The UK Biobank study is the basis for all analyses. Between 2006 and 2010, more than 500,000 participants were recruited across the UK [25]. For the current study, participants who filled out at least one web-based dietary questionnaire for the assessment of previous 24 h dietary intakes (Oxford WebQ) [26] were selected as summarized in Online Resources 2 and 3. The following exclusion criteria were applied: (1) diagnosis of depression before completion of last Oxford WebQ, (2) missing socioeconomic factors (Townsend deprivation index, total household income, ethnic background, highest qualification, or overall health rating), (3) missing data of the physical exam [body mass index (BMI), systolic blood pressure (SBP)], (4) being in the upper 0.1% of total energy and/or carbohydrate intake or total energy intake of 0 kJ/day, (5) malabsorption, and (6) missing lifestyle risk factors (physical activity or smoking status) resulting in a study population of 188,426 participants. The diagnosis malabsorption was assessed at baseline by a verbal interview. Mean (range) age was 56 (39–72) years with 102,575 participants (54.4%) being female. The UK Biobank study was approved by the North West Multicentre Research Ethics Committee and written informed consent was obtained from all participants at baseline [25].

Consumption of sugar and sugar subtypes was calculated based on the Oxford WebQ data with methodology similar to recent studies [27, 28]. In brief, energy and total sugar contents were estimated for each Oxford WebQ questionnaire item based on McCance and Widdowson’s The Composition of Foods and its supplements [26], the UK Data Archive Standard Recipes Database [29], and product labels. The procedure to estimate FS content of the food items is based on Wanselius et al. [30] and summarized in Online Resource 4. FS were divided into FS in beverages and FS in solids. The following sugar-containing beverage subtypes were defined: soda/fruit drinks (the following items were assessed during baseline: carbonated (fizzy) drinks, fruit drinks, J20, squash, cordial, excluding low calorie or diet drinks), pure juice (i.e., fruit and vegetable; indicated as “juice” throughout the manuscript), milk-based drinks (i.e., dairy/yogurt-based smoothies, yogurt drinks, flavoured milk or milkshakes, hot chocolate or other milk-based drinks, excluding plain milk), and sugar added to tea/coffee. For tea/coffee, participants could choose the amount of added sugar per drink as half, one, two, or three teaspoon(s), as well as “varied”. This number of teaspoons was multiplied by the total number of cups of coffee and tea consumed, respectively. Since one teaspoon was the most commonly chosen portion size for added sugar in tea/coffee, this amount was set if participants indicated “varied”. The amount of sugar in tea/coffee was reported in the Oxford WebQ on each occasion the questionnaire was filled out. In total, “varied” was indicated on at least one occasion by 1004 participants for sugar added to tea/coffee. The following sugar-containing solids subtypes were defined: treats (i.e., pastries, candies, chocolate, ice cream, sweetened yoghurt), breakfast cereals (i.e., all food items labelled as “cereal(s)”, porridge, muesli, shreddies; indicated as “cereals” throughout the manuscript), toppings (i.e., table sugar, jam, honey, syrup, peanut butter, chocolate/nut spread, stewed/cooked fruit), and sauces (i.e., all food items labelled as “sauce(s)”, “salad cream”, mayonnaise, ketchup, chutney, salad dressing, pesto, gravy). Standard portion sizes were taken from the UK Food Standards Agency [31] and product labels. For each participant, the intake (g/day) of the specific sugar subtype was calculated by multiplying the frequency of each food item with the estimated content of this sugar subtype in that item in a portion. Intrinsic sugars were calculated as the difference between total sugars and FS. Sugar subtype intake in kJ/day was calculated by multiplying the intake in g/day with 17 kJ/g. Sugar subtype consumption in % total energy (%E) was calculated as follows according to Willett and co-workers [32]: Sugar subtype intake in kJ/day × 100%/total energy in kJ/day. UK Biobank participants could fill out the Oxford WebQ on up to five occasions. For participants who completed more than one questionnaire, the mean %E intake of sugar subtypes was used for all primary analyses.

Linked morbidity data are provided by UK Biobank as the earliest record date and respective health outcome defined with three-character International Statistical Classification of Diseases and Related Health Problems, Tenth Revision (ICD-10) codes [33]. Sources for these morbidity data are self-report at baseline assessment, as well as primary care, inpatient hospital, and death record data [33]. In the current study, the primary outcome was incident depression defined as ICD-10 codes F32 and F33. Follow-up time was calculated by subtracting the date of the baseline assessment from the date of the first diagnosis of depression, loss-to-follow-up, death, or censoring (i.e., December 31, 2021), whichever came first. The shortest duration to diagnosis was used in case of both F32 and F33 diagnoses in a patient.

All data were analysed with R version 4.0.5 [34] as described recently [35, 36]. The hazard ratios (HR) for incident depression were assessed with Cox proportional hazard regression multivariate nutrient density models [32] including %E intake of sugar from different sources and energy intake as penalized cubic splines with their degrees of freedom set to 4. Besides energy intake, models were adjusted for age (split by quintiles), alcohol intake (< 1, 1 to < 8, 8 to < 16, ≥ 16 g/day), BMI (< 18.5, 18.5 to < 25, 25 to < 30, ≥ 30 kg/m2), ethnic background (White, group composed of Mixed, Asian, Black, Chinese, and other), general health status (poor, fair, good, excellent), highest qualification (none of the below, national exams at age 16 years, vocational qualifications or optional national exams at ages 17–18 years, professional, College or University), history of mental illness (yes, no), physical activity [metabolic equivalent of task (MET)-minutes per week derived from the Oxford WebQ; split by quintiles], SBP (split by quintiles), sex (female, male), smoking status (never, previous, current occasional, current < 10, 10–14, 15–19, ≥ 20 cigarettes per day), total household income (< 18, 18 to < 31, 31 to < 52, 52 to < 100, ≥ 100 k£, unknown), and Townsend deprivation index (split by quintiles). Hazard proportionality was assessed for each covariate based on scaled Schoenfeld residuals. All covariates violating the proportional hazard assumption significantly after Holm-adjustment for multiple testing were stratified in the final models.

In each analysis, determination of the nadir of the estimated HR as a function of the intake of a sugar subtype in %E was restricted to the range from zero to the 99%-quantile. To simplify presentations, the HR was then rescaled to a nadir of 1. HR with pointwise 95% confidence intervals (CIs) are shown for all Cox proportional hazard regression models. The analysis of each penalized cubic spline is separated into plin for the linear and pnon−lin for the nonlinear effect as described recently [35].

Several sensitivity analyses were run similarly as described in a recent study [37] to check the robustness of the findings. Reverse causation was considered by excluding participants lost to follow-up or diagnosed with depression within 2 years after baseline (landmark analysis) and by excluding participants who had lost weight unintentionally. To remove implausible energy intake data, participants with under-reporting, i.e., < 1.1 × basal metabolic rate—500 kcal, or over-reporting, i.e., > 2.5 × basal metabolic rate + 500 kcal were excluded from the analysis. Basal metabolic rate was defined according to the Oxford equation [38]. To control for unrepresentative consumption data, participants who reported their previous day´s diet as non-typical on at least one occasion were excluded. To assess whether the portion size “varied” for sugar added to tea/coffee affected the results, all participants indicating “varied” on at least one occasion were removed from the analysis. To focus on nutrient intake closest to baseline assessment, analyses were repeated using the first Oxford WebQ questionnaire only. To apply alternative measures for body composition, waist-to-hip ratio (WHR) and height instead of BMI were used. To further control for residual confounding by dietary factors, a diet quality score was included in the analysis combining five dietary components, i.e., fat, fruit, vegetables, red meat, and processed meat consumption as described by Anderson et al. [37]. Minimum and maximum instead of mean sugar subtype and energy intake levels were assessed in two additional sensitivity analyses to consider lowest and highest consumption levels reported. To assess sex-dependent differences, sensitivity analyses were conducted in females and males separately.

A p value of < 0.05 was regarded as statistically significant in all analyses. If both plin and pnon−lin were non-significant, no further interpretation of the HR-nadir or other individual HR was performed.