A large literature has examined the association between individual-level socio-economic position (SEP) and diet. The findings of this work typically show that socio-economically disadvantaged groups have diets that are least consistent with recommended intakes of foods and nutrients(Reference Metcalf, Scragg and Davis1–Reference De Irala-Estevez, Groth, Johansson, Oltersdorf, Prattala and Martinez-Gonzalez3) and least in accord with dietary guideline messages promoting foods that are high in fibre and low in fat, salt and sugar(Reference Turrell4, Reference Turrell, Hewitt, Patterson, Oldenburg and Gould5). Significantly, the poorer dietary intake of disadvantaged groups contributes in part to their higher rates of mortality and morbidity for chronic disease(Reference James, Nelson and Leather6, Reference Davey Smith and Brunner7).
During the last decade, researchers have increasingly turned their attention to the question of whether place of residence influences diet independently of individual-level factors; and more particularly, whether living in a socio-economically disadvantaged area is associated with a less healthy diet. Our review of this (small) literature suggested that area studies of diet can be broadly divided into two types that reflect the analytical method used: (i) studies that undertake multivariable analyses using both area- and individual-level variables but without the capacity to statistically integrate the two levels (i.e. contextual studies); and (ii) multilevel studies. Six of the former types of study(Reference Diehr, Koepsell, Cheadle, Psaty, Wagner and Curry8–Reference Shohaimi, Welch, Bingham, Luben, Day, Wareham and Khaw13) and five of the latter were identified(Reference Diez-Roux, Nieto, Caulfield, Tyroler, Watson and Szklo14–Reference Giskes, Turrell, van Lenthe, Brug and Mackenbach18), and key aspects of each are summarised in Table 1.
The findings of the contextual studies were reasonably consistent in that they each found some evidence that living in a disadvantaged area was associated with a poorer diet after adjusting for individual-level socio-economic and demographic factors. However, these studies were often based on a small number of areas and, in most cases, the analytical approach did not allow for the partitioning of area- and individual-level sources of variation. Hence it was unclear whether differences in diet between advantaged and disadvantaged areas were due to a composition effect (i.e. the clustering of rich and poor people in rich and poor areas) or the environmental characteristics of the areas per se (i.e. a context effect, possibly reflecting area differences in physical infrastructure, services and facilities). The findings of the multilevel studies, which allow for area- and individual-level variation to be partitioned and quantified, present a somewhat different picture. Of the five identified, only two reported a significant difference in diet between areas after adjustment for individual-level factors(Reference Ecob and Macintyre15, Reference Ball, Crawford and Mishra17). Diez-Roux et al.(Reference Diez-Roux, Nieto, Caulfield, Tyroler, Watson and Szklo14) and Ecob and Macintyre(Reference Ecob and Macintyre15) found that residents of socio-economically disadvantaged areas had poorer diets than those in more advantaged areas, although the findings of the former were weak and often not statistically significant. Area socio-economic status (SES) was not associated with food purchasing behaviour in an Australian study(Reference Turrell, Blakely, Patterson and Oldenburg16) or with dietary intake and food choice in a Dutch study(Reference Giskes, Turrell, van Lenthe, Brug and Mackenbach18).
The present paper contributes to the literature on areas and diet by reporting on a multilevel study that examined the association between area SES and food purchasing behaviour in the Melbourne metropolitan region (Australia) in 2003. The relationship between SES and diet in Australia (and elsewhere) has to date been investigated primarily using ‘quantitative’ dietary indicators such as mean daily intakes of nutrients, nutrient density levels or percentage contribution of food to nutrition and energy(Reference Turrell, Hewitt, Patterson, Oldenburg and Gould5). By contrast, few studies have examined the relationship using ‘qualitative’ indicators such as food purchasing behaviour. Clearly, people need to procure food (which usually means purchase it) before it can be consumed and converted into energy and nutrients, and there are a number of compelling reasons why it is important to better understand the factors that influence the food purchasing choices of different socio-economic groups. First, most people make dietary decisions in relation to food and not nutrients(Reference Crotty, Rutishauser and Cahill19); thus when shopping food choices are more likely to be influenced by factors such as price, availability, taste preference and convenience than by the vitamin and mineral content of the food. Second, research has shown that the type of food people buy influences dietary quality(Reference Shimakawa, Sorlie, Carpenter, Dennis, Tell, Watson and Williams20). Third, food choice differences between socio-economic groups translate into concomitant differences in nutrient intake(Reference Smith and Baghurst21, Reference Syrette, Baghurst and Record22). Fourth, knowing about the factors that influence socio-economic differences in food purchasing is important in assessing the reach and impact of health promotion messages, many of which focus on encouraging people to make healthy food choices when shopping(Reference Dixon, Borland and Segan23–25).
The study investigates whether residents of socio-economically advantaged and disadvantaged areas differ in their purchase of grocery foods, fruits and vegetables. Specifically, three questions are examined:
1. Do areas vary in their food purchasing profiles?
2. To what extent does within-area clustering by individual-level SEP account for any observed differences between areas in their food purchasing profiles?
3. What is the relationship between area SES and food purchasing after adjustment for within-area differences in food purchasing by individual-level SEP?
The present paper is based on data collected as part of the Victorian Lifestyle and Neighbourhood Environment Study (VicLANES), a cross-sectional multilevel investigation of area- and individual-level factors and health-related behaviour. The target population for VicLANES comprised people living in an area extending 20 km from the central business district of Melbourne city, the capital of the state of Victoria.
The sample comprised non-institutionalised residents of private dwellings (households) and Census Collector Districts (CCD). A CCD is the smallest administrative unit used by the Australian Bureau of Statistics to collect census data. In urban areas such as Melbourne, a CCD contains an average of 200 private dwellings which are deemed to be relatively homogeneous in terms of their socio-economic characteristics. Households and CCD were selected using a stratified two-stage cluster design. At the first stage, all CCD in the Melbourne metropolitan area were ranked according to the proportion of households in each CCD with an income of less than $AUS 400 per week. The resultant distribution was stratified into septiles, and a total of fifty CCD were randomly selected from the low- (n 17), middle- (n 16) and high-income (n 17) strata. At the second stage, we used names and addresses on the Australian Electoral Roll to identify all residents aged 18–74 years in each of the fifty CCD. Voting is compulsory in Australia for persons aged 18 years and over, so the electoral roll provides near-complete coverage of the resident adult population. A total of 3995 households were then randomly sampled, and the person within each household who was primarily responsible for most of the food shopping was targeted for data collection.
The household-level data collection within each CCD occurred between September and December 2003, and was conducted using a mail-survey method described by Dillman(Reference Dillman26). A total of 2564 usable surveys were returned to yield a final response rate of 64·2 %.
Area socio-economic status
The septiles forming the sampling strata were used as the basis for measuring area SES. In each of the three strata the average proportion of households earning less than $AUS 400 per week was 7·0 % (range 3·5–8·5 %), 15·3 % (14·4–16·7 %) and 31·0 % (24·1–59·6 %), respectively; these strata were subsequently labelled as high, medium and low SES. The area-level socio-economic characteristics of the three strata were further examined using 2001 census data(27), and they differed markedly in terms of their unemployment rate (4·0 %, 6·6 % and 11·0 %, respectively), the percentage of employees in unskilled and semi-skilled jobs (7·1 %, 13·8 % and 20·7 %, respectively), the percentage of dwellings that were rented from the public housing authority (0·17 %, 1·6 % and 14·5 %, respectively) and the percentage of dwellings with no motor vehicle (3·9 %, 9·6 % and 21·2 %, respectively).
Respondents were asked to provide information about whether they had attained further education since leaving school and, if so, the highest qualification completed. Respondent’s education was subsequently coded as: (i) bachelor degree or higher (the latter included postgraduate diploma, master’s degree or doctorate); (ii) diploma (associate or undergraduate); (iii) vocational (trade or business certificate, or apprenticeship); and (iv) no post-school qualifications.
Respondents who were employed at the time of completing the survey were asked to indicate their job title and then to describe the main tasks or duties they performed. This information was subsequently coded to the Australian Standard Classification of Occupations (ASCO)(28). For the purposes of the present study, the original nine-level ASCO classification was re-coded into three categories: (i) managers/professionals (managers and administrators, professionals and para-professionals); (ii) white-collar employees (clerks, salespersons and personal service workers); and (iii) blue-collar employees (trades-persons, plant and machine operators and drivers, labourers and related workers). A fourth category, ‘not in the labour force’, comprising the retired, unemployed, students and those engaged in home duties on a full-time basis, was also created.
Respondents were asked to indicate their total annual household income (including pensions, allowances and investments) using a fourteen-category measure that was subsequently re-coded into five groups for analysis: (i) $AUS 78 000 or more; (ii) $AUS 52 000–77 999; (iii) $AUS 36 400–51 999; (iv) $AUS 20 800–36 399; and (v) less than $AUS 20 799. Households in categories (iv) and (v) received annual incomes at or below the Australian average in 2000(29).
Age in years (centred), sex and household composition were used as potential confounding variables.
As part of the questionnaire, information was sought about the purchase of grocery items, fruits and vegetables.
Grocery food purchase. This was examined on the basis of fifteen questions, each of which had two or more response options. For example, respondents were asked ‘When shopping for your household, what type of milk do you usually buy?’ The response options included: ‘I do not buy milk’, ‘extra creamy’, ‘full cream’, ‘low-fat/trim’, ‘skimmed/fat-free’, plus others. Multiple responses were permitted for each question. The other fourteen questions were structured in an identical manner and pertained to bread, rice, pasta, noodles, baked beans, tinned fruit, cheese, yoghurt, beef mince, chicken, tinned fish, cooking oils, butter and solid cooking fat. In Australia, dietary authorities recommend that people purchase and consume a variety of foods that are relatively high in fibre and low in fat, salt and sugar(25); consistent with these guidelines, we classified the foods into a recommended and regular category (Table 2). Using this classification, we developed a measure that captured the extent to which peoples’ grocery purchasing patterns were consistent (or not) with dietary guideline recommendations. First, for each food type (e.g. milk), respondents were assigned the value 1 if they reported usually purchasing only the regular option exclusively (and not any recommended options); they were assigned the value 3 if they reported usually purchasing only the recommended option exclusively (and not any regular options); and they were assigned a value of 2 if they reported usually purchasing a mix of regular and recommended options (e.g. full cream and skimmed milk). There was a small number of respondents who reported that they never purchased a particular type of food and these were assigned the value 0. In sum, for each of the fifteen food types, respondents were assigned a value of 0, 1, 2 or 3. Second, an initial food purchasing index was created that involved summing the scores for the fifteen food types, with those scoring 0 being excluded at this point. This initial index had a potential range of 15–45, with 15 denoting people who purchased the regular option for each food type and 45 denoting those who purchased the recommended option for all foods. It is important to note that the respondents included in this initial index reported purchasing all of the fifteen food types. Those scoring 0 for one or more food type(s) were excluded because their final index score would not accurately reflect their purchasing pattern. For example, someone who purchased all fifteen food types and chose the recommended option for each item would score 45, whereas someone who purchased thirteen food types and chose the recommended option for each item would score 39. Clearly, both people have identical purchasing patterns with respect to the dietary guidelines (i.e. they are making the healthier choice for every food item) but this is not reflected in their index score. To deal with this issue, and as a way of including the full sample in the analyses, respondents who reported not buying one or more of the food items were included in the index using the following formula: Index score=a/(15−b). The quantity a represented each respondent’s initial score which was derived by summing the values (1, 2 or 3) for each of the food types. The denominator comprised the constant ‘15’, which represented the number of food types in the index, and the variable b, which represented the number of food types not purchased by the respondent. In effect, the formula calculated a mean food purchasing score for each respondent. Finally, the index was re-scored to range from 0 to 100, with higher scores indicating a purchasing pattern that was more consistent with dietary guideline recommendations (sample mean 47·6, sd 13·4).
Fruit purchasing. This was examined using a question that asked ‘When shopping for fresh fruit, how often do you buy these types?’ The respondent was instructed to include seasonal fruits, but exclude fruit juice, canned and dried fruit. The question item-set consisted of twenty-two fresh fruits selected (mostly) from the FFQ used in the 1995 Australian National Nutrition Survey(Reference McLennan30). For each fruit, respondents were asked to indicate their usual purchasing pattern on the basis of five-point scales: 1 = ‘never buy’, 2 = ‘rarely buy’, 3 = ‘sometimes buy’, 4 = ‘nearly always buy’ and 5 = ‘always buy’.
Using these items we created an index that measured variety of fruit purchased. For each fruit item, respondents reporting ‘never’ or ‘rarely’ buy were scored 0, and those reporting any of the other three options were scored 1. The items were then summed, with the resultant index score for each respondent indicating the variety of fruits purchased (sample mean 14·2, sd 4·1). Importantly, the variety score does not reflect the range of fruits purchased on any particular shopping trip, but rather the types that are purchased at least sometimes over the course of many shopping episodes depending on factors such as seasonality, price and quality. As the variety index was essentially a count-measure and non-normally distributed it was categorised into quartiles, with Q1 denoting high variety and Q4 low variety.
Vegetable purchasing. Respondents were asked to indicate how often they purchased twenty-five vegetables, including fresh and frozen, but excluding canned or dried vegetables. A purchasing index measuring vegetable variety was constructed using an identical format and method to that used for fresh fruit. The mean variety score for vegetables for the sample was 18·5 (sd 4·1).
Table 3 presents descriptive statistics for each of the measures used in this analysis.
From the 2564 questionnaires that were returned, missing data were identified for education (n 106, 4·1 %), occupation (n 83, 3·2 %), income (n 903, 35·2 %), sex (n 4, 0·16 %), age (n 5, 0·20 %) and household composition (n 55, 2·1 %). In total, the proportion of the sample with completely observed data for all the variables examined (complete cases) was 57 %. We have not reported results obtained by analysing only the complete cases because of the potential bias and loss of precision associated with the large proportion of missing income data; instead, we used multiple imputation. We imputed all missing data under a missing at random (MAR) assumption and adopted an inclusive strategy for the imputation model(Reference Collins, Schafer and Kam31–Reference Meng33). Five data sets with imputed values for missing items on each variable were estimated using the command ‘Imputation by Chained Equations (ICE)’ in the STATA statistical software package version 9·2 (Stata Corporation, College Station, TX, USA).
The grocery data were analysed as a two-level random intercept model in STATA. We specified three models that directly addressed the three research questions identified earlier. Model 1 (baseline) quantified the extent of area-level variation in food purchasing behaviour conditional on the confounders. Here, the substantive interest was on the random term which, if significant, indicated that food purchasing patterns differed between the fifty CCD. For this and subsequent models we also calculated an intra-class correlation (ICC) by dividing the between-CCD variance by the total variance, and this is interpreted as the proportion of the total variation in food purchasing behaviour that is between the CCD. Model 2 extends Model 1 by adding education, occupation and income as fixed effects, and examined the extent to which they account for variation in food purchasing between the CCD. Model 3 then extended Model 2 by including the measure of area SES as a fixed effect; here the focus is on whether area SES is associated with food purchasing independently of within-area variation in age, sex, household composition and individual-level SEP.
Variety of fruit and vegetable purchase was examined using a two-level ordered multinomial logit-link model. ‘High’ variety (Q1) was denoted the reference category; hence positive regression coefficients for any of the predictor variables indicate a greater odds of purchasing a lower variety of fruits and vegetables. Three models were specified. Model 1 (baseline) quantified the extent of area-level variation in fruit and vegetable variety conditional on the confounders. Model 2 added education, occupation and income, and Model 3 included area SES. The results are presented as odds ratios and their 95 % confidence intervals.
Table 4 presents the findings of the multilevel analyses which examined the independent contribution of area- and individual-level socio-economic factors to grocery food purchase. In Model 1, the area-level random term was statistically significant (P = 0·033), indicating that the average grocery purchasing score was different (beyond chance) across the fifty CCD. Of the total variability in grocery purchase, 1·5 % occurred between CCD and 98·5 % between individuals. Model 2 adds the fixed (average) effects for education, occupation and income; this attenuated the between-area variation by 59·8 %, and the random term was no longer significant (P = 0·241). Education and income were associated with grocery purchase: respondents with no post-school qualifications and those living in low-income households scored significantly lower on the index. No significant occupational effects were observed. Model 3 adds the fixed effect for area SES and the coefficients indicate that residents of medium- and low-SES areas scored significantly lower on the grocery purchasing index than their counterparts from high-SES areas.
SES, socio-economic status.
Effect was significant: *P ≤ 0·05, **P ≤ 0·01, ***P ≤ 0·001.
†Model 1, baseline model adjusted for age, sex and household composition; Model 2, Model 1 plus education, occupation and income; Model 3, Model 2 plus area SES.
Table 5 presents the findings of the ordered multilevel logistic regression analysis which examined the contribution of area- and individual-level socio-economic factors to variety of fruit and vegetable purchasing. Fruit variety scores were significantly different (P = 0·01) across the fifty CCD (Model 1). After adjustment for education, occupation and income (Model 2), the between-area variation in fruit variety was attenuated by 50·0 % and remained marginally statistically significant (P = 0·06). Respondents with no post-school qualifications had 1·72 (95 % CI 1·25, 2·38) times higher odds of purchasing a lower variety of fruits. The corresponding odds for respondents from low-income families were 1·69 (95 % CI 1·11, 2·57). Model 3 adds the measure of area SES which made no appreciable difference to the between-CCD variation (relative to Model 2) although the random term was no longer statistically significant (P = 0·11). The coefficients for area SES show that residents of low-SES areas had significantly higher odds of purchasing a lower variety of fruits than residents in the high-SES areas (OR = 1·30, 95 % CI 1·00, 1·67). Independent of area SES, respondents with lower levels of education, and residents of lower-income households, had significantly higher odds of purchasing a more limited variety of fruits than their higher status counterparts.
SES, socio-economic status.
† High variety (quartile 1) was denoted the reference category; hence odds ratios greater than 1 indicate an increased likelihood of purchasing a lower variety of fruits and vegetables.
‡ Model 1, baseline model adjusted for age, sex and household composition; Model 2, Model 1 plus education, occupation and income; Model 3, Model 2 plus area SES.
Vegetable variety scores did not differ significantly across the fifty CCD (Model 1) and the inclusion of education, occupation and income further attenuated the CCD variation (Model 2). Respondents with no post-school qualifications had a significantly higher odds of purchasing a lower variety of vegetables relative to those with a bachelor degree (OR = 1·36, 95 % CI 1·08, 1·72). There was no association between vegetable variety and occupation, income or area SES (Model 3).
In metropolitan Melbourne in 2003, area SES was associated with the purchase of grocery foods and fruit variety. Compared with their counterparts in high-SES areas, residents of low-SES areas were less likely to buy groceries that were high in fibre and low in fat, salt and sugar; and they purchased a smaller variety of fruits. These findings are broadly consistent with the results of multilevel studies conducted in the USA(Reference Diez-Roux, Nieto, Caulfield, Tyroler, Watson and Szklo14) and Scotland(Reference Ecob and Macintyre15); however, they are at odds with multilevel research conducted in The Netherlands(Reference Giskes, Turrell, van Lenthe, Brug and Mackenbach18) and Brisbane, Australia(Reference Turrell, Blakely, Patterson and Oldenburg16). Reconciling these differences, and hence being able to generalise about the relationship between area SES and diet, is difficult. In part, these difficulties stem from the limited evidence base (i.e. the small number of multilevel studies) and methodological issues such as differences in the conceptualisation and measurement of diet, the individual-level variables used as confounders, and the number and size of the area units used(Reference Ecob and Macintyre15). The inconsistencies between study findings, however, are likely to be more than a methodological artefact, and may reflect ‘real’ historical, cultural, political, socio-economic and geospatial differences between countries (e.g. USA and Australia) and between regions within the same country (e.g. Brisbane and Melbourne). At present, the mixed findings of the small number of multilevel studies do not provide a sufficiently reliable basis on which to make a general call for area-level public health interventions to improve conditions in deprived areas to facilitate the procurement of foods that are conducive to a healthy diet; rather, any ‘call’ may have to be specific and tailored to each particular geographic and spatial context.
A large literature documents an association between individual-level SEP and diet, and most of this work has focused on socio-economic differences in food and nutrient intakes(Reference Turrell, Hewitt, Patterson, Oldenburg and Gould5). These studies usually find that socio-economically disadvantaged groups have intakes that are consistent with their higher rates of diet-related chronic disease(Reference Metcalf, Scragg and Davis1–Reference De Irala-Estevez, Groth, Johansson, Oltersdorf, Prattala and Martinez-Gonzalez3). To some extent at least, the results of the present food purchasing study extend and complement the findings of the intake studies by showing that those of low SEP are less likely to buy grocery foods that accord with diet-related health promotion messages and dietary guidelines. In addition, low socio-economic groups had significantly higher odds of purchasing a lower variety of fruits and vegetables.
First, survey non-response tends to be higher in disadvantaged areas(Reference Turrell, Sanders, Slade, Marcenes and Spencer34) and among individuals of low SEP(Reference Turrell and Najman35). Non-response in the VicLANES study was 35·8 %; hence the sample probably under-represents the disadvantaged areas and individuals and over-represents the advantaged, and the observed socio-economic differences in food purchasing are likely to be an underestimate of the actual differences in the Melbourne population.
Second, as with most multilevel studies(Reference Boyle and Wilms36), our use of a CCD to represent a neighbourhood was made for reasons of sampling and analytic convenience rather than being underpinned by an explicit theory linking area SES and food purchasing; hence associations among these variables are likely to be underestimated.
Third, our finding of an association between area SES and food purchase might be confounded by individual-level socio-economic factors not included in the models. This said however, we included the three most widely used indicators of a person’s socio-economic characteristics(Reference Dutton, Turrell and Oldenburg37) and, given the correlation among these indicators(Reference Turrell, Hewitt, Patterson and Oldenburg38), it is likely that education, occupation and income were capturing most of the unmeasured influences of other socio-economic factors excluded from the models. Alternatively, it may be that the inclusion of these individual-level measures resulted in ‘over-adjustment’, which argues for the possibility of an even stronger contextual effect on food purchase than was observed here. If education, occupation and household income represent part of the pathway via which area SES influences food procurement, then modelling individual-level socio-economic variables may inappropriately attenuate the variation that is more correctly attributable to area disadvantage(Reference Diez Roux39).
In the Melbourne metropolitan region in 2003, differences between advantaged and disadvantaged areas in their purchasing profiles for grocery foods and fruits, and the ‘healthier’ purchasing in higher-SES areas, suggest that the areas may be differentiated on the basis of food availability, accessibility and affordability, making the purchase of some types of foods more difficult for people living in disadvantaged areas. To date, the between- and within-country (multilevel) evidence linking area disadvantage and diet is both sparse and inconsistent. Methodological issues notwithstanding, this might suggest that area deprivation is not universally associated with poorer access to healthy food. Cummins and Macintyre(Reference Cummins and Macintyre40) reached a somewhat similar conclusion based on their review of the literature on food environments and obesity. A challenge for future area-based dietary research is to identify those ecological characteristics (e.g. urban design, shopping infrastructure, transport services) that promote equality of access to healthy food, and those characteristics that make its attainment difficult.
The study was supported by a grant from the Victorian Health Promotion Foundation (VicHealth). G.T. is supported by a National Health and Medical Research Council (NHMRC) Senior Research Fellowship (No. 390109); R.B. and L.R.T. by an NHMRC Capacity Building Grant; and S.V.S. is supported by a National Institutes of Health Career Development Award (NHLBI K25 HL081275). There are no conflicts of interest. G.T. conceptualised the paper and played the lead role in writing the manuscript and reviewing the literature. R.B. contributed to the data analysis and imputation (revised submission) and to writing the Methods section. L.R.T. contributed to the data analysis and imputation (initial submission) and editing the manuscript. D.J. provided statistical advice, undertook preliminary analysis and contributed to writing and checking the manuscript. S.V.S. contributed to the conceptualisation of the analysis plan, provided statistical advice in relation to the multilevel modelling and edited the manuscript. A.M.K. contributed to the conceptualisation, analysis plan and writing. We gratefully acknowledge Mr Lukar Thornton for his assistance with the missing data imputation and Ms Tania King for her work as Project Manager on VicLANES.