Data from the Dutch contribution to the European Prospective Investigation into Cancer and Nutrition (EPIC) study was used for the analysis. The EPIC study is a population-based, prospective cohort that was initiated in ten European countries to study the role of diet and physical activity in the etiology of cancer and other chronic diseases [50]. The Dutch contribution (EPIC-NL) includes the Prospect cohort and MORGEN cohort and in total consisted of 40,011 participants recruited between 1993 and 1997 [51]. The Prospect cohort comprised 17,357 females aged 49–70 years who participated in a regional breast cancer screening program in the city of Utrecht and its surroundings [52]. The MORGEN cohort included 22,654 males and females aged 20–65 years who were randomly selected from the general population of Amsterdam, Maastricht, and Doetinchem [53, 54]. The EPIC-NL study complies with the guidelines described in the Declaration of Helsinki, and all procedures involving human participants were approved by the Institutional Review Board of the University Medical Centre Utrecht and the Medical Ethical Committee of TNO Nutrition and Food Research. All participants provided their written informed consent [51].

For the present analyses, exclusion criteria were missing dietary information at baseline (n = 217), no informed consent for follow-up of vital status (n = 925), withdrawal of informed consent during follow-up (n = 1), and missing follow-up data of vital status (n = 142). Participants with a history of cancer (n = 1627), diabetes (n = 718), myocardial infarction (n = 427), or stroke (n = 361) at baseline were excluded because their usual reported diet may be influenced by their condition and not reflect their diet before diagnosis. To exclude implausible dietary values that could lead to incorrect analysis of the data, participants in the highest and lowest 0.5% of the reported energy intake to basal metabolic rate ratio (n = 332) were left out as well. Finally, participants with missing information on possible confounders, including BMI (n = 17), educational level (n = 191), or smoking status (n = 23), were excluded. After these exclusions, in total 35,030 participants remained for the complete-case analysis.

Food consumption was measured at baseline (1993–1997) using a self-administered semi-quantitative food frequency questionnaire (FFQ), including questions on the usual frequency of consumption of 77 main food categories during the year before enrolment. The questionnaire allowed estimation of the usual daily dietary intake of 178 food items, and had been validated against twelve 24-h recalls and biomarkers in 24-h urine and serum [55, 56]. Spearman rank correlation coefficients based on estimates of the FFQ and 24-h recalls were 0.51 for potatoes, 0.36 for vegetables, 0.68 for fruits, 0.39 for meat, 0.69 for dairy, 0.76 for sugar and sweet products, and 0.52 for biscuits and pastry in males. Similar results were obtained for females. Nutrient intakes were calculated using the 1996 Dutch Food Composition Table [57].

In order to assess participants’ adherence to a healthy and less healthy plant-based diet, the healthful plant-based diet index (hPDI) and unhealthful plant-based diet index (uPDI) were calculated based on the procedure used by Martínez-González et al. [58] and adapted by Satija et al. [32]. Supplemental Table 1 displays the food groups with included foods and scoring criteria for the two PDI indices. First, 18 food groups were created based on nutrient and culinary similarities within the larger categories of healthy plant-based foods (vegetables, fruit, legumes, whole grains, nuts and seeds, vegetable oils and fats, and tea and coffee), unhealthy plant-based foods (refined grains, potatoes, juices, (sugar) sweetened beverages, and sweets and desserts), and animal-based foods (meat, animal fats, eggs, fish and seafood, dairy products, and miscellaneous animal-based foods). All FFQ items were assigned to the appropriate food group and checked by a research dietician. We distinguished between healthy and unhealthy plant-based foods using the most recent empirical evidence on their associations with several chronic conditions (i.e. obesity, hypertension, lipids, inflammation, type 2 diabetes, cardiovascular disease, and certain cancers) [32]. Alcoholic beverages were not included in the indices due to differential associations with health outcomes, but included as a covariate in the analyses. Mixed dishes that contain substantial amounts of animal sourced ingredients (e.g. soups, pizza, and mayonnaise) were classified as miscellaneous animal-based foods, in accordance with previous studies [30,31,32,33,34].

Second, food consumption in g/day of each food group was calculated for all individuals by summing the consumption of the FFQ items. Within each food group, participants were classified into quintiles according to their consumption of that specific food group after adjusting for total energy intake using the residual method. Based on the quintiles within each food group, participants were given a positive score between 1 (lowest quintile) and 5 (highest quintile) or reverse score between 5 (lowest quintile) and 1 (highest quintile). For the hPDI, healthy plant-based food groups were assigned positive scores and unhealthy plant-based and animal-based food groups were assigned reverse scores. For the uPDI, unhealthy plant-based food groups were given positive scores and healthy plant-based and animal-based food groups were given reverse scores. To obtain the PDI indices, the 18 food group scores for each individual were summed and could range from 18 (lowest possible score) to 90 (highest possible score).

Overall, the hPDI and uPDI give more points to high consumers of healthy and unhealthy plant-based food groups, respectively. It should be noted that high PDI indices do not equal a vegetarian or vegan diet, but rather indicate a relatively high consumption of plant-based foods and/or relatively low consumption of animal-based foods compared to the total study population. The indices are thus dependent on the food consumption in our study population and cannot be directly compared to other populations where the overall consumption of plant-based or animal-based foods may be higher or lower.

The NOVA classification was applied to assess the degree of food processing of the diet [37]. This classification includes four classes: unprocessed/minimally processed foods (MPF), processed culinary ingredients (PCI), processed foods (PF), and ultra-processed foods (UPF). To discriminate different foods between these classes, the NOVA classification takes into consideration the ingredient list of food items and all physical, chemical, and biological methods used during the food production process. An extensive description of the different classes can be found elsewhere [37]. All food items of the FFQ were previously assigned to one of the four classes of the NOVA classification based on the degree of processing [59]. To account for potential changes in food processing over time, three scenarios (lower, middle, and upper bound) were considered when classifying food items. The lower bound scenario encompassed food items that could have been less processed compared to the middle bound scenario and were assigned to a less processed NOVA class, whereas food items that could have been more processed were included in the upper bound scenario and assigned to a more processed NOVA class. The middle bound scenario mostly resembled with the dietary assessment that was conducted in the nineties and was therefore used in all analyses.

The daily consumption in g/2000 kcal of MPF, PCI, PF, and UPF was calculated for each participant by summing the consumption of the FFQ items for each NOVA class and dividing these by the total energy consumed per individual multiplied by 2000. A weight ratio instead of an energy ratio was used to account for food that does not provide energy (e.g. artificially sweetened beverages) and non-nutritional substances related to food processing (e.g. additives). In the primary analysis, alcoholic beverages were included in the NOVA classes to adhere as much as possible to the original classification. Sensitivity analyses in which alcoholic beverages were excluded from the calculation of the NOVA classes did not generate considerably different results (Supplemental Table 2). For the purposes of this study, we focused on UPF consumption from the NOVA classification as this was our main interest.

At baseline, several lifestyle factors were assessed using a general questionnaire, containing questions on demographic characteristics, presence of chronic diseases, and related potential risk factors. Weight and height were measured by trained staff according to standardized protocols [51] and BMI was calculated by dividing weight by height squared (kg/m2). Duration and types of physical activity were assessed with a validated questionnaire [60] and classified according to the Cambridge Physical Activity Index (CPAI) with imputed data for missing values [61]. The CPAI was divided into four different categories: inactive, moderately inactive, moderately active, and active. Smoking was classified as never, former, and current smoker. Educational level was coded as low (lower vocational training or primary school), medium (intermediate vocational training or secondary school), or high (higher vocational training or university).

Environmental impact of the foods consumed was derived from the Dutch Life Cycle Assessment (LCA) food database [62] and has previously been described in more detail [29]. In short, the LCA methodology was applied to quantify the environmental impact for six different indicators (land use, BWC, GHGE, acidification, freshwater eutrophication, and marine eutrophication) throughout the entire foods’ life cycle. All life cycle stages from cradle to grave were included in the analyses, including primary production, processing, primary packaging, distribution, supermarket, retail, storage, preparation by the consumer, and waste or losses. Transport was only included from primary production to supermarket and food waste was calculated by using food group-specific percentages for avoidable and unavoidable food losses throughout the food chain. When production processes led to more than one food product, environmental impact was divided using economic allocation, except for milk where biophysical allocation was applied.

The LCA data were available for 242 foods that were selected based on frequency and quantity of consumption in the Dutch National Food Consumption Survey or its relatively high environmental impact per kg of food. These data were in prior linked to the EPIC-NL FFQ data [63]. For FFQ items of which primary data was not available, extrapolations were carried out by the Dutch National Institute for Public Health and the Environment (RIVM) based on similarities in type of food, production method, and ingredient composition. The LCAs are based on current production practices and are assumed to be equal in the nineties when food consumption was assessed. Since a previous study measured high correlations between GHGE and several environmental indicators (land use, acidification, fresh water eutrophication, and marine eutrophication), with exception of BWC, this study focused on GHGE and BWC to assess environmental impact [29]. BWC refers to the total volume of water sourced from surface and groundwater, as defined by Hoekstra et al. [64]. For each participant, the daily GHGE (kg CO2-eq) and daily BWC (L) were calculated in absolute and standardized amounts. Standardized amounts were divided by total energy intake and expressed per 2000 kcal.

Vital status of all participants was obtained through linkage with the municipal population registries of the Netherlands. Participants were followed up over time until death from any cause, migration, loss to follow-up, or were censored. All-cause mortality was defined as death from any cause after study inclusion. Follow-up was completed on December 31st, 2014.

Interaction between the PDI indices and UPF consumption in the associations with all-cause mortality and environmental impact were evaluated following the recommendations of Knol and VanderWeele [65]. First, participants were divided into sixteen dietary categories based on quartiles of the hPDI and quartiles of UPF consumption. The most extreme quartiles were mainly of interest: (1) low hPDI score, high UPF consumption (Q1hPDI/Q4UPF), (2) low hPDI score, low UPF consumption (Q1hPDI/Q1UPF), (3) high hPDI score, high UPF consumption (Q4hPDI/Q4UPF), and (4) high hPDI score, low UPF consumption (Q4hPDI/Q1UPF). For example, the Q1hPDI/Q4UPF category represented diets that were low in healthy plant-based foods (hPDI ≤ 50 points) and high in UPF (UPF > 433 g/2000 kcal), and was taken as a reference for all analyses. The same procedure was applied to the scores of the uPDI.

Cox proportional hazard models were used to estimate hazard ratios (HRs) with 95% confidence intervals (CIs) for the associations between the sixteen dietary categories and all-cause mortality. Risks are also presented per 10-point increase in the PDI indices stratified by quartiles of UPF consumption. Person-years were calculated from the date of study inclusion to the date of death or the end of follow-up, whichever came first. Confounders were considered in two separate models. Model 1 was cox-stratified for age and adjusted for sex and total energy intake. Model 2 was additionally adjusted for educational level, smoking status, physical activity level, and alcohol consumption. The proportional hazard assumption was checked using the Schoenfeld residuals test, but no evidence for violation of this assumption was found (all P > 0.05). Linear trends were tested by assigning median values to each quartile and entering this as continuous variables in the models. We evaluated additive interaction by estimating the relative excess risk due to interaction (RERI) based on continuous exposure variables [66, 67]. The hPDI was recoded into a risk factor by multiplying with − 1 for correct calculation of the RERI. A RERI of 0 means a lack of significant interaction on the additive scale.

To investigate associations between the dietary categories and daily GHGE and BWC, multiple linear regression models were used to calculate differences in adjusted mean GHGE and BWC with 95% CIs for the different dietary categories. Mean differences are also presented per 10-point increase in the PDI indices stratified by quartiles of UPF consumption. These analyses were adjusted for age, sex, and total energy intake (Model 1). Assumptions of the linear model, including multicollinearity, homoscedasticity, independence and normality of the residuals, and linearity of the association, were checked by calculating VIF-values (all < 2) and generating various plots (scale-location, Q-Q, and residuals vs. fitted values). Interaction on the additive scale was evaluated by testing the product term (β3) of the PDI indices and UPF consumption [66]. This estimate of interaction is not the same as RERI, but rather reflects the change in absolute values instead of a change in relative risks. A product term of 0 means a lack of significant additive interaction.

All statistical analyses were performed using R software (version 3.5.0). A two-sided P-value of < 0.05 was considered statistically significant.