Study population
The UKB is a prospective cohort study of around 500,000 individuals aged 40–69 years enrolled between 2006 and 2010 across 22 centers located in England, Scotland, and Wales. At recruitment, information on socio-demographic characteristics, lifestyle factors, diet, anthropometry, and biological samples were collected [23]. Participants were followed from recruitment until the earliest of cancer, death, loss to follow-up, or end of the study period (between February 2020 and March 2021 depending on center).
EPIC is a prospective cohort with approximately 521,000 adults mostly aged 35–69 years at enrolment (between 1992 and 2000) from 23 research centers across 10 European countries (Denmark, France, Germany, Greece, Italy, Norway, Spain, Sweden, the Netherlands, and the UK) [24]. At recruitment, participants completed questionnaires covering socio-demographic, lifestyle, diet, and reproductive factors and anthropometric measurements and blood samples were also collected [24]. Participants were followed from recruitment until end of follow-up (i.e., last date of center- and event-specific ascertainment of CVD, T2D, or cancer, whichever came first), death, loss to follow-up, or end of the study [25].
As shown in Additional File 1: Figures S1 and S2, we excluded participants who had cancer, CVD, or T2D prior to enrolment in both UKB and EPIC. The rationale to exclude participants with a history of CVD or T2D was to avoid potential reverse causation (i.e., CVD or T2D affecting BMI at recruitment). We further excluded participants with missing values in any covariate (~ 21% and 3% in UKB and EPIC, respectively). All analyses were performed in a sample restricted to participants with no missing data (complete-case analysis). In EPIC, participants from France, Greece, and Norway were excluded due to the lack of data on incident events of CVD or T2D. Last, we excluded participants from Sweden due to uncertain dates for a majority of T2D diagnoses.
Anthropometry
Weight and height were measured by trained staff in both cohorts using comparable procedures [24]. In UKB, height was measured to the nearest centimeter using a Seca 202 stadiometer, and body weight to the nearest 0.1kg using a Tanita BC-418 body composition analyser [23]. In EPIC, body measurements were obtained using a standard protocol in all centres, except in Oxford (UK) where measurements were self-reported [26]. Depending on study center, height was measured to the nearest 0.1, 0.5, or 1.0 cm and weight to the nearest 0.1 kg [27]. BMI was calculated as weight/height2 (kg/m2) and categorized according to WHO definitions [1] into overweight/obesity (BMI ≥ 25 kg/m2) and obesity (BMI ≥ 30 kg/m2).
Ascertainment of cardiometabolic diseases
Incident cases of both CVD and T2D were coded using the 10th Edition of the International Classification of Diseases (ICD-10). In both cohorts, CVD was defined as a composite of ischemic heart diseases (I20-I25), atrial fibrillation (I48), and cerebrovascular disease (I60-I69), and T2D was defined as E11 (Additional File 1: Table S1) [28].
In the UKB, cases for both CVD and T2D were identified via linked hospital admissions records (primary diagnosis). The inpatient hospital data were obtained through linked medical records, mapped across England, Scotland, and Wales using the Hospital Episode Statistics in England, Scottish Morbidity Record, and Patient Episode Database for Wales.
In EPIC, the diagnoses of CVD were ascertained within the framework of the EPIC-Heart study using active follow-up through questionnaires, medical records, hospital morbidity registers, contact with medical professionals, retrieving and assessing death certificates, or verbal autopsy [29]. T2D cases were identified within the framework of the EPIC-Interact through a review of several sources including self-report, linkage to primary care registers, secondary care registers, medication use, hospital admissions, and mortality data, depending on the center [28]. Both EPIC-Heart and EPIC-InterAct were designed as case-cohort studies nested in the full EPIC cohort; however, the nested case-cohort design of these studies was not used in the current analysis.
Ascertainment of cancer
The outcomes of interest were the occurrence of any first primary malignant cancer (excluding non-melanoma skin cancer and in situ cancers), combined as overall cancer, and obesity-related cancers. Obesity-related cancers were defined as meningioma, multiple myeloma, adenocarcinoma of the esophagus, and cancers of the thyroid, postmenopausal breast, gallbladder, stomach, liver, pancreas, kidney, ovary, uterus, colon, and rectum (colorectal) [3]. Type of incident cancer cases were coded according to ICD-10 and information on tumor morphology and histology was ascertained using the 3rd Edition of the International Classification of Diseases for Oncology (ICD-O-3), detailed in Additional File 1: Table S1. In UKB, data on cancer diagnoses were provided by NHS Digital and Public Health England for participants from England and Wales and by NHS Central Register (NHSCR) for participants residing in Scotland and were ascertained through cancer registries. In EPIC, cases were identified by linkage to cancer registries for Denmark, Italy, the Netherlands, Spain, and the UK, and through a combination of health insurance records, cancer pathology registries, and active follow-up in Germany.
Statistical analysis
We used Cox proportional hazards regression to estimate cause-specific hazard ratios (HRs) and 95% confidence intervals (CIs) for associations with obesity-related cancers and all-cancers per 1 standard deviation (SD) increment in BMI (~ 5 kg/m2). The entry time was age at recruitment and the exit time was age at first primary cancer diagnosis, end of follow-up, loss to follow-up, or death, whichever occurred first. Deaths from any cause were modelled as a censored observation. Follow-up for CVD and T2D, i.e., the CMD, also started at age at recruitment and the exit time was the same as described for cancer.
Our base model was stratified by center of recruitment, age (5-year categories), and sex and adjusted for educational level, alcohol consumption, smoking status, height, physical activity, diet score, and in women additionally for use of hormone therapy and menopausal status (Additional File 1: Table S2). The menopausal status variable was allowed to change from pre- to post-menopausal among women, who turned 55 years during follow-up [30].We used a directed acyclic graph (DAG) [31] to identify confounding variables (Additional File 1: Figure S3). Our “overall adjusted model” further included CVD and T2D and their duration (time-varying accounting for non-linearity through natural splines) (Additional File 1: Figure S4). CVD and T2D were modelled as a time-varying variable with four categories (1: neither CVD nor T2D; 2: T2D; 3: CVD; and 4: T2D and CVD). Our main model further included a multiplicative interaction term between BMI (continuous) and the CMD variable (time-varying categorical) and estimates for the four categories were reported. We performed a likelihood ratio test to evaluate the multiplicative interaction between BMI and the CMD variable by comparing a model with the interaction term to a nested model without the interaction term.
We next evaluated the separate and joint associations of overweight and/or obesity (BMI ≥ 25 kg/m2 or BMI ≥ 30 kg/m2) and CMD (T2D, CVD, or both) with the risk of obesity-related cancer and total cancer. We created three variables each with four exclusive categories of combinations of overweight (and alternatively obesity) and in turn T2D, CVD, and both CMDs: (1) BMI < 25 kg/m2, without CMD of interest (reference); (2) BMI < 25 kg/m2, with CMD of interest; (3) BMI ≥ 25 kg/m2, without CMD of interest; and (4) BMI ≥ 25 kg/m2, with CMD of interest (joint effect). Based on this categorization, we quantified additive interaction for each of the three variables with the relative excess risk due to interaction (RERI) (as recommended in the STROBE statement [32, 33]) of the joint effect. The RERI was estimated as RERIRR = RR11 – RR10 – RR01 + 1, with RR11 the relative risk of being exposed to both factors (e.g., overweight/obesity and T2D), RR10 exposed to one of the factors (overweight/obesity) and RR01 to the other one (e.g., T2D). Estimations of 95% CI were based on the delta method described by Hosmer and Lemeshow [34]. A RERI of 0 means a lack of additive interaction. The model was adjusted for the same variables as our base model and further adjusted for the two binary variables of the other two CMD statuses not studied [35].
All models were fitted in each of the two cohorts separately and results then combined using random-effects meta-analysis [36]. The proportion of total variability due to between-study heterogeneity was assessed by I2.
All analyses were also performed stratified by sex. To investigate potential biases, the following sensitivity analyses were carried out. First, to investigate uncontrolled confounding and/or collider bias due to conditioning on CMD, we considered non-obesity-related cancers as a negative-control outcome, for which a null association of obesity with non-obesity-related cancer was expected. Second, we additionally adjusted our main models for CMD treatment (use of metformin and/or statins) to assess its impact on risk estimates. The data on CMD treatment were available only in the UKB. Third, to assess residual confounding by smoking, we ran the main models among never smokers only. Lastly, we also evaluated potential collinearity between BMI and height by replacing height with residuals of height, which we computed by regressing BMI on height. Comparing our main models to a model with adjustment for residuals of height instead of height by likelihood ratio test indicated that both models were equivalent.
Statistical tests were two-sided, and p-values < 0.05 were considered statistically significant. All analyses were performed using R version 4.1.2. Epi and InteractionR packages were used for the main analyses.
Rohit Malhotra is a medical expert and health journalist who offers evidence-based advice on fitness, nutrition, and mental well-being. His articles aim to help readers lead healthier lives.
