In this cross-sectional study, data from respondents in the CIVISANO project was used, a mixed-methods study investigating the role of objective and perceived environmental factors on physical activity and eating behavior among adults residing in two medium-sized peri-urban municipalities, Duffel and Herselt, in Flanders, Belgium [40]. The study was approved by the Medical Ethics Committee of Ghent University Hospital (BC-248 09260) and conducted in accordance with the recommendations of the Belgian Data Protection Authority. All the respondents provided informed consent.

The study population consisted of respondents to the questionnaire part of the project. The questionnaire included items on sociodemographics, eating behavior, health behavior, and perceptions of the local environment. It was primarily based on the Local Health Interview Survey (Local HIS) of 2019 [41]. Variables included in the Local HIS 2019 were derived from the Belgian National Health Interview Survey of 2018 [42]. Additionally, items from the Flemish version of the Sustainable Prevention of Obesity through Integrated Strategies Project (SPOTLIGHT) and the Perceived Nutrition Environments Measures Survey (NEMS-P) were included to assess the food environment [43, 44]. To increase accessibility, the questionnaire was read and redacted by Wablieft, a Flemish organization that advocates the accessibility and comprehensibility of the Dutch language for underserved groups [45]. Full details on the questionnaire and CIVISANO project can be found elsewhere [40].

Study sample

The inclusion criteria for respondents were age between 25 and 65 years and residing in Duffel or Herselt. In the study, an overrepresentation of respondents with lower socioeconomic status was intended. Therefore, active recruitment, similar to time-location-based sampling, was used. The locations (e.g., food banks/distributions, neighborhoods with a higher concentration of government-assisted social housing, private rentals, and remedial schools.) in which people with lower socioeconomic status were overrepresented were compiled into a list, and these locations were randomly visited by volunteers during the recruitment period (i.e., between May and November 2021). The volunteers offered respondents the option to fill in the questionnaire themselves using a tablet, or guided the respondents through the questionnaire using an interview approach. In conjunction, other sampling strategies were also employed, such as posting QR codes to fill in the questionnaire on traditional (local) media and social media and making the questionnaire available in local places that were not visited during the active recruitment, such as libraries and medical offices.

Dependent variable – eating behavior

Respondents completed several short statements regarding their regular consumption frequency of key indicators of healthy and unhealthy food groups, such as fruits (excluding juices), vegetables (excluding juices and potatoes), sugar-sweetened beverages (excluding light and zero beverages), sweet and salty snacks, and fast food. These statements were based on the nutritional part of the Belgian National Health Interview Survey [42]. Statements included “How often do you eat or drink the following?” with examples of food and/or drinks provided to each group. The response options were: “never”, “less than once a week”, “1 to 3 times a week”, “4 to 6 times a week, “once a day”, and “more than once a day”. Subsequently, the responses were converted into numerical values corresponding to the times per day, as defined by Van Mierlo et al. (2021) and Haubrock et al. (2011) [46, 47]. In addition, fruit and vegetable consumption frequencies were combined into one variable. Resulting in a total of four outcome variables: fruit-and vegetable consumption frequency (FV), fast-food consumption frequency (FF), snack consumption frequency (SN) and sugar-sweetened beverages consumption frequency (SSB).

Explanatory variables – food environment

The assessment of the food environment was conducted at the individual level for each respondent and was divided into objective and perceived domains. For the objective domain, multiple objective measures, that is, the proximity to healthy, unhealthy, and fast-food outlets, the density of healthy, unhealthy, and fast-food outlets in the 500m buffer, the density of healthy, unhealthy, and fast-food outlets in the 1000m buffer and the mRFEI in the 500m and 1000m buffer were assessed. To calculate these, respondents were asked to localize the intersection nearest to their home address when filling in the questionnaire. The nearest intersection, instead of the home address, was used to protect respondents’ privacy. Using ArcGIS Pro, the location of each respondent was linked to information regarding all food outlets in the municipalities. Data on these outlets were obtained from the Locatus 2020 database, supplemented with data on local (farmers) markets, farm stores, and community gardens. Solely, food outlets selling food as a primary function were included and outlets selling food as a secondary function, e.g. cinema’s or sport facilities with vending machines, were excluded. Food outlets were classified as healthy, neutral or unhealthy based on the opinion of an expert committee consisting of food policy experts and nutritionists from Flanders. From the unhealthy category, fast food outlets were extracted as separate metrics. Healthy food outlets were defined as outlets that primarily sell healthy foods, i.e. greengrocers, fishmongers, farmers’ markets, stores selling nuts and organic stores. Neutral outlets were defined as outlets selling a mix of healthy and unhealthy food and drinks, i.e. supermarkets, mini-supermarkets, poulterers, cheese shops, bakeries, hotels with restaurants, lunchrooms and full-service restaurants. Unhealthy food outlets were defined as outlets that primarily sold unhealthy foods, i.e. cafés, pancake restaurants, butchers, flan restaurants, ice cream parlors, confectionary and convenience stores. Fast food outlets were defined as outlets selling fast food to take-away and/or eat in, i.e. fast food outlets (sit-in and delivery) and grillroom/shawarma outlets.More information about the classification of food outlets can be found in Additional File 2 and in the study by Smets et al. (2022) [22]. In line with the classification, suggested by the expert committee, it was decided to classify supermarkets as neutral in this study, even though multiple studies have classified them as healthy (e.g. Thornton et al., 2012;Clary et al., 2016) [48, 49]. This was based on our research questions, which were formulated to evaluate the consumption frequency of multiple nutritional categories, not solely fruits and vegetables but also fast-food, snacks, and sugar-sweetened beverages, in relation to healthy and unhealthy food outlets. Unlike many studies that primarily focus on fruit and vegetable intake in relation to accessibility (e.g. Evans et al., 2012; Pessoa et al., 2015) [50, 51]. Additionally, although supermarkets are crucial points of access for fruits and vegetables, studies have shown that supermarkets may not universally be deemed as healthy food outlets. For instance findings from the International Network for Food and Obesity/NCDs Research, Monitoring, and Action Support (INFORMAS) (Vandevijvere et al., 2018) and a Belgian study from Vandevijvere et al. (2023) suggest that the ratio between healthy and unhealthy foods and drinks in supermarkets, is geared towards unhealthy foods [52, 53]. This is in line with in-store measurements, which have been conducted between February and May 2021, as part of this project. Which assessed the in-store food environment in a sample of six supermarkets across four different chains in the municipalities participating in this study. Overall, in the visited supermarkets the ratio between healthy/unhealthy foods was found to be 0.45, indicating that for every 10m of shelf length of unhealthy foods there was 4.5m of healthy foods. Based on the above-mentioned findings it was therefore decided to retain the classification of supermarkets as neutral.

Buffers of 500m and 1000m were used to calculate the density. These buffer sizes were chosen based on previous studies conducted internationally and in Ghent (Belgium) as part of the ‘International Physical Activity and Environment Network’ (IPEN) which recommend the use of street network buffers of 500m and 1000m around respondents’ residences to develop a standardized spatial definition of a ‘neighborhood’ that can be used to compare results between countries [54, 55]. In addition, two different buffer sizes were used to account for potential variations in the food environment and travel behaviors of respondents to food outlets [56].

Proximity measures were defined as the shortest road network distance in meters to the nearest healthy, unhealthy, and fast-food outlets and were calculated for each respondent. In addition, the modified Retail Food Environment Index (mRFEI), which is the ratio of healthy food outlets to the total number of food outlets, was also calculated as part of the objective measures. The mRFEI score is a continuous variable ranging from 0 to 1. A lower score indicated a more unhealthy food environment [57]. A score of 0 indicated that there were no healthy food outlets in an area. Because the mRFEI could only be calculated for respondents with food outlets in their buffers, respondents without healthy, neutral and unhealthy food outlets in their buffers were excluded (n = 71).

The mRFEI was calculated using the following formula:

$$mRFEI=\frac&\# healthy\,food\,outlets + \# neutral\,food\,outlets \\&+ \# unhealthy\,food\,outlets\end}$$

The perceived domain was based on respondents’ perceptions of five statements regarding their local food environment, which were based on the NEMS-P questionnaire [44]. The statements included: “Fresh fruits-and vegetables are easily available in my neighborhood”, “Fresh fruits- and vegetables are cheap to buy in my neighborhood”, “Fresh fruits-and vegetables in my neighborhood are of good quality”, “Fast-food is easily available in my neighborhood” and “Fast-food is cheap to buy in my neighborhood”.

Covariates

Respondents’ age and gender identity were included as covariates in all the models. Age was measured in years and rounded off to the nearest whole number. Respondents’ gender identity was determined by asking respondents to indicate if they identified as “male,” “female” or “prefer not to answer.” SES was used as an interaction term to assess its moderating effect on the associations between objective and perceived food environments and eating behavior. All respondents were classified as having a lower socioeconomic status (LSES) or higher socioeconomic status (HSES). LSES was based on meeting at least one of the following criteria: (a) low level of education (no tertiary level), (b) no current paid employment within the household, and (c) net family income below the national minimum income (that is, € 1625.72 gross per month per person in 2021), taking family size into account, (d) perceived financial difficulties (= difficult to very difficult to make ends meet on a monthly basis), or (e) low perceived socioeconomic status (lower than or equal to five on the MacArthur scale). The MacArthur scale measures subjective socioeconomic status by asking respondents to place themselves on a hypothetical ladder relative to others in their group [58]. All variables used as covariates were obtained from the questionnaire and were self-reported.

Statistical analysisStudy population

Descriptive statistics were used to describe the characteristics of the total sample and were stratified by SES. Continuous variables are presented as means and standard deviations (SD). Categorical variables are presented as frequencies.

Differences in food environment along SES

Independent t-tests, chi-square, or Mann-Whitney U analysis tests were conducted to evaluate the hypothesized differences in domains in the food environment between SES groups.

Food environment, eating behavior and SES

Linear regression models were used to test associations between the objective and perceived food environment domains and the four eating behavior outcomes (i.e., FV, FF, SN, and SSB consumption frequency), adjusted for of age and sex. Before running the models the assumptions for linear regression were first tested. During this assessment, the assumptions were checked for using the original- and log-transformed values of the dependent variables to determine whether this resulted in more robust conclusions. Since there were no large differences when comparing the plots showing residuals vs. fitted values and both showed homoscedasticity, the original values of the dependent variables were used because it is known that log-transformed variables are more difficult to interpret [59]. Afterwards, we tested whether SES had a moderating effect on the association between domains of the food environment and eating behavior. Six multivariable models with interaction terms, that is, one model per domain (i.e., objective and perceived) for each eating behavior outcome under study were created. As indicators of the perceived domain were not available for SN and SSB consumption frequencies, only objective indicators were tested for these outcomes. Table S1 in Additional file 1 shows the variables tested for each eating behavior outcome. Stratified analysis of the SES groups was conducted when statistically significant interactions between domains of the food environment and eating behavior were observed. If no statistically significant interaction term was observed, the effect on the whole population was assessed and the analysis repeated without including an interaction term. These analyses were adjusted for the five variables used to construct SES (i.e., education, employment, income, subjective SES, and subjective financial difficulties). This was done to ensure that no confounding factors occurred when analyzing the main effects. The outcomes of the analysis of the main effects are shown in Table 1 in Additional File 5. Statistical tests were two-sided, and differences or associations were considered statistically significant at p < 0.05. All analysis were performed using RStudio.