Study design
UK Biobank recruited more than 500,000 participants (aged 37–73 years, 56.3% were women) between 2006 and 2010 [16]. Participants attended one of 22 assessment centres across England, Scotland, and Wales, where they completed a self-administered, touch-screen questionnaire and face-to-face interviews [17, 18]. After excluding participants with a prevalent cancer diagnosis at baseline (n = 41,460), those with missing data for exposures and covariates (n = 21,064), and participants who were classified as underweight (n = 2629), 437,393 participants were finally included in the study. The outcomes defined for this study were incidence and mortality of overall cancer and 24 specific cancers. Of the 24 cancers, 17 were relevant to both men and women, two were specific to men (testicular and prostate cancer), and five were specific to women (breast, endometrium, uterine, cervix and ovary). The exposures were six adiposity-related markers, including BMI, WC, WHR, waist-to-height ratio (WHtR), hip circumference (HC), and BF%. The covariates were sociodemographic factors (age, ethnicity, education, and Townsend deprivation), smoking status, dietary intake (red meat, processed meat, fruit and vegetables, oily fish, and alcohol), physical activity, and sedentary behaviour. Additional cancer-specific covariates were added for women-related cancer (hormonal replacement, ages at first live birth, last live birth, and at menarche). Additionally, sun exposition was added as a covariate for melanoma cancer, and for lung, oesophageal, and oral cancer, we restricted the analysis to never smoker only. Association between adiposity markers and cancer mortality is likely the combined effect of adiposity’s association with incident cancer, and adiposity’s association with cancer fatality among cancer patients.
Procedures
Date of death was obtained from death certificates held within the National Health Service Information Centre (England and Wales) and the National Health Service Central Register Scotland (Scotland). Date and cause of hospital admissions were obtained through record linkage to Health Episode Statistics (England and Wales) and Scottish Morbidity Records (Scotland). Detailed information about the linkage procedures can be found at http://content.digital.nhs.uk/services. At the time of analysis, mortality data were available up to 01 June 2020. Mortality analysis was therefore censored at this date or date of death, whichever occurred earlier. Hospital admission data were available until 31 March 2017 for Scotland and Wales and until 01 June 2020 for England, resulting in analyses of incident outcomes being censored at this date or the date of relevant hospitalisation or death, whichever occurred earlier. We defined incident cancer as fatal or nonfatal events. The International Classification of Diseases, 10th revision (ICD-10), was used to define the following 27 cancers: overall cancer (C00–C97, D37, D48), oral (lip, pharynx and larynx) (C00–C14), oesophagus (C15) upper oesophagus (C15.0, 15.1, 15.3, and 15.4), stomach (C16) stomach cardia (C16.0), stomach non cardia (C16.1–16.6), colorectal (C18, C19, and C20), colon proximal (C 18.0–18.5), colon distal (C18.6, C18.7), colon (C18.0-C18.9), rectum (C19–C20), liver (C22), gallbladder (C23), pancreas (C25), lung (C34), malignant melanoma (C43), breast (C50), uterine (C54–C55), cervix (C53), endometrium (C54), ovary (C56), prostate (C61), testis (C62), kidney (C64-C65), bladder (C67), brain (C71), thyroid (C73), lymphatic and haematopoietic tissue (C81–C96), non-Hodgkin lymphoma (C82–C85), multiple myeloma (C90), and leukaemia (C91–C95).
The exposures were six adiposity-related markers (BMI, WC, WHR, WHtR, HC, and BF%) measured by trained staff using standardised protocols across the assessment centres at baseline. Height was measured to the nearest centimetre, using a Seca 202 stadiometer, and body weight to the nearest 0.1 kg, using a Tanita BC-418 body composition analyser. BMI was calculated as weight (kg) divided by height (m) squared and classified into the following categories: underweight (< 18.5 kg/m2), normal weight (18.5 to < 25 kg/m2), overweight (25 to < 30 kg/m2), and obese (> 30 kg/m2) [19].
BF% was measured using the Tanita BC-418 MA body composition analyser (fat mass divided by the total body mass).
The natural indent was used to measure WC (the umbilicus was used if the natural indent could not be observed) and used to determine central obesity (WC ≥ 88 cm for women and WC ≥ 102 cm for men). HC was recorded at the widest part of the hips. WHR and WHtR are the ratios of the waist-to-hip circumference and waist circumference to height, respectively.
Age, sex, ethnicity, smoking status, diet (portions of fruits and vegetables, red and processed meat, and oily fish) and alcohol intake (daily, 2–4 times a week, once or twice a week, 1–3 time a month, special occasions and never), sun exposition (do not go out in the sunshine, rarely, sometimes, most of the time, always), and female-specific factors were self-reported at the baseline assessment by touch-screen questionnaire. Townsend area deprivation index was derived from the postcode of residence using aggregated data on unemployment, car and homeownership, and household overcrowding [20]. Educational qualification was self-reported. Physical activity level over a typical week was self-reported using the International Physical Activity Questionnaire and reported as metabolic equivalent of task (MET) per week [21]. Time spent on discretionary sedentary behaviours was derived from the questionnaire and included time spent in front of a TV or computer or driving during leisure time. Further details of these measurements can be found in the UK Biobank online protocol (http://www.ukbiobank.ac.uk).
Statistical analyses
Cox proportional hazard models were used to estimate hazard ratios (HR) and 95% confidence intervals for each adiposity marker (BMI, WC, BF%, WHR, WHtR, and HC) separately with incidence and mortality for 24 cancers and all-cause cancer. Duration of follow-up was used as the timeline variable. The exposure variables were fitted separately on penalised cubic splines to investigate non-linear associations between each adiposity exposure and the outcomes. Penalised spline is a variation of basis spline [22]. Non-linearity was tested by likelihood ratio tests. To compare the associations between cancer across different adiposity markers, all adiposity exposures were standardised by sex and HR were expressed per 1-standard deviation increment (1-SD was equivalent to BMI units of 4.2 and 5.1 kg/m2, WC 11.3 and 12.5 cm, WHR 0.07 and 0.07, WHtR 6.5 and 7.9, HC 7.6 and 10.4 cm, BF% 5.8 and 6.9%, and BFI 2.6 and 3.8 kg/m2 for men and women, respectively). Participants with prevalent cancer at the baseline assessment were excluded from the study (n = 41,406). Underweight participants were also excluded from the study (n = 2629). In addition, a landmark analysis was performed to reduce the potential for reverse causality, with follow-up commencing 2 years after recruitment. The association between adiposity and oesophageal, oral, and lung cancer was restricted to participants who reported being never smokers, to avoid reverse causation bias. For breast cancer, all analyses were stratified by menopausal status. Additional sensitivity analyses were performed including underweight people and adding height as a covariate.
Population attributable fractions (PAFs), assuming causality, were calculated based on the BMI distribution of Health Surveys of England, Scotland, and Wales in 2018 [23,24,25] and the HRs derived from this study using the standard formula with 95% confidence interval (CI) and P values estimated using bootstrapping (formula shown in Additional file 1: Figure S1) [26].
To compare cancer risk discrimination between BMI and the remaining five adiposity markers, we calculated Harrell’s C-index (the probability of concordance between observed and predicted responses) for a model that included the adiposity marker and covariates (age, ethnicity, deprivation, education, smoking, alcohol consumption, intakes of fruit and vegetables, red and processed meat, oily fish, physical activity, and sedentary behaviours). The model with BMI was defined as baseline model. The C-indices of the baseline model and the C-index difference between other adiposity model and the baseline model were reported. The variance of the C-indices was calculated using the formula as described previously [27]. These were then used to calculate confidence intervals and P values using normal approximation.
Competing risk due to non-cancer mortality was handled using a cause-specific model [28]. Participants who died due to non-cancer causes were marked as censored at their date of death. This approach was used instead of the sub distribution proportional hazards model because there is no evidence that the competing events influence the risk of cancer events, and because the current study aims to investigate associations rather than absolute risk.
Finally, because of potentially inflated type I errors due to multiple tests, all analyses were corrected for multiple testing using Holm’s method [29], which performed similarly to Bonferroni’s method while retaining higher statistical power [30]. The multiple testing corrected P value are denoted as Padj for P value for testing overall significance against no association, and Pnonlinear for P value testing non-linearity.
All analyses were adjusted for age, sex, ethnicity, deprivation, education, smoking, alcohol consumption, intakes of fruit and vegetables, red and processed meat, oily fish, physical activity, and sedentary behaviours. Additionally, women-related cancer was further adjusted for hormonal replacement, age at menarche, and age at first and last live birth. Prostate cancer was additionally adjusted for family history of prostate cancer, and melanoma was further adjusted for sun exposure. All analyses were performed using R Statistical Software, version 3.6.2, with the package survival and pifpaf.