Study population
The E3N study is a large ongoing prospective cohort study involving 98,995 French women born between 1925 and 1950, primarily insured through the Mutuelle Générale de l’Education Nationale (MGEN). Recruitment began in 1990 with a baseline self-administered questionnaire and informed consent, follow-up questionnaires were dispatched every 2–3 years. The average response rate was approximately 83%, with an overall loss to follow-up of only 3% since 1990. The study was approved by the French National Commission for Data Protection and Privacy [23].
Between 1993 and 2018, detailed information was collected at various time points, including anthropometric characteristics, health status, diet, reproductive history, hormonal treatments, smoking habits, alcohol consumption, or physical activity [23].
For this study, we included 74,522 women who completed the dietary questionnaire in 1993 (Q3, baseline for the present study). The E3N cohort consisted of middle-aged women with a median age of 51 in 1993, making the inclusion of pregnant women highly unlikely due to their age and the timeframe of the study. Participants were excluded if follow-up ended at Q3, if baseline height or weight was missing, or if weight data were missing on consecutive questionnaires. To avoid under or over reporting of dietary intake, 1327 participants in the top or bottom 1% of the ratio of energy intake to energy requirement were excluded [24, 25]. Finally, to assess the risk of obesity and overweight, participants with prevalent obesity or overweight at baseline were excluded. Consequently, the study population size varied depending on the case variable considered (obesity, overweight, or weight gain of more than 10 kg) (Fig. 1).
Fig. 1
Assessment of food consumption
Dietary data were estimated in 1993 using a validated semi-quantitative food frequency questionnaire, which contained 208 food items. The dietary information collected included details on breakfast, morning snack, aperitif before lunch, lunch, afternoon snack, pre-dinner aperitif, dinner, and after dinner snack. Women’s daily intake of nutrients, such as fat intake, was then calculated based on data from the French Information Center on Food Quality (CIQUAL) [26]. The validity and reproducibility of the dietary questionnaire have been tested and previously described [27].
Assessment of dietary intake of PBDEs
Data on PBDEs food contamination were obtained from the 2nd French Total Diet Study (TDS2) conducted by the French Agency for Food, Environmental and Occupational Health and Safety (ANSES) [5, 6, 28, 29]. Each food sample analysed in TDS2 was categorised and described according to the nomenclature used in the French national food consumption study INCA2 (Étude Individuelle Nationale des Consommations Alimentaires 2) and that PBDE concentration values for each sample were reported individually. The TDS2 database is open source [30]. A total of approximately 20,000 food products were collected across eight French regions between June 2007 and January 2009 and combined into 1352 composite samples representative of typical French dietary habits. Eight PBDEs were measured in food corresponding to the main known sources of exposure, which include meat, fish, eggs and dairy products [31]. In the present study, PBDE values in food items below the limit of detection were replaced by 0 (lower-bound scenario). The E3N database for food consumption and the ANSES database for food contaminant concentrations were subsequently merged, as described elsewhere [32]. The daily mean dietary intake of each PBDE (BDE-28, BDE-47, BDE-99, BDE-100, BDE-153, BDE-154, BDE-183 and BDE-209; ng/day) was estimated by multiplying the mean daily quantities consumed of each food component by the values of contamination of the corresponding food component [10]. The intake of dioxins and polychlorinated biphenyls (PCBs) was collected and estimated using the same approach as for PBDEs [33].
Identification of obesity, overweight and weight gain
In the E3N cohort, height and weight were self-reported by participants at each E3N questionnaire. The BMI is defined as the body weight divided by the square of the body height (kg/m2) [33]. A validation study involving 152 women from the Paris center of the cohort found strong correlations between self-reported and technician-measured anthropometric factors, with coefficients of 0.94 for weight and 0.92 for BMI [34].
In the present study,
Obesity: is defined as BMI ≥ 30 kg/m 2 [35].
Overweight: is defined as BMI ≥ 25 kg/m 2 [35].
Weight gain: is defined as weight gain of more than 10 kg compared with the weight reported at baseline
Covariates
Adjustment variables included in the analyses described below were defined by Directed Acyclic Graph (DAG) (supplementary Fig. 1) in order to assess the total effect of PBDEs intake on the risk of obesity, overweight or weight gain.
Information on educational level (duration <12 years; 12–14 years; >14 years) was collected at the first questionnaire sent in 1991 (Q1). Information on smoking status (non-smoker; former smoker; current smoker), parity (nulliparous; one or two children; more than 3 children), silhouette at puberty (very thin, thin, medium, wide, very wide), menopausal status and recent use of menopausal hormone therapy (MHT) (pre-menopaused, menopaused with recent use of MHT, i.e. less than a year ago; menopaused without recent use of MHT, no information) and utilisation of contraceptive pill (ever/never) was obtained from the second questionnaire sent in 1992 (Q2). In addition, information on physical activity (continuous, in metabolic equivalents of task (MET)-hour/week), daily intake of alcohol (continuous, in g/day), daily intake of lipids (continuous, in g/day), daily intake of fatty acids (continuous, in g/day) and daily total energy intake (continuous, in kcal/day) was obtained from the dietary and non-dietary questionnaires sent in 1993 (Q3). The Programme National Nutrition Santé (PNNS) adequacy score is a composite dietary index based on 13 components, including seven adequacy components (fruits and vegetables, nuts, legumes, whole-grain foods, milk and dairy products, fish and seafood, and added fats) and six moderation components (red meat, processed meat, sugary foods, sweet-tasting beverages, alcoholic beverages, and salt), reflecting adherence to French dietary recommendations. A low score, indicating poor adherence to recommendations, has been associated with increased risk of mortality and type 2 diabetes [36].
The modal value (the most frequently occurring value, for categorical variables) or the median (for continuous variables) was used to impute missing values for covariates with less than 5% missing data, in order to reflect the most common or central values in the population. In the present study, missing data for education level (3.4%) were imputed as 12–14 years; missing data for parity (0.7%) were imputed as one or two children; missing data for contraceptive pill use (0.5%) were imputed as ever use; missing data for silhouette at puberty (3.7%) were imputed as thin; and missing data for smoking status (0.8%) were imputed as never smoking. Missing data for physical activity (0.6%) were imputed using the median value within the entire population (37.97 MET-hours/week).
Statistical analyses
Participants’ baseline characteristics were described according to quartile groups of total PBDE intakes, and separately among cases and non-cases. Also, the proportion of each PBDE congener in the total dietary intake of PBDEs in the study population and the correlations between each PBDE were described.
Cox proportional hazard regression models were used to estimate hazard ratios (HRs) and 95% confidence intervals (95% CI) for associations between the dietary intake of PBDEs and the risk of obesity, overweight and more than 10 kg weight gain. Adjustment variables included in the models were selected using the DAG (supplementary Fig. 1). For Cox proportional hazards models, the proportional hazards assumption was assessed using Schoenfeld residuals and no significant violations were detected. All models were adjusted on age as time scale. Age at exit was the age at which the participant developed obesity (or overweight or gained 10 kg), or the age at the last completed questionnaire before death or the end of follow-up; or age at the last completed questionnaire for lost to follow-up, whichever occurred first.
The first model was adjusted for age as time scale. Model 2 was further adjusted on education level, smoking status, silhouette at puberty, alcohol consumption (g/day), total energy intake without alcohol (kcal/day), and physical activity (MET-hours/week). Model 3, the main model, including all the variables identified by the DAG, was additionally adjusted on BMI at baseline, contraceptive pill use, parity and menopausal status and recent use of menopausal hormone therapy.
In the Cox model, dietary intake of total PBDEs was assessed both as continuous variables and as categorical variables into quartiles. The first quartile group for PBDEs dietary intake was used as reference. For the continuous exposure variables, standardisation was performed by dividing the intake of PBDEs by its respective standard deviation (SD), thereby estimating the HR per one SD increment. To assess linear trends across categories of PBDEs dietary intake, the median value was allocated to each category and subsequently employed as a continuous variable within the models.
Spline functions were used in model 3 to assess any non-linear association between PBDEs and risk of obesity, overweight or weight gain. Variables were modelled using Restricted Cubic Splines (RCS) with four knots (0.05th, 0.35th, 0.65th and 0.95th) [37]. Since BMI at baseline did not meet the assumptions of log-linearity and proportionality of hazards, it was included in the model by stratifying the baseline hazard function into quintiles of baseline BMI (20.40 kg/m2, 21.72 kg/m2, 23.11 kg/m2, 25.10 kg/m2). This approach allowed for accounting for BMI at baseline without making assumptions about its relationship with the outcome.
Sensitivity analyses
Sensitivity analyses were performed adjusting model 3 for PNNS adequacy score, to disentangle the effects of exposure to food contaminants from those of overall quality of the diet. The total fat intake, polyunsaturated fatty acids (PUFA) intake, and n-3 polyunsaturated fatty acids (n-3 PUFA) intake were separately adjusted to distinguish the effect of exposure to food contaminants from those of fat intake (model 3). Model 3 was also adjusted for the sum of dioxins and dioxin-like polychlorinated biphenyls (Dioxins+DL-PCBs) and non-dioxin like PCBs (NDL-PCBs) to adjust for the effect of exposure to other food contaminants associated with the risk of weight gain [33]. In order to account for differences in energy intake, a residual contaminant model (adjusted for energy) was used based on Model 3. In this model, total PBDEs intakes were regressed on total energy intake, and the residuals from this regression were used as the intake variables [38]. Model 3 was additionally adjusted separately on the fish, dairy and meat consumption to investigate potential residual confounding from the diet. To minimise potential reverse causality and reduce the influence of early outcome-related dietary changes, we conducted a 5-year lag analysis. In this analysis, participants who developed obesity, became overweight, or gained more than 10 kg during the first five years of follow-up after Q3 were excluded (Model 3). The first five years were selected because dietary changes occurring shortly after baseline could be influenced by early disease-related weight changes, and excluding this period helps to better capture the temporal relationship between diet and subsequent weight outcomes.
Subgroup analyses
Subgroup analyses were conducted based on the median follow-up time to evaluate long-term effects of PBDE dietary intake. In the first analysis, follow-up was stopped at the median, and in the second analysis, participants with follow-up periods shorter than median follow-up times were excluded. In order to evaluate potential effect modification, interaction analyses were performed between silhouette at puberty (very thin and thin, medium, large and very large) and PBDEs intake, as well as between baseline BMI and PBDEs intake. Subgroup analyses were performed only when the interaction tests were statistically significant.
All statistical analyses were performed with SAS software version 9.4 (SAS institute) and R version 4.3.1. All tests were two-sided and we considered P < 0.05 to be statistically significant.
Dining and Cooking