Study population
EPIC-NL is the Dutch contribution to the European Prospective Investigation into Cancer and Nutrition (EPIC), and consists of the MORGEN and Prospect cohorts for which recruitment took place between 1993 and 1997. The design and rationale EPIC-NL have been described previously [30]. Briefly, the MORGEN cohort includes 22,654 men and women aged 20–69 years who were randomly sampled from three towns across the Netherlands (Maastricht, Doetinchem and Amsterdam). Prospect includes 17,357 women aged 49–70 from Utrecht and vicinity who participated in the national Dutch breast cancer screening program. The study was conducted in accordance with the Declaration of Helsinki and the institutional review boards of the University Medical Center Utrecht and Medical Ethics Committee of TNO Nutrition and Food Research approved the study. Written consent was obtained from all participants.
In 2015, participants who responded to a follow-up questionnaire on electromagnetic radiation in 2011, who were still alive and residing in the Netherlands in 2015, and who provided informed consent (n = 13,421 from Prospect and MORGEN Amsterdam and Maastricht only; the Doetinchem cohort was not invited) were invited to fill out a food frequency questionnaire (FFQ) for which the response rate was 62.7 % (n = 8,409). From this population, we excluded those with extreme energy intake defined as those within the lower and upper 0.5 % of the ratio between energy intake and basal metabolic rate (n = 84). We further excluded participants who did not return the general questionnaire at follow-up (n = 94). Consequently, the population for analysis comprised 8,231 participants.
Exposure assessment – relative FFR exposure
Data on relative FFR exposure was obtained through linkage of the 6-digit postal code from home address in 2015 with a commercial retail database from Locatus. Locatus is a commercial company that collects data (e.g., location, type) on food outlets in the Netherlands based on systematic field audits, which are conducted regularly (e.g., once per year for busy shopping areas and once per two to three years for less busy shopping areas). Additionally, field audits are complemented by surveys and telephone interviews with retailers, assuring up-to-date data. A recent validation study showed good to excellent agreement between the Locatus data and research field audits [31]. Although the participants included in this study were originally recruited in Amsterdam, Utrecht and Maastricht (1993–1997), approximately 40 % of study participants included in the present study still lived in the recruitment areas, whereas the remaining study participants had moved elsewhere (Additional File 1).
From the Locatus database, we extracted data on food retailers operating in 2015. FFR were defined as traditional fast-food restaurants, grill-rooms and take-away outlets. Using PCRaster Python (pcraster.eu), we calculated a relative measure of FFR exposure in a 400, 1000 and 1500 street-network buffer (e.g., the distance someone can cover using the street-network) around the home of residence. This was done by dividing the densities of FFR by the density of all food retailers in the corresponding buffer.
Outcome assessment – diet quality and BMI
In 2015, dietary intake was assessed with a validated semi-quantitative food-frequency questionnaire, the FFQ-NL 1.0 [32]. The FFQ assessed habitual consumption of 160 food items in the preceding year, through questions on consumption frequency and consumed amounts. Estimated food group intake was validated against an average of 2.7 telephone-based 24-hour recalls. Spearman correlation coefficients between estimates of the FFQ and estimates from the 24-hour recalls were 0.66 for fruits, 0.61 for bread and bread products, 0.38 for meat, 0.29 for vegetables, 0.27 for fish, 0.20 for nuts, seeds and snacks, and 0.13 for legumes. Average daily energy intake was estimated using the Dutch food composition table from 2011 [33].
In order to assess diet quality, we calculated scores on the Dutch Healthy Diet 2015 index (DHD15-index) [34]. Studying overall diet quality in contrast to isolated food groups or nutrients is nowadays preferred in nutritional epidemiology, as overall dietary patterns capture synergistic properties of individual foods and are more likely to affect health/weight status as compared to consumption of individual foods [35]. The DHD15-index reflects adherence to the Dutch dietary guidelines as issued by the Health Council in 2015. The index consists of fifteen food groups for which participants are allocated points based on absolute intakes of the respective food groups, resulting in a score ranging from 0 to 150 (Additional file 2). For the present analysis three out of the fifteen original food groups were excluded in the calculation of the DHD15-score. First, the coffee component was dropped from the score as the FFQ did not differentiate between types of coffee (filtered vs. unfiltered). Secondly, we excluded the sodium component from the DHD15-score, as sodium intake was not reliably captured with the FFQ. Third, we excluded the alcohol component from the score since we deemed it to be inappropriate in the context of our research question as alcohol is usually not sold in FFR. Moreover, we did not have data on type of cereal product (wholegrain vs. refined) except for bread. Therefore, the scoring of the wholegrain component was based on bread only, with an intake equal to or above 90 g receiving the maximum score of 10 points, and a proportionate decrease in points with decreased intake to the point where participants are assigned with 0 points when consuming no wholegrain bread. Taken together, the DHD15-score could range between 0 and 120 with higher scores indicating better adherence to the dietary guidelines and thus better diet quality.
Lastly, we used Body Mass Index (BMI) as a measure of overweight and obesity. Weight and height were self-reported in the follow-up questionnaire in 2015. BMI was calculated by dividing the weight by the height squared. Participants were categorized as normal weight (BMI < 25 kg/m2), overweight (25 ≤ BMI < 30), or obese (BMI ≥ 30).
Covariate assessment
At baseline, participants completed a general questionnaire providing data on age, sex and educational level. Given the high proportion of older women in EPIC-NL, individual educational level may not be representative of women’s socioeconomic position when for example their partner is more highly educated. Therefore, we included the highest attained household educational level of the participant or the partner as a covariate in our analyses. Household educational level was categorized into low (primary education or intermediate vocational education), moderate (higher secondary education), and high (higher vocational education or university). At follow-up, participants provided information and smoking status, which was was categorized into never, former and current.
Data on neighborhood level socioeconomic position (SEP) in 2014 and 2016 – based on neighborhood income, educational level and job status – was obtained from the Netherlands Institute for Social Research and linked to the address information. This provided a continuous summary measure of neighborhood SEP for each participant based on their address, with higher scores indicating higher neighborhood SEP. The score ranges from approximately − 8 to 3. The summary measure of neighborhood level SEP in 2014 and 2016 was averaged to approximate neighborhood level SEP in 2015. Data on urbanicity of the neighborhood in 2015 was obtained from Statistics Netherlands, providing a categorical variable for each participant indicating very high level of urbanisation (≥ 2,500 addresses per km²), high level of urbanisation (1,500–2,500 addresses per km²), moderate level of urbanisation (1,000–1,500 addresses per km²), low level of urbanisation (500–1,000 addresses per km²), and very low level of urbanisation (< 500 addresses per km²). Data were linked to the EPIC-NL database through 4-digit postal codes and GWB-codes (Gemeente Wijk Buurt codes, or City Neighborhood Area codes) for the neighborhood level SEP and level of urbanisation, respectively.
Statistical analyses
Baseline characteristics are displayed as means with standard deviations (SDs) for normally distributed variables, medians and interquartile range (IQR) for non-normally distributed variables, and frequencies and percentages for categorical variables, across quartiles of relative FFR exposure.
We performed multiple imputation on missing data (n = 422 for smoking; n = 193 for level of urbanisation; n = 92 for BMI; n = 22 for household educational level, n = 16 for neighborhood SEP), using age, sex, cohort, physical activity, smoking status, educational level, neighborhood SEP, level of urbanisation, DHD-15 score, and BMI as predictor variables and using 20 imputation sets.
Given the fact that recruitment took place across three cities in the Netherlands, we tested for a multilevel-structure by including a random intercept and random slope for recruitment area, with a variance component covariance pattern. The model without the random slope showed better model fit based on lowest AIC. The Wald statistic for the random intercept for recruitment area was non-significant (p = 0.32), indicating that accounting for clustering of participants was not necessary. We performed multivariable linear regression, with quartiles of relative FFR exposure as the independent variable and DHD15-scores as the dependent variable in order to allow for categorical comparisons between participants with varying relative FFR exposure. Models were adjusted for confounding variables based on previous literature. Model 1 was adjusted for age at follow-up, sex, and cohort. Model 2 was additionally adjusted for smoking, household educational level, and energy intake. Model 3 was additionally adjusted for neighborhood level SEP and level of urbanisation. Model 4 was additionally adjusted for the total count of food outlets in the corresponding buffer. We checked the assumptions of linearity and homoscedasticity by plotting the standardized residuals against standardized predicted values. The plots indicated that there was no evidence of non-linearity or heteroscedasticity. We checked the assumption of multicollinearity by examining the correlation coefficients among predictor variables, and the corresponding variance inflation factors and tolerance values, which showed no indication of multicollinearity.
We performed multinomial logistic regression for weight status (overweight and obesity vs. normal weight) as dependent variable, with the lowest quartile of relative FFR exposure as the reference category. Model structure was similar as to the linear regression analysis.
We checked effect modification by household educational level, neighborhood SEP, and level of urbanisation in both the linear and multinomial logistic regression analysis by including an interaction-term between the continuous variable of relative FFR exposure and the potential effect-modifier in fully adjusted models. We considered a p-value of < 0.20 to be indicative of possible effect modification. In sensitivity analyses, we excluded participants who lived at their address for < 1 year and examined the associations in strata of age. For the analyses on weight status, we also conducted two additional sensitivity analyses: one using four instead of three BMI categories as outcome variable (adding an underweight category, defined as BMI < 18.5), and another excluding energy intake as confounder from the model since it might well be an intermediate in the relative FFR exposure – weight status pathway.
Statistical analyses were performed using IBM SPSS Statistics 24 (IBM Analytics, United States of America, New York). A p-value of < 0.05 was considered to be statistically significant.