The Brazilian Longitudinal Study of Adult Health (in Portuguese, ‘Estudo Longitudinal de Saúde do Adulto,’ ELSA-Brasil) is a multicenter prospective occupational cohort aiming primarily to address risk factors for the development and progression of chronic diseases, particularly cardiovascular diseases and diabetes, over a long-term follow-up [15]. Ethics committees of each institution approved the research protocol, and subjects gave written consent to participate in each visit.

Between August 2008 and December 2010, we recruited 15105 active or retired, non-pregnant civil servants, aged 35–74 years, from public institutions of higher education and research located in six Brazilian capital cities (Salvador, Belo Horizonte, Rio de Janeiro, São Paulo, Vitoria, and Porto Alegre), and applied a series of questionnaires as well as laboratory and clinical examinations [15,16,17]. Participants returned twice to the study sites (2012–2014 and 2017–2019) for further investigation. Additionally, they have responded to annual telephone surveillance since 2009.

Among the 15105 participants enrolled, we excluded those with prevalent diabetes at baseline (n = 2429), with implausible food intake (< 600 kcal/d or > 6000 kcal/d) (n = 197), who died (n = 321), were lost to follow-up (n = 1361), had missing data on variables of interest (n = 484), or had bariatric surgery between visits (n = 111). The final sample had 10202 participants (Additional file 1: Fig. S1).

All measurements followed standardized protocols and regular quality control assessments [18]. In each visit, after an overnight fast, we measured weight, height, and waist circumference following internationally standardized protocols and defined body mass index (BMI) as weight (kg)/height (m)2. We also obtained a fasting blood sample by venipuncture and conducted a standardized 2-h 75-g oral glucose tolerance test (WHO 1999). Plasma glucose was measured using hexokinase and HbA1c by high-pressure liquid chromatography (Bio-Rad, certified by the National Glycohemoglobin Standardization Program).

We interviewed participants using structured questionnaires to ascertain age, sex, self-declared race/color, educational achievement, family income, previous medical history, smoking (current and previous), alcohol consumption, physical activity, and family history of diabetes.

Food consumption was evaluated at baseline through a previously validated food frequency questionnaire, with 114 food items [19]. For each item, we obtained the frequency of consumption in the last year (with eight response options: ‘more than 3 times/day’, ‘2–3 times/day’, ‘once daily’, ‘5–6 times/week’, ‘2–4 times/week’, ‘once/week’, ‘1–3 times/month’ and ‘never/almost never’) and the number of portions consumed, using standardized portion sizes. We then calculated the daily amount consumed for each food item in grams by multiplying its portion number, weight, and frequency. Next, we estimated the nutritional composition and energy using the University of Minnesota Nutrition Data System for Research (NDSR) software. Finally, we imputed the 99th percentile value for a food item when that food's value was above the 99th percentile of its distribution.

We grouped food items according to the extent and purpose of their industrial processing (the NOVA classification): (i) non- or minimally processed foods and culinary ingredients, (ii) processed foods, and (iii) ultra-processed foods [1]. We aggregated non- or minimally processed foods and culinary ingredients into a single group because our food frequency questionnaire did not distinguish culinary ingredients from the food items that included them (Fig. 1).

Box of ELSA-Brasil food subgroups according to the NOVA classification, based on the degree of food industrial processing

In secondary analyses we classified artificially sweetened natural juices, coffee, or tea as UPFs (rather than non- or minimally-processed foods and culinary ingredients), given our previous findings suggesting increased diabetes risk in certain groups [20].

We ascertained diabetes based on laboratory measurements taken during visits and reports from the annual telephone follow-up surveillance. We defined as diabetes cases those who: (i) reported a medical diagnosis of diabetes or current use of medication for diabetes, or (ii) had a laboratory measurement reaching the thresholds for fasting plasma glucose (FPG) (≥ 7.0 mmol/L; 126 mg/dL), 2 h post-load glucose (PG) (≥ 11.1 mmol/L; 200 mg/dL), or HbA1c (≥ 48 mmol/mol; 6.5%) (WHO 2006; ADA 2014). We excluded prevalent diabetes at baseline and ascertained incident diabetes at follow-up visits based on these criteria. We additionally included new cases reporting a diagnosis of diabetes on at least two annual telephone interviews after the last clinic visit.

We describe participant characteristics and outcomes using absolute and relative frequencies for categorical variables and mean, standard deviation or median and 25th–75th percentiles for continuous variables. To assess the statistical significance of differences between means or proportions, we employed ANOVA and Chi-square tests, respectively.

We characterized UPF consumption in two ways: first, using mean consumption in grams per day (g/day), expressed as a mean difference of 150 g/day, which represents approximately a 10% difference in consumption in our sample; and second, creating quartiles of consumption (g/day). We used grams instead of kcal to express quantity because some foods and beverages in the UPF group do not provide energy.

We examined the shape of the association along the continuum of UPFs using restricted cubic splines and tested the non-linearity of the associations [21]. We estimated relative risks (RR) and 95% confidence intervals (95% CI) using robust Poisson regression to investigate the associations of UPF intake with incident diabetes. Progressively adjusted models included: in model 2, age (in years), sex (male or female), race/color (white, brown, black, Asian, or Indigenous), school achievement (less than elementary, elementary, secondary or college/university), per capita family income (in Brazilian currency, reais), family history of diabetes (yes or no), smoking (never, former or current), physical activity (in MET minutes/week), and alcohol consumption (in grams/week); in model 3a, model 2 plus energy intake (in kilocalories/day); in model 3b, model 2 plus hypertension (yes or no); and in model 3c, model 2 plus BMI (in kg/m2). We considered model 2 as our final adjustment model; models 3a-c permit a comparison of our results with those of other studies. We drew directed acyclic graphs to understand the variable relationships (Additional file 1: Figure S2).

We tested potential multiplicative interactions by adding multiplicative terms to models. The interactions considered were intermediate hyperglycemia (yes or no), self-reported recent diet change (yes or no), sex (male or female), fruit and vegetable consumption (in grams/day), and (model 3c) with BMI (kg/m2). We assessed multicollinearity between variables, setting a limit of 2 for the variance inflation factor. The overall fit of the model to the data was assessed using the method described by Hosmer and Lemeshow [22].

To assess the robustness of our findings, we performed several sensitivity analyses, based on model 2, by (i) reclassifying natural juices and coffee/tea with added artificial sweeteners from the group of non- or minimally processed foods and culinary ingredients to the UPF group, (ii) alternatively quantifying UPFs as a proportion of grams relative to the total daily grams consumed, (iii) including red meat intake, (iv) fruits and vegetable intake, (v) saturated fat intake, (vi) sugar intake, (vii) fiber intake, and (viii) diet quality as adjustment variables [23], (ix) including 16 cases of incident diabetes who died and were thus excluded in the original analyses, and, finally, (x) performing multiple imputation to permit the addition in analyses of participants with missing values in covariates.

We also evaluated the association between the consumption of specific UPF subgroups and the incidence of diabetes using model 2 for increments of 50 g/d and one standard deviation in each group. We conducted all analyses with the statistical software package SAS Studio® (SAS OnDemand for Academics).