to 85% of the between-person variance for each energy and nutrient intake in the middle-aged population [
24]. The reason for the underestimation may be that the foods and beverages listed in the short questionnaire may not sufficiently reflect those habitually consumed by the target population. However, the current studies had a higher EI/TEE estimated by FFQ and DLW than that in the previous studies (
Table 5) [
8,
9]. These differences may not be explained due to the differences in the number of food items included in FFQs, but may be because of the differences in sex, race/ethnicity, and BMI. Therefore, these population attributes need to be considered when comparing studies reporting EI by FFQ. For the individual ranking, the EI estimated by DR and FFQ correlated modestly with the TEE measured by DLW (
Table 3). Interestingly, we found no significant correlation between EI and TEE with any of the EI assessment methods in women. A previous study has reported that, among middle-aged Japanese adults, women tend to demonstrate a weaker correlation between EI and TEE than men [
19]. Both studies in
Table 5 included only women, while, in the current study, 54% of the participants were men, which perhaps explains the higher EI/TEE of the current study than the studies presented in
Table 5.
In addition, there was a significant difference between the ratio of EI estimated by DR to TEE and that estimated by FFQ to TEE in participants aged <75 years old but not ≥75 years old (
Table 2). Moreover, participants aged ≥75 years old tended to have a higher correlation between TEE measured by DLW and EI assessed by FFQ than participants aged <75 years old (
Table 3). We previously reported that the number of food items consumed daily tended to decrease in older people [
5]. Therefore, we speculated that a lower number of food items may be sufficient to estimate the energy and nutrient intake using an FFQ in older populations, as the number of food items affecting inter-individual variability consumed by older people seemed smaller than that of middle-aged adults. This is perhaps the reason why, even though the FFQ has limited food items compared with the DR, there was no difference in the results obtained from DR and FFQ in the present study.
The EI estimated by self-reporting dietary assessment methods has been reported to be associated with variables such as age [
7,
8,
9], sex [
10,
11], BMI [
8,
9,
11,
12], smoking status [
8,
11], educational attainment [
8,
11], socioeconomic status [
8,
13], and social desirability [
13], all of which can generate reporting bias. We included these variables in our multiple regression analysis and showed that age, sex, BMI, and EI were related to TEE measured by DLW (
Table 4). Previous studies reported that the reporting bias of dietary intake error was associated with BMI, age, race/ethnicity, annual income, physical activity, and diet change intervention (
Table 5). Some epidemiological studies have reported that, for individuals with a high BMI, the EI calculated using self-reporting dietary assessment methods tends to be underestimated [
8,
9,
11,
12]. We have shown that there tends to be a negative (but not quite significant) correlation between BMI and EI/TEE calculated using an uncalibrated EI estimated by FFQ (
Table 6). Notably, this error may not only decrease the accuracy of the group’s mean EI but also induce the misclassification of individual ranking. This makes EI estimated by self-reporting dietary assessment methods difficult to use in nutritional epidemiological studies. Previous works have reported that calibrating EI using TEE measured by the DLW method using multiple regression analysis improved the measurement error from self-reporting dietary assessment methods [
8,
9]. Similarly, we have also shown that calibration lowers the correlation coefficient between BMI and EI/TEE (
Table 6). Recently, some nutritional epidemiological studies have reported that EI is associated with the incidence of diabetes [
42] and cancer [
43] only when EI is calibrated. Based on these reports, we could speculate that the calibration of EI attenuates measurement errors due to reporting bias. Therefore, this approach is helpful to assess the relationship between EI and disease in nutritional epidemiological studies because the self-reporting bias problem in EI assessment makes it unsuitable for this type of study.
Our study has several limitations. First, we assumed perfect nourishment balance conditions in all participants to be able to use the DLW method. However, since we did not confirm this assumption, it is possible that not every participant had a perfect nourishment balance. Second, we could not develop a calibration equation of macronutrients by other biomarkers including urinary samples and serum concentrations. It is a definite limitation of the study that the calibration is only useful for studies focused on energy intake, but not useful for studies that are focused on the role of a specific macronutrient in the incidence or prevalence of disease. Third, there was an interval of 2–3 months between the TEE measurement by the DLW method and the survey with FFQ. In addition, we could not follow a multipoint method for the measurement of TEE by the DLW method. These experimental conditions might have led to random and systematic errors because participants may have modified their usual activity. Finally, this study only included participants who consented to participate in the physical check-up examination in the Kyoto–Kameoka study. Such participants might be more health conscious than those who did not. In addition, this study included a smaller number of participants than in previous studies (
Table 5). Therefore, the distribution of EI and TEE may not reflect that of the general population because these participants were not randomly sampled from the general Japanese older population.
DR faces the problem of measurement error, and there is also a large burden on the participant. However, in our study, in comparison with the measured TEE by the DLW method, the energy intake estimated from DR tended to have a higher correlation coefficient than that from FFQ. Therefore, DR may invariably be important in nutritional epidemiology. Better-quality studies will help assess the precision and accuracy of energy intake estimated using recovery biomarkers such as DLW. Energy intake estimated by self-reporting dietary assessment methods including FFQ tends to be underestimated. The calibration technique using DLW for energy intake estimated by FFQ may be helpful to assess the relationship between EI and disease in nutritional epidemiological studies.