How do socio-economic status, perceived economic barriers and nutritional benefits affect quality of dietary intake among US adults?

Abstract

Background

Socio-economic factors may affect diet quality, perhaps differentially across gender and ethnicity. The mechanism of this association is still largely unknown.

Objectives

We examined the independent effects of socio-economic status (SES), perceived barrier of food price (PBFP), and perceived benefit of diet quality (PBDQ) on diet quality indicators and indices (DQI_j,k), across gender and ethnicity. Additionally, we estimated the mediation proportion of the effect of SES on DQI_j,k through PBFP and PBDQ.

Methods

Data from two cross-sectional surveys, the Continuing Survey of Food Intakes by Individuals (CSFII) and Diet and Health Knowledge Survey (DHKS) 1994–96 were used. Our sample consisted of 4,356 US adults aged 20–65 years. With principal components analysis, SES (an index) was measured using household income per capita and education, and PBDQ was measured using an 11-item scale. PBFP was defined as the ratio of importance of food price score relative to nutrition. DQI_j,k were assessed by a set of indicators and two indices including the Healthy Eating Index.

Results

The associations between SES, PBFP, PBDQ, and DQI_j,k varied significantly across gender and ethnic groups. PBFP acted as a mediator in the association between SES and selected DQI_j indicators, namely energy, fat intake, sodium, and simple sugar consumption (mediation proportion>10%), but not PBDQ.

Conclusions

SES, PBFP and PBDQ all affect dietary intake, and vary by ethnicity and gender. Positive effect of SES on DQI_j,k may be mediated by PBFP but not PBDQ which is an independent protective factor. Nutrition education is important to promote healthy eating.

INTRODUCTION

Poor quality of dietary intake has been associated with various adverse health conditions including overweight and obesity (Kennedy et al., 2001, Guo et al., 2004), coronary heart disease (Srinath Reddy and Katan, 2004), stroke (Ding and Mozaffarian, 2006), hypertension (Appel, 2000), inflammation (Ford et al., 2005) as well as cancer and all-cause mortality (Kant et al., 2000, Mai et al., 2005). Food choice is thought to be influenced by a number of environmental and individual factors. Environmental factors may include the changing nature of food supply, increased reliance on foods consumed away from home, food advertising, marketing and promotion as well as food pricing (Popkin et al., 2005). Among individual factors, a person may prefer some types of foods over others after considering taste and palatability, as well as convenience and health effects. However, perceived importance of food price may also act on food choice with marked consequences on diet quality, a pattern thought to be more salient among the poor segment of the population. In fact, based on a number of studies conducted recently, cost of food was negatively associated with dietary quality, low cost diets tending to have increased energy density and poor nutrient adequacy (Bowman and Vinyard, 2004, Schroder et al., 2006, Drewnowski and Darmon, 2005b, Drewnowski and Darmon, 2005a, Drewnowski, 2004).

Moreover, previous studies have shown that SES was positively related to indicators and indices of diet quality (Groth et al., 2001, Bodnar and Siega-Riz, 2002, Friel et al., 2003, Hulshof et al., 2003, Giskes et al., 2004). In our present study, we hypothesized that this positive relationship varies across ethnic and gender groups in the United States. We expect that, in general, minority groups such as African-Americans are less responsive to SES than Whites and women to have better outcomes with increased SES than men. We also test independent effect of perceived barriers of food choices, i.e., food prices relative to nutritional aspects. This hypothesis was tested in a recent study conducted in Brisbane City in Australia (Turrell and Kavanagh, 2006). Higher barrier among the poor was suggested in a recent qualitative study of fifty-six Australian women with low, middle and upper area-level SES (Inglis et al., 2005). Another determinant that may be related to socioeconomic disparity in diet quality is individual perceived health benefits of diet quality (Turrell and Kavanagh, 2006, Moser et al., 2005, Gittelsohn et al., 2006). In the present study, using nationally representative data, the independent associations of each of these factors with diet quality is estimated and compared between genders and racial groups, after taking into account potential confounding factors in the association. Finally, food price barrier and attitude towards nutritional quality are assessed as mediators between SES and diet quality.

SUBJECTS AND METHODS

Study Design and Subjects

Survey methods

Data from the US Department of Agriculture (USDA) Continuing Survey of Food Intakes by Individuals (CSFII, 1994–96) and the Diet and Health Knowledge Survey (DHKS, 1994–96) were used (US Department of Agriculture ARS, 1994–96). A nationally representative multi-stage stratified sample of 16,103 non-institutionalized persons aged 0 to 90 years residing in the United States contained information about dietary intake (by one or two nonconsecutive, multiple-pass 24-hour recalls that were 3 to 10 days apart); socioeconomic, demographic and health parameters. DHKS is a supplementary survey of the CSFII, collected dietary and health knowledge data on a subset of adults. The two surveys allow linking individuals’ dietary and health knowledge and attitudes to their food intakes. The DHKS collected data from individuals age 20 years and older who provided at least 1 day of dietary intake information in CSFII (one person per household). The interviews of the CSFII and DHKS were separated by at least 1 week. In our present analyses, only those respondents to the DHKS with 2 complete days of intake data in CSFII were included and we used average intakes over these two days to define the main outcome variables.

Study Sample

Among the 16,103 respondents of all ages who completed CSFII 1994–96, 9,872 aged 20 years or older had complete data on day 1 of recall. Out of those, only 5,765 subjects (one person per household) were sampled to complete the DHKS survey. We further excluded those over the age of 65 years (n=1,319) and those who completed only one 24-hour dietary recall (n=90). As a result, our final sample consisted of 4,356 individuals (2,219 men and 2,137 women) who completed both surveys (CSFII and DHKS).

Measures

Outcome variables

1. Dietary Intake and diet quality indicators (DQI_j)

Dietary intake was elicited from subjects participating in the CSFII 1994–96 survey using two 24-hour dietary recalls. Based on responses which uncovered types of foods consumed during these two days along with their portion sizes, nutrient intake was estimated using food composition tables that were designed specifically to be used for this survey. In addition, CSFII 1994–96 re-grouped the foods consumed by broader categories. Average dietary intakes of foods and nutrients from the two-day 24-hour recalls were considered. At a first stage, we chose to look at separate diet quality indicators that are commonly used to construct global indices (Dennis et al., 2003). These included energy (kcal/day), energy density (kilocalories per 100 grams of foods consumed), total fat and saturated fat intake (as percent of total calories), dietary cholesterol (mg), sodium intake (mg) and intake of simple sugars (grams), the excess of which are believed to reduce diet quality. Energy density was defined as the total kilocalories of energy obtained per 100 grams of total food amount consumed. The total food amount consumed included all food and all alcoholic and non-alcoholic beverages in the form reported consumed (i.e., includes water present in beverages such as tea, coffee, cocoa, fruit drinks from dry mixes, and milk drinks), and excluded water such as tap water or bottled water drunk separately. We also considered intake of fruits and vegetables (grams), fiber (grams), calcium (milligrams) and dairy products (grams), adequate intake of which are believed to improve diet quality.

2. Diet quality Indices (DQI_k)

To assess the overall quality of diet, we applied two widely known diet quality indices, namely the Health Eating Index (HEI) (McCullough et al., 2000, Fung et al., 2005) and the Alternate Mediterranean Diet Score (aMED) (Trichopoulou et al., 1995, Trichopoulou et al., 2003, Fung et al., 2005). Appendix A shows the criteria for scoring points on each of these two indices as well as the food groups and nutrients considered to create them. For many of the food group criteria, serving estimates rather than grams were used as made available by USDA website http://www.barc.usda.gov/bhnrc/foodsurvey/. A HEI and aMED score was calculated for each subject, respectively. Our analysis shows a moderate agreement between the HEI and aMED scores, (spearman correlation r=0.50, P<0.05) a kappa=0.15, P<0.05 and % agreement of 32.2% (>expected agreement by chance of 20%) for age- and gender-specific quintiles.

Explanatory variables

1. Perceived benefit of diet quality (PBDQ)

One diet and health-related knowledge scale (“importance of diet component selection”) was selected for subjects participating in the CSFII/DHKS. This scale consisted of 11 questions which were initiated by the following cue: “To you personally, is it very important (score of 4), somewhat important, not too important, or not at all important (score of 1) to”: a) use salt or sodium only in moderation b) Choose a diet low in saturated fat, c) Choose a diet with plenty of fruits and vegetables, d) Use sugars only in moderation?, e) Choose a diet with adequate fiber, f) Eat a variety of foods, g) Maintain a healthy weight, h) Choose a diet low in fat, i) Choose a diet low in cholesterol, j) Choose a diet with plenty of breads, cereals, rice, and pasta; k) Eat at least two servings of dairy products daily.

Using principal components analysis (Sharma, 1996), we extracted a single component score by imposing an eigenvalue criteria >1.0, as well as examining the scree plot. The component which we designated as “perceived benefit of diet quality” (PBDQ) explained around 40% of the variance in the 11 items, which loaded almost equally on this component (loadings ranged from 0.23 and 0.35).

2. Perceived barrier of food prices (PBFP)

Respondents to the DHKS were asked to identify factors they consider important when buying a food. This included safety, taste, ease of preparation (convenience), how well the food keeps (freshness), nutrition and food price. Responses to each question were measured on a 4-point likert scale from 1 “not at all important” to 4 “very important”. We focused on the last two questions (i.e., nutrition and food price) and computed an index of perceived importance of price relative to the importance of nutrition by simply dividing the score of the first by the score of the second and multiplying by 100. Because of the varying gap between these scores, the variable was regrouped into an ordinal one centered at 100 (0=equally important) and ranging between −5 ( x =25) and +5 ( x =400), where $x=\frac{{\mathit{Score}}_{\mathit{price}}}{{\mathit{Score}}_{\mathit{nutrition}}}\times 100$. Both PBFP and PBDQ were self-perceived by the DHKS respondent.

3. Socio-economic status (SES)

Our main exposure variable was measured by a composite z-score of completed educational level (in years) and household income per capita (in $1000). The SES index score was computed using principal components analysis to capture the maximum amount of variance in both measures (Sharma, 1996). The component extracted explained 69% of the variance of both measured variables and loading on both education and income was equal to 0.71. Ordinary Least-square (OLS) bivariate regression analysis showed that an increase in the SES z-score by 1 SD corresponded on average to an increase in household income per capita by $10,093 and in years of education by 2.05. Our analysis shows in this sample the correlation between income and education was 0.4. It has been shown by others that looking at the separate association with one SES measure while adjusting for the others constitutes over-adjustment. Data reduction using a factor analytic approach such as PCA may resolve part of the problem (Onwujekwe, 2005, Hatcher, 1994). Thus, in our analysis we chose to use the created SES index score to estimate the overall association with SES, although we conducted analysis using income and education as two separate variables as well (results not presented).

4. Other covariates

Covariates included in our statistical models as potentially confounding variables were: age (in years), gender, ethnicity (Whites vs. African-American), 1990 Census geographic regions (Northeast, Midwest, South, and West) and degree of urbanization of the geographical area in which households were selected (Metropolitan Statistical Area-central city, MSA- suburban, and rural). In addition, we included food choice factors as they would have an influence on the final outcome (diet quality) without necessarily being affected by SES and we wished to hold them constant while looking at the effect of perceived barrier of food price and benefit of diet quality on the outcome. These factors were measured by asking the question: “When buying a food, what is important to you?” and they included taste, convenience, safety and freshness. Each variable was on a scale of “1” being not important at all to “4” very important. In stratified models, ethnic group and gender were considered as effect modifiers.

Statistical Analysis

To describe the characteristics of the study population (demographic, geographic, socio-economic, diet quality and knowledge) by stratifying variables (ethnic groups and gender), proportions (%) were calculated for categorical variables and mean and standard error of the mean (SEM) computed for continuous variables. Because of the complexity of the sampling design, we adjusted for the complex survey design effect on precision of means and proportions by using survey-related commands. In particular, we specified PSU and stratum for each observation, as well as appropriate weights corresponding to three-year and two-day 24-hour recalls for CSFII/DHKS sample, in order to accurately estimate variances and obtain nationally representative estimates. In order to test hypotheses of differences in means or proportions across categories of another variable and adjust for the survey design effect, we assessed overlap between 95% confidence intervals (CIs) across these categories. In addition, locally weighted regression (LOWESS) was used as a non-parametric smoothing technique to show relationship between continuous variables with a bandwidth of 0.5 (Schimek, 2000).

Stratified ordinary least square (OLS) multivariate regression analysis was carried out to test whether the effects of socio-economic status on diet quality vary across gender and ethnic categories. Design complexity was also taken into account in this analysis. Effect modification by ethnicity or gender was assessed based on a statistically significant interaction term introduced into the OLS model. In all tests, a p-value less than 0.05 was considered as statistically significant. These analyses were conducted using STATA release 9 (STATA, 2003).

To estimate the direct and indirect effects of SES on diet quality indicators and indices as well as to compute the mediation proportion through PBFP and PBDQ, separately we conducted structural equations modeling using AMOS version. 5.0 (Byrne, 2001), in which a set of simultaneous equations were used (Ditlevsen et al., 2005). Although no criteria were available for the mediation proportion, we considered a value of 10% or more as statistically significant mediation. Estimates are based on an age and sex-adjusted models. The models were set up as shown in Figure 1.

FIGURE 1.

Structural Equations Model: Mediation through PBFP and PBDQ

Note: Based on these models, the main parameters of interest were: Effect_direct = α₁₁; Effect_Indirect = α₁₂ × α₁₃; Effect_total = α₁₂ × α₁₃ + α₁₁; $\text{Mediation}\text{proportion}=\left|\frac{{\mathit{Effect}}_{\mathit{indirect}}}{{\mathit{Effect}}_{\mathit{total}}}\right|\times 100$

RESULTS

Characteristics of study population

The un-weighted study sample of 4,356 comprised 2,137 women and 523 African-American subjects, with an overall mean age of 40 years (Table 1). While participants were mostly concentrated in MSA-suburban areas (46.7%), African Americans had a higher proportion in MSA-central city regions than whites (66.4% vs. 26.3%). Educational attainment, per capita income and SES factor scores were significantly higher among whites and among men (p<0.05).

TABLE 1.

Characteristics of study population by ethnicity and gender; CSFII/DHKS 1994–96

		Ethnic group		Gender
	All (n=4,356)	Whites (n=3,487)	African-American (n=523)	Men (n=2,219)	Women (n=2,137)
Demographics	( Mean ± SEM )
Age	40.0 ±0.33	40.6 ±0.34	39.5 ±0.85	39.6 ±0.5	40.4 ±0.3
Urbanization	( proportion ± SEM )
MSA-central city 1	32.8 ±1.8	26.3 ±2.1	66.4 ±4.4*	32.3 ±2.5	33.3 ±1.8*
MSA-suburban	46.7 ±1.8	51.4 ±2.1	24.6 ±3.9	47.0 ±2.5	46.4 ±1.9
Rural	20.4 ±1.2	22.3 ±1.1	8.9 ±2.6	20.6 ±1.8	20.3 ±1.1
Region	( proportion ± SEM )
Northeast	19.3 ±2.1	18.9 ±2.4	23.4 ±4.0*	18.9 ±2.0	19.6 ±2.5
Midwest	22.1 ±1.1	24.0 ±1.2	20.7 ±3.3	22.3 ±1.7	21.8 ±1.1
South	36.1 ±2.8	36.7 ±3.2	49.7 ±4.6	35.8 ±3.0	36.4 ±3.0
West	22.5 ±2.4	20.4 ±2.5	6.2 ±1.3	22.9 ±2.7	22.2 ±2.5
Socio-economic factors	( Mean ± SEM )
Education (years)	13.3 ±0.1	13.6 ±0.1	12.6 ±0.2*	13.3 ±0.1	13.3 ±0.1
Household Income per capita(1000$)	16.8 ±0.3	18.1 ±0.3	11.4 ±0.6*	17.4 ±0.4	16.2 ±0.4
SES factor score2	0.11 ±0.04	0.22 ±0.03	−0.34 ±0.06*	0.14 ±0.04	0.08 ±0.04
Food choice factors3	( Mean ± SEM )
Taste	3.84 ±0.01	3.83 ±0.01	3.88 ±0.03	3.79 ±0.02	3.88 ±0.02
Convenience	3.14 ±0.02	3.11 ±0.03	3.39 ±0.06*	3.10 ±0.03	3.18 ±0.03
Safety	3.83 ±0.01	3.82 ±0.02	3.92 ±0.02*	3.76 ±0.02	3.90 ±0.01*
Freshness	3.46 ±0.02	3.39 ±0.02	3.85 ±0.05*	3.41 ±0.03	3.52 ±0.02
Nutrition	3.59 ±0.02	3.56 ±0.02	3.72 ±0.04*	3.49 ±0.02	3.69 ±0.02*
Price	3.27 ±0.02	3.23 ±0.03	3.51 ±0.08*	3.22 ±0.03	3.32 ±0.03
Perceived Barrier of Food Price (PBFP)	−0.43 ±0.04	−0.47 ±0.0	−0.28 ±0.12*	−0.39 ±0.05	−0.48 ±0.05*
Perceived benefit of diet quality (PBDQ)	( Mean ± SEM )
	−0.08 ±0.05	−0.11 ±0.05	−0.14 ±0.12	−0.41 ±0.08	0.24 ±0.06*

¹MSA: Metropolitan Statistical Area.

²SES factor score is a z-score estimated using principal components analysis of education (years) and household income per capita ($1000).

³These factors were measured by asking the question: “When buying a food, what is important to you?” and they included taste, convenience, safety, freshness, nutrition and price. Each variable was on a scale of “1” being not important at all to “4” very important.

^*p<0.05 for null hypothesis that proportions or means are equal across gender or ethnicity strata: 95% confidence intervals are non-overlapping. STATA survey commands do not allow for hypothesis testing for complex sampling designs.

Food choice factors reflected what was considered as important for individuals when buying a food. In general, African-Americans voiced as important most of these factors and to a greater extent than whites (p<0.05). Only taste scores were similar between racial groups. Safety and nutrition were of greater importance among women than among men. The PBFP score reflects the perceived value of food price relative to the importance of nutrition. On average, subjects valued nutrition more than food price and this is shown by the overall negative PBFP score. However, African Americans saw food price more important than whites (PBFP scores of −0.28 vs. −0.47, P<0.05). Women tended to put more weight on nutrition when compared to men.

PBDQ score reflects the importance placed by individual subjects on meeting general recommended dietary guidelines in terms of food groups and nutrients that would improve diet quality. This score which, like SES, was derived by principal components analysis method, was similar across ethnic groups but was significantly higher among women compared to men (0.24 vs. −0.41, p<0.05).

Associations between SES, perceived barrier of food prices (PBFP) and perceived benefit of diet quality (PBDQ)

Figure 2 is a smoothed graphical representation, using the LOWESS technique, of the association between SES, PBFP and PBDQ scores, stratified by gender and ethnicity. We noticed an almost linear inverse association between PBFP and SES for both genders (r=−0.19, p<0.001) and among whites (r=−0.21, p<0.001). For African Americans, PBFP was inversely associated with SES at lower SES scores, but then reversed its trend afterward yielding a U-shaped pattern (r=−0.11, p<0.01). PBDQ had an inconsistent and non-linear relationship with SES (r=0.01, p=0.32), although patterns were U-shaped among men and white subjects.

FIGURE 2.

Association between socio-economic status (SES), perceived barrier of food price (PBFP) and perceived benefit of diet quality (PBDQ) scores: stratified by gender and ethnicity; CSFII/DHKS 1994–96*

*Smoothing of the curves was done through locally weighted regression models (LOWESS) with a bandwidth of 0.50.

Diet quality indicators and indices: ethnic and gender differentials

Table 2 presents unadjusted means of diet quality indicators and indices across ethnic and gender groups. In general, HEI and aMED indices reflected a significantly poorer diet quality among African-Americans compared to Whites and a better quality among women compared to men (p<0.05). Separate indicators of interest for diet quality gave similar results. Racial disparities were noted for energy density, fat as percent of energy intake, cholesterol, fiber intake as well as intake of dairy products. Gender differentials were noted for all indicators, with the exception of simple sugars (p>0.05). Women had lower energy, energy density, total fat, saturated fat, cholesterol and sodium intake than men. However, men had higher intake of fruits and vegetables, fiber, calcium and dairy products.

TABLE 2.

Distribution of diet quality indicators and indices by ethnicity and gender among US adults; CSFII/DHKS 1994–96

		Ethnic group		Gender
	All (n=4,356)	Whites (n=3,556)	African-American (n=535)	Men (n=2,219)	Women (n =2,137)
Dietary Quality Indicators (DQIj)	( Mean ±SEM )
Energy (kcal)	2,085 ±25	2,074 ±21	2,142 ±168	2,515 ±46	1,670 ±15*
Energy density (kcal/100g)	93.9 ±0.6	91.5 ±0.8	109.4 ±2.3*	97.5 ±0.9	90.4 ±1.0*
Total fat (% kcal)	33.4 ±0.2	33.5 ±0.2	35.1 ±0.4*	34.0 ±0.3	32.8 ±0.3*
Saturated fat (%kcal)	11.1 ±0.1	11.2 ±0.1	11.5 ±0.2	11.4 ±0.1	10.7 ±0.1*
Cholesterol (mg)	275 ±5	263 ±5	335 ±24*	338 ±8	214 ±3*
Sodium (mg)	3500 ±50	3496 ±41	3565 ±305	4219 ±88	2809 ±27*
Fiber (g)	16.1 ±0.2	16.1 ±0.2	13.8 ±0.6*	18.5 ±0.3	13.8 ±0.3*
Calcium (mg)	763.1 ±11.1	781.8 ±9.7	662.5 ±70.0	900.8 ±20.5	630.5 ±9.5*
Simple sugars (g)	22.0 ±1.1	22.9 ±1.3	21.9 ±3.1	23.7 ±1.9	20.5 ±0.9
Fruits and vegetables (g)	372 ±7	368 ±7	346 ±17	398 ±8	346 ±9*
Dairy products (g)	212 ±5.8	220 ±5	156 ±25*	243 ±11	181 ±5*
Diet Quality Indices (DQIk)	( Mean ± SEM )
Healthy Eating Index (HEI)	63. 2±0.2	63.8 ±0.3	57.6 ±0.7*	62.5 ±0.3	63.9 ±0.3*
Alternate Mediterranean Diet Score (aMED)	3.5 ±0.0	3.6 ±0.0	3.1 ±0.1*	3.4 ±0.0	3.6 ±0.1*

^*P<0.05 for null hypothesis means are equal across gender or ethnicity strata: 95% confidence intervals are non-overlapping.

Diet quality indicators and indices: the influence of SES

Table 3 presents a set of OLS multivariate linear regression models showing independent effects of SES, PBFP and PBDQ scores on diet quality indicators and indices. Models with SES as the main exposure controlled for demographics and food choice factors. Models with PBFP and PBDQ as main exposures controlled additionally for SES score.

TABLE 3.

Effects¹ of socio-economic status (SES), perceived barrier of food price (PBFP) and perceived benefit of diet quality (PBDQ) on diet quality indicators and diet quality indices among US adults stratified by ethnicity and gender; CSFII/DHKS 1994–96

			Ethnic group		Gender
		All (n=4,356)	Whites (n=3,556)	African-American (n =535)	Men (n=2,219)	Women (n=2,137)
		β̂ ± SEE
Dietary Quality Indicators (DQIj)
Energy (kcal)
SES2	model 1	6.16 ±17.93	4.79 ±17.14	22.84 ±72.33	−10.20 ±29.22	29.77 ±13.85*
PBFP3	model 2	13.77 ±19.00	12.45 ±15.63	148.38 ±110.11	29.52 ±31.00	−11.66 ±16.28
PBDQ2	model 3	−31.02 ±11.45*	−37.84 ±12.26*	−20.03 ±52.09	−33.73 ±15.69*	−18.65 ±8.40
Energy density (kcal/100g)
SES	model 1	−0.34 ±0.47	0.42 ±0.51	−3.45 ±1.67*	−0.92 ±0.69	0.36 ±0.61
PBFP	model 2	−0.06 ±0.60	0.17 ±0.60	2.01 ±1.68	0.43 ±0.79	−0.68 ±0.73
PBDQ	model 3	−0.07 ±0.45	−0.04 ±0.57	−1.82 ±1.24	0.03 ±0.70	−0.35 ±0.37
Total fat (% kcal)
SES	model 1	−0.57 ±0.15*	−0.71 ±0.17*	0.30 ±0.42a	−0.63 ±0.19*	−0.49 ±0.21*
PBFP	model 2	0.12 ±0.13	0.11 ±0.15	0.42 ±0.46	0.18 ±0.16	0.05 ±0.19
PBDQ	model 3	−0.34 ±0.08*	−0.38 ±0.10*	−0.42 ±0.21*	−0.28 ±0.10*	−0.48 ±0.12*,b
Saturated fat (%kcal)
SES	model 1	−0.33 ±0.06*	−0.43 ±0.07*	0.00 ±0.18a	−0.34 ±0.07*	−0.31 ±0.08*
PBFP	model 2	0.05 ±0.05	0.02 ±0.06	0.20 ±0.16	0.05 ±0.06	0.05 ±0.08
PBDQ	model 3	−0.16 ±0.03*	−0.17 ±0.04*	−0.23 ±0.07*	−0.13 ±0.04*	−0.22 ±0.05*
Cholesterol (mg)
SES	model 1	−16.36 ±3.51*	−14.98 ±3.77*	−15.58 ±8.67	−18.53 ±4.83*	−13.35 ±4.02*
PBFP	model 2	−0.16 ±3.41	−0.45 ±3.47	18.35 ±19.55	−1.32 ±5.59	1.56 ±2.86
PBDQ	model 3	−3.81 ±2.48	−5.59 ±2.51*	3.03 ±10.69	−2.93 ±2.94	−4.70 ±2.98
Sodium (mg)
SES	model 1	11.91 ±34.57	13.93 ±29.01	33.85 ±132.29	−32.29 ±61.56	69.40 ±27.27*
PBFP	model 2	58.36 ±28.42*	31.75 ±24.72	347.59 ±199.13	81.11 ±49.11	21.96 ±21.99
PBDQ	model 3	−34.76 ±17.77*	−51.42 ±17.63*	6.24 ±72.13	−42.05 ±24.00	−9.46 ±15.88
Fiber (g)
SES	model 1	0.86 ±0.24*	1.07 ±0.20*	1.09 ±0.49*	0.72 ±0.38*	0.95 ±0.17*,b
PBFP	model 2	−0.52 ±0.23*	−0.21 ±0.14	0.19 ±0.41	−0.35 ±0.27	−0.66 ±0.34
PBDQ	model 3	0.37 ±0.13*	0.33 ±0.15*	0.06 ±0.35	0.32 ±0.18	0.44 ±0.09*
Calcium (mg.)
SES	model 1	21.82 ±7.97*	18.92 ±8.18*	4.49 ±32.47	10.44 ±12.79	36.45 ±9.49*
PBFP	model 2	−1.94 ±8.50	−7.55 ±8.04	56.46 ±51.28	3.89 ±13.67	−10.52 ±8.61
PBDQ	model 3	7.74 ±4.12	8.91 ±4.12*	−0.44 ±17.09	5.85 ±4.88	14.18 ±5.44*
Simple sugars (g)
SES	model 1	−0.34 ±0.66	−0.24 ±0.70	−3.26 ±2.56	−0.29 ±0.99	−0.20 ±0.83
PBFP	model 2	0.20 ±0.70	−0.38 ±0.79	3.12 ±2.41	0.05 ±0.84	0.33 ±0.97
PBDQ	model 3	−0.89 ±0.67	−1.28 ±0.89	0.68 ±1.21	−0.74 ±0.93	−1.12 ±0.76
Fruits and vegetables (g)
SES	model 1	45.60 ±4.78*	47.07 ±5.12*	29.41 ±19.72	56.80 ±8.66*	34.20 ±6.29*
PBFP	model 2	2.30 ±4.58	4.04 ±4.99	9.32 ±13.77	10.44 ±6.24	−7.88 ±5.44b
PBDQ	model 3	10.93 ±3.19*	11.81 ±3.55*	11.04 ±13.04	6.39 ±3.69	19.88 ±2.94*, b
Dairy products (g)
SES	model 1	5.75 ±3.85	0.05 ±4.01	0.46 ±17.19	3.43 ±6.23	8.40 ±4.63
PBFP	model 2	4.64 ±3.92	2.76 ±3.55	18.33 ±18.29	8.93 ±6.65	−0.96 ±4.64
PBDQ	model 3	5.12 ±2.66	7.65 ±2.57*	2.16 ±7.75	2.82 ±3.13	9.86 ±3.33*
Diet quality indices
Healthy Eating Index (HEI)
SES2	model 1	2.36 ±0.19*	2.47 ±0.22*	1.36 ±0.71	2.57 ±0.31*	2.12 ±0.26*
PBFP3	model 2	−0.18 ±0.14	−0.00 ±0.17	−0.81 ±0.50	0.01 ±0.22	−0.42 ±0.24
PBDQ2	model 3	0.64 ±0.13*	0.75 ±0.16*	0.49 ±0.41	0.47 ±0.16*	1.02 ±0.15*,b
Alternate Mediterranean Diet score (aMED)4
SES	model 1	0.35 ±0.03*	0.38 ±0.03*	0.24 ±0.10*	0.34 ±0.04*	0.36 ±0.03*,b
PBFP	model 2	−0.01 ±0.02	−0.00 ±0.03	−0.05 ±0.05	−0.01 ±0.03	−0.01 ±0.04
PBDQ	model 3	0.09 ±0.03*	0.09 ±0.03*	0.13 ±0.05*	0.07 ±0.04	0.13 ±0.02*

Note: Model 1: Controlling for age, region, urbanization, taste, convenience, safety and freshness scores. Additional control for gender and ethnicity for non-stratified models, control for gender in ethnicity-stratified models, control for ethnicity is gender-stratified models. Models 2 and 3: Same controls as model 1 + SES.

¹Ordinary Least Square (OLS) multivariate regression model with no adjustment for complex design effect.

²Per z-score unit increase in SES or PBDQ.

³Per one unit increase in PBFP on a scale between −5 and 5.

⁴aMED and HEI scores had a Spearman correlation of 0.50 (p<0.001).

^aIn a separate analysis including an interaction term, this interaction term with ethnic group (White vs. African American) is significant, P< 0.05.

^bIn a separate analysis including an interaction term, this interaction term with gender group (Female vs. Male) is significant, P< 0.05.

^*P<0.05 for null hypothesis that β=0.

In the total study population, SES was positively associated with diet quality particularly in terms of reduced consumption of fat, saturated fat and cholesterol and higher intakes of fiber, fruits and vegetables as well as calcium (p<0.05). Ethnic disparities in these relationships were also noted whereby African-Americans showed non-significant associations between SES level and fat and saturated fat intakes, respectively, whereas Whites showed a significant inverse association between these two variables (P<0.05 for interaction term SES×ethnicity). No other significant interactions with ethnicity and SES were noted. In terms of gender interactions, SES was positively associated with fiber intake and the association was stronger among women than among men (P<0.05 for interaction term SES×gender).

Moreover, SES was positively associated with diet quality indices in the total population and among all subgroups. Only one significant interaction was noted between SES and gender when aMED score was considered as the diet quality index. This is shown clearly in Figure 3. Nevertheless, the figure also suggests that ethnic gaps in the effect of SES on HEI were larger than gender gaps. Partial R² for SES in our model using HEI as the outcome variable was 0.05 (i.e., 5% of the variation in HEI could be explained by SES), and between SES and aMED was 0.06. Thus, only a small proportion of the variation in dietary quality could be explained by SES. The large part is due to the influence of other variables. Our analysis also suggests that income has a stronger influence than education. The adjusted R² in the model with HEI as the outcome and included both income and education (i.e., HEI=α+β₁×income+β₂×education) was 0.12. It became 0.09 when only education was kept in the model (i.e., HEI=α + β×education) and stayed close to 0.11 when income only was kept. A similar pattern was observed in models with aMED as the outcome.

FIGURE 3.

Association between socio-economic status (SES) and diet quality indices (HEI and aMED): stratified by gender and ethnicity; CSFII/DHKS 1994–96*

*Smoothing of the curves was done through locally weighted regression models (LOWESS) with a bandwidth of 0.50.

Diet quality indicators and indices: the influence of PBFP and PBDQ

As for PBFP, placing weight on price over nutrition was positively associated with sodium and inversely related to fiber intake in the total population (P<0.05). There was only one significant interaction with gender in the case of fruit and vegetable intake (P<0.05 for interaction term PBFP×gender), although associations across gender were non-significant.

PBDQ on the other hand, was an independent protective factor in the total population for many diet quality indicators and indices, namely energy, total fat, saturated fat, fiber, fruits and vegetables, HEI and aMED. Gender acted as an effect modifier in the case of fat, fruits and vegetable intakes as well as HEI (P<0.05 for interaction term PBDQ×gender). In all cases, women had a higher quality diet than men with increased PBDQ score. Ethnicity did not act as an effect modifier of the relationship between PBDQ and DQI_j,k, independently of SES.

Testing direct and indirect effects of SES on diet quality: structural equation models

Table 4 presents the total, direct and indirect effects of SES on diet quality indicators and indices. The effects presented were age and sex-adjusted. The mediation proportion through PBFP varied between less than 1% for energy density, calcium and dairy products and over 10% for energy, total fat and sodium intakes as well as consumption of simple sugars. Hence, in the latter cases, PBFP had some mediation in the relationship between SES and DQI. As for PBDQ, mediation did not exist or was small with proportions ranging between 0.9% for fiber and 5.6% for simple sugars. Thus, a large proportion of the association between SES and dietary intakes could not be explained by the PBFP and PBDQ measures.

TABLE 4.

Total, direct and indirect effects¹ of SES on diet quality indicators and indices among US adults: CSFII/DHKS 1994–96

	→ PBFP2				→ PBDQ3
	Total	Direct	Indirect	Mediation Proportion (%)4	Total	Direct	Indirect	Mediation Proportion (%)4
Dietary Quality Indicators (DQIj)
Energy (kcal)	37.3	42.7	−5.4	14.5	37.1	38.5	−1.4	3.8
Energy density (kcal/100 g)	−1.21	−1.22	0.01	0.8	−1.21	−1.22	0.01	0.8
Total fat (% kcal)	−0.59	−0.53	−0.06	10.0	−0.59	−0.57	−0.02	3.4
Saturated fat (%kcal)	−0.26	−0.24	−0.02	7.7	−0.26	−0.25	−0.01	3.8
Cholesterol (mg)	−15.9	−15.6	−0.3	2.0	−15.9	−15.7	−0.2	1.2
Sodium (mg)	58.3	67.6	−9.4	16.1	58.1	60.1	−1.9	3.3
Fiber (g)	1.07	1.01	0.05	4.7	1.07	1.06	0.01	0.9
Calcium (mg)	40.45	40.28	0.16	0.4	40.56	40.48	0.08	0.2
Simple sugars (g)	0.72	0.80	−0.09	12.5	0.71	0.75	−0.04	5.6
Fruits and vegetables (g)	43.6	42.0	1.6	3.7	43.7	43.1	0.6	1.4
Dairy products (g)	13.9	14.0	−0.1	0.7	13.9	13.7	0.2	1.4
Dietary Quality Indices (DQIk)
HEI	2.69	2.60	0.09	3.3	2.69	2.65	0.04	1.5
aMED	0.38	0.37	0.01	2.6	0.38	0.38	0.00	0.0

Abbreviations: PBFP, Perceived Barrier of Food Price; PBDQ, Perceived Benefit of Diet Quality; SES, Socio-Economic Status.

¹Age and gender-adjusted effects in the total study population.

²→PBFP: Structural Equations Model (SEM) in which mediator is perceived barrier of food price (PBFP).

³→PBDQ: SEM in which perceived benefit of diet quality (PBDQ).

⁴ $\text{Mediation}\text{proportion}(\%)=\left|\frac{{\mathit{Effect}}_{\mathit{indirect}}}{{\mathit{Effect}}_{\mathit{total}}}\right|\times 100$

DISCUSSION

This study is one of few to assess the impact of SES on diet quality among US adults and test whether the association varies by ethnicity and gender, and is the first assess the effect of perceived barriers of food price (PBFP) and benefits of diet quality (PBDQ) on dietary intake, as well as conduct mediational analyses, using nationally representative data. Our study has several main findings. First, there are considerable ethnic and gender differentials in the associations between SES, PBFP, PBDQ, on one hand and dietary intake, on the other. Second, socio-economic constraints on individuals and households can lead to poorer diet quality. Third, PBFP appeared to increase sodium intake while reducing fiber intake, independently of SES. Fourth, PBDQ was directly related to better nutritional behavior, including lower energy, percentage of energy intake from fat and saturated fat and higher consumption of fiber, fruits and vegetables. Fifth, while SES and PBDQ independently improved HEI and aMED indices, PBFP which is inversely related to SES did not significantly affect them. Finally, PBFP but not PBDQ acted as a mediator in the relationship between SES and some diet quality indicators such as energy, fat and sodium intake.

Previous studies have indicated a positive relationship between SES and diet quality indicators and indices. For example, a cross-sectional study conducted in the Netherlands among 6,957 men and women aged 19 years or older over a period of 10 years and using 2 day dietary records showed that a higher SES was associated with higher intake of vegetable protein, dietary fiber and most micronutrients and a lower fat intake. However, differences between social classes were small after taking into account other factors such as energy and alcohol intake. Other studies found similar patterns in various adult populations (Hulshof et al., 2003, Groth et al., 2001, Bodnar and Siega-Riz, 2002, Friel et al., 2003). While all these studies focused on describing the relationship between SES and diet quality or dietary patterns, few tried researched mechanisms by which SES influenced dietary outcomes. A recent study conducted among 1,003 Australian adults suggested that socio-economic differences in dietary knowledge represented part of the pathway through which educational attainment exerts an influence on diet; and food purchasing differences by household income were related to diet in part via food-cost concern (Turrell and Kavanagh, 2006). While this study focused on food purchasing behavior, we looked at actual dietary intake among individuals. Our findings however differed in that PBFP was strongly influenced by SES in a way that its impact on dietary outcomes was not independent of social class. In addition, knowledge and attitude about diet and health as measured by PBDQ was not linearly associated with SES, unlike findings from other studies (Turrell and Kavanagh, 2006, Parmenter et al., 2000), which made this variable an independent protective factor for diet quality. Our structural equations modeling confirmed that mediation particularly for PBDQ was non-existent while PBFP acted as a mediator in particular for energy, fat and sodium intake as well as in the consumption of simple sugars (mediation proportion >10%).

Our study has several strengths in making contributions to the existing literature regarding the determinants of people’s eating behaviors. First, it extended findings from previous reports that showed a positive relationship between income, education and diet quality (Hulshof et al., 2003, Groth et al., 2001, Bodnar and Siega-Riz, 2002, Friel et al., 2003). Second, we made use of a large nationally representative dataset with a wealth of social, demographic, psychosocial and nutritional information. Third, we examined and found considerable ethnic and gender differentials in the associations between SES, perceived barriers of food choices among US adults. Moreover, we explored the complex mediation relationships between the study variables using structural equations modeling.

Despite its strengths, our study has its limitations. First, it is based on cross-sectional data, which generally, does not allow for ascertainment of temporality. However, socioeconomic position among adults is often determined in early adulthood and only very weakly affected by people’s present dietary intake. In addition, attitudinal variables are often antecedents to dietary behaviors. Second, while some of the structural equations modeling results are compelling, they should be interpreted with caution. In fact, these models are based on a number of assumptions, which include multivariate normality of measured variables and no effect modification by the mediating variable (Ditlevsen et al., 2005, Kaufman et al., 2005). Third, linear regression models assumed linearity in the association between our main exposure and outcome variables. To verify the linearity assumption, we re-ran analyses using tertiles of SES, PBFP and PBDQ instead of their continuous forms (for Tables 3 and 4). In some cases, the linearity assumption was not verified and a threshold effect was noted at the third tertile. A separate further study is needed to fully examine the complexity of the relations. Fourth, to test the ethnicity- and gender-differences in the effects of SES and perceived barriers of food choices, we examined whether there were overlap between the related 95% CIs in order to take into account of the complex sample design effects. Likely this approach is over-conservative. On the other hand, some researchers have argued that adjustment should be made (e.g., using smaller p values) when making multiple comparisons (Bender and Lange, 2001), which we chose not to perform, but others disagree as to its importance and its use of a frequentist approach (Perneger, 1999, Rothman, 1990, Greenland and Robins, 1991). Finally, there is a multitude of ways by which one can measure the mediating factors as well as SES and the outcome. For this reason, it is possible that measuring them differently may have yielded different results. Finally, we only contrasted Whites to African-Americans due to the relatively small sample sizes of other ethnic groups.

Our findings have several policy implications in light of the contextual setting. First, because of significant differences in the United States in the costs of healthy and unhealthy foods (Drewnowski and Darmon, 2005a, Drewnowski and Darmon, 2005b), our findings suggest that low SES may cause a significant food cost barrier which in turn reduces the quality of the diet, particularly in terms of energy, fat, sodium and simple sugars. Hence, perceptions go in line with actual reality and it is therefore important to make healthy foods accessible to the poor segments of the adult population by lowering their price and increasing the price of the unhealthy foods. We may also conclude that promoting positive attitudes towards benefits of healthy diets can be effective in improving diet quality in the whole population for both genders and all ethnicities. On the other hand, increasing SES, which in some cases works through lowering perceived barrier of food price, can also improve diet quality, especially for Whites and among women. However, future studies, in particular based on longitudinal data, should try to uncover the complexity of the relationship between SES, nutritional knowledge, attitudes and perceptions and dietary behavior and choices.

APPENDIX A

1. Health Eating Index (HEI)

	Criteria1
Components	<50y	≥50y	Score
Grains	9.1 servings/d	7.4 servings/d	10; 1 point less for each 10% less than intake required for full score Range: 0–10
Vegetables	4.2 servings/d	3.5 servings/d	Same as above
Fruit	3.2 servings/d	2.5 servings/d	Same as above
Milk	2.0 servings/d	2.0 servings/d	Same as above
Meat	2.4 servings/d	2.2 servings/d	Same as above
Total fat	≤30% energy	≤30% energy	10
	31–44% of energy	31–44% of energy	5
	≥45% of energy	≥45% of energy	0
Saturated fat	≤10% of energy	≤10% of energy	10
	11–14% of energy	11–14% of energy	5
	≥15% of energy	≥15% of energy	0
Cholesterol	<300 mg	<300 mg	10
	301–449 mg	301–449 mg	5
	≥450 mg	≥450 mg	0
Sodium	≤2,400 mg	≤2,400 mg	10; 1 point less for each 10% less intake required for full score
Variety	Top 10% intake of sum of unique foods	Top 10% intake of sum of unique foods	Same as above

¹Based on 2,200 kcal for the <50 y category and 1,900 kcal for the ≥51 y category.

Source: (McCullough et al., 2000, Fung et al., 2005)

2. Alternate Mediterranean Diet Score (aMED)¹

Food groups	Criteria for 1 point
Vegetables	Greater than median intake (servings/d)
Legumes	Greater than median intake (servings/d)
Fruit	Greater than median intake (servings/d)
Nuts	Greater than median intake (servings/d)
Whole grains	Greater than median intake (servings/d)
Red and processed meat	Less than median intake (servings/d)
Fish	Greater than median intake (servings/d)
Ratio of monounsaturated to saturated fatty acids	Greater than median intake (servings/d)
Ethanol	5–25 g/d

Sources: (Trichopoulou et al., 2003, Trichopoulou et al., 1995, Fung et al., 2005)

¹Scores may range between 0 and 10.