Sustained Weight Loss and Risk of Breast Cancer in Women 50 Years and Older

A Pooled Analysis of Prospective Data

Lauren R. Teras; Alpa V. Patel; Molin Wang; Shiaw-Shyuan Yaun; Kristin Anderson; Roderick Brathwaite; Bette J. Caan; Yu Chen; Avonne E. Connor; A. Heather Eliassen; Susan M. Gapstur; Mia M. Gaudet; Jeanine M. Genkinger; Graham G. Giles; I-Min Lee; Roger L. Milne; Kim Robien; Norie Sawada; Howard D. Sesso; Meir J. Stampfer; Rulla M. Tamimi; Cynthia A. Thomson; Shoichiro Tsugane; Kala Visvanathan; Walter C. Willett; Anne Zeleniuch-Jacquotte; Stephanie A. Smith-Warner


J Natl Cancer Inst. 2020;112(9):929-937. 

In This Article


Study Population

The DCPP is an international consortium of prospective cohort studies established in 1991 to examine associations between dietary factors and cancer risk among women with no history of cancer (except nonmelanoma skin cancer).[15] The current analysis includes 10 cohorts from the United States, Australia, and Asia (Table 1) that had at least three surveys with reported [or measured[16,17]] body weights and breast cancer incidence data.[16–24] Data were harmonized across studies. Each study was approved by its respective institutional review board. The present study was restricted to women aged 50 years and older at first weight measure. Women who were alive and cancer free at the start of follow-up (between 1996 and 2004 depending on the individual study) were considered for the analytic cohort (n = 232 733). Women who did not report updated weight information (n = 48 039) or reported extreme values for BMI, weight, or height (BMI ≥ 59 kg/m2, height <1.2 m, height ≥2.0 m, or weight ≥225 kg; n = 3809) were excluded from the study population.

Case Ascertainment

Incident invasive breast cancer cases were confirmed by medical records or cancer registry linkage. Some studies identified additional cases through linkage to mortality registries. Estrogen receptor (ER) status was available for 89.7% of cases.

Weight Change Assessment

Weight change was assessed using surveys at three time points before breast cancer follow-up began (Figure 1). On average, the weight change assessment period was 10 years total, including two approximately 5-year intervals. For a given interval, stable weight was defined as plus or minus 2 kg. Amount of weight change was assigned based on the first interval, whereas the second interval was used only to determine whether the initial weight loss and/or gain was sustained. Weight loss in the first interval was categorized as follows: more than 2 to 4.5 kg, more than 4.5 to less than 9 kg, or more than equal 9 kg lost. Weight gain was categorized as follows: more than 2 to 4.5 kg, more than 4.5 to less than 9 kg, 9 to less than 13.5 kg, or more than equal 13.5 kg gained. These cut points were selected as those that could examine the most extreme amounts of weight change, with as much granularity as possible, while maintaining adequate sample size. We then evaluated whether the weight change that occurred during interval 1 was sustained, somewhat sustained, or not sustained during interval 2. A 24-category variable was used to describe weight change patterns across the two intervals (Table 2). The referent group was women with stable weight across both intervals. Sustained weight loss was weight lost in the first interval that was not regained in the second interval. Likewise, sustained weight gain was weight gain in the first interval that was not subsequently lost. The remaining categories included women whose weight changes showed more fluctuation between interval 1 and interval 2. Women with the same starting and ending weight could end up in different categories depending on whether they gained, lost, or maintained the same weight during the first interval.

Figure 1.

Schematic of weight loss intervals in relation to breast cancer risk among women aged 50 years and older in the Pooling Project of Prospective Studies of Diet and Cancer. BMI = body mass index; IQR = interquartile range (25th percentile–75th percentile).

Statistical Analysis

Data from all studies were aggregated into a single dataset. Follow-up time began after the last weight assessment and ended at the date of incident invasive breast cancer diagnosis unless the women were first censored for death, loss to follow-up, or administrative end of follow-up. Hazard ratios (HR) and 95% confidence intervals (CI)[25] were estimated using Cox proportional hazards regression.[26] Tight control for cohort-of-origin, age, and calendar year at the start of breast cancer follow-up was done by including these variables in the STRATA statement of the model. We further controlled for baseline BMI (normal = 18.5–<25 kg/m2; overweight = 25–<30 kg/m2; obese = ≥30 kg/m2, baseline physical activity (PA; low or no, medium, high as defined by each study), and postmenopausal hormone replacement therapy (HT) use (current at the start of breast cancer follow-up, not current) because we a priori expected them to be associated with both our exposure and outcome and we wanted to assess associations of weight change and breast cancer beyond the impact of BMI, PA, and HT. Furthermore, we found that several of the weight-change hazard ratios changed by more than 10% when controlling for these variables (Table 2). We also evaluated the following additional factors as potential covariates: height, age at first birth, parity, oral contraceptive use, history of benign breast disease, education, race, smoking status, and alcohol intake. We chose, however, the more parsimonious model because there was very little change to the hazard ratios when controlling for these additional variables (Supplementary Table 2, available online). The proportional hazards assumption was assessed using a likelihood ratio test, and no violations were detected.

A variety of subgroup and sensitivity analyses were conducted: stratifying weight change associations by HT, BMI, and PA; examining ER positive and negative tumors separately; recalculating weight-change hazard ratios after dropping one study at a time; limiting studies to the United States only; and examining percent weight change. In sensitivity analyses, we analyzed each study separately and then pooled the study-specific risk estimates using a random effects model. In these analyses, heterogeneity between cohorts was examined using the Q statistic[27] and I2 index.[28] All P values were based on two-sided tests and considered statistically significant if the P value was less than .05. All 95% confidence intervals were considered statistically significant if the null value of 1.00 was included. Analyses were conducted using SAS version 9.4 (Cary, NC).