J Clin Epidemiol Vol. 50, No. 8, pp. 925-937, Copyright 0 1997 Elsevier Science Inc.
0895-4356/97/$17.00 PII SO895.4356(97)00098-X
1997
ELSEVIER
Development and Testing of a Seven-Day Dietary Recall James R. Hebert, I,* ha S. Ockene,’ Thomas G. Hurley ,’ Rose Luippold,’ Arnold D. Well,3 Morton G. Hamatz,3 and other members of the Dietary Assessment Working Group of the Worcester Area Trial for Counseling in Hyperlipidemia (WATCH) 4 ‘DIVISION OF PREVENTIVE AND BEHAVIORAL MEDICINE, MASSACHUSETTS 01655, ‘DIVISION OF CARDIOVASCULAR WORCESTER, MASSACHUSETTS 01655, )DEPARTMENT MASSACHUSETTS 01003, 4T~~
UNIVERSITY OF MASSACHUSETTS MEDICAL SCHOOL, WORCESTER, MEDICINE, UNIVERSITY OF MASSACHUSETTS MEDICAL SCHOOL, OF PSYCHOLOGY, UNIVERSITY OF MASSACHUSETTS, AMHERST, DIETARY ASSESSMENT WORKING GROUP
ABSTRACT.
Using multiple 24.hr recalls (24HR) we tested the Seven Day Dietary Recall (7DDR) developed to assess nutrient exposures, especially lipids, in dietary interventions and other clinical trials requiring measurement of effect over moderate time periods. A total of 261 individuals in three studies completed a 7DDR at the end of a 3- to 5-week period during which 3 to 7 24HR were telephone-administered on randomly selected days. One of these studies and data from one additional study (total n = 678) allowed us to test the ability of the 7DDR to predict serum lipid changes in an intervention setting. In correlation and linear regression analyses, high levels of agreement between 7DDR and 24HR were obtained. For total energy: r = 0.67 and b = 0.69, and for total fat intake (g/day): I = 0.67 and b = 0.80. When 7 days of 24HR were available agreement tended to be higher. For total energy: r = 0.69 and b = 0.95, and for total fat (g/day): r = 0.71 and b = 1.04. Data derived from the 7DDR and fit to the Keys and Hegsted equations closely predicted actual changes in total serum cholesterol (within 15% and lo%, respectively). The 7DDR is a relatively easily administered,-sensitive method to assess short-term changes in dietary f,t consumption in individuals. J CLIN EPIDEMIOL 50;8:925-937, 1997. 0 1997 Elsevier Science Inc.
KEY WORDS.
Dietary
fats, nutrition
assessment,
nutrition
INTRODUCTION Dietary
assessment
include:
historical
tionnaires
(FFQ),
methods used in epidemiologic studies methods, especially food frequency quesused
to assess
long-term
habitual
dietary
intake; 24-hour diet recalls (24HR), used to assess intakes of groups or, with multiple administrations, to assess intake of individuals over the period of administration; and food diaries (FD), similar to the 24HR in terms of the assessment ‘Address for correspondence: Dr. James R. Hebert, Division of Preventive and Behavioral Medicine, University of Massachuserts Medical School, 55 Lake Avenue North, Worcester, MA 01655 The Dietary Assessment Working Group of the Worcester Area Trial for Counseling m Hyperlipidemia (WATCH) consisted of all authors named above plus Linda Rider of the Arizona Department of Health Services, Office of Nutrition, Phoenix, Arizona; Gordon Saperia and Barbara Mullins of the Fallon Clinic, Worcester, Massachusetts; Judy C. C. Phillips of the Department of Cancer Epidemiology and Cancer Conrrol, Dana Farber Cancer Institute, Boston, Massachusetts; Kenneth Fletcher of the Department of Psychiatry, University of Massachusetts Medical School, Worcester, Massachusetts; and Sarah Ellis, Philip A. Merriam, and Judith K. Ockene of rhe Diwsion of Prevenuve and Behavioral Medicme, University of Massachusetts Medical Center, Worcester, Massachusetts. Accepted for publication on 13 May 1997.
surveys,
cognition,
recall,
serum
cholesterol
time frame and ultimate use of the data, but requiring a written record of the foods as they are eaten. Two advantages shared by the 24HR and the FD are that they do not require the subject to provide an average estimate of intake and that they are open-ended so that foods that may not be listed on an FFQ can be reported [l]. In addition, the 24HR can be randomly administered, thus eliminating biases related to rehearsing or changing diet. The FD can provide a very complete and precise account for careful, conscientious, and well-informed participants. The fact that both the 24HR and FD have been used to assess the accuracy of various FFQs [2] is evidence of the widespread agreement that these methods have the potential for very accurate and precise characterization of the diet of individuals. However, they are not without their problems.
Both
require
very
active
subject
involvement
and
a
high level of compliance in order to work reasonably well. A trained nutritionist must devote a great deal of time either to interview subjects or to check records for errors and enter
data.
Also,
it is necessary
to collect
many
(23 days for even the most easily characterized
days
of data
nutrient,
fat
926
J. R. Hebert et al.
[3-61) to estimate even moderate-term (less than 3 months) dietary exposure. The 24HR requires very good short-term memory and the FD requires that the subject not alter diet on the days it is being recorded. For even a small-scale study, use of these methods is very expensive, usually costing well in excess of $30 per subject per day. In comparing dietary fat data derived from the FFQ with those from multiple days of either records or recalls, the correlation of nutrient scores is generally around 0.50 [2]. Results for other nutrients tend to show even lower levels of agreement [7] owing to higher levels of intra-person variability [8-l 11. Despite its limitations, the FFQ has become the de facto method of choice in large-scale epidemiologic studies of chronic disease. For most studies requiring estimation of nutrient intake, the practical disadvantages of multiple administrations of the 24HR and FD, including cost, simply preclude their use. Pharmacologic or dietary intervention trials present a category of studies in which alternatives to the FFQ may best serve the needs of the study. Trials of pharmacologic agents, such as those aiming to lower serum cholesterol levels, may require nutrient data in order to control for potential confounding or to model for interaction between study agents and specific nutrients. Because the effect may be expressed over a short time, typically several weeks, diet should be measured over a much shorter interval than that which is typically of interest in most epidemiologic studies. Dietary intervention trials, most of which aim to modify intake of specific nutrients [12-E], typically have similar requirements to those of drug trials with respect to the duration of the intervention effect and the interval over which it is expressed. In this paper we report on the development of an instrument to assess short-term intake of fat in individuals and tests of its validity in several different settings. MATERIALS
AND
METHODS
This section describes both the development and the testing of the seven day dietary recall (7DDR). The testing of the 7DDR, which occurred in three studies separate from the parent study for which the 7DDR was developed, is described after the motivation and procedures for its development are presented. Development
of the Sewen Day Dietary
Recall
The 7DDR was developed to assess dietary change in the Worcester Area Trial for Counseling in Hyperlipidemia (WATCH), a randomized controlled trial of a physiciandelivered nutrition counseling intervention for hyperlipidemit patients [16]. As with many diet intervention trials, the WATCH sought to produce a relatively rapid (i.e., ~4 weeks), though durable, change in dietary intake that could be detected and expressed in daily intake of specific nutrients. Also like many diet intervention trials, intake of fat
was a focus of the intervention. In lipid outcome trials, such as the WATCH, change in fat consumption is assumed to be related, on average, to change in serum cholesterol concentration [5,17-191. In trials related to other endpoints, such as cancer, dietary fat often is the nutrient of interest partly because of its putative direct effect on carcinogenesis [20] and in part because change in fat intake may be related to changes in intake of other nutrients thought to affect risk [12,21]. The WATCH intervention was targeted at dietary fat, which has relatively low intra-person, short-term variability [3-61 and tends to remain stable in healthy free-living people. Therefore, we were able to focus the assessment instrument on a relatively short period of time, i.e., the previous week. Because of this short-term focus, it was conceivable that the assessment method could rely, at least in part, on episodic memory, thus obviating some problems with reliance on habitual memory and long-term averaging associated with frequency methods [22-251. Our review of the literature on memory of health-related events and diet indicated that the reference period preferably should be not more than two weeks and optimally about one week [23,25], owing primarily to the very high intrusion rate (i.e., the spurious report of foods as eaten that had not really been eaten) evident in longer-term dietary self-reports. So as to prompt episodic memory and thereby decrease reliance on generic or habitual memory, a recall grid representing all seven days of the prior week and all possible food-encounter times was developed. This grid was included as the second page of the questionnaire packet, immediately after the instruction page. Subjects were asked to count food exposures and not to estimate frequencies, a potential source of error in the FFQ [25]. In pilot-testing, subjects reported that the grid was very helpful in recalling the specific foods eaten over the previous week; an encouraging sign because the goal was to have respondents rely as heavily as possible on episodic, rather than generic, memory. Subjects were asked to select a single portion size estimate from a list that encompassed a 4-fold increase from the smallest to the largest portion size, rather than the usual 2-fold range. So as to reduce the probability of biases resulting from subjects identifying an entire category as consisting of highfat or low-fat foods, listed foods were organized into eight categories encompassing foods with varying fat contents (e.g., low-fat and full-fat dairy products), and without labeling that could lead subjects to conclude all items in a category had similar fat content. This reduces the likelihood of participants identifying categories of foods as either “good” or “bad.” In reviewing dietary data for the U.S. as a whole [26,27] and for our region (unpublished data) it was apparent that we could account for more than 95% of lipid consumption with a list of about 120 foods. The resulting 7DDR is a selfadministered instrument that asks questions concerning the specific number of times consumed and portion sizes of 118
A Seven-Day Diet Assessment
927
food categories or individual foods and 13 beverage items consumed during the previous seven days. The 14-member Dietary Assessment Working Group that designed the 7DDR included four psychologists, four dietitians, a health educator, two cardiologists, two statisticians, and a nutritional epidemiologist who reviewed and modified the form for content validity and clarity. In pilot-testing, it was evident that most subjects were limited more by their ability to use numbers and deal with the complexity of their diets than by reading ability pr se. Therefore, an introductory page was developed that orients subjects to measures that can be taken to help with recall and that provides instructions on the use of the form and examples of the manner in which it should be filled out. In addition to the food items specified on the questionnaire, additional space is provided for foods consumed but not listed. For all such non-listed foods, we applied the following schema in determining nutrient content: (1) We attempted to match the response to one of the 13 1 items listed (in over half the instances the food listed represented an oversight by the subject in not selecting a good match from the list provided); (2) If no good match was found on the 7DDR list, we used the Nutrition Data System (NDS) from the Nutrition Coordinating Center (NCC) at the University of Minnesota to identify a match; (3) In the under 1% of instances where no match was found, we referred to the list of “missing” foods generated from our many thousands of 24HR interviews conducted using the NDS for which the NCC had provided us with the food’s nutrient content. To aid in identification of portion sizes eaten, an eight-page (4 pages, double-sided) booklet of two dimensional food models shown full-sized, as well as a paper ruler superimposed on each page, accompanies the 7DDR. Prompts to recall foods eaten between meals or at meetings, parties, or restaurants are included. Other items in the 7DDR include questions regarding difficulty of completion of the 7DDR; exercise by type, frequency, and duration; level of relative fitness; aspirin use; and vitamin supplement consumption by type, dose, and duration. The questionnaire takes approximately 25 min, on average, for study subjects to complete. The form is optically scanned.
Selection of the 24HR
as the Criterion
Measure
In a previous study by our group we compared an average of 14 days of 24HR to 12 days of FD and a single FFQ. We found very high correlations for all nutrients in comparing the 24HR and FD-derived scores but observed much lower agreement between the FFQ-derived and the FD- or 24HRderived nutrient scores [28-301. As in the work of Posner et al. [31], we found that the multiple 24HR method produces nutrient estimates with smaller total variance than multiple-day FD. Unlike the 24HR, the FD allows the subject to alter diet prior to providing consumption data. Buzzard et al. [15] found that the FD overestimated fat reduction in a
low-fat intervention by 42% relative to telephone-administered 24HR and concluded that randomly administered 24HR may be preferable to multiple-day FD in diet intervention studies. Using potassium and sodium excretion as standards, Forster et al. [32] also found that repeated diet records may not accurately represent usual diet in an intervention trial.
Study Design The subjects for the diet analysis portion of this study come from three sources: (1) The WATCH External Validation Study (WEVS) consisted of a volunteer population of 42 healthy subjects, some of whom were enrolled at the Fallon Clinic, a health maintenance organization in Worcester County, and others who were among the staff at the University of Massachusetts Medical Center (UMMC). WEVS participants were not subjects in the WATCH study. The study was conducted over 5 weeks from October 20 to November 27, 1991. In this study, subjects completed a baseline 7DDR and then were called seven times for the 24HR over a 3week period to include each day of the week for every participant. Participants were told only that they would be interviewed on seven days but not the specific days on which they would be called. We rescheduled the call on the same day the following week if the participant could not be reached, but subjects were not informed of the call day. All calls were completed in the three-week time frame. After all seven 24HRs were completed, subjects were asked to complete a second 7DDR. Of the 42 subjects completing baseline measures, 36 had complete data for analysis. (2) In the Ross Laboratories Nutrient Bar Study, 39 hyperlipidemic subjects were enrolled into a trial to test the effect of a lipid-lowering nutrient supplement bar. Data used here represent the stabilization period before patients received either diet counseling or the supplement bar. This portion of the study occurred from May to June 1994 and was identical to study 1 in every way except that only 4 days of 24HR were collected (1 randomly selected weekend day and 3 randomly selected weekdays). Following this stabilization period bloods were drawn for a lipid profile. (3) The Seasons Study was designed to examine the causes of seasonal variation in serum cholesterol levels. In this purely observational study healthy subjects were recruited from the Fallon Clinic, an HMO in Worcester, MA. They are followed for one year and are called for a 24HR on three (1 weekend and 2 weekdays) days just after being enrolled and in the final three weeks of each quarter. At the beginning of the study and at the end of each quarter, a 7DDR is completed. Data for the 7DDR validation study are from the end of the first study quarter (n = 182) for subjects who had been recruited from December 1994 to September 1995. The 7DDR was designed to measure fat intake in dietary
928
J. R. Hebert
TABLE 1. Baseline characteristics of the 7DDR Validation Study population and overall by the study of origination, Worcester, MA, 1991-1995
Overall Characteristic Gender Male
FemaIe Education Some high school High school graduate Vocational/trade Some college/Assoc. Bachelors degree
deg.
Graduate school Marital status Married
Other Race White Non white
Age
(yr)
WEVS” %
n
%
n
%
%
114 147
44 56
12 24
::
16 23
41 59
86 100
46 54
1 9
3 23
5 43
6 68 9 75 44 54
2:
0 16
4:
2; 17 21
1: 3 6
2: 9 17
1: 7 9
3: 18 23
5; 34 39
3 24 5 28 19 21
185 73
72 28
26 10
72 28
30 9
77 23
129 54
70 30
239 17
93 7
31 4
89 11
39 0
100 0
169 13
93 7
Mean
SD
Mean
SD
Mean
SD
Mean
SD
50.5
12.4
50.2
14.9
54.8
6.2
49.7
12.7
Nutrient
Descriptive
on
both
serum
diet intervention lipids
and
dietary
trials
in text.
The WATCH study itself (n = 1278) represents an opportunity to examine how well, on average, the 24HR- and 7DDR-derived estimates of group-level changes in nutrient consumption predict changes in total serum cholesterol level. Similarly, we can analyze data from the Ross Study diet intervention period where subjects were placed on a low-fat diet for 6 weeks. Both WATCH and rhe Ross Study are explicit diet intervention trials of the type for which the 7DDR was designed.
information
low-fat
n
the proportional intake of individual foods of which the composite food grouping is comprised (e.g., different kinds of cookies that comprise the general grouping called “cookies”). This is necessary to assure that foods comprising a composite food are represented in the approximate proportions eaten in the population. A 7DDR nutrient database was constructed based on the NDS database (Version 2.3) and applied to the foods items listed on the 7DDR. A detailed description of the process of developing the 7DDR and its nutrient database may be requested from the first author (J.R.H.).
provide
trials. Two
(n = 261) Seasons’
n
“Refers to the WATCH External Validation Study, as described bRefers to the Ross Study, as described in text. ‘Refers to the Seasons Study, as described in text
fat intervention
Ross’
et al.
fat.
Computation
Each 24HR was completed using the NDS developed by the Nutrition Coordinating Center (NCC), University of Minnesota, Minneapolis, Minnesota (Food Database version 2.3 for WATCH and Ross and version 2.7 for Seasons) [33]. Changes in nutrient composition of foods are minimal as versions change, and virtually nonexistent for macronutrients, which are the focus of this work. Brand name or composite foods are broken down into single ingredients that are identified by a five-digit NDS food code. The 7DDR contains typical portion sizes and frequency of consumption data for 118 (57 individual and 61 composite) food items and 13 beverage items. Nutrient data for composite foods are computed from a series of 24HR in the WEVS population. These 24HR-derived data form the basis for assigning
Data Analyses
Summary statistics for categorical background variables shown in Table 1 were computed using Proc FREQ in SAS [34]. Descriptive statistics were computed on age of the population (Table 1) and for the 14 nutritional parameters shown in Table 2 and Table 3, using Proc UNIVARIATE in SAS [35]. Although we computed statistics and made comparisons for over 30 nutrients, a more limited list reflecting the assessment goals of the parent study and lipid interventions in general are presented here. Results based on the larger list of nutrients may be requested from the first author (J.R.H.). All nutrient data presented here were normally distributed. Some of the micronutrients, not shown here, are not normally distributed, as often occurs with such data [11,21].
A Seven-Day
Diet
929
Assessment
TABLE 2. Mean estimates of daily nutrient intakes and difference Validation Study, Worcester, MA, 1991-1995 (n = 261)
24HR’ Nutrient
Mean
Energy (kcal/d) Total fat (g/d) Total fat (% kcal) Total SFA (g/d) Total MFA (g/d)d Total PFA (g/d) Palmitic acid (g/d) Oleic acid (g/d) Linoleic acid (g/d) Linolenic acid (g/d) Cholesterol (mg/d) Dietary fiber (g/d) Alcohol (g/d) Alcohol (% kcal)
1882
69.0 32.2 24.3 26.0 13.3 12.7 24.1 11.7 1.2 238.2 15.3 6.4 2.4
scores from the 7DDR
7DDR’ SD
Mean
(SD) the
SD
1908 81.4 37.6 26.6 31.0 18.1 14.3 28.3 16.2 1.5 225.1 12.5 7.7 2.6
649 33.0 6.8 13.0 12.8 6.9 6.5 12.0 6.3 0.6 139.1 5.9 12.8 5.0
“Values shown are the means and standard deviations the two assessment instruments. bThe difference score is derived from subtracting derived nutrient score. ‘Total of all saturated fatty acids (sfa). dTotal of all monounsaturated fatty acids (mfa). ‘Total of all polyunsaturated fatty acids (pfa).
Differenceb
725 41.1 8.7 13.3 16.2 11.2 7.1
24HR-derived
663 37.1 7.5 12.3 15.0 10.7 6.5 14.3 9.8 0.9 123.6 5.6 7.6 3.5
::ti 1.6 4.3 4.5 0.4 -13.1 -2.8 1.3 0.3
10.3 0.9 108.0 5.5 12.3 4.0 (and
SD
26 12.4 5.4 2.3
15.0
for the nutrient
multiple
Mean
units
shown)
nutrient
score
as derived from
from
the 7DDR-
TABLE 3. Regression and correlation analyses comparing scores from multiple 24-hour recall interviews with those derived from the 7DDR, the 7DDR Validation Study, Worcester, MA, 1991-1995”
Nutrient Energy Total Total Total Total Total Palmitic Oleic
(kcal/d) fat (g/d) fat (% kcal) SFA (g/d) MFA (g/d) PFA (g/d) acid (g/d) acid (g/d) Linoleic acid (g/d) Linolenic acid (g/d) Cholesterol (mg/d) Dietary fiber (g/d) Alcohol (g/d) Alcohol (% kcal)
Unadjusted bb SE’ 0.72 0.80 0.87 0.70 0.82 0.71 0.76 0.76 0.71 0.67 0.51 0.57 0.85 0.77
0.05 0.06 0.07 0.05 0.06 0.07 0.05 0.06 0.07 0.06 0.04 0.04 0.04 0.04
Adjusted bd
SE’
re
0.69 0.80 0.86 0.68 0.81 0.71 0.74 0.75 0.72 0.65 0.48 0.56 0.76 0.70
0.06 0.06 0.07 0.05 0.06 0.07 0.05 0.07 0.08 0.07 0.05 0.04 0.04 0.04
0.67 0.67 0.62 0.68 0.65 0.53 0.69 0.62 0.53 0.54 0.63 0.65 0.75 0.77
(95% CI) (0.60-0.73) (0.60-0.73) (0.54-0.69) (0.61-0.74) (0.57-0.71) (0.44-0.61) (0.62-0.75) (0.54-0.69) (0.44-0.61) (0.45-0.62) (0.55-0.70) (0.57-0.71) (0.69-0.80) (0.71-0.81)
“Comparisons are between the nutrients scores derived from the multiple 24 hour diet recalls (24HR) collected over approximately a three-week measurement period and those derived from the 7DDR administered at the end of the assessment period. bb, the regression coefficient, is obtained by fitting the 7DDR-derived nutrient score as the dependent variable and the 24HR-derived score as the independent variable using Proc GLM in SAS. All results are based on the orthogonal model (Type III sums of squares) not adjusted for any covariates, weighted analyses by the number of 24HR contributing to the nutrient score and the reciprocal of the sample size for the substudy from which the data were obtained. ‘SEI, is the standard error of the respective regression coefficient. “b, the regression coefficient, is obtained by fitting the 7DDR-d erived nutrient score as the dependent variable and the 24HR-derived score as the independent variable using Proc GLM in SAS. All results are based on the orthogonal model (Type 111 sums of squares), which controlled for age and sex, weighted analyses by the number of 24HR contributing to the nutrient score and the reciprocal of the sample size for the substudy from which the data were obtained, and computed the effect shown as though it entered the model Last. ‘Pearson correlation coefficient is shown along with its 95% confidence interval; (lower bound, upper bound).
930
J. R. Hebert et al.
Concurrent
Reliability
Analyses
Data used for these analyses include the multiple 24HR from the first available study interval, and the 7DDR from the end of that first interval for each of the three studies. This assured that we have data more typical of epidemiologic studies where multiple assessments, and the concomitant “training” that results, are not the norm. Although correlation is the conventional means for comparing dietary data, it is an incomplete description of the relation between two sets of measures [36]. Examination of the differences between nutrient scores derived from the two methods and the use of regression analyses, though encountered less often in the dietary assessment literature, better describe the relation between two sets of measures [36-381. Therefore, we also examined difference scores and used linear regression (the general linear model) to assess agreement in the two sets of measures, as has been suggested by our group [2] and others [37-391. We then fit two separate regression models for each nutrient. The first simply regressed the 7DDRderived nutrient score on the 24HR-derived nutrient score weighted by the number of days of 24HR and the reciprocal of the study size (so as to reduce the influence of the Seasons Study, which is about four times as large as either of the two other studies). The second regression model also controlled for age and sex, which are themselves major determinants of total dietary intake. Both regression models were fit using Proc GLM [35] in SAS. We computed Pearson correlations and 95% confidence limits between the multiple 24HR-derived nutrient scores with those derived from the 7DDR using Proc CORR in SAS [37]. Correlation analyses were unweighted. Upper and lower bounds on the confidence limits of the correlation coefficient employed the Fisher’s Z-transformation [40]. In addition to the analyses of the combined data, all statistical procedures were repeated, stratifying by study. For these analyses no weighting was employed. For the WEVS we analyzed all seven days of data as well as only the first 4 days of 24HR in order to examine the effect of the number of 24HR. To assess the possible effects of training as well as proximity of the 24HR relative to the 7DDR, we repeated the overall regression and correlation analyses in the WEVS where we had two or more 24HR per week. For these analyses we stratified by week of the study. Using
the Nutrient
Data
to Predict
Serum Lipid Data
We were fortunate to have access to data from two explicit low-fat intervention studies (total n with paired data = 678) on changes in serum lipid values and corresponding changes in proportion of calories from total fat; saturated (sfa) , monounsaturated (mfa) , and polyunsaturated (pfa) fatty acids; and cholesterol intake as estimated by both 24HR and the 7DDR. For such group-level changes the use of even a single 24HR is appropriate, provided the sample
size is sufficient [5,17]. For the Ross Study, all 39 participants had paired lipid, 7DDR, and 24HR data representing the beginning and the end of the Step I diet period. For the overall WATCH, we had access to a subset of 513 subjects who had both paired serum and 7DDR-data and a subset of 639 individuals who had pairs of lipid and 24HR data. For both studies we were able to fit both 24HR-derived and 7DDR-derived data on changes in nutrient intake to the Keys and Hegsted equations [ 17,19,41,42] to assess the level of agreement with actual change in total serum cholesterol concentrations. We used the current form of the Keys equation: dy = 1.35 (2dS - dP) + 1.5 AZ; where dy = change in serum cholesterol (mg/dl), d.S and dP = change in dietary intake of saturated and polyunsaturated fatty acids expressed as percentages of calories, and dZ = (x7.’ - x:.5) where x1 and xb are the dietary cholesterol intakes of the one-year and baseline diets in mg/lOOO kcal; and the Hegsted equation: Ay = 2.16 AS - 1.65 AP + 0.176 AC; where AC is change in cholesterol intake in mg/lOOO kcal [42]. For these analyses, we obtained the least squares mean controlling for age and gender using the general linear model. For the WATCH, the model took into account the nested nature of that study. RESULTS Baseline characteristics of the overall study population and its three component substudies are given in Table 1. Of the 261 study participants available for analysis, 147 were women. Because the 7DDR was aimed at fat intake and omitted some non-fat sources of energy, it underestimated total energy intake. The 24HR may underestimate total energy intake because of omissions of some foods not containing fat [25,43]. In both the 24HR and 7DDR data there were large differences in nutrient consumption by gender (data not shown). For example, from the 24HR total energy and fat were 2174 kcal/d and 81.6 g/d, respectively, for men and 1656 kcal/d and 59.2 g/d, respectively, for women (0 < 0.0001 for both comparisons). From the 7DDR, the values were 2146 kcal/d and 92.8 g/d, respectively, for men and 1723 kcal/d and 72.5 g/d, respectively, for women (p < 0.0001 for both comparisons). With one exception, the gender differences in consumption of specific nutrients could be explained by the higher total energy consumption in men. For example, percent of energy as fat was 38.1% and 37.2% for men and women, respectively, based on the 7DDR (p = 0.44). Alcohol was the only factor for which we saw a large and significant gender difference that persisted after control for total energy (e.g., 13.8 g/d and 3.0 g/d for men and women, respectively, based on the 7DDR (p < 0.0001) and 4.5% of energy and 1.2% of energy, respectively (p < 0.0001). Gender differences tended to be very consistent across methods. In only one instance was the difference across methods by gender significant (i.e., al-
A Seven-Day
Diet
931
Assessment
cohol consumption in men was overestimated by 2.7 g/d on the 7DDR versus the 24HR in comparison to an overestimation in women of only 0.2 g/d (p = 0.01). These results indicate that for energy-contributing constituents besides alcohol, total consumption, and not proportional intake, differed between the sexes and that the 24HR and 7DDR captured the gender difference about equally well. Table 2 shows the mean daily intakes for 14 nutritional parameters used in analyses (i.e., 11 nutrients by weight, total energy intake, and total fat and alcohol expressed as a percentage of total energy consumption). As we have shown previously in comparing multiple days of 24HR to food frequency questionnaire-derived data for energycontributing components [28,30], the 24HRs had the smallest overall variance. For dietary components that do not contribute materially to total energy intakes, such as cholesterol and fiber (for which there typically is large day-to-day variation in intake [lo,1 1,44]), the 7DDR produced daily intake estimates with a smaller variance than those derived from the multiple 24HR. A pattern evident in examining difference scores is that the 7DDR tended to give larger estimates for total fat and fatty acids than did the 24HR. Results of the regression analyses are shown in Table 3. First we show the results of the models without controlling for any covariates. Each of these models fits the ‘IDDRderived nutrient score as the dependent variable and the corresponding 24HR-derived score as an independent variable. We then fit models that include age and sex as covariates. The regression coefficients shown are based on the orthogonal m&e1 (Type III sum of squares) in SAS [35], which computes each effect as though it were entered last. Regression coefficients were consistently below 1.00. Because of the large sample size, they were statistically different from p = 1 .OO at the nominal type I error rate (a) of 0.05. Because of their familiarity to readers of the dietary assessment literature [2,38] and their direct applicability to researchers planning to use the same instrument in a similar population, we also present the Pearson correlation coefficients in Table 3. We obtained relatively high correlation coefficients for those study-related lipid parameters on which we focused development of the 7DDR (i.e., those lipid parameters that were targeted by the WATCH intervention and those of diet interventions in general). The correlation coefficient for fiber also was fairly high. When we conducted stratified analyses by study of origin, we obtained the results shown in Fig. 1, for total energy intake, and in Fig. 2, for total fat intake. As reported earlier and as shown in Figs 1 and 2 and Table 4 [28], we obtained regression coefficients much closer to 1 .OO in analyses conducted on 7 days of data from the WEVS. The Seasons Study, which had only three days of 24HR, had the poorest agreement. In the WEVS, correlation coefficients shown in Table 4 were closer to 1 .O by an average of about 0.05 units
compared with those computed in the combined data (Table 3). Table 4 also contains results of regression and correlation analyses restricted to four days of 24HR for the WEVS population. We found no clear temporal pattern when we compared data from each of the three weeks of a subject’s involvement in the WEVS. Of the 14 parameters, 6 were highest in the first week and 5 were highest in the week just prior to the 7DDR. Three were highest in the middle week. Of the 10 fat-related nutrient scores, 5 were highest in the first week, and 5 were highest in the last week. In the WATCH diet intervention in which we observed a significant decrease of 9.9 mg/dl in total cholesterol, the 7DDR-derived nutrient scores predicted a decrease of 8.4 mg/dl (about 15% too low) according to the Keys equation [11,45] and 9.1 mg/dl (about 8% too low) according to the Hegsted equation [11,19]. In the 24HR-pair subset, the observed reduction was 8.3 mg/dl. The 24HR-derived nutrient scores fit to the Keys equation predicted a decrease of 15.2 mg/dl (about 84% too high). The Hegsted equation predicted a decrease of 17.0 mg/dl (about 105% too high). For the Ross Study, we observed a decrease of 2 1.1 mg/ dl(7.9% of the baseline value of 267.0 mg/dl) in total cholesterol concentration. 7DDR data fit to the Keys equation predicted a decrease of 28.0 mg/dl (about 33% too high) while 24HR-data predicted a decrease of 32.3 mg/dl (about 53% too high). Data fit to the Hegsted equation resulted in a better fit: for the 7DDR, the predicted decrease was 17.9 mg/dl (about 15% too low) and the 24HR decrease was 28.6 mg/dl (about 35% too high).
DISCUSSION By using data from three study populations, with only one study designed as an explicit validation (WEVS), results probably are more broadly applicable to situations in which the 7DDR normally would be used. The Ross Study was similar in many respects to the WATCH Study for which the instrument was designed. Both were intervention trials aimed at lowering dietary fat and serum lipids in a group of hyperlipidemic patients. One motivation for presenting correlation coefficients is to provide a means for comparison with other studies. As we predicted earlier [2], in our separate analyses of the three datasets (see Fig. 1 and 2) we found that populations with different distributions of the parameters of interest produced different values for the correlation coefficient. Unlike correlation analysis, the use of multiple linear regression allows us the opportunity to assess the extent to which, using our standard measure (i.e., the 24HR-derived nutrient scores), we could predict the 7DDR-derived scores. As Freedman et al. found [46] in their data, our regression coefficients in the combined data were consistently less than 1.00. When we performed similar analyses in the separate datasets we found
4000
a
i
0
1000
2000
g 3000 GM Y d
g
g t! 63 p:
5000
6000
0
FIGURE
1000
1.
Scatterplot
I
-
I
I
-
Energy Intake
I
-*
I
l -/
from 24HR
3000
n
(kcalld) (n = 261).
4000
of 7DDR and 24 hr estimates of energy intake
2000
n
I
5000
WEVS
6000
8 g It: Y 3 5 3 k
Lz
5
0
50
100
150
200
250
300
0
Season
Ross
n
-
wvs
l
FIGURE 2. Scatterplot
.
Fat Intake
I
-
from 24HR
150 (g/d)
200
of 7DDR and 24 hr estimates of total fat intake
100
I
n-
n
(II = 261).
250
R*=O.15
y = 0.52x + 46.1
300
---I
934
J.
R. Hebert et al.
TABLE 4. Results of regression and correhtion analysis, effect of number of available days of 24HR from the WATCH bernal Validation Study, Worcester, MA, 1991 (n = 36)
7 days
Energy (kcal/d) Total fat (g/d) Total fat (% kcal) Total sfa (g/d) Total mfa (g/d) Total pfa (g/d)
4 days
b’
(S&J6
rc
b”
Wdb
0.95 1.04
(0.19) (0.20)
0.69 0.71 0.75 0.71 0.71 0.58
0.80 0.95 0.94 0.91 0.97 0.71
(0.18) (0.18) (0.20) (0.17) (0.18) (0.18)
1.01
(0.17)
0.96 1.08 0.81
(0.17) (0.20) (0.23)
“The regression coefficient obtained by regressing the 7DDR-derived ent score controlling for age and gender. ‘This is the standard error of the regression coefficient. ‘This is the Pearson correlation coefficient based on comparing
that the regression coefficient, b, was much closer to the ideal value of /? = 1.00 in the WEVS. This is not surprising because the WEVS, with seven days of 24HR, was designed explicitly as a validation study. Still, we thought it would be useful to combine data from the two other studies, partly to obtain a more robust estimate of concordance and partly to provide data from study populations in some ways more typical of those available to epidemiologic studies. The fact that correlation and regression coefficients were closer to 1.00 in the WEVS than in the Ross or Seasons studies may be due to the unique volunteer aspect of that study group or may be due to the availability of seven days of 24HR. Data shown in Table 4 indicate that it probably is due to a combination of these factors. Measures of concordance were lower for 4 days than for 7 days of 24HR. However, restricting the WEVS analyses to only 4 days of 24HR produced regressions lines that: were closer to p = 1.00 than for the 4 days of 24HR in the Ross. In the WEVS there was an overlap in the reference period of the two assessments: the ‘I-day period covered by the 7DDRs represented approximately one-third of the actual days during the three weeks the 24HR interviews were conducted. The WEVS was the only one of the three studies that had a sufficiently large number of days of 24HR to examine the “proximity of days” effect. Because the WEVS had no explicit intervention, it is unlikely that a real change in diet occurred during the course of the three weeks of participant involvement as a result of an explicit attempt at dietary change. When we compared data separately from each of the 3 weeks of a participant’s involvement we found that there was no clear relationship between the temporal proximity of the 24HR to the 7DDR and the size of the correlation coefficient. This indicates that the relatively high correlations were not simply an artifact of timing, and that if training is an explanation, it must have occurred quickly (i.e., in the first week). The use of the Keys and Hegsted equations to assess validity of the nutrient scores derived from the dietary assessment methods is a novel feature of this study. Rarely do
nutrient
the 24HR-
rc
0.66 0.71 0.68 0.69 0.71 0.58
score on the 24HR-derived
and 7DDR-derived
nutri-
scores.
studies evaluating dietary assessment techniques have access to biochemical data to assist in judging validity. Because of considerable variation in individuals’ responses of serum cholesterol to dietary lipids it is unreasonable to expect changes in individual nutrient intake to predict individual changes in serum cholesterol [5]. As the original work of Hegsted [ 191 and Keys [17,45] showed, when careful measurement of diet can be made, those data will accurately predict blood lipid changes on a group level. We found that the 7DDR-derived data predicted actual aggregate change in serum cholesterol levels better than did the 24HR. Our results compare favorably with those in a recent report by Schaefer et al. [47] in which they combined data from five feeding trials (where dietary intake was measured more accurately than by self-report) of the NCEP Step 2 diet and examined the effect on LDL cholesterol. They reported that for three of the five studies (total n = 120), actual change in LDL cholesterol was within 20% of that predicted using the Hegsted formula. They further stated that their data indicated that when baseline LDL cholesterol levels are elevated the Hegsted formula underestimates the change in LDL cholesterol. In the current study, the WATCH and Ross study populations were comprised of subjects with elevated cholesterol levels. Fitting the 7DDR data to the Hegsted equation yielded estimated changes in serum cholesterol 8% low for the WATCH data and 15% low for the Ross data. The equivalent. predicted change using the 24HR data was 105% too high for WATCH and 35% too high for Ross. The differences observed between the WATCH and Ross studies are intriguing. We can speculate as to the cause of these differences: (1) Being derived from a much smaller sample, the Ross results are less stable. An analysis of the type carried out here depends on the stability of the group mean values of dietary fat and serum lipids before and after the intervention. Given the wide variability of serum cholesterol response to dietary change [5], a large sample (n > 500 in WATCH) would be much less subject to random variation in cholesterol response than would a much smaller
A Seven-Day Diet Assessment
sample (n = 39 in Ross); (2) The much longer intervention period in the WATCH (l-year versus 6-week for the Ross) lead to greater diet stabilization; (3) The intervention in the Ross was much more intensive than in the WATCH. This may have created a larger demand to do well in adhering to the recommended diets and led to a bias in reporting on the 7DDR relative to the 24HR. Because successful dietary change is expected to be associated with other healthful behaviors such as physical activity [48], it might be reasonable to expect that dietary change alone would underestimate the change in serum cholesterol. We recognize that the use of a recall period ranging up to 7 days and averaging 84 hours in length mixes specific recall with that of habitual recall. Although the 7DDR was designed primarily to be a recall-based approach, it does include elements of the frequency method to the extent that subjects who cannot elicit a specific memory of a food encounter are encouraged to enter their “usual” eating behavior. In this respect, we are secondarily reliant on habitual memory but not on conventional frequency reporting, because we make no attempt to have the subject conduct a long-term averaging. We believe that our memory prompting grid helps respondents to recall specific food encounters. The available evidence indicates that the major drop-off in actual match rates of foods for a two-week reference period occurs after two days but the intrusion rate (spurious recall of foods that were not eaten) achieves its maximum by the end of the second week [25]. Clearly, additional work is needed in the area of cognitive processing of dietary information in order to understand this more fully. It is conventionally accepted that dietary assessments can be dichotomized into daily methods (mainly the 24HR) or frequency methods [25,49-511. As pointed out by Smith [25], the method (e.g., recall) and the reference period are orthogonal attributes of dietary assessment. Avoiding the confounding that normally exists between the method and its reference period allows consideration of cognitive issues that condition reporting of dietary intake as well as unconventional combinations of method and reference period that better address the epidemiologic issues at hand. The 7DDR may do this by focusing on a period where episodic memory can operate and by using methods to increase the likelihood of this happening. By limiting the time interval for which data are to be recalled to one week, we provide an opportunity for subjects to respond in terms of actual counts of items consumed rather than relying solely on a computation of rates of consumption, another potential source of error [25]. Ideally, dietary assessment methods aimed at establishing long-term habitual patterns of intake will eliminate intraperson sources of variability. Thus, total fluctuation for nutrients, even for fat which has relatively low intraperson variability, should be lower on instruments such as the FFQ than for even a moderate number of days of FD or 24HR. The fact that this is generally not the case indicates that
935
there are other sources of error in the use of such historical methods [2]. In an unrelated study we previously found that the multiple 24HR approach produces the smallest overall variance [29,30] compared to multiple-day FD or an FFQ, indicating the multiple 24HR have the lowest total error. Here, we found that the 7DDR-derived scores had variances only slightly larger than the multiple 24HR-derived measures for total energy and fat components and they were somewhat smaller for cholesterol, fiber, and alcohol. This is a rough indicator that the 7DDR is capturing relevant aspects of the differences between individuals without excessive noise, or intraperson variability. Although the 7DDR and the 24HR are both recall methods, the method of administration and the time frame of reference are different. Because so little is known about the true nature of error in self-assessed diet, we are not convinced that errors obtained in the 7DDR and 24HR are unduly correlated. Though autocorrelation between recall methods seems logical on its face, a specific mechanism has not been postulated [42]. In recent work by our group on a completely unrelated dataset we found that nutrient scores from recalls and records were interchangeable, though the recalls had lower total error (unpublished data). Clearly, specific errors of memory could account for correlation. However, our analysis of days of 24HR (by week over a 3week period before the 7DDR) in the WEVS indicates that temporal proximity of the recall, and therefore defects in memory that vary with time, do not explain the high level of agreement we observed. We believe that the 24HR is sufficiently superior to the FD to warrant its choice for use in reliability and calibration studies. Clearly, future work on this instrument should entail comparisons with FD data and other bioIogica1 markers besides cholesterol. Our primary motivation for developing the 7DDR and the major impetus for the FFQ is the computation of nutrient scores for individuals. Nutrient scores typically are used as a criterion for assessing the utility of a dietary assessment method, which assess foods eaten and not nutrients per se. Although estimating nutrient intakes is necessary for epidemiologic studies, including intervention trials, nutrient scores provide little insight into the specific issues involved in the psychology of dietary reporting [25]. This should be a focus of future work in this area. Although the 7DDR performed very well in comparison to other methods in the three study populations described here it is subject to the same concerns about generalizability to demographically dissimilar study populations that exist for all dietary assessment methods [52]. For any close-ended dietary assessment, one factor affecting generalizability is that the foods listed may be inappropriate to the population being studied. The list of foods on the 7DDR correspond to those contributing to the nutrients of interest in the general U.S. population [26,27] supplemented by foods identified as important in the local population. It lists about 50%
936
more items than commonly used FFQs, so the food list and concomitant nutrient derivations should not limit generalizability beyond what would be expected for an FFQ in a similar population. Another factor affecting generalizability is the level of literacy, cognitive capability, and commitment of time and concentration necessary to complete the form. Although we attempted to make the 7DDR readily understandable to a wide range of individuals, it should be noted that only 2% of the entire validation study population had less than a high school education and the vast majority (93%) were white. The fact that the 7DDR asks the subject to provide counts of food eaten rather than averages should make it a cognitively simpler task to complete than an FFQ. Even though our population may be typical of many intervention trials, it is important to note that if the 7DDR is to used in a population of different educational status and cultural background, it will require retesting in the appropriate context, perhaps with prior modification of the food list to include different food sources of the target nutrients.
within-subject variation 1991; 54: 464-470.
10.
11.
12.
13.
14.
CONCLUSION
This work was supported by grants HL 44492 and HL 52745 from the National Heart Lung and Blood Institute and funds from the Division of Preventive and Behawk& Medicine.
16.
17.
18.
19.
References 1. Hebert JR, Miller DR. Methodologic considerations for investigating the diet-cancer link. Am J Clin Nutr 1988; 47: 1068-1077. 2. Hebert JR, Miller DR. The inappropriateness of conventional use of the correlation coefficient in assessing validity and reliability of dietary assessment methods. Eur J Epidemiol 1991; 7: 339-343. 3. El Lazy M. Dietary variability and its impact on nutritional epidemiology. J Chronic Dis 1983; 36: 237-249. 4. Nelson M, Black AE, Morris JA, Cole TJ. Betweenand within-subject variation in nutrient intake from infancy to old age: Estimating the number of days required to rank intakes with desired precision. Am J Clin Nutr 1989; 50: 155-167. 5. Jacobs DJ, Anderson J, Blackbum H. Diet and serum cholesterol: Do zero correlations negate the relationship? Am J Epidemiol 1979; 110(l): 77-87. 6. Hebert JR, Backlund JYC, Engle A, Barone J, Biener K. Intraand inter-person sources of variability in fat intake in a feeding trial of 14 men. Eur J Epidemiol 1990; 6: 55-60. 7. Block G, Hartman AM. Issues in reproducibility and validity of dietary studies. Am J Clin Nutr 1989; 50: 1133-1138. 8. Tarasuk V, Beaton GH. The nature and individuality of
intake.
et al.
Am
Nutr
J Clin
9. de BackerG. Methodologicalconsiderations on intra- versus
15.
In summary, the 7DDR appears to have utility for the measurement of fat in clinical trials where the intervention periods are relatively short. The 7DDR performs well in comparison to other methods using correlation as the criterion of comparison. We believe that this is due, in large part, to a special focus on cognitive issues in the development of this dietary assessment method.
in energy
J. R. Hebert
20.
21.
22.
23. 24.
25.
26.
27.
interpopulation correlation studies. Ann Nutr Metabol 1991; 35: 84-88. Beaton GH, Milner J, Corey P, et al. Source of variance in 24.hour dietary recall data: Implications for nutrition study design and interpretation. Am J Clin Nutr 1979; 32: 25462559. Willett W, Sampson L, Stampfer M, et al. Reproducibility and validity of a semiquantitative food frequency questionnaire. Am J Epidemiol 1985; 122: 51-65. Hebert JR, Harris DR, Sorensen G, Stoddard AM, Hunt MK, Morris DH. A worksite nutrition intervention: Its effects on the consumption of cancer-related nutrients. Am J Public Health 1993; 83: 391-394. Hebert JR, Hurley TG, Hsieh J, et al. Determinants of plasma vitamins and lipids: The Working Well Study. Am J Epidemiol 1994; 140: 132-147. Kristal A, Shattuck A, Williams A. Food frequency questionnaires for diet intervention research. In: Proceedings of the 17th National Nutrient Databank Conference-Baltimore, MD, June 7-9, 1992. Washington, DC: International Life Sciences Institute; 1994; 110-125. Buzzard IM. Estimating compliance bias and other sources of error in assessing dietary intake in a low fat diet intervention study. In: Proceedings of the Second International Conference on Dietary Assessment Methods. Boston, MA: 1995; 13. Ockene IS, Hebert JR, Ockene JK, et al. Effect of physician delivered nutrition counseling and a structured office practice on diet and serum lipid measurements in a hyperlipidemic population: The Worcester Area Trial For Counseling in Hyperlipidemia (WATCH). Am J Prev Med 1996; 12: 252258. Keys A, Anderson JT, Grande F. Prediction of serum-cholesterol response of man to changes in fats in the diet. Lancet 1957; 7003: 959-966. Keys A, Anderson JT, Grande F. Serum cholesterol response to changes in the diet-III. Differences among individuals. Metabolism 1965; 14: 766-775. Hegsted DM, McGandy RB, Myers ML, Stare FJ. Quantitative effects of dietary fat on serum cholesterol in man. Am J Clin Nutr 1965; 17: 281-295. National Academy of Sciences. National Academy of Sciences Report on Diet and Health. Nutr Rev 1989; 47: 142149. Hebert JR, Stoddard AM, Harris DR, et al. Measuring the effect of a worksite-based nutrition intervention on food consumption. Ann Epidemiol 1993; 3: 629-635. Dwyer JT, Krall EA, Coleman KA. The problem of memory in nutritional epidemiology research. J Am Diet Assoc 1987; 87: 1509-1512. Jobe JB, Mingay DJ. Cognitive research improves questionnaires. Am J Public Health 1989; 79: 1053-1055. Dwyer JT, Gardner J, Halvorsen K, Krall EA, Cohen A, Valadian I. Memory of food intake in the distant past. Am J Epidemiol 1989; 130: 1033-1046. Smith AF. Cognitive psychological issues of relevance to the validity of dietary reports. Eur J Clin Nutr 1993; 47(SuppI. 2): S6-S18. Block G, Dresser CM, Hartman AM, Carroll MD. Nutrient sources in the American diet: Quantitative data from the NHANES II survey: II. Macronutrients and fats. Am J Epidemiol 1985; 122: 27-39. Block G, Rosenberger WF, Patterson BH. Calories, fat and
A Seven-Day
28.
29.
30.
31.
32.
33.
34. 35. 36.
37.
38.
39.
40.
Diet
Assessment
cholesterol: Intake patterns in the US population by race, sex and age. Am J Public Health 1988; 78: 1150-l 155. Hebert JR, Ockene IS, Botelho L, Luippold R, Merriam P, Saperia G. Development and validation of a seven-day diet recall. In: Proceedings of the American Public Health Association 120th Annual Meeting. Washington, DC; 1992. Hebert JR, Miller DR, Barone J, Engle A. Assessment of change in fat intake in an intervention study. In: Proceedings of the American Public Health Association-l 19th Annual Meeting. Atlanta, GA: APHA; 1991. Miller DR, Hebert JR, Barone J, Engle A. Comparison of dietary assessment methods: Advantages of random 24 hour recalls. In: Proceedings of the American Public Health Association119th Annual Meeting. Atlanta, GA: APHA; 1991. Posner BM, Martin-Munley SS, Smigelski C, et al. Comparison of techniques for estimating nutrient intake: The Framingham Study. Epidemiol 1992; 3: 171-177. Forster JL, Jeffery RW, VanNatta M, Pirie P. Hypertension prevention trial: Do 24-h food records capture usual eating behavior in a dietary change study? Am J Clin Nutr 1990; 51(2): 253-257. Buzzard IM, Price KS, Warren RA. Considerations for selecting nutrient-calculation software: Evaluation of the nutrient database. Am J CIin Nutr 1991; 54: 7-9. SAS. SAS User’s Guide. Cary, NC: SAS Institute; 1996. SAS. SAS User’s Guide-Statistic Vers. 6.5. Cary, NC: SAS Institute; 1995. Miettinen OS. Theoretical Epidemiology: Principles of Occurrence Research in Medicine. New York: John Wiley &. Sons, Inc.; 1985. Borrelli R, Cole TJ, DiBiase G, Contaldo F. Some statistical considerations on dietary assessment methods. Eur J Clin Nutr 1989; 43: 453-463. Bellach B. Remarks on the use of Pearson’s correlation coefficient and other association measures in assessing validity and reliability of dietary assessment methods. Eur J Clin Nutr 1993; 47(Suppl. 2): S42-S45. Borrelli R. Collection of food intake data: A reappraisal of criteria for judging the methods. Br J Nutr 1990; 63: 411417. Snedecor GW, Cochran WG. Statistical Methods. 8th ed. Ames, IA: Iowa State University Press; 1989.
937
41.
42. 43.
44.
45.
46.
47.
48.
49.
50.
5 1.
52.
Keys A, Anderson JT, Grande F. Serum cholesterol response to changes in the diet-I. Iodine value of dietary fat versus 2SP. Metabolism 1965; 14: 7747-7758. Willett W. Nutritional Epidemiology. Oxford: Oxford University Press; 1990. Smith AF. Cognitive Processes in Long-Term Dietary Recall. Washington, DC: U.S. Government Printing Office, 1991. [Series 6, 4. DHHS publication (PHS) 91 1079.1 Basiotis PP, Welsh SO, Cronin FJ, Kelsay JL, Mertz W. Number of days of food intake records required to estimate individual and group nutrient intakes with defined confidence. J Nutr 1987; 117: 1638-1641. Keys A, Anderson JT, Grande F. Serum cholesterol response to changes in the diet--II. The effect of cholesterol in the diet. Metabolism 1965; 14: 759-775. Freedman LS, Carroll RJ, Wax Y. Estimating the relation between dietary intake obtained from a food frequency questionnaire and true average intake. Am J Epidemiol 1991; 134: 310-320. Schaefer EJ, Lamon-Fava S, Ausman LM, et al. Individual variability in lipoprotein cholesterol response to national cholesterol education program step2 diets. Am J Clin Nutr 1997; 65(3): 823-830. Matthews CE, Hebert JR, Ockene IS, Saperia G, Merriam PA. The inter-relation between leisure time physical activity and selected dietary variables. Medicine and Science in Sports and Exercise 1996. (In press) Dwyer JT. Assessment of dietary intake. In: Shils M, Young V, Eds. Modem Nutrition in Health and Disease. Philadelphia: Lea & Febiger; 1988: 887-905. Sempos CT. Invited commentary: Some limitations of semiquantitative food frequency questionnaires. Am J Epidemiol 1992; 135: 1127-1132. Sempos CT, Briefel RR, Flegal KM, Johnson CL, Murphy RS, Wotecki CE. Factors involved in selecting a dietary survey methodology for national nutrition surveys. Aust J Nutr Diet 1992; 49: 96-101, 103-104. Bingham SA, Nelson M. Assessment of food consumption and nutrient intake. In: Margetts BM, Nelson M, Eds. Design Concepts in Nutritional Epidemiology. New York: Oxford University Press; 1991: 153-167.