- Email: [email protected]
A Novel Approach to Selecting and Weighting Nutrients for Nutrient Profiling of Foods and Diets
RESEARCH Original Research
A Novel Approach to Selecting and Weighting Nutrients for Nutrient Profiling of Foods and Diets Joanne E. Arsenault, PhD, RD; Victor L. Fulgoni III, PhD; James C. Hersey, PhD; Mary K. Muth, PhD
ARTICLE INFORMATION Article history: Accepted 16 July 2012
Keywords: Nutrient profiling Dietary quality Dietary intake Copyright © 2012 by the Academy of Nutrition and Dietetics. 2212-2672/$36.00 doi: 10.1016/j.jand.2012.08.032
ABSTRACT Background Nutrient profiling of foods is the science of ranking or classifying foods based on their nutrient composition. Most profiling systems use similar weighting factors across nutrients due to lack of scientific evidence to assign levels of importance to nutrients. Objective Our aim was to use a statistical approach to determine the nutrients that best explain variation in Healthy Eating Index (HEI) scores and to obtain -coefficients for the nutrients for use as weighting factors for a nutrient-profiling algorithm. Design We used a cross-sectional analysis of nutrient intakes and HEI scores. Participants Our subjects included 16,587 individuals from the National Health and Nutrition Examination Survey 2005-2008 who were 2 years of age or older and not pregnant. Main outcome measure Our main outcome measure was variation (R2) in HEI scores. Statistical analyses Linear regression analyses were conducted with HEI scores as the dependent variable and all possible combinations of 16 nutrients of interest as independent variables, with covariates age, sex, and ethnicity. The analyses identified the best 1-nutrient variable model (with the highest R2), the best 2-nutrient variable model, and up to the best 16-nutrient variable model. Results The model with 8 nutrients explained 65% of the variance in HEI scores, similar to the models with 9 to 16 nutrients, but substantially higher than previous algorithms reported in the literature. The model contained five nutrients with positive -coefficients (ie, protein, fiber, calcium, unsaturated fat, and vitamin C) and three nutrients with negative coefficients (ie, saturated fat, sodium, and added sugar). -coefficients from the model were used as weighting factors to create an algorithm that generated a weighted nutrient density score representing the overall nutritional quality of a food. Conclusions The weighted nutrient density score can be easily calculated and is useful for describing the overall nutrient quality of both foods and diets. J Acad Nutr Diet. 2012;112:1968-1975.
N
UTRIENT PROFILING OF FOODS IS THE SCIENCE OF ranking or classifying foods based on their nutrient composition. A nutrient profiling algorithm can calculate a numerical score or be used to assign a symbol that indicates nutritional quality. Nutrient profiling has been used to educate consumers,1 rate foods on the front of food packages and store shelves,2 and define healthy foods that can be advertised to children.3 The development of an algorithm to profile foods involves several decisions, such as the nutrients to include, the unit of measure to base the nutrient amounts (per 100 kcal or serving size), the nutrient intake standards to assess the nutrient contents, and how to weigh the contribution of nutrients to the score. Previous research has explored these topics,4-6 although the selection of which nutrients best describe the nutritional quality of a food is still somewhat arbitrary. Most nutrient profiling systems, such as the Nutrient Rich Foods Index,7 weigh nutrients equally because of a lack of clear scientific basis to quantify the strength of associations between single nutrients and 1968
JOURNAL OF THE ACADEMY OF NUTRITION AND DIETETICS
health benefits. One algorithm that uses weighting factors for nutrients, the Overall Nutritional Quality Index algorithm for the NuVal system, incorporates proprietary weights that represent the relative impact of a nutrient based on prevalence and severity of a given health condition associated with a nutrient based on expert judgment.8,9 Nutrient profiling systems have been validated against an overall index of dietary quality, such as the 2005 Healthy Eating Index (HEI), which is a tool developed by the US Department of Agriculture to measure diet quality based on recommendations for foods and nutrients to consume or avoid from the 2005 Dietary Guidelines for Americans (DGA).2,7 Using dietary intake data of individuals from the National Nutrition Health and Examination Survey (NHANES), the NuVal system’s Overall Nutritional Quality Index scores of foods consumed explained 29% of the variance in HEI scores of diets2 and the Nutrient Rich Foods Index scores explained 45% of variance in HEI scores.7 The Nutrient Rich Foods Index explored various nutrient combinations, including only “nega© 2012 by the Academy of Nutrition and Dietetics.
RESEARCH tive” nutrients (ie, nutrients that have a negative impact on health and should be limited) and a combination of negative and 6 to 15 “positive” nutrients (ie, that have a positive impact on health and should be encouraged). The validation results demonstrated that the algorithm with nine positive nutrients and three negative nutrients explained the most variation in HEI scores.7 The objective of this study is to develop a nutrient-profiling algorithm with weighting factors using a statistical linear regression approach to determine the nutrients that best explain HEI scores. Our goal was to maximize the explanatory power of HEI scores and achieve a higher level of explanatory power than has been seen with nutrient-profiling algorithms developed previously. The determination of nutrients to include in our study is based on a growing body of scientific information and the latest dietary recommendations (ie, the 2010 DGA). This research can be useful for development of point-of-purchase nutrition labeling systems, education of consumers, and assessment of food or overall dietary intakes.
METHODS Identification of Nutrients or Food Components for Algorithm The nutrients considered were selected based on various criteria. We included both positive nutrients, which are associated with health benefits and should be encouraged, and negative nutrients, which are associated with obesity and chronic disease when consumed in excess and should be limited in the diet. Selection of positive nutrients was based on nutrients of concern in the US diet specified in the 2010 DGA report (ie, vitamin D, calcium, potassium, iron, vitamin B-12, folate, and fiber)10 and the 2005 DGA report (ie, magnesium, vitamins A, C, and E).11 In addition, unsaturated fat was included because of its recognized importance and recommendation in the 2010 DGA to replace saturated fatty acids with polyunsaturated and monounsaturated fatty acids. We also included protein because of its dietary importance, even though it is consumed in adequate amounts by the majority of the population.11 Negative nutrients considered were saturated fat and sodium because of their known effects on chronic disease risk and their overconsumption in the US diet, as identified by both the 2005 and 2010 DGA.10,11 We also considered added sugars, a food component that contributes to excess energy intakes10 and is implicated in obesity and chronic diseases.12 For simplicity, hereafter the term nutrients refers to the 15 nutrients plus the food component added sugars. A statistical approach was used to determine which of the 16 nutrients of interest to include in a food scoring algorithm by assessing which nutrients best explain variation in scores of overall dietary quality among the US population. We used the HEI as an overall score of dietary quality.13 Dietary intake data from NHANES 2005-2008 was used to calculate nutrient intakes and HEI scores. Data were from day 1 dietary records of 16,587 individuals who were 2 years of age or older, not pregnant, and had reliable intake records, defined as all relevant variables associated with the dietary recall containing a value.14 Written informed consent was obtained for all participants and the survey protocol was approved by the Research Ethics Review Board at the National Center for Health Statistics. HEI scores were calculated using methods described by Guenther and colleagues.15 Regression analyses December 2012 Volume 112 Number 12
were conducted with HEI scores as the dependent variable and all possible combinations of the 16 nutrients as independent variables, with covariates age, sex, and ethnicity. Nutrient intake values were expressed as a percentage of recommended intake levels per 100 kcal.16 The daily recommended intake values were primarily from the Daily Value or Daily Reference Value.17 For unsaturated fat, the denominator was derived from the total fat Daily Reference Value of 30% of total kilocalories minus the saturated fat Daily Reference Value of 10% total kilocalories; therefore, 20% of total energy from unsaturated fat or 44 g was used as the basis for unsaturated fat. There is no Daily Value for added sugars, so we used the recommendation of 10% of total energy as a basis.18 Values for added sugars were obtained from the MyPyramid Equivalents Database, 2.0 for US Department of Agriculture Survey Foods, 2003–2004,19 and values for new foods in 2005-2008 were determined by matching to similar foods. The MAXimum R2 option in some SAS procedures allows for examining every possible combination of variables of interest. However, a MAXimum R2 option is not available in the SAS SurveyReg procedure used to analyze complex survey data, such as NHANES; therefore, a macro was developed to assess every possible combination of the 16 nutrients while accounting for the survey weights in NHANES (SAS version 9.2; SAS Institute). This method identifies the best one-variable model producing the greatest percentage of variation explained (R2), the best two-variable model, and so forth. The MAXimum R2 approach, unlike other standardized stepwise methods, evaluates the possible switching of the order of variables entered into the model, which can affect the model results. These analyses resulted in evaluation of 65,536 separate regression models, which is equivalent to 2 to the power of 16 (nutrients). From each set of models that contained a specified number of nutrient variables (1 through 16), the model with the highest adjusted R2was chosen as the best model. Adjusted R2 values were used to compare the various models because the number of variables in the model affects the R2. The adjusted R2 adjusts for the number of explanatory terms in the model including nutrients and the other covariates age, sex, and ethnicity.
Calculation of a Weighted Nutrient Density Score The basis for deciding which nutrients to include in the scoring algorithm was to choose the lowest-nutrient number model for which variation in HEI scores R2reached a plateau. The regression coefficients from this model were used as weighting factors in the scoring algorithm. The algorithms calculate a score based on the nutrient value of foods per 100 kcal divided by the Daily Recommended Intake value of the nutrient, summing these values for all of the positive nutrients and subtracting the sum of the negative nutrients. Positive nutrients were capped at 100% of Daily Value to avoid undue influence by one nutrient, lack of strong evidence for intakes above daily requirements, and to discourage food manufacturers from overly fortifying foods in order to yield high scores.4,20
Testing of the Weighted Nutrient Density Score with Population Subgroups Previous nutritional scoring algorithms have used the HEI as a standard of validation.2,7 Because our algorithm was develJOURNAL OF THE ACADEMY OF NUTRITION AND DIETETICS
1969
JOURNAL OF THE ACADEMY OF NUTRITION AND DIETETICS
RESEARCH
1970
Table 1. -coefficients from linear regression models of nutrient intakes on Healthy Eating Index scores using 1-day dietary intakes of 16,587 participants from National Health and Nutrition Examination Survey 2005-2008a No. of nutrient Adjusted Vitamin Vitamin variables R2 Protein Fiber E D Calcium Iron
Unsaturated Vitamin Vitamin Folic Vitamin Saturated Added Potassium fat Magnesium A C acid B-12 fat Sodium sugar
1
0.37
5.32b
2
0.43
4.67b
⫺1.80b
3
0.52
b
⫺2.39b
4
0.55
3.59
3.55b b
0.58
1.42
3.60
6
0.61
1.59b
3.74b
0.63
b
b
7
1.38
b
1.40
3.13
9
0.65
1.29b
3.19b
0.65
b
b
b
b
1.14
b
0.65
1.31
1.78b b
b
1.00 0.60b
0.67b b
2.88
c
0.82
b
b
0.90
b
c
0.37
2.50b
0.37b
b
b
2.69
1.25
2.97
0.43
0.61
0.73
2.64
12
0.66
1.27b
3.09b
0.38b
0.73b
0.96b
2.68b
13
0.66
b
1.33
b
3.07
b
0.48
b
0.72
b
b
14
0.66
1.31b
3.11b
0.36b
0.68b
0.66
b
b
b
b
⫺0.26 0.96
b
⫺0.26 0.97
16
0.66
1.34
b
1.35
3.08
b
3.07
0.43 f
0.24
b
0.44
0.68 0.68
⫺0.33c 0.94b e e
b b
0.32
b
0.66
0.98
b
2.51
11
15
⫺2.35
⫺1.32
⫺0.79b
⫺2.73b
⫺1.41b
⫺0.65b
⫺3.10
⫺1.37
⫺0.56b
⫺2.95
b
⫺1.34
⫺0.52b
b
2.41
b
b
⫺0.54e
2.67
⫺0.53
2.68b
⫺0.49e
e
December 2012 Volume 112 Number 12
b
⫺0.50
b
⫺0.56
2.68 2.60
e e
⫺0.82b b
b
3.51
8 10
⫺2.43b
1.24b
b
5
⫺0.89b
b
b
⫺2.90b
⫺1.27b
⫺0.52b
b
⫺2.85
⫺1.36
⫺0.47b
b
⫺2.83
b
⫺1.31
⫺0.48b
⫺2.91b
⫺1.33b
⫺0.49b
⫺2.88
b
⫺1.33
⫺0.48b
⫺2.93b
⫺1.31b
⫺0.49b
b
0.50
0.33
0.41
0.31b
0.40b
b
b
b
⫺0.17
d
b
b
0.31
0.48
0.25e
0.30b
0.58b
e
b
b
⫺0.14
⫺2.91
⫺1.32
⫺0.48b
b
⫺0.15
⫺2.89
⫺1.32
⫺0.48b
0.25
f
0.22
0.30
b
0.29
0.60 0.59
e e
b b
b b
a Linear regression models used the maximum R2 option to determine, first, the best 1-nutrient variable model that produces the greatest percentage of variation explained (R2); then, repeated to yield the best 2-nutrient variable model, up to the best 16-nutrient variable model; additional covariates in all models included age, sex, and ethnicity. b P⬍0.0001. c P⬍0.001. d P⬍0.01 e P⬍0.05. f Pⱖ0.05.
RESEARCH
Figure 1. Algorithm for a weighted nutrient density score of foods. Weighting factors were derived from linear regression analysis of nutrient intakes on Healthy Eating Index (HEI) scores using dietary intake data from National Health and Nutrition Examination Survey 2005-2008. The amount of each nutrient is divided by the Daily Value of recommended intakes. Positive nutrients were capped at 100% of recommended intake.
Table 2. Results of linear regression models of weighted nutrient density scores on Healthy Eating Indexa Mean WNDSbⴞSEc
n
-coefficient (SE)
P value
R2 (%)
16,587
7.4⫾0.3
1.00 (0.02)
⬍0.0001
64.76
Children (2-18 y)
6,706
5.1⫾0.2
1.12 (0.02)
⬍0.0001
64.48
Adults (ⱖ19 y)
9,881
8.2⫾0.4
0.96 (0.02)
⬍0.0001
65.54
Older adults (ⱖ50 y)
4,792
10.6⫾0.3
0.96 (0.02)
⬍0.0001
66.81
Non-Hispanic white
6,700
7.0⫾0.4
1.00 (0.02)
⬍0.0001
66.07
Non-Hispanic black
4,110
6.3⫾0.3
0.99 (0.02)
⬍0.0001
60.05
Mexican Americans
3,781
9.9⫾0.3
0.99 (0.03)
⬍0.0001
60.11
ⱕ1.85
7,353
6.8⫾0.5
1.00 (0.01)
⬍0.0001
64.63
⬎1.85
8,175
7.5⫾0.3
1.00 (0.02)
⬍0.0001
64.87
1,262
4.9⫾0.5
1.02 (0.06)
⬍0.0001
65.00
996
4.8⫾0.6
1.11 (0.04)
⬍0.0001
69.17
4,448
5.2⫾0.3
1.15 (0.02)
⬍0.0001
63.44
Obese (ⱖ30)
3,457
7.6⫾0.4
0.97 (0.03)
⬍0.0001
64.19
Overweight (ⱖ25 and ⬍30)
3,316
8.7⫾0.4
0.95 (0.03)
⬍0.0001
65.05
Normal weight (⬍25)
2,965
8.0⫾0.5
0.98 (0.03)
⬍0.0001
67.70
Normal (⬍130 mg/dL)f
4,132
7.1⫾0.3
0.95 (0.04)
⬍0.0001
64.10
Elevated (ⱖ130 mg/dL)
1,450
7.4⫾0.6
0.95 (0.03)
⬍0.0001
65.10
Overall Age groups
Ethnicity
Poverty income ratio
d
Child BMI -for-age status Obese (ⱖ95th percentile) Overweight (ⱖ85th and ⬍95th percentile) Normal weight (⬍85th percentile) Adult BMI status
e
LDL status
a Regression models were adjusted for complex sampling of National Health and Nutrition Examination Survey and included covariates for age, sex, and ethnicity, except regression models by age group included covariates for sex and ethnicity and models by ethnicity included age and sex. b WNDS⫽weighted nutrient density scores. c SE⫽standard error. d BMI⫽body mass index. e LDL⫽low-density lipoprotein. f To convert mmol/L cholesterol to mg/dL, multiply mmol/L by 38.7.To convert mg/dL cholesterol to mmol/L, multiply mg/dL by 0.0259. Cholesterol of 5.00 mmol/L ⫽ 193 mg/dL.
oped based on the model that best explained variance in HEI scores, it cannot be validated using the HEI. Instead, we verified that the R2 from a model that regressed HEI scores on algorithm scores of foods consumed by individuals in NHANES was the same as the R2 from the model chosen from the MAXimum R2 analyses. In addition, we conducted separate verification analyses for several population subgroups December 2012 Volume 112 Number 12
and compared R2 from these models. Age groups included children 2 to 18 years, adults ⱖ19 years, and adults ⱖ50 years. Ethnic groups included non-Hispanic white, non-Hispanic black, and Mexican American. For income status, the variable representing the poverty income ratio, a ratio of family income to the poverty threshold level, was categorized as ⱕ1.85 or ⬎1.85. Weight status for adults was defined as obese (body JOURNAL OF THE ACADEMY OF NUTRITION AND DIETETICS
1971
RESEARCH
Figure 2. Weighted nutrient density scores (WNDS) of foods by major food groupings. WNDS for all foods consumed by 16,587 participants on day of dietary recall in the National Health and Nutrition Examination Survey 2007-2008. Maximum, minimum, and median scores are presented for all 4,059 foods and for each major food grouping as specified by the US Department of Agriculture food coding scheme. mass index [BMI] ⱖ30), overweight (BMI ⱖ25 and ⬍30), and normal (BMI ⬍25).21 Weight status for children was defined as obese (BMI-for-age ⱖ95th percentile), overweight (BMIfor-age ⱖ85th and ⬍95th percentile), and normal (BMI-forage ⬍85th percentile).22 Lipid status was defined as normal low-density lipoprotein (⬍130 mg/dL [3.37 mmol/L]) and elevated low-density lipoprotein (ⱖ130 mg/dL [3.37 mmol/L]).23
RESULTS Results of the MAXimum R2 analyses with the best 1- through 16-nutrient models are shown in Table 1. Most of the P values were significant (P⬍0.05). The best one-nutrient variable model contained fiber, explaining 36.6% of the variance in HEI scores. The highest R2 two-nutrient variable model included fiber and saturated fat, and the highest R2 three-nutrient variable model included fiber, saturated fat, and added sugars. Fiber was included in all of the best models for each number of nutrient variables, and saturated fat was included in all but the one-nutrient model. There was a consistent increase in adjusted R2 up to eight nutrient variables, explaining approximately 65% of the variation in HEI scores. These eight nutrients were protein, fiber, calcium, unsaturated fat, vitamin C, saturated fat, sodium, and added sugars. To increase the explanatory power by only one percentage point to 66% would require three additional nutrients—vitamin D, potassium, and folic acid. For efficiency of the algorithm, we focused on the first eight nutrients. The algorithm created from the regression coefficients of the nutrients from the best eight nutrient model is depicted in Figure 1. Results of regression analyses using the algorithm developed from the eight-nutrient model are presented in Table 2. The algorithm explained more of the variance in HEI scores for some subpopulations than others. The R2 of the model with older adults was higher than models with all adults or children only. Older adults also had more nutrient-dense diets; the weighted mean score was higher than the other age 1972
JOURNAL OF THE ACADEMY OF NUTRITION AND DIETETICS
groups. Mexican Americans had a higher weighted mean score than the other ethnic groups, but that model had a lower R2 than the model with non-Hispanic whites. A more nutrient-dense diet as measured by weighted mean score does not equate to a higher R2 because the R2 is driven by better agreement between algorithm scores and HEI scores across all levels of the HEI. Models with data for overweight children had higher R2 than models with normal or obese children. There was little difference in R2 in models by poverty status or lowdensity lipoprotein status. The weighted nutrient density score of 4,059 foods that were consumed by participants in NHANES 2007-2008 on the first dietary intake day ranged from to ⫺1,189 to 325. Legumes, vegetables, and fruits had the highest median scores of major food groups (Figure 2). Some foods had extreme values, particularly low-calorie foods, such as condiments and pickled vegetables, which are high in sodium, because the nutrients represent the amounts per 100 kcal. Examples of weighted nutrient density score are presented in Table 3. The relative ranking of food scores within categories was as expected. For example, low-fat dairy foods scored higher than their higher-fat counterparts, roasted chicken scored higher than hot dogs, and almonds had higher scores than peanut butter.
DISCUSSION We developed a weighted nutrient density scoring algorithm that derived weighting factors for nutrients from a novel application of a standard statistical approach using linear regression of nutrient intakes of the US population to predict overall dietary quality as measured by the HEI. The analyses yielded the best 1- through 16-nutrient models, and we selected the model with 8 nutrients as the one with the minimal number of nutrients that most efficiently maximized R2, explaining 65% of the variation in HEI scores. This algorithm included positive weighting factors for protein, unsaturated December 2012 Volume 112 Number 12
RESEARCH Table 3. Selected examples of weighted nutrient density scores of foods consumed by participants in the National Health and Nutrition Examination Survey (NHANES) 20052008
Table 3. (continued) WNDSa
Grain products Bread, wheat or cracked wheat
21.9
Nutri-Grain cereal bar (Kellogg’s)
20.3
55.0
Waffle, plain
16.4
Milk, cow’s, fluid, skim or nonfat
54.7
Bread, white
9.5
Milk, cow’s, fluid, 1% fat
28.4
Cookie, peanut butter
7.2
Yogurt, plain, low-fat
27.8
Corn flakes (Kellogg’s)
3.9
Milk, chocolate, low-fat
19.1
Lasagna with meat
2.9
Cheese, cottage, low-fat (1% to 2%)
14.5
Cookie, brownie, without icing
0.8
Yogurt, fruit, low-fat, with low-calorie sweetener
14.3
Pizza, cheese, regular crust
11.7
Pretzels, hard
⫺0.4
7.6
Macaroni or noodles with cheese
⫺1.2
1.4
Cake, yellow, with icing
⫺6.9
WNDSa
Dairy foods Yogurt, plain, nonfat
Yogurt, fruit variety, nonfat Milk, cow’s, fluid, 2% fat Yogurt, fruit variety, low-fat
0.1
Milk, chocolate, whole
⫺3.1
Fruit
Cheese, Mozzarella, part skim
⫺4.5
Orange, raw
Milk, cow’s, fluid, whole
⫺6.6
Orange juice, with calcium added
91.8
Yogurt, plain, whole
⫺9.4
Avocado, raw
82.1
115.2
Ice cream, fat-free, chocolate
⫺10.3
Apple, raw
65.5
Yogurt, fruit variety, whole
⫺11.9
Plum, raw
58.1
Ice cream, regular, chocolate
⫺32.3
Orange juice, from frozen
48.8
Cheese, natural, cheddar or American
⫺36.0
Banana, raw
45.4
Meats, poultry, fish, and eggs
Watermelon, raw
42.1
Chicken, breast, roasted, no skin
37.9
Applesauce, stewed apples, unsweetened
36.0
Salmon, baked or broiled
34.1
Grapes, raw
29.5
Beef, roast, roasted, lean only
21.3
Peach, cooked or canned, drained solids
22.8
Chili con carne with beans
20.0
Apple juice
21.0
18.1
Raisins
20.2
Pork chop, broiled or baked, lean only Egg, whole, boiled Ground beef, ⬍80% lean, cooked Frankfurter, wiener, or hot dog
5.0
Vegetables
⫺0.9
Spinach, raw
215.7
⫺26.7
Broccoli, raw
163.5
White potato, chips
33.5
74.5
Tomato and vegetable juice
26.0
Pinto, beans, dry, cooked
66.3
18.9
Pork and beans
44.1
White potato, french fried, from frozen, deep fried
Mixed nuts
43.1
White potato, baked, peel not eaten
17.9
Peanut butter, reduced fat
31.9
Olives
17.2
Peanut butter
31.3
Salsa, red, not homemade
Legumes and nuts Almonds
Pickles, dill
Grain products
⫺18.3 ⫺240.9
Fiber One chewy bar (General Mills)
80.6
Fats, oils, and dressings
Cheerios (General Mills)
49.5
Olive oil
30.8
Bread, multigrain
41.3
Mayonnaise, regular
20.9
Oatmeal, cooked, regular
35.5
Margarine
12.8
Popcorn, popped in oil, unbuttered
25.9
(continued on next page)
(continued)
December 2012 Volume 112 Number 12
JOURNAL OF THE ACADEMY OF NUTRITION AND DIETETICS
1973
RESEARCH Table 3. Selected examples of weighted nutrient density scores of foods consumed by participants in the National Health and Nutrition Examination Survey (NHANES) 20052008 (continued) Fats, oils, and dressings French dressing
WNDSa 6.5
French dressing, reduced calorie
⫺22.1
Butter, stick, salted
⫺90.4
Italian dressing, reduced calorie, fat-free
⫺115.1
Sweets and beverages Coffee, made from ground, regular
48.6
Fruit juice drink, low-calorie, with high vitamin C
48.0
Tea, herbal
14.9
Soft drink, cola-type, sugar-free Fruit flavored drink, from powdered mix
8.1 ⫺3.0
Hard candy
⫺17.1
Soft drink, cola-type
⫺24.8
Milk chocolate candy, plain
⫺35.8
a
WNDS⫽weighted nutrient density scores.
fat, fiber, calcium, and vitamin C and negative weighting factors for saturated fat, sodium, and added sugars. Comparing nutrient scores of foods to overall diet quality scores has been suggested as a validation tool and a method to compare profiling algorithms for their ability to predict nutrient quality.4,6 Previous nutrient density⫺based scoring algorithms predicted 45%7 and 29%2 of the variance in HEI scores of participants in NHANES. Our model was based on weighting factors from regressions of nutrient intakes on HEI scores and performed better than the Nutrient Rich Food Index7 and the Overall Index of Nutrient Quality.8 The Nutrient Rich Food Index used no weighting factors and the Overall Index of Nutrient Quality, which is the basis for the NuVal system, used proprietary weighting factors based on the perceived relative strength of the association between nutrients and health conditions. Few nutrient scoring algorithms have incorporated weighting because of the lack of a strong evidence base for decisions on how weighting factors should be determined. Although it seems sensible to base weighting factors on the association of the nutrient with important health outcomes, assigning an actual specific value that represents the impact of that nutrient on health outcomes is not straightforward and is dependent on the interpretation of available evidence in the literature. Our model is unique because it incorporates unsaturated fat, based on the new suggestion in the 2010 DGA that Americans should replace saturated fatty acids with polyunsaturated and monounsaturated fatty acids. To our knowledge, only one other nutrient scoring algorithms has included unsaturated fatty acids; the NuVal system includes n-3 fatty acids and a fat-quality score based on the percentage of calories derived from unsaturated fatty acids in its algorithm.9 The inclusion of unsaturated fat provides nutritious foods, such as 1974
JOURNAL OF THE ACADEMY OF NUTRITION AND DIETETICS
nuts, with a high food score rather than being penalized for the relatively high total fat content. The advantage of creating a weighted nutrient density score for foods based on the prediction of HEI score is that the algorithm can be used to score the nutritional quality of diets as well as individual foods. Calculating the HEI score can be cumbersome for many registered dietitians or researchers because the score uses information on intake of foods in MyPyramid equivalent servings, such as the number of cup equivalents of fruits and ounce equivalents of whole grains. Most nutrient databases do not contain information on the content of MyPyramid equivalents for foods and would need to be assigned for each food using the US Department of Agriculture MyPyramid Equivalents Database, which contains information on foods consumed in the NHANES 2003-2004 dietary survey.19 A diet score based on nutrients can easily be calculated using most nutrient databases. Our algorithm is transparent and uses seven nutrients and added sugars. Added sugar is the only component that would need to be calculated for foods of interest, but can be obtained for most food items from MyPyramid Equivalents Database. Our methodology has some limitations. Our selection of nutrients was based primarily on recent federal dietary guidance and would need to be reassessed as new evidence emerges that influences future federal guidance. We could have included many more nutrients; however, the number of models tested in the MAXimum R2 analyses would have been prohibitive, requiring many thousands more regression analyses. Our 8-nutrient model explained similar variance in dietary quality as the 16-nutrient model; therefore, inclusion of additional nutrients would be unlikely to improve the algorithm. In addition, although the HEI is the best measure of overall dietary quality currently available, it has some limitations. The HEI is based on dietary recommendations in the MyPyramid and the 2005 DGA, which are no longer in effect as federal guidance. The 2010 DGA recommends incorporating unsaturated fats in place of saturated fats in the diet, and no longer emphasizes some nutrients that were noted as shortfalls in the diets for the 2005 DGA, such as vitamin A and C. Our algorithm testing took into account the recommendation for unsaturated fat and this nutrient remained significantly related to HEI scores. Also, the relationship between HEI and our nutrient scoring algorithm depends somewhat on the influence of the various components on the HEI score. For example, the HEI validation study conducted by Guenther and colleagues13 demonstrated that certain components of the HEI have more influence on the score, that is, calories from solid fats, alcohol, added sugars, and fruits. We found that our algorithm resulted in high scores for fruits and better agreement with HEI scores when added sugar was present in the algorithm, which might be, in part, an artifact of the influence these components have on the HEI. With the 2010 DGA, the HEI can be updated and it will be important to re-evaluate the algorithm.
CONCLUSIONS The weighted nutrient density score is a novel nutrient profiling algorithm that incorporates weighting of nutrients based on overall dietary quality. The weighted nutrient density score algorithm explained 65% of the variation in HEI scores of diets of Americans, which is more than other previous nutrient scoring December 2012 Volume 112 Number 12
RESEARCH algorithms. The weighted nutrient density score can be easily calculated with information on seven nutrients and added sugars, and can be used to describe overall nutrient quality of foods or diets. The weighted nutrient density score was applicable to various subpopulations of Americans, although it would be warranted to test the algorithm with other populations with more diverse dietary patterns.
References 1.
Mobley AR, Kraemer D, Nicholls J. Putting the Nutrient-Rich Foods Index into practice. J Am Coll Nutr. 2009;28(4):427S-435S.
2.
Katz DL, Njike VY, Rhee LQ, Reingold A, Ayoob KT. Performance characteristics of NuVal and the Overall Nutritional Quality Index (ONQI). Am J Clin Nutr. 2010;91(suppl):1102S-1108S.
3.
Rayner M, Scarborough P, Lobstein T. The UK Ofcom Nutrient Profiling Model: Defining ‘healthy’ and ‘unhealthy’ foods and drinks for TV advertising to children. Oxford, UK: British Heart Foundation Health Promotion Research Group, University of Oxford; 2009. http://www. publichealth.ox.ac.uk/bhfhprg/publicationsandreports/publications/ bhfhprgpublished/nutrientprofilingmodel. Accessed September 16, 2011.
4.
Drewnowski A, Fulgoni V. Nutrient profiling of foods: Creating a nutrient-rich food index. Nutr Rev. 2008;66(1):23-39.
5.
Drewnowski A, Maillot M, Darmon N. Should nutrient profiles be based on 100 g, 100 kcal or serving size? Eur J Clin Nutr. 2009;63(7): 898-904.
6.
Tetens I, Oberdorfer R, Madsen C, de Vries J. Nutritional characterisation of foods: Science-based approach to nutrient profiling. Eur J Nutr. 2007;46(suppl 2):4-14.
7.
Fulgoni VL III, Keast DR, Drewnowski A. Development and validation of the nutrient-rich foods index: A tool to measure nutritional quality of foods. J Nutr. 2009;139(8):1549-1554.
8.
Katz DL. ONQI Version 1 Abbreviated Reference Manual. Derby, CT: Griffin Hospital; 2007.
9.
Katz DL, Njike VY, Faridi Z, et al. The stratification of foods on the basis of overall nutritional quality: The Overall Nutritional Quality Index. Am J Health Promot. 2009;24(2):133-143.
10.
Dietary Guidelines Advisory Committee. Report of the Dietary Guidelines Advisory Committee on the Dietary Guidelines for Americans, 2010, to the Secretary of Agriculture and the Secretary of Health and Human Services. Washington, DC: US Department of Agriculture, Agricultural Research Service; 2010.
11.
US Department of Health and Human Services, US Department of
Agriculture. Dietary Guidelines for Americans, 2005. 6th ed. Washington, DC: US Government Printing Office; 2005. 12.
Malik VS, Popkin BM, Bray GA, Despres JP, Hu FB. Sugar-sweetened beverages, obesity, type-2 diabetes mellitus, and cardiovascular disease risk. Circulation. 2010;121(11):1356-1364.
13.
Guenther PM, Reedy J, Krebs-Smith SM, Reeve BB. Evaluation of the Healthy Eating Index-2005. J Am Diet Assoc. 2008;108(11):1854-1864.
14.
Centers for Disease Control and Prevention. NHANES 2005-2006 Data Documentation. Dietary Interview (Dr1IFF_D, DR2IFF_D, DR1TOT_D, and DR2TOT_D). Atlanta, GA: Centers for Disease Control and Prevention; 2008.
15.
Guenther PM, Reedy J, Krebs-Smith SM, Reeve BB, Basiotis PP. Development and Evaluation of the Healthy Eating Index-2005: Technical Report: Center for Nutrition Policy and Promotion, US Department of Agriculture. Available at http://www.cnpp.usda.gov/ HealthyEatingIndex.htm; 2007.
16.
Code of Federal Regulations. Title 21: Food and Drugs, Part 101—Food Labeling, Subpart A—General Provisions. § 101.12 Reference amounts customarily consumed per eating occasion; 2010 October.
17.
Code of Federal Regulations. Title 21: Food and Drugs, Part 101—Food Labeling, Subpart A—General Provisions. § 101.9 Nutrition labeling of food; 2010 October.
18.
World Health Organization. Diet, Nutrition and the Prevention of Chronic Disease. Report of a Joint WHO/FAO Expert Consultation. WHO Technical Report Series 916. Geneva, Switzerland: World Health Organization; 2003.
19.
Bowman SA, Friday JE, Moshfegh A. MyPyramid Equivalents Database, 2.0 for USDA Survey Foods 2003-2004. Beltsville, MD: US Department of Agriculture, Agricultural Research Service, Food Surveys Research Group; 2008. http://www.ars.usda.gov/ba/bhnrc/fsrg. Accessed November 25, 2010.
20.
Kennedy E, Rasca P, Dallal G, et al. Alternative approaches to the calculation of nutrient density. Nutr Rev. 2008;66(12):703-709.
21.
Expert Panel on the Identification Evaluation and Treatment of Overweight in Adults. Clinical guidelines on the identification, evaluation, and treatment of overweight and obesity in adults: Executive summary. Am J Clin Nutr. 1998;68(4):899-917.
22.
Ogden CL, Flegel KM. Changes in terminology for childhood overweight and obesity. Natl Health Stat Report. 2010;June 25(25):1-5.
23.
National Cholesterol Education Program, National Institutes of Health. Third Report of the National Cholesterol Education Program (NCEP) Expert Panel on Detection, Evaluation, and Treatment of High Blood Cholesterol in Adults (Adult Treatment Panel III). NIH Publication No. 02-5216. Bethesda, MD: National Institutes of Health; 2002.
AUTHOR INFORMATION J. E. Arsenault is a nutrition policy analyst, J. C. Hersey is a senior scientist, and M. K. Muth is a program director, RTI International, Research Triangle Park, NC. V. L. Fulgoni is senior vice president, Nutrition Impact, LLC, Battle Creek, MI. Address correspondence to: Joanne E. Arsenault, PhD, RD, Food and Nutrition Policy Research Program, RTI International, 3040 Cornwallis Road, Research Triangle Park, NC 27709. E-mail: [email protected]
STATEMENT OF POTENTIAL CONFLICT OF INTEREST No potential conflict of interest was reported by the authors.
FUNDING/SUPPORT This work was supported by contract HHSP23320095651WC with the Department of Health and Human Services, Office of Assistant Secretary for Planning and Evaluation.
December 2012 Volume 112 Number 12
JOURNAL OF THE ACADEMY OF NUTRITION AND DIETETICS
1975
A Novel Approach to Selecting and Weighting Nutrients for Nutrient Profiling of Foods and Diets Joanne E. Arsenault, PhD, RD; Victor L. Fulgoni III, PhD; James C. Hersey, PhD; Mary K. Muth, PhD
ARTICLE INFORMATION Article history: Accepted 16 July 2012
Keywords: Nutrient profiling Dietary quality Dietary intake Copyright © 2012 by the Academy of Nutrition and Dietetics. 2212-2672/$36.00 doi: 10.1016/j.jand.2012.08.032
ABSTRACT Background Nutrient profiling of foods is the science of ranking or classifying foods based on their nutrient composition. Most profiling systems use similar weighting factors across nutrients due to lack of scientific evidence to assign levels of importance to nutrients. Objective Our aim was to use a statistical approach to determine the nutrients that best explain variation in Healthy Eating Index (HEI) scores and to obtain -coefficients for the nutrients for use as weighting factors for a nutrient-profiling algorithm. Design We used a cross-sectional analysis of nutrient intakes and HEI scores. Participants Our subjects included 16,587 individuals from the National Health and Nutrition Examination Survey 2005-2008 who were 2 years of age or older and not pregnant. Main outcome measure Our main outcome measure was variation (R2) in HEI scores. Statistical analyses Linear regression analyses were conducted with HEI scores as the dependent variable and all possible combinations of 16 nutrients of interest as independent variables, with covariates age, sex, and ethnicity. The analyses identified the best 1-nutrient variable model (with the highest R2), the best 2-nutrient variable model, and up to the best 16-nutrient variable model. Results The model with 8 nutrients explained 65% of the variance in HEI scores, similar to the models with 9 to 16 nutrients, but substantially higher than previous algorithms reported in the literature. The model contained five nutrients with positive -coefficients (ie, protein, fiber, calcium, unsaturated fat, and vitamin C) and three nutrients with negative coefficients (ie, saturated fat, sodium, and added sugar). -coefficients from the model were used as weighting factors to create an algorithm that generated a weighted nutrient density score representing the overall nutritional quality of a food. Conclusions The weighted nutrient density score can be easily calculated and is useful for describing the overall nutrient quality of both foods and diets. J Acad Nutr Diet. 2012;112:1968-1975.
N
UTRIENT PROFILING OF FOODS IS THE SCIENCE OF ranking or classifying foods based on their nutrient composition. A nutrient profiling algorithm can calculate a numerical score or be used to assign a symbol that indicates nutritional quality. Nutrient profiling has been used to educate consumers,1 rate foods on the front of food packages and store shelves,2 and define healthy foods that can be advertised to children.3 The development of an algorithm to profile foods involves several decisions, such as the nutrients to include, the unit of measure to base the nutrient amounts (per 100 kcal or serving size), the nutrient intake standards to assess the nutrient contents, and how to weigh the contribution of nutrients to the score. Previous research has explored these topics,4-6 although the selection of which nutrients best describe the nutritional quality of a food is still somewhat arbitrary. Most nutrient profiling systems, such as the Nutrient Rich Foods Index,7 weigh nutrients equally because of a lack of clear scientific basis to quantify the strength of associations between single nutrients and 1968
JOURNAL OF THE ACADEMY OF NUTRITION AND DIETETICS
health benefits. One algorithm that uses weighting factors for nutrients, the Overall Nutritional Quality Index algorithm for the NuVal system, incorporates proprietary weights that represent the relative impact of a nutrient based on prevalence and severity of a given health condition associated with a nutrient based on expert judgment.8,9 Nutrient profiling systems have been validated against an overall index of dietary quality, such as the 2005 Healthy Eating Index (HEI), which is a tool developed by the US Department of Agriculture to measure diet quality based on recommendations for foods and nutrients to consume or avoid from the 2005 Dietary Guidelines for Americans (DGA).2,7 Using dietary intake data of individuals from the National Nutrition Health and Examination Survey (NHANES), the NuVal system’s Overall Nutritional Quality Index scores of foods consumed explained 29% of the variance in HEI scores of diets2 and the Nutrient Rich Foods Index scores explained 45% of variance in HEI scores.7 The Nutrient Rich Foods Index explored various nutrient combinations, including only “nega© 2012 by the Academy of Nutrition and Dietetics.
RESEARCH tive” nutrients (ie, nutrients that have a negative impact on health and should be limited) and a combination of negative and 6 to 15 “positive” nutrients (ie, that have a positive impact on health and should be encouraged). The validation results demonstrated that the algorithm with nine positive nutrients and three negative nutrients explained the most variation in HEI scores.7 The objective of this study is to develop a nutrient-profiling algorithm with weighting factors using a statistical linear regression approach to determine the nutrients that best explain HEI scores. Our goal was to maximize the explanatory power of HEI scores and achieve a higher level of explanatory power than has been seen with nutrient-profiling algorithms developed previously. The determination of nutrients to include in our study is based on a growing body of scientific information and the latest dietary recommendations (ie, the 2010 DGA). This research can be useful for development of point-of-purchase nutrition labeling systems, education of consumers, and assessment of food or overall dietary intakes.
METHODS Identification of Nutrients or Food Components for Algorithm The nutrients considered were selected based on various criteria. We included both positive nutrients, which are associated with health benefits and should be encouraged, and negative nutrients, which are associated with obesity and chronic disease when consumed in excess and should be limited in the diet. Selection of positive nutrients was based on nutrients of concern in the US diet specified in the 2010 DGA report (ie, vitamin D, calcium, potassium, iron, vitamin B-12, folate, and fiber)10 and the 2005 DGA report (ie, magnesium, vitamins A, C, and E).11 In addition, unsaturated fat was included because of its recognized importance and recommendation in the 2010 DGA to replace saturated fatty acids with polyunsaturated and monounsaturated fatty acids. We also included protein because of its dietary importance, even though it is consumed in adequate amounts by the majority of the population.11 Negative nutrients considered were saturated fat and sodium because of their known effects on chronic disease risk and their overconsumption in the US diet, as identified by both the 2005 and 2010 DGA.10,11 We also considered added sugars, a food component that contributes to excess energy intakes10 and is implicated in obesity and chronic diseases.12 For simplicity, hereafter the term nutrients refers to the 15 nutrients plus the food component added sugars. A statistical approach was used to determine which of the 16 nutrients of interest to include in a food scoring algorithm by assessing which nutrients best explain variation in scores of overall dietary quality among the US population. We used the HEI as an overall score of dietary quality.13 Dietary intake data from NHANES 2005-2008 was used to calculate nutrient intakes and HEI scores. Data were from day 1 dietary records of 16,587 individuals who were 2 years of age or older, not pregnant, and had reliable intake records, defined as all relevant variables associated with the dietary recall containing a value.14 Written informed consent was obtained for all participants and the survey protocol was approved by the Research Ethics Review Board at the National Center for Health Statistics. HEI scores were calculated using methods described by Guenther and colleagues.15 Regression analyses December 2012 Volume 112 Number 12
were conducted with HEI scores as the dependent variable and all possible combinations of the 16 nutrients as independent variables, with covariates age, sex, and ethnicity. Nutrient intake values were expressed as a percentage of recommended intake levels per 100 kcal.16 The daily recommended intake values were primarily from the Daily Value or Daily Reference Value.17 For unsaturated fat, the denominator was derived from the total fat Daily Reference Value of 30% of total kilocalories minus the saturated fat Daily Reference Value of 10% total kilocalories; therefore, 20% of total energy from unsaturated fat or 44 g was used as the basis for unsaturated fat. There is no Daily Value for added sugars, so we used the recommendation of 10% of total energy as a basis.18 Values for added sugars were obtained from the MyPyramid Equivalents Database, 2.0 for US Department of Agriculture Survey Foods, 2003–2004,19 and values for new foods in 2005-2008 were determined by matching to similar foods. The MAXimum R2 option in some SAS procedures allows for examining every possible combination of variables of interest. However, a MAXimum R2 option is not available in the SAS SurveyReg procedure used to analyze complex survey data, such as NHANES; therefore, a macro was developed to assess every possible combination of the 16 nutrients while accounting for the survey weights in NHANES (SAS version 9.2; SAS Institute). This method identifies the best one-variable model producing the greatest percentage of variation explained (R2), the best two-variable model, and so forth. The MAXimum R2 approach, unlike other standardized stepwise methods, evaluates the possible switching of the order of variables entered into the model, which can affect the model results. These analyses resulted in evaluation of 65,536 separate regression models, which is equivalent to 2 to the power of 16 (nutrients). From each set of models that contained a specified number of nutrient variables (1 through 16), the model with the highest adjusted R2was chosen as the best model. Adjusted R2 values were used to compare the various models because the number of variables in the model affects the R2. The adjusted R2 adjusts for the number of explanatory terms in the model including nutrients and the other covariates age, sex, and ethnicity.
Calculation of a Weighted Nutrient Density Score The basis for deciding which nutrients to include in the scoring algorithm was to choose the lowest-nutrient number model for which variation in HEI scores R2reached a plateau. The regression coefficients from this model were used as weighting factors in the scoring algorithm. The algorithms calculate a score based on the nutrient value of foods per 100 kcal divided by the Daily Recommended Intake value of the nutrient, summing these values for all of the positive nutrients and subtracting the sum of the negative nutrients. Positive nutrients were capped at 100% of Daily Value to avoid undue influence by one nutrient, lack of strong evidence for intakes above daily requirements, and to discourage food manufacturers from overly fortifying foods in order to yield high scores.4,20
Testing of the Weighted Nutrient Density Score with Population Subgroups Previous nutritional scoring algorithms have used the HEI as a standard of validation.2,7 Because our algorithm was develJOURNAL OF THE ACADEMY OF NUTRITION AND DIETETICS
1969
JOURNAL OF THE ACADEMY OF NUTRITION AND DIETETICS
RESEARCH
1970
Table 1. -coefficients from linear regression models of nutrient intakes on Healthy Eating Index scores using 1-day dietary intakes of 16,587 participants from National Health and Nutrition Examination Survey 2005-2008a No. of nutrient Adjusted Vitamin Vitamin variables R2 Protein Fiber E D Calcium Iron
Unsaturated Vitamin Vitamin Folic Vitamin Saturated Added Potassium fat Magnesium A C acid B-12 fat Sodium sugar
1
0.37
5.32b
2
0.43
4.67b
⫺1.80b
3
0.52
b
⫺2.39b
4
0.55
3.59
3.55b b
0.58
1.42
3.60
6
0.61
1.59b
3.74b
0.63
b
b
7
1.38
b
1.40
3.13
9
0.65
1.29b
3.19b
0.65
b
b
b
b
1.14
b
0.65
1.31
1.78b b
b
1.00 0.60b
0.67b b
2.88
c
0.82
b
b
0.90
b
c
0.37
2.50b
0.37b
b
b
2.69
1.25
2.97
0.43
0.61
0.73
2.64
12
0.66
1.27b
3.09b
0.38b
0.73b
0.96b
2.68b
13
0.66
b
1.33
b
3.07
b
0.48
b
0.72
b
b
14
0.66
1.31b
3.11b
0.36b
0.68b
0.66
b
b
b
b
⫺0.26 0.96
b
⫺0.26 0.97
16
0.66
1.34
b
1.35
3.08
b
3.07
0.43 f
0.24
b
0.44
0.68 0.68
⫺0.33c 0.94b e e
b b
0.32
b
0.66
0.98
b
2.51
11
15
⫺2.35
⫺1.32
⫺0.79b
⫺2.73b
⫺1.41b
⫺0.65b
⫺3.10
⫺1.37
⫺0.56b
⫺2.95
b
⫺1.34
⫺0.52b
b
2.41
b
b
⫺0.54e
2.67
⫺0.53
2.68b
⫺0.49e
e
December 2012 Volume 112 Number 12
b
⫺0.50
b
⫺0.56
2.68 2.60
e e
⫺0.82b b
b
3.51
8 10
⫺2.43b
1.24b
b
5
⫺0.89b
b
b
⫺2.90b
⫺1.27b
⫺0.52b
b
⫺2.85
⫺1.36
⫺0.47b
b
⫺2.83
b
⫺1.31
⫺0.48b
⫺2.91b
⫺1.33b
⫺0.49b
⫺2.88
b
⫺1.33
⫺0.48b
⫺2.93b
⫺1.31b
⫺0.49b
b
0.50
0.33
0.41
0.31b
0.40b
b
b
b
⫺0.17
d
b
b
0.31
0.48
0.25e
0.30b
0.58b
e
b
b
⫺0.14
⫺2.91
⫺1.32
⫺0.48b
b
⫺0.15
⫺2.89
⫺1.32
⫺0.48b
0.25
f
0.22
0.30
b
0.29
0.60 0.59
e e
b b
b b
a Linear regression models used the maximum R2 option to determine, first, the best 1-nutrient variable model that produces the greatest percentage of variation explained (R2); then, repeated to yield the best 2-nutrient variable model, up to the best 16-nutrient variable model; additional covariates in all models included age, sex, and ethnicity. b P⬍0.0001. c P⬍0.001. d P⬍0.01 e P⬍0.05. f Pⱖ0.05.
RESEARCH
Figure 1. Algorithm for a weighted nutrient density score of foods. Weighting factors were derived from linear regression analysis of nutrient intakes on Healthy Eating Index (HEI) scores using dietary intake data from National Health and Nutrition Examination Survey 2005-2008. The amount of each nutrient is divided by the Daily Value of recommended intakes. Positive nutrients were capped at 100% of recommended intake.
Table 2. Results of linear regression models of weighted nutrient density scores on Healthy Eating Indexa Mean WNDSbⴞSEc
n
-coefficient (SE)
P value
R2 (%)
16,587
7.4⫾0.3
1.00 (0.02)
⬍0.0001
64.76
Children (2-18 y)
6,706
5.1⫾0.2
1.12 (0.02)
⬍0.0001
64.48
Adults (ⱖ19 y)
9,881
8.2⫾0.4
0.96 (0.02)
⬍0.0001
65.54
Older adults (ⱖ50 y)
4,792
10.6⫾0.3
0.96 (0.02)
⬍0.0001
66.81
Non-Hispanic white
6,700
7.0⫾0.4
1.00 (0.02)
⬍0.0001
66.07
Non-Hispanic black
4,110
6.3⫾0.3
0.99 (0.02)
⬍0.0001
60.05
Mexican Americans
3,781
9.9⫾0.3
0.99 (0.03)
⬍0.0001
60.11
ⱕ1.85
7,353
6.8⫾0.5
1.00 (0.01)
⬍0.0001
64.63
⬎1.85
8,175
7.5⫾0.3
1.00 (0.02)
⬍0.0001
64.87
1,262
4.9⫾0.5
1.02 (0.06)
⬍0.0001
65.00
996
4.8⫾0.6
1.11 (0.04)
⬍0.0001
69.17
4,448
5.2⫾0.3
1.15 (0.02)
⬍0.0001
63.44
Obese (ⱖ30)
3,457
7.6⫾0.4
0.97 (0.03)
⬍0.0001
64.19
Overweight (ⱖ25 and ⬍30)
3,316
8.7⫾0.4
0.95 (0.03)
⬍0.0001
65.05
Normal weight (⬍25)
2,965
8.0⫾0.5
0.98 (0.03)
⬍0.0001
67.70
Normal (⬍130 mg/dL)f
4,132
7.1⫾0.3
0.95 (0.04)
⬍0.0001
64.10
Elevated (ⱖ130 mg/dL)
1,450
7.4⫾0.6
0.95 (0.03)
⬍0.0001
65.10
Overall Age groups
Ethnicity
Poverty income ratio
d
Child BMI -for-age status Obese (ⱖ95th percentile) Overweight (ⱖ85th and ⬍95th percentile) Normal weight (⬍85th percentile) Adult BMI status
e
LDL status
a Regression models were adjusted for complex sampling of National Health and Nutrition Examination Survey and included covariates for age, sex, and ethnicity, except regression models by age group included covariates for sex and ethnicity and models by ethnicity included age and sex. b WNDS⫽weighted nutrient density scores. c SE⫽standard error. d BMI⫽body mass index. e LDL⫽low-density lipoprotein. f To convert mmol/L cholesterol to mg/dL, multiply mmol/L by 38.7.To convert mg/dL cholesterol to mmol/L, multiply mg/dL by 0.0259. Cholesterol of 5.00 mmol/L ⫽ 193 mg/dL.
oped based on the model that best explained variance in HEI scores, it cannot be validated using the HEI. Instead, we verified that the R2 from a model that regressed HEI scores on algorithm scores of foods consumed by individuals in NHANES was the same as the R2 from the model chosen from the MAXimum R2 analyses. In addition, we conducted separate verification analyses for several population subgroups December 2012 Volume 112 Number 12
and compared R2 from these models. Age groups included children 2 to 18 years, adults ⱖ19 years, and adults ⱖ50 years. Ethnic groups included non-Hispanic white, non-Hispanic black, and Mexican American. For income status, the variable representing the poverty income ratio, a ratio of family income to the poverty threshold level, was categorized as ⱕ1.85 or ⬎1.85. Weight status for adults was defined as obese (body JOURNAL OF THE ACADEMY OF NUTRITION AND DIETETICS
1971
RESEARCH
Figure 2. Weighted nutrient density scores (WNDS) of foods by major food groupings. WNDS for all foods consumed by 16,587 participants on day of dietary recall in the National Health and Nutrition Examination Survey 2007-2008. Maximum, minimum, and median scores are presented for all 4,059 foods and for each major food grouping as specified by the US Department of Agriculture food coding scheme. mass index [BMI] ⱖ30), overweight (BMI ⱖ25 and ⬍30), and normal (BMI ⬍25).21 Weight status for children was defined as obese (BMI-for-age ⱖ95th percentile), overweight (BMIfor-age ⱖ85th and ⬍95th percentile), and normal (BMI-forage ⬍85th percentile).22 Lipid status was defined as normal low-density lipoprotein (⬍130 mg/dL [3.37 mmol/L]) and elevated low-density lipoprotein (ⱖ130 mg/dL [3.37 mmol/L]).23
RESULTS Results of the MAXimum R2 analyses with the best 1- through 16-nutrient models are shown in Table 1. Most of the P values were significant (P⬍0.05). The best one-nutrient variable model contained fiber, explaining 36.6% of the variance in HEI scores. The highest R2 two-nutrient variable model included fiber and saturated fat, and the highest R2 three-nutrient variable model included fiber, saturated fat, and added sugars. Fiber was included in all of the best models for each number of nutrient variables, and saturated fat was included in all but the one-nutrient model. There was a consistent increase in adjusted R2 up to eight nutrient variables, explaining approximately 65% of the variation in HEI scores. These eight nutrients were protein, fiber, calcium, unsaturated fat, vitamin C, saturated fat, sodium, and added sugars. To increase the explanatory power by only one percentage point to 66% would require three additional nutrients—vitamin D, potassium, and folic acid. For efficiency of the algorithm, we focused on the first eight nutrients. The algorithm created from the regression coefficients of the nutrients from the best eight nutrient model is depicted in Figure 1. Results of regression analyses using the algorithm developed from the eight-nutrient model are presented in Table 2. The algorithm explained more of the variance in HEI scores for some subpopulations than others. The R2 of the model with older adults was higher than models with all adults or children only. Older adults also had more nutrient-dense diets; the weighted mean score was higher than the other age 1972
JOURNAL OF THE ACADEMY OF NUTRITION AND DIETETICS
groups. Mexican Americans had a higher weighted mean score than the other ethnic groups, but that model had a lower R2 than the model with non-Hispanic whites. A more nutrient-dense diet as measured by weighted mean score does not equate to a higher R2 because the R2 is driven by better agreement between algorithm scores and HEI scores across all levels of the HEI. Models with data for overweight children had higher R2 than models with normal or obese children. There was little difference in R2 in models by poverty status or lowdensity lipoprotein status. The weighted nutrient density score of 4,059 foods that were consumed by participants in NHANES 2007-2008 on the first dietary intake day ranged from to ⫺1,189 to 325. Legumes, vegetables, and fruits had the highest median scores of major food groups (Figure 2). Some foods had extreme values, particularly low-calorie foods, such as condiments and pickled vegetables, which are high in sodium, because the nutrients represent the amounts per 100 kcal. Examples of weighted nutrient density score are presented in Table 3. The relative ranking of food scores within categories was as expected. For example, low-fat dairy foods scored higher than their higher-fat counterparts, roasted chicken scored higher than hot dogs, and almonds had higher scores than peanut butter.
DISCUSSION We developed a weighted nutrient density scoring algorithm that derived weighting factors for nutrients from a novel application of a standard statistical approach using linear regression of nutrient intakes of the US population to predict overall dietary quality as measured by the HEI. The analyses yielded the best 1- through 16-nutrient models, and we selected the model with 8 nutrients as the one with the minimal number of nutrients that most efficiently maximized R2, explaining 65% of the variation in HEI scores. This algorithm included positive weighting factors for protein, unsaturated December 2012 Volume 112 Number 12
RESEARCH Table 3. Selected examples of weighted nutrient density scores of foods consumed by participants in the National Health and Nutrition Examination Survey (NHANES) 20052008
Table 3. (continued) WNDSa
Grain products Bread, wheat or cracked wheat
21.9
Nutri-Grain cereal bar (Kellogg’s)
20.3
55.0
Waffle, plain
16.4
Milk, cow’s, fluid, skim or nonfat
54.7
Bread, white
9.5
Milk, cow’s, fluid, 1% fat
28.4
Cookie, peanut butter
7.2
Yogurt, plain, low-fat
27.8
Corn flakes (Kellogg’s)
3.9
Milk, chocolate, low-fat
19.1
Lasagna with meat
2.9
Cheese, cottage, low-fat (1% to 2%)
14.5
Cookie, brownie, without icing
0.8
Yogurt, fruit, low-fat, with low-calorie sweetener
14.3
Pizza, cheese, regular crust
11.7
Pretzels, hard
⫺0.4
7.6
Macaroni or noodles with cheese
⫺1.2
1.4
Cake, yellow, with icing
⫺6.9
WNDSa
Dairy foods Yogurt, plain, nonfat
Yogurt, fruit variety, nonfat Milk, cow’s, fluid, 2% fat Yogurt, fruit variety, low-fat
0.1
Milk, chocolate, whole
⫺3.1
Fruit
Cheese, Mozzarella, part skim
⫺4.5
Orange, raw
Milk, cow’s, fluid, whole
⫺6.6
Orange juice, with calcium added
91.8
Yogurt, plain, whole
⫺9.4
Avocado, raw
82.1
115.2
Ice cream, fat-free, chocolate
⫺10.3
Apple, raw
65.5
Yogurt, fruit variety, whole
⫺11.9
Plum, raw
58.1
Ice cream, regular, chocolate
⫺32.3
Orange juice, from frozen
48.8
Cheese, natural, cheddar or American
⫺36.0
Banana, raw
45.4
Meats, poultry, fish, and eggs
Watermelon, raw
42.1
Chicken, breast, roasted, no skin
37.9
Applesauce, stewed apples, unsweetened
36.0
Salmon, baked or broiled
34.1
Grapes, raw
29.5
Beef, roast, roasted, lean only
21.3
Peach, cooked or canned, drained solids
22.8
Chili con carne with beans
20.0
Apple juice
21.0
18.1
Raisins
20.2
Pork chop, broiled or baked, lean only Egg, whole, boiled Ground beef, ⬍80% lean, cooked Frankfurter, wiener, or hot dog
5.0
Vegetables
⫺0.9
Spinach, raw
215.7
⫺26.7
Broccoli, raw
163.5
White potato, chips
33.5
74.5
Tomato and vegetable juice
26.0
Pinto, beans, dry, cooked
66.3
18.9
Pork and beans
44.1
White potato, french fried, from frozen, deep fried
Mixed nuts
43.1
White potato, baked, peel not eaten
17.9
Peanut butter, reduced fat
31.9
Olives
17.2
Peanut butter
31.3
Salsa, red, not homemade
Legumes and nuts Almonds
Pickles, dill
Grain products
⫺18.3 ⫺240.9
Fiber One chewy bar (General Mills)
80.6
Fats, oils, and dressings
Cheerios (General Mills)
49.5
Olive oil
30.8
Bread, multigrain
41.3
Mayonnaise, regular
20.9
Oatmeal, cooked, regular
35.5
Margarine
12.8
Popcorn, popped in oil, unbuttered
25.9
(continued on next page)
(continued)
December 2012 Volume 112 Number 12
JOURNAL OF THE ACADEMY OF NUTRITION AND DIETETICS
1973
RESEARCH Table 3. Selected examples of weighted nutrient density scores of foods consumed by participants in the National Health and Nutrition Examination Survey (NHANES) 20052008 (continued) Fats, oils, and dressings French dressing
WNDSa 6.5
French dressing, reduced calorie
⫺22.1
Butter, stick, salted
⫺90.4
Italian dressing, reduced calorie, fat-free
⫺115.1
Sweets and beverages Coffee, made from ground, regular
48.6
Fruit juice drink, low-calorie, with high vitamin C
48.0
Tea, herbal
14.9
Soft drink, cola-type, sugar-free Fruit flavored drink, from powdered mix
8.1 ⫺3.0
Hard candy
⫺17.1
Soft drink, cola-type
⫺24.8
Milk chocolate candy, plain
⫺35.8
a
WNDS⫽weighted nutrient density scores.
fat, fiber, calcium, and vitamin C and negative weighting factors for saturated fat, sodium, and added sugars. Comparing nutrient scores of foods to overall diet quality scores has been suggested as a validation tool and a method to compare profiling algorithms for their ability to predict nutrient quality.4,6 Previous nutrient density⫺based scoring algorithms predicted 45%7 and 29%2 of the variance in HEI scores of participants in NHANES. Our model was based on weighting factors from regressions of nutrient intakes on HEI scores and performed better than the Nutrient Rich Food Index7 and the Overall Index of Nutrient Quality.8 The Nutrient Rich Food Index used no weighting factors and the Overall Index of Nutrient Quality, which is the basis for the NuVal system, used proprietary weighting factors based on the perceived relative strength of the association between nutrients and health conditions. Few nutrient scoring algorithms have incorporated weighting because of the lack of a strong evidence base for decisions on how weighting factors should be determined. Although it seems sensible to base weighting factors on the association of the nutrient with important health outcomes, assigning an actual specific value that represents the impact of that nutrient on health outcomes is not straightforward and is dependent on the interpretation of available evidence in the literature. Our model is unique because it incorporates unsaturated fat, based on the new suggestion in the 2010 DGA that Americans should replace saturated fatty acids with polyunsaturated and monounsaturated fatty acids. To our knowledge, only one other nutrient scoring algorithms has included unsaturated fatty acids; the NuVal system includes n-3 fatty acids and a fat-quality score based on the percentage of calories derived from unsaturated fatty acids in its algorithm.9 The inclusion of unsaturated fat provides nutritious foods, such as 1974
JOURNAL OF THE ACADEMY OF NUTRITION AND DIETETICS
nuts, with a high food score rather than being penalized for the relatively high total fat content. The advantage of creating a weighted nutrient density score for foods based on the prediction of HEI score is that the algorithm can be used to score the nutritional quality of diets as well as individual foods. Calculating the HEI score can be cumbersome for many registered dietitians or researchers because the score uses information on intake of foods in MyPyramid equivalent servings, such as the number of cup equivalents of fruits and ounce equivalents of whole grains. Most nutrient databases do not contain information on the content of MyPyramid equivalents for foods and would need to be assigned for each food using the US Department of Agriculture MyPyramid Equivalents Database, which contains information on foods consumed in the NHANES 2003-2004 dietary survey.19 A diet score based on nutrients can easily be calculated using most nutrient databases. Our algorithm is transparent and uses seven nutrients and added sugars. Added sugar is the only component that would need to be calculated for foods of interest, but can be obtained for most food items from MyPyramid Equivalents Database. Our methodology has some limitations. Our selection of nutrients was based primarily on recent federal dietary guidance and would need to be reassessed as new evidence emerges that influences future federal guidance. We could have included many more nutrients; however, the number of models tested in the MAXimum R2 analyses would have been prohibitive, requiring many thousands more regression analyses. Our 8-nutrient model explained similar variance in dietary quality as the 16-nutrient model; therefore, inclusion of additional nutrients would be unlikely to improve the algorithm. In addition, although the HEI is the best measure of overall dietary quality currently available, it has some limitations. The HEI is based on dietary recommendations in the MyPyramid and the 2005 DGA, which are no longer in effect as federal guidance. The 2010 DGA recommends incorporating unsaturated fats in place of saturated fats in the diet, and no longer emphasizes some nutrients that were noted as shortfalls in the diets for the 2005 DGA, such as vitamin A and C. Our algorithm testing took into account the recommendation for unsaturated fat and this nutrient remained significantly related to HEI scores. Also, the relationship between HEI and our nutrient scoring algorithm depends somewhat on the influence of the various components on the HEI score. For example, the HEI validation study conducted by Guenther and colleagues13 demonstrated that certain components of the HEI have more influence on the score, that is, calories from solid fats, alcohol, added sugars, and fruits. We found that our algorithm resulted in high scores for fruits and better agreement with HEI scores when added sugar was present in the algorithm, which might be, in part, an artifact of the influence these components have on the HEI. With the 2010 DGA, the HEI can be updated and it will be important to re-evaluate the algorithm.
CONCLUSIONS The weighted nutrient density score is a novel nutrient profiling algorithm that incorporates weighting of nutrients based on overall dietary quality. The weighted nutrient density score algorithm explained 65% of the variation in HEI scores of diets of Americans, which is more than other previous nutrient scoring December 2012 Volume 112 Number 12
RESEARCH algorithms. The weighted nutrient density score can be easily calculated with information on seven nutrients and added sugars, and can be used to describe overall nutrient quality of foods or diets. The weighted nutrient density score was applicable to various subpopulations of Americans, although it would be warranted to test the algorithm with other populations with more diverse dietary patterns.
References 1.
Mobley AR, Kraemer D, Nicholls J. Putting the Nutrient-Rich Foods Index into practice. J Am Coll Nutr. 2009;28(4):427S-435S.
2.
Katz DL, Njike VY, Rhee LQ, Reingold A, Ayoob KT. Performance characteristics of NuVal and the Overall Nutritional Quality Index (ONQI). Am J Clin Nutr. 2010;91(suppl):1102S-1108S.
3.
Rayner M, Scarborough P, Lobstein T. The UK Ofcom Nutrient Profiling Model: Defining ‘healthy’ and ‘unhealthy’ foods and drinks for TV advertising to children. Oxford, UK: British Heart Foundation Health Promotion Research Group, University of Oxford; 2009. http://www. publichealth.ox.ac.uk/bhfhprg/publicationsandreports/publications/ bhfhprgpublished/nutrientprofilingmodel. Accessed September 16, 2011.
4.
Drewnowski A, Fulgoni V. Nutrient profiling of foods: Creating a nutrient-rich food index. Nutr Rev. 2008;66(1):23-39.
5.
Drewnowski A, Maillot M, Darmon N. Should nutrient profiles be based on 100 g, 100 kcal or serving size? Eur J Clin Nutr. 2009;63(7): 898-904.
6.
Tetens I, Oberdorfer R, Madsen C, de Vries J. Nutritional characterisation of foods: Science-based approach to nutrient profiling. Eur J Nutr. 2007;46(suppl 2):4-14.
7.
Fulgoni VL III, Keast DR, Drewnowski A. Development and validation of the nutrient-rich foods index: A tool to measure nutritional quality of foods. J Nutr. 2009;139(8):1549-1554.
8.
Katz DL. ONQI Version 1 Abbreviated Reference Manual. Derby, CT: Griffin Hospital; 2007.
9.
Katz DL, Njike VY, Faridi Z, et al. The stratification of foods on the basis of overall nutritional quality: The Overall Nutritional Quality Index. Am J Health Promot. 2009;24(2):133-143.
10.
Dietary Guidelines Advisory Committee. Report of the Dietary Guidelines Advisory Committee on the Dietary Guidelines for Americans, 2010, to the Secretary of Agriculture and the Secretary of Health and Human Services. Washington, DC: US Department of Agriculture, Agricultural Research Service; 2010.
11.
US Department of Health and Human Services, US Department of
Agriculture. Dietary Guidelines for Americans, 2005. 6th ed. Washington, DC: US Government Printing Office; 2005. 12.
Malik VS, Popkin BM, Bray GA, Despres JP, Hu FB. Sugar-sweetened beverages, obesity, type-2 diabetes mellitus, and cardiovascular disease risk. Circulation. 2010;121(11):1356-1364.
13.
Guenther PM, Reedy J, Krebs-Smith SM, Reeve BB. Evaluation of the Healthy Eating Index-2005. J Am Diet Assoc. 2008;108(11):1854-1864.
14.
Centers for Disease Control and Prevention. NHANES 2005-2006 Data Documentation. Dietary Interview (Dr1IFF_D, DR2IFF_D, DR1TOT_D, and DR2TOT_D). Atlanta, GA: Centers for Disease Control and Prevention; 2008.
15.
Guenther PM, Reedy J, Krebs-Smith SM, Reeve BB, Basiotis PP. Development and Evaluation of the Healthy Eating Index-2005: Technical Report: Center for Nutrition Policy and Promotion, US Department of Agriculture. Available at http://www.cnpp.usda.gov/ HealthyEatingIndex.htm; 2007.
16.
Code of Federal Regulations. Title 21: Food and Drugs, Part 101—Food Labeling, Subpart A—General Provisions. § 101.12 Reference amounts customarily consumed per eating occasion; 2010 October.
17.
Code of Federal Regulations. Title 21: Food and Drugs, Part 101—Food Labeling, Subpart A—General Provisions. § 101.9 Nutrition labeling of food; 2010 October.
18.
World Health Organization. Diet, Nutrition and the Prevention of Chronic Disease. Report of a Joint WHO/FAO Expert Consultation. WHO Technical Report Series 916. Geneva, Switzerland: World Health Organization; 2003.
19.
Bowman SA, Friday JE, Moshfegh A. MyPyramid Equivalents Database, 2.0 for USDA Survey Foods 2003-2004. Beltsville, MD: US Department of Agriculture, Agricultural Research Service, Food Surveys Research Group; 2008. http://www.ars.usda.gov/ba/bhnrc/fsrg. Accessed November 25, 2010.
20.
Kennedy E, Rasca P, Dallal G, et al. Alternative approaches to the calculation of nutrient density. Nutr Rev. 2008;66(12):703-709.
21.
Expert Panel on the Identification Evaluation and Treatment of Overweight in Adults. Clinical guidelines on the identification, evaluation, and treatment of overweight and obesity in adults: Executive summary. Am J Clin Nutr. 1998;68(4):899-917.
22.
Ogden CL, Flegel KM. Changes in terminology for childhood overweight and obesity. Natl Health Stat Report. 2010;June 25(25):1-5.
23.
National Cholesterol Education Program, National Institutes of Health. Third Report of the National Cholesterol Education Program (NCEP) Expert Panel on Detection, Evaluation, and Treatment of High Blood Cholesterol in Adults (Adult Treatment Panel III). NIH Publication No. 02-5216. Bethesda, MD: National Institutes of Health; 2002.
AUTHOR INFORMATION J. E. Arsenault is a nutrition policy analyst, J. C. Hersey is a senior scientist, and M. K. Muth is a program director, RTI International, Research Triangle Park, NC. V. L. Fulgoni is senior vice president, Nutrition Impact, LLC, Battle Creek, MI. Address correspondence to: Joanne E. Arsenault, PhD, RD, Food and Nutrition Policy Research Program, RTI International, 3040 Cornwallis Road, Research Triangle Park, NC 27709. E-mail: [email protected]
STATEMENT OF POTENTIAL CONFLICT OF INTEREST No potential conflict of interest was reported by the authors.
FUNDING/SUPPORT This work was supported by contract HHSP23320095651WC with the Department of Health and Human Services, Office of Assistant Secretary for Planning and Evaluation.
December 2012 Volume 112 Number 12
JOURNAL OF THE ACADEMY OF NUTRITION AND DIETETICS
1975