Weight change in middle adulthood and risk of cancer in the European Prospective Investigation into Cancer and Nutrition (EPIC) cohort

Obesity is a risk factor for several major cancers. Associations of weight change in middle adulthood with cancer risk, however, are less clear. We examined the association of change in weight and body mass index (BMI) category during middle adulthood with 42 cancers, using multivariable Cox proportional hazards models in the European Prospective Investigation into Cancer and Nutrition cohort. Of 241 323 participants (31% men), 20% lost and 32% gained weight (>0.4 to 5.0 kg/year) during 6.9 years (average). During 8.0 years of follow‐up after the second weight assessment, 20 960 incident cancers were ascertained. Independent of baseline BMI, weight gain (per one kg/year increment) was positively associated with cancer of the corpus uteri (hazard ratio [HR] = 1.14; 95% confidence interval: 1.05‐1.23). Compared to stable weight (±0.4 kg/year), weight gain (>0.4 to 5.0 kg/year) was positively associated with cancers of the gallbladder and bile ducts (HR = 1.41; 1.01‐1.96), postmenopausal breast (HR = 1.08; 1.00‐1.16) and thyroid (HR = 1.40; 1.04‐1.90). Compared to maintaining normal weight, maintaining overweight or obese BMI (World Health Organisation categories) was positively associated with most obesity‐related cancers. Compared to maintaining the baseline BMI category, weight gain to a higher BMI category was positively associated with cancers of the postmenopausal breast (HR = 1.19; 1.06‐1.33), ovary (HR = 1.40; 1.04‐1.91), corpus uteri (HR = 1.42; 1.06‐1.91), kidney (HR = 1.80; 1.20‐2.68) and pancreas in men (HR = 1.81; 1.11‐2.95). Losing weight to a lower BMI category, however, was inversely associated with cancers of the corpus uteri (HR = 0.40; 0.23‐0.69) and colon (HR = 0.69; 0.52‐0.92). Our findings support avoiding weight gain and encouraging weight loss in middle adulthood.


| INTRODUCTION
Obesity is an acknowledged risk factor for the development of major cancers of the digestive system (oesophagus [adenocarcinoma], gastric cardia, colon and rectum, liver, gallbladder, pancreas), the female reproductive system (postmenopausal breast, corpus uteri, ovary), the thyroid, renal-cell carcinoma, meningioma and multiple myeloma. [1][2][3] Body mass index (BMI) at a single time point, usually at study recruitment, is the most commonly used measure of obesity. 1 Given that in a cancer-free middle-aged population, neither excess muscularity nor sarcopenia would be particularly prominent, BMI attained at cohort entry would primarily reflect the state of the adipose depots at this time point. Nevertheless, this could not distinguish between a lifelong fat excess and a more recent fat accumulation. Weight change over time, on the other hand, may reflect age-related metabolic changes and may also be more relevant from a public health perspective, as it may clarify whether lifestyle modifications in a particular period of life could influence the risk of cancer.
From a developmental point of view, middle adulthood represents a transitional period between early and later life, during which weight reaches peak levels and changes relatively slowly. 4 While genetic factors determining energy balance would likely present earlier in life, during adolescence or early adulthood, lifestyle and hormonal factors, especially peri-menopausal hormonal changes in women, would likely determine weight change during middle adulthood. Middle adulthood also precedes the loss of lean mass, a major contributor to weight loss in later life. 5 This raises the question whether weight loss during middle adulthood can mitigate the influence of fat accumulated during early adulthood and whether fat accumulated during middle adulthood can further increase the risk of cancer.
Studies examining the association of short-term weight change in middle adulthood with cancer risk, however, are limited and inconclusive. Published reports have addressed mainly colorectal cancer, postmenopausal breast cancer, or endometrial cancer, with only a limited number examining cancers at other locations or a wider range of cancers in a single study and several focusing only on men or women. A common constraint has been the limited number of cases, especially for less frequent cancer types, precluding some studies from reporting on individual cancer sites (see Supplementary Table S1 and Table S2 for summary of references).
Our aim in the current study was to examine in a large cohort, the European Prospective Investigation into Cancer and Nutrition (EPIC), the association of prospectively evaluated short-term changes in weight and BMI category during middle adulthood with the risk of cancer development in the most common tumour sites and the major morphological subtypes.

| Study population
EPIC is a well-established, prospective, multicentre cohort examining the association of nutrition and lifestyle with cancer and other chronic diseases. 6 Participants, mostly aged 40-70 years, from 10 European countries were recruited between 1991 and 1999. In our study, we excluded 280 001 participants due to missing information on weight or confounders, extreme anthropometry or a prevalent cancer at the second weight assessment (details shown in Figure 1), in accordance with previous reports. 7,8 We additionally restricted the analysis to participants in the age range 40 to 70 years between baseline and the second weight assessment, in order to focus on weight changes during middle adulthood as opposed to changes in early adulthood or in the elderly.

| Anthropometric assessments
Anthropometric characteristics were assessed twice: at baseline and after a mean follow-up for weight change of 6.9 years. Weight was mainly measured and adjusted for clothing at baseline and was selfreported at the second assessment (see details in Supplementary Methods). 7,8 Average annual weight change, that is, weight change rate (kg/year), was calculated by subtracting weight at baseline from weight at the second assessment and dividing by the years between

What's new
Obesity is well known as a risk factor for multiple cancers.
What about gaining or losing weight mid-life? Here, the authors investigated the association between cancer and change in weight and BMI category during mid-life. Among 241,323 people, about a third gained weight and 20% lost weight during the study. Independent of starting weight, gaining weight was associated with several obesity-related cancers including cancers of the gallbladder, uterus, ovary, kidney, thyroid, breast after the menopause and in men pancreas. Losing weight was inversely associated with obesityrelated cancers overall, and specifically colon and uterine cancer. The authors conclude that public health interventions to support weight loss in middle age could help reduce cancer incidence. the two assessments, to account for the difference in the time interval between the centres. BMI was calculated as weight/ height 2 (kg/m 2 ).

| Cancer ascertainment
The outcome of interest was first primary cancer diagnosed after the second weight assessment. We defined cancer types, subtypes and morphologies according to the International Classification of Diseases for Oncology, as specified in Supplementary Table S3 and Reference   9. Participants diagnosed with a second (or third) cancer, as well as those with cancers with unconfirmed or behavioural codes other than 3 (malignant, primary site) were censored at the date of diagnosis of the first cancer. We defined breast cancer as premenopausal when the diagnosis was before 55 years of age in women premenopausal at the second weight assessment. We defined postmenopausal breast cancer as those diagnosed at age 55 years or later, irrespective of menopausal status at the second weight assessment, censoring women with breast cancer diagnosed before age 55 years. The group of obesity-related cancers included oesophageal adenocarcinoma, colorectal cancer (overall), cancers of the stomach (overall), liver (overall), pancreas, kidney, breast (postmenopausal), ovary, corpus uteri (overall), thyroid and multiple myeloma. Age restrictions (n = 55 127) age at baseline <40 years or age at the second weight assessment ≥70 years F I G U R E 1 Flow diagram of participants included in the current study. Superscript "a" indicates the percentage from the number of participants per country or centre in the total cohort; "b" indicates that the weight at baseline was measured in 68.8% of participants, except in France and Norway, where weight and height were self-reported, and in Oxford (United Kingdom), where correcting equations were used for self-reported weight (see details in Supplementary Methods); "c" indicates that the weight at the second assessment was self-reported in most centres, except Umea (Sweden) and part of the cohort from Bilthoven (Netherlands), where weight was measured (4.6%) and Oxford, where correcting equations were used for self-reported weight (7.8%); "n" is the number of participants; "ca" is the number of cancer cases; the exclusion criteria were applied sequentially, that is, each excluded participant was counted only once, in a single exclusion step 2.4 | Assessment of lifestyle and personal history Participants completed detailed questionnaires on lifestyle, diet and, in women, menstrual and reproductive history and use of exogenous hormones at both weight assessments. Variables were harmonised to enable compatibility between EPIC centres. 6 Supplementary Figure S1 shows the dichotomisation rules for menopausal status. We used more recent updates for incident cancer cases and lifestyle factors compared to the earlier EPIC reports on short-term weight change and risk of colorectal and breast cancer. 7,8 We further examined change in BMI category, defined according to the World Health Organisation as normal weight (NW, 18.5 to <25 kg/ m 2 ), overweight (OW, 25 to <30 kg/m 2 ) or obese (OB, ≥30 kg/m 2 ). We compared maintaining OW or OB BMI category at both assessments to maintaining NW BMI category as reference. We further compared changing the baseline BMI category to a higher or lower BMI category at the second weight assessment to maintaining the corresponding baseline BMI category as reference. We performed these comparisons by repeating the same model three times, using each of the maintaining BMI category groups as reference, and have shown only the comparisons of interest. Due to very small numbers, we excluded participants with BMI < 18.5 kg/m 2 (n = 4043) and those changing between NW and OB BMI categories (n = 559).

| Statistical analysis
We estimated hazard ratios (HR) (95% confidence intervals [CIs]) using delayed-entry Cox proportional hazards models, that is, entry was conditional on surviving to the start of cancer follow-up. The underlying time scale for survival analysis was age in years. The origin of time was age zero, that is, participants were considered at risk from birth, even though they were not observed until entering the cohort.
Entry time was age at the second weight assessment, which was the start of cancer follow-up. Exit time was age at diagnosis of the first incident cancer, or death, or last complete follow-up, whichever occurred first. Models with weight change as exposure were adjusted for baseline BMI (per 5 kg/m 2 increment), as this may influence associations with subsequent weight change. All models were adjusted for the time interval between the two weight assessments, to account for differences in total weight change.
We additionally stratified all models by study centre, sex (except for sex-specific cancers) and age at the second weight assessment in 5-year categories (one category below 50 years) and adjusted for major risk factors for cancer and weight change and potential confounders (see rational for selection in Supplementary Table S4): height, energy intake (log-transformed), fruit and vegetable consumption (log-transformed), attained education, smoking status and intensity, alcohol consumption, physical activity index and for women also the major determinants of oestrogen levels: menopausal status and indicators of ever use of exogenous oestrogens, that is, oral contraceptives and hormone replacement therapy (HRT) (categories are listed in Table 1). To enable comparability, we used the same set of adjustment variables for all cancer sites. Height, energy intake, fruit and vegetable consumption and education were assessed at baseline and the remaining covariates at the second weight assessment, complementing missing information with baseline assessments (Supplementary Table S5). To account for information missing at both time points, we performed multiple sequential imputations using chained equations (function mi impute in STATA-13) and created m = 5 imputed datasets (Supplementary Table S6). To account for variability within and between imputations, we derived the estimates of coefficients and standard errors using Rubin's combination rules (function mi estimate in Stata 13.0 11 ). We considered as stronger evidence for association P < .001, which corresponds to Bonferroni correction for 50 comparisons (the approximate number of examined cancer types), and a weaker evidence for association a P-value between .05 and .001.
For cancers observed in both sexes, we explored further heterogeneity by sex because some cancers have sex-specific incidence and some published studies include only men or women. We examined separately subgroups of men and women, additionally adjusting for menopausal status and use of oral contraceptives and HRT in women.
In sensitivity analyses, we excluded the first 2 years of follow-up, to mitigate possible reverse causality. To examine the influence of adjustment, we derived unadjusted HR estimates retaining only the stratification by study centre, sex (except for sex-specific cancers) and age.
We used R version 3.6.1 12 for management of data and results, and STATA-13 for statistical analyses. 11 3 | RESULTS

| Characteristics of study participants
Our study comprised 241 323 participants (31.3% men), with a mean age at baseline of 51.5 years. During a mean weight follow-up of 6.9 years, 20.0% experienced weight loss and 32.2% weight gain >0.4 to 5.0 kg/year (Table 1). Fewer participants experienced weight change to higher (13.0%) or lower (6.8%) BMI category at the second assessment (Supplementary Table S7 Used as covariates in women; all covariates were derived from questionnaires at the second weight assessment, except from education, energy intake and fruit and vegetable consumption, which were derived from questionnaires at baseline; n (%), number of individuals (percentage from total number in category or from total cohort size and cancer cases).

| Associations between weight change and cancer risk independent of baseline BMI
The main analyses are presented in Table 2

| Associations between change in BMI category and cancer risk
The main analyses are presented in Table 3  at both assessments was positively associated with obesity-related cancers overall and individually with oesophageal adenocarcinoma, HCC, cancers of the colon, gallbladder and bile ducts, pancreas, kidney, postmenopausal breast, ovary, corpus uteri and thyroid, but not Abbreviations: CI, confidence interval; HR, hazard ratio; SCC, squamous cell carcinoma; premenopausal, breast cancer diagnosed at age < 55 years in women premenopausal at the second weight assessment; postmenopausal, breast cancer diagnosed at age ≥ 55 years, irrespective of menopausal status at the second weight assessment; obesity-related cancers, oesophageal adenocarcinoma, cancers of the stomach (overall), colorectum (overall), liver (overall), pancreas, kidney, postmenopausal breast, ovary, corpus uteri (overall), thyroid and multiple myeloma. a HR estimates were obtained from Cox proportional hazards models including weight change as a continuous variable (interpreted as the risk associated with weight gain per one kg/year increment), stratified by study centre, sex (except for sex-specific cancers) and age at the second weight assessment and adjusted for baseline body mass index (per 5 kg/m 2 increment), height, education, energy intake, fruit and vegetable consumption (assessed at baseline), as well as for smoking status and intensity, physical activity, alcohol consumption and for female cancers also menopausal status (except premenopausal cancer), ever using oral contraceptive and hormone replacement therapy (at the second assessment) and time interval between the two weight assessments. b The models included weight change as a categorical variable and compared weight loss or weight gain categories to stable weight (−0.4 to 0.4 kg/year) as reference, with stratification and adjustments as in footnote a. *P < .05. **P < .001. postmenopausal, breast cancer diagnosed at age ≥ 55 years, irrespective of menupausal status at the second weight assessment; obesity-related cancers, oesophageal adenocarcinoma, cancers of the stomach, colorectum, liver, pancreas, kidney, postmenopausal breast, ovary, corpus uteri, thyroid and multiple myeloma; HR estimates were obtained from Cox proportional hazards models with change in BMI category between the two assessments (categorical) as exposure, stratification by study centre, sex (except for sex-specific cancers) and age at the second weight assessment and adjustment for height, education, energy intake, fruit and vegetable consumption (assessed at baseline), smoking status and intensity, physical activity, alcohol consumption and for women menopausal status, ever using oral contraceptive and hormone replacement therapy (at the second assessment) and time interval between the two assessments; participants with BMI <18.5 kg/m 2 and those changing between normal weight and obese (n = 4602, 1.9%)

T A B L E 3 Change in body mass index category in relation to cancer risk
were excluded from the analysis; HR estimates for categories with less than 10 cases are not shown; Groups maintaining stable overweight (OW-OW) or stable obese (OB-OB) at both assessments were compared to the group maintaining stable normal weight (NW-NW), while the groups with weight change to a higher or lower BMI category were compared to the group maintaining the corresponding baseline BMI category as follows: weight gain from normal weight to overweight (NW!OW) was compared to stable normal weight (NW-NW); weight gain from overweight to obese (OW!OB) and weight loss from overweight to normal weight (OW!NW) were compared to stable overweight (OW-OW); weight loss from obese to overweight (OB!OW) was compared to stable obese (OB-OB).
*P < .05.**P < .001. The evidence for all observed associations with change in weight or BMI category was considered weak, with P-values between .05 and .001.  Tables S12 and S13).

| DISCUSSION
We report positive associations of weight gain in middle adulthood with obesity-related cancers overall and individually with cancers of the gallbladder and bile ducts, postmenopausal breast, corpus uteri and thyroid, which were independent of associations with baseline BMI. Compared to maintaining NW, maintaining OW or OB BMI category was positively associated with most obesity-related cancers. Gaining weight to a higher BMI category was also positively associated with obesity-related cancers overall and specifically with cancers of the kidney, postmenopausal breast, ovary and corpus uteri, while losing weight to a lower BMI category was inversely associated with obesity-related cancers overall and specifically with cancers of the colon and corpus uteri. Thus in breast adipose tissue, leptin stimulates aromatase activity, which contributes to higher oestrogen levels and both promote breast cancer development, 19 while in HCC cell lines, oestradiol induces apoptosis and opposes the oncogenic actions of leptin. 20 Furthermore, obesity is not only accompanied by higher oestrogen levels but also by gonadal dysfunction, with higher testosterone levels in OB women and lower in OB men and changes in opposite directions with weight loss. 21 An additional constraint is the lack of studies specifically examining the dynamics of adipose-derived factors after weight changes in middle adulthood.
A notable pattern in our findings was the positive association of weight change in middle adulthood specifically with cancers promoted by oestrogens, either female specific or with higher incidence in women, despite the fact that maintaining OW or OB BMI category was positively associated with most obesity-related cancers. In agreement with our findings, observational studies have consistently reported positive associations of weight gain in middle adulthood with cancers of the postmenopausal breast 7,10,22,23 and corpus uteri 10,22,24 and inverse associations with weight loss, especially when intentional, 25 sustained, 26 or after bariatric surgery. 27 No associations, however, have previously been reported with ovarian cancer. 10,25 An involvement of weight change in oestrogen-driven cancers of the female reproductive system is not surprising, given that adiposederived oestrogens gain prominence in postmenopausal women, when gonadal oestrogen production decreases. 13 Nevertheless, although ERα activation promotes cancer development in the breast, ovary and corpus uteri, differences exist in the local regulation of ER expression, 28 oestrogen signalling 29 and aromatase transcription. 30 Leptin signalling pathways also differ. 31 44 Further, oestradiol suppresses in a dose-dependent manner the proliferation of oesophageal SCC cells 45 and, in women, HRT is inversely associated while menopause is positively associated with oesophageal SCC. 46 Furthermore, female mice have shown higher skin tumour resistance compared to male mice, with ovariectomy resulting in overexpression of ERα, suppression of ERβ and increased susceptibility to skin SCC, comparable to male mice. 47 In vitro, increased ERβ expression or stimulation with oestrogen agonists inhibits proliferation of SCC cells and promotes squamous cell differentiation. 48 It is thus possible that adipose-derived oestrogens mitigate the risk of tumours with SCC morphology.
Supplementary Discussion includes comments on cancers with sex-specific or limited evidence for association with weight change and a comparison of the current with previous EPIC studies on weight change and cancer risk.
A major strength of our study is the prospective assessment of weight. EPIC is also a large multicentre cohort, including both men and women from several European countries, with a variety of lifestyles and dietary patterns and a sizeable number of incident cancer cases. This provided large statistical power for analyses of the most common cancer sites, although statistical power was limited for the less common cancer sites, especially for change in BMI category and for the sex-specific analyses.
A major limitation of our study is that the second weight assessment was mostly self-reported, which can result in underestimating weight, especially in heavier individuals. 49 Therefore, weight gain may have been underestimated and weight loss overestimated, with no obvious way to anticipate in which direction associations were biased. 50 Although for most centres we had no centre-specific equations to predict measured from self-reported weight, examining the application of correcting equations has shown that they improve the distribution of BMI but do not remove the bias in the estimates for associations between self-reported BMI and different disease outcomes. 51 Nevertheless, BMI categories based on measured and selfreported weight were in good agreement in EPIC-Norway (Cohen's kappa = 0.73). 52 A further limitation was that there were only two weight assessments and we could not explore fluctuations in weight and weight cycling, that is, alternating gain and loss, which have been associated with a higher risk of cancer. 27 Furthermore, we focused on change in weight and BMI category, whereas other anthropometric parameters, for example, central rather than overall adiposity, could be relevant for cancer aetiology. 1 The large proportion of missing values at the second assessment of waist circumference limited this evaluation in EPIC. Finally, we could not distinguish between intentional and unintentional weight loss, which can have different effects, 25 but we examined the possibility of reverse causality by excluding participants with less than 2 years follow-up after the second weight assessment. The caveats of this approach are that removing individuals could potentially introduce selection bias, while removing cancer cases would reduce statistical power and subclinical cancer development may take longer than 2 years.

| CONCLUSIONS
Our findings confirm a positive association between maintaining OW or OB BMI in middle adulthood and most obesity-related cancers. In addition, independent of baseline BMI, weight gain in middle adulthood was positively associated with obesity-related cancers overall and specifically with cancers of the gallbladder and bile ducts, postmenopausal breast, corpus uteri and thyroid. Weight gain to a higher BMI category was also positively associated with obesity-related cancers overall and specifically with cancers of the kidney, postmenopausal breast, ovary, corpus uteri and in men pacreas, while weight loss to a lower BMI category was inversely associated with obesity-related cancers overall and specifically with cancers of the colon and corpus uteri. Our observations support public health interventions in middle adulthood advocating maintenance of BMI in the NW category, avoidance of weight gain, and weight loss when BMI is high, in order to reduce the risk of some obesity related cancers.
necessarily represent the decisions, policy or views of the International Agency for Research on Cancer/World Health Organisation.

DATA AVAILABILITY STATEMENT
For information on how to submit an application for gaining access to EPIC data and/or biospecimens, please follow the instructions at http://epic.iarc.fr/access/index.php.