Article Text

A multilevel analysis of socioeconomic (small area) differences in household food purchasing behaviour
1. G Turrell1,
2. T Blakely2,
3. C Patterson1,
4. B Oldenburg1
1. 1School of Public Health, Queensland University of Technology, Brisbane, Queensland, Australia
2. 2Department of Public Health, Wellington School of Medicine and Health Sciences, University of Otago, Wellington, New Zealand
1. Correspondence to:  Dr G Turrell  School of Public Health, Queensland University of Technology, Victoria Park Road, Kelvin Grove, Brisbane, Queensland, Australia, 4059; g.turrellqut.edu.au

## Abstract

Study objective: To examine the association between area and individual level socioeconomic status (SES) and food purchasing behaviour.

Design: The sample comprised 1000 households and 50 small areas. Data were collected by face to face interview (66.4% response rate). SES was measured using a composite area index of disadvantage (mean 1026.8, SD = 95.2) and household income. Purchasing behaviour was scored as continuous indices ranging from 0 to 100 for three food types: fruits (mean 50.5, SD = 17.8), vegetables (61.8, 15.2), and grocery items (51.4, 17.6), with higher scores indicating purchasing patterns more consistent with dietary guideline recommendations.

Setting: Brisbane, Australia, 2000.

Participants: Persons responsible for their household’s food purchasing.

Main results: Controlling for age, gender, and household income, a two standard deviation increase on the area SES measure was associated with a 2.01 unit increase on the fruit purchasing index (95% CI −0.49 to 4.50). The corresponding associations for vegetables and grocery foods were 0.60 (−1.36 to 2.56) and 0.94 (−1.35 to 3.23). Before controlling for household income, significant area level differences were found for each food, suggesting that clustering of household income within areas (a composition effect) accounted for the purchasing variability between them.

Conclusions: Living in a socioeconomically advantaged area was associated with a tendency to purchase healthier food, however, the association was small in magnitude and the 95% CI for area SES included the null. Although urban areas in Brisbane are differentiated on the basis of their socioeconomic characteristics, it seems unlikely that where you live shapes your procurement of food over and above your personal characteristics.

• socioeconomic status
• diet
• health inequalities
• BFS, Brisbane food study
• SSD, statistical sub-division
• CCD, census collectors district
• IRSD, index of relative socioeconomic disadvantage
• ABS, Australian Bureau of Statistics
• SES, socioeconomic status

## Statistics from Altmetric.com

A large literature shows that socioeconomic groups differ in their rates of mortality and morbidity for cardiovascular disease, type 2 diabetes, and many cancers, with the socioeconomically disadvantaged experiencing the poorest health.1,2 Diet plays a part in the onset and progression of these degenerative conditions,3–5 and it is increasingly believed that dietary differences between socioeconomic groups contribute in part to their different health profiles for chronic disease.6,7

Most studies investigating the relation between socioeconomic status (SES) and diet have focused on individual level factors. Sampled individuals are grouped on the basis of similar socioeconomic characteristics such as occupation, education, or income, and these groupings are compared in terms of their dietary behaviours or food and nutrient intakes. Studies of this type often show that socioeconomically disadvantaged groups are least likely to engage in behaviours that accord with healthy eating messages,8,9 and they are more likely to have food and nutrient intake profiles that parallel their higher rates of diet related disease.10,11

During the past decade, researchers have increasingly called for a greater focus on the potential contribution of environments and places in terms of shaping and circumscribing the health related behaviour of people.12–14 It is argued that an improved understanding of the determinants of behaviour, and by extension, more effective approaches to advancing health, will necessarily require studies that consider the person, their context or setting (for example, neighbourhood, work, family), and interactions between these. Dietary studies of this type have been conducted in Britain,15–19 Finland,20,21 and the USA.22–24 Despite differences in analytical method, and heterogeneity of area unit, sample size, or how diet was measured, each study found evidence that area characteristics might influence diet independent of individual level characteristics. The findings of some of these studies, however, are challengeable, as they were based on statistical methods that did not allow for the partitioning of area and individual level sources of variation (that is, between contextual and compositional effects). Less open to challenge are the findings of multilevel studies, which do allow for this partitioning, and of the few that have examined area variations in diet, each has provided suggestive evidence that both individual and contextual factors separately influence diet.19,23,25 Specifically, these studies showed that residents of socioeconomically disadvantaged areas had poorer dietary intakes after adjusting for individual level SES, suggesting that unmeasured features of the wider social and physical environment in disadvantaged areas acted to hinder the procurement and consumption of a healthy diet.

British and US researchers have identified a number of possible explanations for dietary differences between urban areas that vary in their socioeconomic characteristics. Firstly, some socioeconomically disadvantaged areas are underserved by large supermarkets,12,25,26 and as a result, residents are disproportionately reliant on smaller shops, which typically stock a limited range of foods, their prices are higher, and fresh food is often of a lesser quality. Secondly, socioeconomically disadvantaged groups sometimes experience difficulties accessing large (and often distant) shopping facilities because they lack private transport, or live in areas where public transport is inadequate or non-existent,27–29 which also increases the likelihood that a greater amount of food is purchased from smaller local shops. Thirdly, healthy foods (that is, those consistent with dietary guideline recommendations) have been found to be less readily available in shops located in socioeconomically disadvantaged areas, and also more expensive than their less healthful equivalents.30,31

In this paper, we add to the international evidence base about context effects on diet by examining small area variations in food purchasing behaviour among residents of Brisbane City, Australia. Specifically, we use multilevel modelling to determine whether there is variation between socioeconomically different areas in the purchase of fruits, vegetables, and grocery foods after controlling for personal and household sociodemographic characteristics. Significant area level variation independent of individual and household level factors would raise the possibility that urban regions in Australia are differentiated on the basis of food availability, accessibility, and affordability, making the procurement of healthy food difficult for socioeconomically disadvantaged groups. In the US and Britain two societal level processes have probably contributed to area variations in diet. Firstly, these countries have witnessed markedly increased spatial segregation of their populations along social and economic lines.32–34 Secondly, this increasing socioeconomic polarisation appears to have been accompanied by concomitant changes to the structure and organisation of the food retail industry, such that supermarkets and large stores have disinvested in, and relocated from urban disadvantaged areas to regions characterised by large population size and density, higher average incomes, and reduced operating costs.35,36 While urban areas in Australia are also socially and economically segregated,37 the nature and extent of this separation appears qualitatively different (that is, less extreme) than that observed elsewhere. In addition, this country has not seemingly undergone similar changes to the food retailing industry. As a result, it remains an open question whether or not urban areas in Australia are differentiated in their dietary behaviours in ways that are found in the US and Britain.38

## METHODS

The data were collected as part of the 2000 Brisbane food study (BFS). Details of the study’s scope and coverage, its research design, sampling procedures, data collection methods, and representativeness have been published elsewhere.39 Only a brief overview is provided here.

### Sample design

The BFS was conducted in the Brisbane City Statistical Sub-Division (SSD). The sample comprised 1000 households and 50 census collectors districts (CCD), and was selected using a stratified two stage cluster design. A CCD is the smallest administrative unit used by the Australian Bureau of Statistics (ABS) to collect census data. As at 1996, the Brisbane SSD consisted of 1517 contiguous CCD, each containing an average of 200 occupied private dwellings. Stratification consisted of ranking the CCD on the basis of each area’s index of relative socioeconomic disadvantage (IRSD) score. A CCD’s IRSD score is derived by the ABS using principal components analysis, and it reflects the overall level of socioeconomic disadvantage of each area measured on the basis of attributes such as low income, low educational attainment, high levels of public sector housing, high unemployment, and jobs in relatively unskilled occupations.40 The IRSD scores used in this study were calculated from data collected in the 1996 Australian census. The distribution of IRSD scores was subsequently divided into 10 strata (deciles) and five CCD were selected from each of the strata using systematic without replacement probability proportional to size sampling. The spatial and socioeconomic characteristics of the 50 CCD are presented in figure 1. As would be predicted from the stratification process, the sampled CCD differed markedly on all key socioeconomic indicators.

Figure 1

Sampled census collectors districts (CCDs) in the Brisbane statistical sub-division and their socioeconomic characteristiscs.

Stage 2 involved selecting 1000 private dwellings from the 50 CCD (20 dwellings on average per CCD), and this was undertaken using simple random sampling. Given the focus of the study, we interviewed the person within each dwelling who was primarily responsible for most of the food shopping. A final response rate of 66.4% was achieved.39

The individual level data collection within each CCD occurred between September and December 2000, and was conducted on the basis of face to face interviews. Interviews lasted an average of one hour, and respondents were offered a small financial gratuity (AUS$10). The interview sought information on food purchasing choices, factors influencing choice, shopping practices, subjective perceptions of food availability and food prices, food expenditure, food and nutrition knowledge, and the sociodemographic characteristics of the respondent and other household head (if a couple household). Although the data were collected from a single individual, the interview questions elicited information about food purchasing patterns for the household as a whole. ### Measures Area level SES for each CCD was measured using its IRSD score (see above). Individual level SES was measured by the study participant’s estimate of total household income (including pensions, allowances, and investments) collected as a 14 category variable and subsequently re-coded into four categories for analysis: (1) less than AUS$20 799, (2) $20 800–36 399, (3)$36 400–51 999, and (4) \$52, 000 or more. Households in categories 1 and 2 received incomes at or below the Australian average as at 2000, and those in categories 3 and 4, above the average.41 Household income was used as the socioeconomic indicator for three reasons. Firstly, income is a well established and important determinant of dietary quality, and affects directly a family’s ability to afford and procure food.42 Secondly, household income was likely to capture the socioeconomic characteristics of all people living in the household (reflecting individual level incomes, and to some extent education and occupation) and therefore presumably embodied most of the within household socioeconomic processes influencing food choice. Thirdly, it seemed appropriate (substantively and analytically) to examine the relation between SES and food purchasing using variables that were each measured at the same level (that is, household), thus improving model specificity and fit.

Foods purchased for each household were classified into two broad groups: grocery items (including meat and chicken), and fruit and vegetables. Grocery purchasing was examined on the basis of 16 questions, each of which had two or more response categories. For example, respondents were asked: “When you go shopping, what type of bread do you usually buy?” The response options included: I do not buy bread, white, wholemeal, multigrain, white high in fibre, rye, soy and linseed, plus others. Multiple responses were permitted for each question. The other 15 questions were structured in an identical manner and pertained to rice, pasta, baked beans, fruit juice, tinned fruit, milk, cheese, yoghurt, beef mince, chicken, tinned fish, vegetable oil, margarine, butter, and solid cooking fat. In Australia, health promotion and education campaigns43 directed at disseminating dietary guideline messages44 recommend that people purchase and consume a variety of nutritious foods that are comparatively high in fibre, and low in fat, salt, and sugar. In keeping with these campaigns, we classified respondents’ food purchasing choices into a “recommended” and “regular” category (table 1).

Table 1

Classification of grocery food types into “recommended” and “regular” categories*

Purchasing patterns for each grocery food type were then scored as follows. Respondents were categorised as never purchasing the food (scored 0), as purchasing the regular option exclusively (scored 1), as purchasing a variety of food that included both the recommended and regular options (scored 2), or as purchasing the recommended option exclusively (scored 3). The food types were then summed to form a purchasing index, and using an approach described elsewhere,8,45 the index scores were adjusted to account for the fact that some people did not purchase particular foods. This index was then scaled to range from range from 0 to 100, with high scores being indicative of greater compliance with dietary guideline recommendations.

Fruit purchasing information was elicited 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, tinned fruit, and dried fruit. The question item set consisted of 19 fruits selected from the food frequency questionnaire used in the 1995 Australian National Nutrition Survey.46 Respondents were asked to indicate their usual fruit purchasing pattern on the basis of five point scales that ranged from never buy (scored 0) to always buy (scored 4). A fruit purchasing index was created by summing the items, and scoring the measure to range from 0 to 100. Higher scores indicated that respondents regularly purchased many different types of fruits when shopping for their household (that is, a high score was obtained by reporting “Always” or “Nearly always” for most of the fruits listed). In addition, high scores were consistent with two of the Australian Dietary Guideline recommendations, namely, “Eat a wide variety of nutritious foods”, and “Eat plenty of... vegetables (including legumes) and fruits”.44

Vegetable purchasing behaviour was measured using an identical format and method to that used for fruit. Respondents were asked to indicate how often they purchased 21 vegetables (including fresh and frozen, but excluding tinned or dried) using five point items. These were subsequently summed to form an index and re-scored to range from 0 to 100, with higher scores being interpreted in the same way that was outlined for fruit purchase.

### Analysis

Table 2 presents descriptive statistics for each of the measures used in this analysis. Of the 1000 households interviewed for the BFS, 24 declined to answer the income question, four did not know the income of other people in their household, and two provided insufficient information for their food purchasing behaviours to be reliably assessed. Each of these respondents was excluded, resulting in a final useable sample of 970.

Table 2

Descriptive statistics for the fixed effect variables and food purchasing indices

The data were analysed as a two level random intercept variance components model, using MLwiN version 2.1c.47 Three models were specified for each food purchasing behaviour. Firstly, a null model, comprising individuals (level 1) nested in CCDs (level 2) with no predictor variables in the fixed part of the model. Substantive interest for the null model focuses on the CCD level random term, which if significant (indicated using χ2), suggests between area variation in food purchasing behaviour. For the null (and all other) models the intraclass correlation was calculated to estimate the percentage of total variance in food purchasing behaviour that was between the CCD (the remaining percentage is between individual variation). The null model was subsequently extended to include fixed effects for age, sex, and household income (model 2) and then the exposure of interest in this study: area socioeconomic disadvantage (model 3). The effect size for the area SES variable was expressed as a two standard deviation increase in area SES, which is equivalent to the difference in area SES score between the median values for the top and bottom quartiles of the area SES index. Improvements in the fit of the three nested models due to the successive inclusion of the fixed effect variables were assessed using the deviance statistic.

## RESULTS

Table 3

Area and individual level effects on fruit purchasing (random intercept models)*

### Key points

• In the US and Britain, area level socioeconomic status is associated with food and nutrient intake and dietary behaviour independent of individual level socioeconomic characteristics

• Within Brisbane City, Australia, there is no convincing association between area level socioeconomic status and food purchasing behaviour

• Much of the apparent association of area socioeconomic status with food purchasing in Brisbane was attributable to confounding by household income, thus the clustering of household income within areas (a composition effect) accounted for the food purchasing variability between them

• Despite urban areas in Brisbane being differentiated in their socioeconomic characteristics, this does not seem to influence the procurement of healthy food, which is in contrast with that found in other countries

Tables 4 and 5 present the equivalent results for vegetable and grocery purchasing respectively. The null models for both vegetable and grocery purchase showed that no statistically significant variation was evident at the CCD level: vegetables (χ2 = 0.613, p = 0.433), grocery foods (χ2 = 0.581, p = 0.445). In other words, apart from non-systematic sampling fluctuations, there were no differences in the purchasing scores among the 50 areas. The inclusion of the fixed terms for age, sex, and household income (model 2) showed that these factors significantly improved the fit of each model (results for deviance tests not reported). For both vegetable and grocery purchasing, average index scores were significantly higher for older persons, women, and residents of high income households. Area SES was only weakly related with the purchase of vegetables and grocery foods (model 3). A two standard deviation increase on the area SES measure was associated with a 0.60 unit increase on the vegetable purchasing index (95% CI −1.36 to 2.56) and a 0.94 unit increase on the grocery index (95% CI −1.35 to 3.23). The inclusion of area SES produced no statistically significant improvement in the fit of the models for vegetable and grocery purchasing. For models that included area SES but not household income, a two standard deviation increase in area SES was associated with a 1.86 unit increase on the vegetable index (95% CI 0.00 to 3.73) and a 3.22 unit increase on the grocery purchasing index (95% CI 1.04 to 5.39).

Table 4

Area and individual level effects on vegetable purchasing (random intercept models)*

Table 5

Area and individual level effects on grocery purchasing (random intercept models)*

## DISCUSSION

Multilevel studies conducted in the US and Britain have found evidence in support of contextual or neighbourhood socioeconomic effects on diet independent of individual level factors.19,23,25 Typically, residents of socioeconomically disadvantaged areas have poorer diets than those in more advantaged areas. Our study in the Brisbane metropolitan region suggests that small area variation in the purchase of fruit, vegetables, and grocery foods mainly reflect spatial differences in the socioeconomic composition of the people living in the areas. Much of the apparent association of area SES with food purchasing was attributable to confounding by household income. After controlling for household income, and the age and sex of respondents, a two standard deviation increase on the area SES measure produced a modest increase of 2.01 units on the fruit purchasing index, with the 95% confidence limits including zero (−0.49 to 4.50), and very small unit increases for the vegetable (0.60, 95% CI −1.36 to 2.56) and grocery indices (0.94, 95% CI −1.35 to 3.23). A two standard deviation change in area SES was equivalent to the difference in score between the median values for the top and bottom quartiles of the area SES measure, enabling an approximate comparison with the effect sizes between the high and low categories of household income. This comparison shows that the area SES effect for fruit purchasing was only about 25% of the household income association, and about 10% of the income association for vegetable and grocery purchase.

Importantly, our results and conclusions about the likely limited effect of area SES on food purchasing behaviour in Brisbane needs to be considered against a number of study limitations. Firstly, (and with the benefit of hindsight), our study was seemingly under-powered to detect statistically significant contextual effects. This notwithstanding however, the association of area SES with each outcome variable was in the expected direction, thus while a larger study may have found statistically significant area effects due to increased precision, it is unlikely that a larger sample would have found a substantially increased strength of association between area SES and food purchasing behaviour. Secondly, we only controlled for one individual level socioeconomic factor as a potential confounder (that is, income), which argues against there being any true contextual effect. If we had controlled for other (potentially confounding) individual level socioeconomic factors such as occupation or education, then it is likely that the already weak to moderate area SES effect would have further reduced to the null. Thirdly, it is possible that our study was adversely influenced by selection or information bias, although we are uncertain of the probable magnitude and direction of this bias. As with most multilevel studies52,53 our areal units were selected for reasons of sampling and analytical convenience rather than for reasons that were hypothesised to influence food purchasing behaviour, and this would probably underestimate area SES associations. Furthermore, non-differential misclassification bias of food purchasing would probably result in an underestimate of the area SES association and the (confounding) income association. In short, the net effect of measurement error in our multilevel study (and multivariable models generally) is unclear.54–56 Fourthly, the inclusion of individual level covariates in multilevel analyses may result in over-control, which argues for the possibility of a true contextual effect on food purchasing behaviour in Brisbane. Household income, for example, may in part depend on where you live or on cumulative small area effects over the lifecourse. Given each of these limitations, the finding of no significant area SES effect needs to be viewed circumspectly, and further research in a variety of settings is required before more definitive conclusions can be reached.

There is now a large body of Australian and international research that has examined the relation between individual level SES and diet, with diet most often being measured on the basis of food and nutrient intake.8 These studies usually find that socioeconomically disadvantaged groups have intakes that are least in accord with minimal risk for the onset of chronic disease.10,11,57,58 The individual level results of the BFS adds to this research, and shows that low income households were less likely to purchase foods consistent with recommendations promulgated in diet related promotion messages. For each food type, purchasing score was graded across the income categories, suggesting a high degree of income sensitivity to the purchase of healthy food.

In summary, this first known Australian multilevel study of diet found little evidence that food purchasing behaviour in Brisbane was influenced by area level socioeconomic disadvantage. Thus despite the fact that major urban areas in this country are differentiated on the basis of their social and economic characteristics37 this does not seem to be sufficient to shape and circumscribe the procurement of food. It seems that what matters most in Brisbane City in terms of food purchasing behaviour is the socioeconomic characteristics of individuals and their households, rather than the socioeconomic characteristics of the areas in which they live. This Australian finding seems to be in contrast with countries like the US and Britain, where the nature and extent of spatial segregation along social and economic lines is large enough to be detectable in people’s dietary behaviour.

View Abstract

## Footnotes

• Funding: funding for this research was provided by a National Health and Medical Research Council Project Grant (no 101217). Dr Turrell is supported by a National Health and Medical Research Council/National Heart Foundation Career Development Award (CR 01B 0502). Dr Blakely is supported by funding from the New Zealand Health Research Council and Ministry of Health.

• Conflicts of interest: none declared.

## Request permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.