Participants of this relatively large (n
1011) calibration study were representative of the parent cohort with respect to gender, age, education, BMI and dietary patterns. Because of our initial selection by church, subjects also were geographically spread throughout the USA and Canada. As a reference measure the present study obtained six 24 h recalls, whereas other cohorts with large calibration samples typically used fewer recalls(4,11,12)
or a 4 d food record(13)
that did not systematically combine weekdays and weekends.
Intake estimates of most nutrients from both FFQ and recall data generally were higher in whites compared with blacks. However, blacks reported higher intakes of protein, animal protein, very-long-chain fatty acids and vitamin D. This is not surprising, given the greater proportion of non- and semi-vegetarian eating patterns followed by this minority population in the study. As expected, nutrient intake estimates derived from FFQ tended to be higher than those from 24 h recalls. This probably results when people are asked to recall the frequency of intake of a large number of foods in an FFQ and thus tend to overestimate their actual intake(9)
. In a study that assessed bias of food frequency-based measures of fruit and vegetable intakes, Kristal and colleagues compared intake estimates between recalls (or food records) ν
. a brief questionnaire with only seven items or an FFQ with thirty or more items(14)
. They found that estimated intakes were lower from the brief questionnaire, but higher from the longer questionnaire when compared with recalls or food records.
Despite differences between the recalls and FFQ in the current study, deattenuated correlation coefficients were ≥0·40 for forty-three of fifty-one nutrients examined in both ethnicities combined. Estimates from the FFQ of individual fatty acids, total long-chain fatty acids and fibre were moderately to highly correlated, and the micronutrients (except Na, Cu and dietary vitamin E) were also generally highly correlated with recall data. This shows that the AHS-2 FFQ has the ability to provide relatively good estimates of these nutrients.
Deattenuated correlation coefficients with values less than 0·40 included energy, total protein, total carbohydrate (in blacks), non-fibre carbohydrate (in blacks), Na, dietary vitamin E (in whites) and β-carotene (in blacks). Energy intakes in both races were lower compared with others(4,12,15–18)
. For carbohydrate, our results in blacks and whites were lower or within the range of those observed by other investigators(12,15,17,19)
Our deattenuated validity correlations for total protein were within the range of(4,20)
than those reported in other validation studies. It was particularly noticeable in our study that the individual protein components (animal and vegetable protein) that contribute to this composite nutrient had greatly higher coefficients than total protein. In examining the results from the FFQ and recall data, it appears that the questionnaire over-estimates the amount of vegetable protein intake, but underestimates the amount of animal protein compared with the recalls. The most likely explanation is the differential scaling of animal and plant sources of protein in our FFQ, which includes a long list of plant proteins (in part due to the attention we give to meat analogues) but a comparatively shorter list of animal protein foods. In calculating the deattenuated correlation coefficient for total protein, we noted that the variance of animal protein (APRO) from recalls (315·9) was much greater than the variance of animal protein from the FFQ (211·34), yet the variance of vegetable protein (VPRO) from recalls (165·24) was much smaller than the variance of vegetable protein from the FFQ (325·13).
We resolved this problem by putting the animal and vegetable protein FFQ values on the same scale as each other and the recall values. In a separate analysis, we first calculated for both APRO and VPRO E(recalls|FFQ), summed the two estimates and then correlated this with recall values for total protein. Our correlations for total protein subsequently improved from 0·28 to 0·54 in all subjects combined. This interesting observation suggests that the product-sum method of calculating FFQ intakes may cause distortion at times, depending on the form of the FFQ, and we plan to investigate this issue more fully for other variables (e.g. total energy).
In the current validation study energy-adjusted uncorrected correlations improved after correction for attenuation in the recalls, as others have found. We also observed improved correlations in general when supplements were added, which emphasizes not only the importance of assessing both dietary and supplemental sources of micronutrients, but also that supplements contribute importantly to total nutrient intake in this population.
values reported by Johansson et al
address the same issue as the bias factor (λ
– 1) results in the current study. They illustrate quite dramatically the severe proportional biases that result even when validity correlations are well within values commonly held to be acceptable in FFQ work. For example, a univariate regression where uncorrected energy-adjusted PUFA from the FFQ in white subjects (deattenuated validity correlation of 0·72) was used as the independent variable would produce a β
coefficient that was biased downwards by 58 %.
When comparing our validation results by race, we found that the deattenuated energy-adjusted validity correlations of several nutrients in our study were smaller in blacks than whites. This is consistent with the reports of others(4,13,18,20,22)
. Presumably these may be explained by somewhat lesser educational attainment(13,18)
on average among blacks or less familiarity with research studies of this sort. In an older biracial sample, investigators also reported lower validity in the very old (aged 79 years and older) compared with those aged 68–78 years(20)
. In our calibration study, a greater proportion of whites achieved higher education, but they were some-what older compared with blacks.
We provide in this report the validation of an unusually comprehensive list of micronutrients and fatty acids, and with few exceptions find moderate to high validity. Thus we expect to be able to use most of these results to good effect in both traditional analyses of disease risk and analyses incorporating recall- or biomarker-guided measurement error correction.