Public Health Nutrition: 7(6), 745–749 DOI: 10.

1079/PHN2004599

Selection of food items for inclusion in a newly developed

food-frequency questionnaire
Iris Shai1,2,*, Danit R Shahar2, Hillel Vardi3 and Drora Fraser2,3
1
Harvard School of Public Health, Departments of Nutrition and Epidemiology, 665 Huntington Avenue, Boston,
MA 02115, USA: 2The S Daniel Abraham International Center for Health and Nutrition, Ben-Gurion University of the
Negev, Beer-Sheva, Israel: 3Epidemiology and Health Services Evaluation Department, Faculty of Health Sciences,
Ben-Gurion University of the Negev, Beer-Sheva, Israel

Submitted 1 October 2003: Accepted 7 January 2004

Abstract
Objectives: To highlight the differences between the food list required in a food-
frequency questionnaire (FFQ) to rank people by their intake and the food items that
contribute to absolute intake, and to discuss possible applications.
Methods: We conducted a nutritional survey among 1173 adults using an adapted 24-
hour recall questionnaire.
Statistical analysis: To develop an FFQ, we analysed the 24-hour recall survey data by
performing a stepwise multiple regression after grouping conceptually similar food
items into 175 food groups.
Results: In total, 126 food groups were included in the developed FFQ in order to
explain at least 80% of the variance in the consumption of each of 27 nutrients. The
nutrients that were explained by a few food groups were vitamin A (one food group),
alcohol (two), b-carotene (two), vitamin E (three) and cholesterol (five). Nutrients
that were explained by a large number of food groups were energy (37 food groups),
potassium (31), magnesium (31), dietary fibre (30), phosphorus (31) and sodium (29).
Using energy intake as an example, soft drinks were the best between-person energy
classifiers, while providing only 2.4% of the total energy intake. Wine, seeds and nuts,
which contributed highly to the variance, were minor energy contributors. In contrast,
milk, sugar, fried chicken/turkey breast or whole chicken/turkey, which explained
little of the variation in the population, were major energy contributors. Keywords
Conclusions: Developing an FFQ on the basis of common foods may not explain the Food-frequency questionnaire
between-person variation required for ranking individual intake in diet–disease 24-Hour recall
studies. Producing lists of ‘discriminating items’ can be a useful application in Between-person variation
developing mini-FFQs for selected nutrients. Absolute intake

A food-frequency questionnaire (FFQ) is designed to the FFQ to rank people by their intake and the food items
identify foods that discriminate between individual that contributed to the absolute intake, as well as
nutrient intakes1,2. One approach to developing an FFQ demonstrate several practical uses of this knowledge.
is to base it upon a population survey, as done in The
Netherlands3, Germany4, Spain5 and the USA6. The Methods
development of the food list can be based on stepwise
multiple regression analyses1,2,7. As demonstrated pre- Dietary assessment and the food database
viously7 – 9, the number of foods needed to explain a given A full description of the Negev Nutritional Study (NNS) has
proportion of variability in nutrient intake is less than the been given elsewhere10,11. To summarise briefly, we
number needed to explain the same proportion of generated a random sample population aged 35 years and
absolute intake. above in the Negev (the southern region of Israel), using
The combination of foods that contributes most to voters registries. Participants were interviewed at home
between-person variability is unique and characteristic for regarding their dietary intake using a modified version of
each specific population. Therefore, it is necessary to the US Department of Agriculture’s 24-hour recall
determine these key foods in each different country and questionnaire, with additional questions added to address
culture. We conducted a population-based dietary survey general health and eating habits. We then developed a
in order to develop an FFQ in Israel. In this paper we food database and data-entry system for the 24-hour recall
highlight the differences between the food list required in questionnaires12 based on the logic of the US Food

*Corresponding author: Email irish@bgumail.bgu.ac.il q The Authors 2004

746 I Shai et al.
Information Analysis System13. The system was designed vitamin E (three groups) and cholesterol (five groups). On
as bilingual (English and Hebrew). Local foods and recipes the other hand, nutritional components that required a
were collected during the NNS. large number of foods were energy (37 food groups),
potassium (31 groups), magnesium (31 groups), dietary
Development of the FFQ fibre (30 groups), phosphorus (31 groups) and sodium (29
The complete process of developing our FFQ has been groups). The number of items required to explain different
described elsewhere14. In short, the results of the NNS levels of between-person variation for selected nutrients is
were used to develop an FFQ using the following steps. presented in Fig. 1 along with the number of foods required
First, conceptually similar foods were aggregated into for each level of between-person variability. For energy
groups based on their fat and caloric content per portion intake, two items explained 20% of the between-person
eaten. Fruits and vegetables were aggregated based on variability, seven items explained 40%, 17 items explained
their vitamins and minerals content per portion. After- 60% and 37 items explained a total of 80% of the between-
wards, in stepwise multiple regression analyses15, foods person variation. For calcium, 20% of the between-person
important in contributing to between-person variation variability was explained by only one item, 40% by three
were identified. Items were entered into the model based items, 60% by eight items and 80% by 20 items.
on their specific nutrient content and the frequency of Tables 1 and 2 demonstrate the differences between
intake in the population. Foods in the model that contribution of food groups to between-person variation
explained at least 80% of the between-person variability and total intake in the population. The earlier the stage of
were considered for the final questionnaire. The entry into the regression, the higher is the contribution to
procedure described was undertaken for each of 27 variance in consumption (R 2) of the food group to the
nutrients including energy, folic acid, calcium, vitamin E, nutritional component.
dietary fibre, etc. The contribution to the total nutrient We detected 37 food groups that accounted for 80% of
intake was also calculated for each of the foods and the variance in energy consumption, the top 20 of which
recipes included in the final questionnaire. are presented in Table 1. Most variance in energy intake
(20%) in the population resulted from the consumption of
Statistical analysis carbonated beverages and fruit-flavoured drinks (not
Statistical analyses were conducted using SPSS version 10 including diet). Dry wine, seeds and nuts, which entered
(SPSS Inc., Chicago, IL, USA). After aggregation of all in early stages to the regression (4, 5 and 9, respectively),
reported items into 175 groups, the means of all foods were minor energy contributors (rank 72, 25 and 34,
included in the aggregated group were used in the model respectively). However, 3% fat milk and sugar, which
weighted by the frequency of consumption of that item. entered in the later stages to the regression (16 and 17,
Stepwise regression models were used to assess the respectively), were a major energy contributors (rank 2
contribution of each food item to the total intake and the and 3, respectively). Similarly, fried chicken, turkey breast
between-person variability, with individual food items as and whole chicken/ turkey which entered almost at the
the independent variables and total nutrient intake as the end step of the 37 food items (31 and 33, respectively)
dependent variable. were major energy contributors (rank 4 and 5, respect-
ively) (data not shown).
Results A total of 11 items were required to explain 80% of the
between-person variation in folic acid intake (Table 2).
During the study, 1465 24-hour recall questionnaires were
obtained from 1173 participants (650 women and 523
men). This group reflects those who remained after 9.9%
refusals, 2.7% language problems and 3.4% deaths in the
original sample. Body mass index (BMI) was
27.0 ^ 4.9 kg m22 for women and 26.5 ^ 4.1 kg m22 for
men (mean ^ standard deviation (SD)). Mean ^ SD of
reported energy intake was 1498.7 ^ 875.5 kcal for
women and 1989.4 ^ 1138.5 kcal for men. The women
reported 20.0 ^ 9.8 food items, 2.8 ^ 0.9 meals and
6.1 ^ 4.6 snacks; and the men reported 21.0 ^ 10.8 food
items, 2.8 ^ 0.9 meals and 6.2 ^ 5.4 snacks (mean ^ SD).
An average of 19 foods (range 1–37) was needed to
account for the consumption of nutritional components in
the 27 regression models. The nutritional components
explained by a few food groups were vitamin A (one food Fig. 1 Number of food items required to explain different levels of
group), alcohol (two groups), b-carotene (two groups), between-person variation for selected nutrients
Selection of food items for an FFQ 747
Table 1 Contributions of the top 20 specific grouped foods to between-person variation of energy intake (using stepwise multiple
regression analysis*) and to total energy intake

Contribution to
between-person variation Contribution to total intake
Stage of entry
2
into regression Grouped foods Cumulative R Percentage (rank†) Cumulative percentage

1 Carbonated drinks, non-dietetic 0.11 2.37 (7) 2.37

2 Drinks, fruit-flavoured 0.19 2.51 (5) 4.88
3 Bread 0.24 9.33 (1) 14.21
4 Dry wine 0.29 0.34 (72) 14.55
5 Sunflower, watermelon, pumpkin seeds 0.33 1.18 (25) 15.73
6 Plain cookies 0.36 2.11 (12) 17.84
7 Dry cakes 0.38 1.46 (19) 19.30
8 Vegetable oils 0.41 2.45 (7) 21.75
9 Nuts, almonds, pistachios 0.43 0.87 (34) 22.62
10 French fries 0.46 1.39 (20) 24.01
11 Mixed meat dishes 0.48 2.32 (10) 26.33
12 Hard cheeses 0.50 1.15 (26) 27.48
13 Pita bread 0.52 1.97 (13) 29.45
14 Potatoes (cooked, baked, mashed) 0.54 2.33 (9) 31.78
15 Cooked yellow/white rice with additions 0.56 2.27 (11) 34.05
16 3% fat milk 0.57 3.04 (3) 37.09
17 Sugar 0.60 3.27 (2) 40.36
18 Individual puff pastry filled with cheese, potato 0.62 0.88 (32) 41.24
19 Cheese/cream cakes 0.63 0.66 (43) 41.90
20 Peanuts 0.64 0.27 (81) 42.17

* Dependent variable: total energy intake, independent variable: grouped foods.

† Rank: the order of contributors to total energy intake from higher to lower.

The significant prominent group was internal organs of foods, regressed from 1173 survey interviewees’ nutri-
poultry/meat, which contributed almost half of the tional data. In the Israeli population, soft drinks are the
explained variance. Other foods contributed only between best between-person energy classifiers, while providing
1% and 8% of the variance. Sweetened fruit juice and only 2.4% of the total energy intake. About half of the
breakfast cereals were good classifiers (stage 5 and 7, energy variation (46%) was explained by soft drinks, wine,
respectively, in the model) but minor contributors to total regular bread, seeds and mixed nuts, cakes, cookies,
folic acid intake (rank 47 and 58, respectively). vegetable oil, French fries, beef dishes and hard cheese.
However, these 10 food groups accounted for only 24% of
Discussion the total energy intake.
Fewer foods are required to explain the between-
The purpose of this study was to highlight differences person variance in energy consumption than the larger
between the foods needed to explain the between-person number required for evaluating total intake. Stryker et al.8
variability as compared with those contributing to absolute studied the correlation between the absolute contribution
nutrient intake. We designed an FFQ that incorporates 126 of foods to energy and their contribution to the variance of

Table 2 Contributions of the entire specific grouped foods to between-person variation of folic acid (using stepwise multiple regression
analysis*) and to total folic acid intake

Contribution to
between-person variation Contribution to total intake
Stage of entry
2
into regression Grouped foods Cumulative R Percentage (rank†) Cumulative percentage

1 Internal organs of poultry/meat 0.47 5.10 (4) 5.1

2 Juice, fruit-flavoured 0.55 5.29 (2) 10.39
3 Falafel in pita bread 0.61 1.89 (13) 12.28
4 Sunflower, pumpkin, watermelon seeds 0.65 2.71 (7) 14.99
5 Sweetened fruit juice (nectar) 0.69 0.54 (47) 15.53
6 Cooked dried legumes 0.72 2.15 (10) 17.68
7 Breakfast cereals 0.74 0.45 (58) 18.13
8 Vegetable soup (home-made) 0.76 3.50 (5) 21.63
9 Bread 0.78 8.66 (1) 30.29
10 Fresh cucumber 0.79 1.91 (12) 32.20
11 Orange or grapefruit 0.80 1.57 (17) 33.77

* Dependent variable: total folic acid intake, independent variable: grouped foods.
† Rank: the order of contributors to total folic acid intake from higher to lower.
748 I Shai et al.
consumption. In the American population, cake was nutrients. For example, when designing a preventive
pointed out as the best classifier of energy consumption cardiology intervention study, one could obtain a list of
(a contribution of 19.1% to the variance between people). the foods most pertinent (e.g. cholesterol, fats, etc.) and
In order to reach a cumulative R 2 ¼ 0.50, questions then develop and use a mini-checklist of the best-
regarding frequency of intake of cake, butter, hard discriminating items, thereby saving time and expense
cheeses, whole milks and margarine intake were compared with the more complete traditional FFQs. The
sufficient, although these foods contributed only 14% of technique is also applicable when using a different
the total energy. In the analysis during the development of method, e.g. a 24-hour recall for dietary assessment, in that
an FFQ in Denmark7, only three foods – potatoes, butter one could add a list of the foods responsible for the major
and barley bread – were inquired in order to explain half between-person variation following the open question-
of the variance in energy consumption. These same three naire. This list could include, for our own population,
jointly contributed 18.3% of the total energy. Ninety per contributors such as soft drinks for energy, mixed meat
cent of the variance in Denmark was explained by 25 dishes for saturated fat and internal organs for folic acid.
foods only, a less diverse list than our own. Similar Some limitations of the regression analysis merit
findings concerning the fewer number of foods required consideration1. Because the original food list may include
to identify variance were found by Byers et al.9 and Block several hundred potential food variables, some will enter
et al.6. The longer list of 37 foods required for reaching the as ‘statistically significant’ variables on the basis of chance
level of R 2 ¼ 0.80 in our population is probably due to alone. Therefore, the sample should be large, between
ethnic variability, characterised by different dietary 1000 and 2000 subjects rather than a few hundred.
patterns among people from different ethnic origins. As However, even with a large sample, a few foods that are
we have previously shown14, dietary intakes of energy and associated with the relevant items may occasionally make
vitamin E are significantly lower among European – a moderate contribution to the variation, but not make
American born subjects and calcium intake is significantly sense in terms of containing the nutrient being predicted.
lower among Asian –African born subjects. Differences Thus, nutritional experts rather than statisticians should
between ethnic groups were seen in the list of foods, make the final decision for inclusion of foods when
which explained the between-person variability for developing FFQs.
energy. For most nutrients, fewer items were needed to
explain the between-person variation in the group of Acknowledgements
people born in Israel. Thus, an extended list of food items
explaining energy intake variation might reflect the multi- We thank the S Daniel Abraham International Center for
ethnic population. Health and Nutrition for its support of this project.
The contribution of food items to the consumption of
nutritional components differs according to the specific
References
nutritional component. Vitamins, in particular, were
explained by a small number of foods. For example, 80% 1 Willett W, ed. Food frequency methods. In: Nutritional
of the variance in the consumption of vitamin E Epidemiology. New York: Oxford University Press, 1998;
was explained by three groups only: sunflower/pumpkin/ 69 –91.
watermelon seeds, vegetable oils and mixed nuts. On the 2 Block G. A review of validation of dietary assessment
methods. American Journal of Epidemiology 1982; 115:
other hand, the between-person variation of common 492 – 505.
minerals, such as potassium, nitrate, magnesium and 3 Ocke MC, Bueno-De-Mesquita HB. The Dutch EPIC FFQ. l.
phosphorus, was explained by the largest number of foods, Description of the questionnaire, and relative validity and
as found in other populations7. In our survey, only 11 food reproducibility for food groups. American Journal of
Epidemiology 1997; 26: S37 –48.
groups were required to explain the between-person 4 Stefanie BT, Ina H, Heiner B. Reproducibility and relative
variation in folic acid. The food group internal organs of validity of food group intake in a FFQ developed for the
poultry/meat, which explained 47% of the variance, can German part of the EPIC project. American Journal of
categorise folic acid intake almost exclusively. This can be Epidemiology 1997; 26: S59 –70.
5 Martin-Moreno JM, Boyle P, Gorgojo L, Maisonneuve P,
explained by the folic acid content of the foods. While it Fernandez-Rodriguez JC, Salvini S, et al. Development and
ranges from 5 to 60 mg per 100 g of food for most items, liver validation of a FFQ in Spain. International Journal of
is a rich source of folic acid (770 mg per 100 g). If we Epidemiology 1993; 22(3): 512 – 9.
combine this fact with the great variation of liver 6 Block G, Hartman AM, Naughton D. A reduced dietary
questionnaire: development and validation. Epidemiology
consumption frequency in the population, we get the 1990; 1: 58– 64.
primary indicator for the classification of folic acid 7 Overvad K, Tjonneland A, Haraldsdottir J, Ewertz M, Jensen
consumption in this population, by exclusive use of liver OM. Development of a semiquantitative food frequency
questionnaire to assess food, energy and nutrient intake in
consumption patterns.
Denmark. International Journal of Epidemiology 1991;
A potential application for producing lists of ‘discrimi- 20(4): 900 –5.
nating items’ is to develop specific mini-FFQs for selected 8 Stryker WS, Salvini S, Stampfer MJ, Sampson L, Colditz GA,
Selection of food items for an FFQ 749
Willett WC. Contributions of specific foods to absolute intake 12 Shai I, Vardi H, Shahar RD, Azrad BA, Fraser D. Adaptation of
and between-person variation of nutrient consumption. international nutrition databases and data-entry system tools
Journal of the American Dietetic Association 1991; 91: to a specific population. Public Health Nutrition 2003; 6(3):
172 – 8. 401 –6.
9 Byers T, Marshall J, Fiedler R, Zielezny M, Graham S. 13 Food Intake Analysis System, Version 3. University of Texas
Assessing nutrient intake with an abbreviated dietary Health Science Center at Houston School of Public Health.
interview. American Journal of Epidemiology 1985; 122: US Department of Agriculture, Human Nutrition Information
41– 50. Service, 1996.
10 Shahar D, Shai I, Vardi H, Fraser D. Dietary intake and eating 14 Shahar DR, Shai I, Vardi H, Brener-Azrad A, Fraser D.
patterns of elderly people in Israel: who is at nutritional risk? Development of a semi quantitative food frequency
European Journal of Clinical Nutrition 2003; 57(1): 18– 25. questionnaire (FFQ) to assess dietary intake of multiethnic
11 Bilenko N, Shahar D, Shai I, Weitzman S, Fraser D. populations. European Journal of Epidemiology 2003; 18(9):
Prevalence and characteristics of myocardial infarction, 855 –61.
diabetes and hypertension in the adult Jewish population: 15 Willett W, Sampson L. Reproducibility and validity of a
results from the Negev Nutritional Study. Harefuah 2003; semiquantitative food frequency questionnaire. American
142(1): 17 –21. Journal of Epidemiology 1985; 122: 51 –65.