Association between Body Mass Index and Cancer Survival in a Pooled Analysis of 22 Clinical Trials

In 2012, an estimated 69% of the U.S. adult population was overweight (body mass index [BMI] ≥ 25 kg/m²), and 35% of the population was obese (BMI > 30 kg/m²; ref. 1). Overweight and obesity are not only risk factors for multiple cancers but are now potentially prognostic factors associated with cancer survival. Given the high prevalence of overweight and obesity in the United States, and the potential for BMI to be an important factor associated with survival outcomes, we need a much better understanding of cancer patient populations at risk for poor BMI-related outcomes. Once this is known, it will be possible to develop and test interventions to improve cancer prognosis among specific high-risk populations.

Higher BMI at the time of a cancer diagnosis has been associated with poorer survival for specific cancer types, including breast cancer (2–5), prostate cancer (6–8), leukemia (9), and oral cancer (10). However, the data are inconsistent for breast cancer (11) and pancreatic cancer (12, 13), and no associations have been observed in ovarian cancer (14), colorectal cancer (15), lymphoma (14), pancreatic cancer, or esophageal cancer (16). In contrast, higher BMI at diagnosis has been associated with better prognosis in acute myeloid leukemia (17), multiple myeloma (18), head and neck cancers (19), and among women with non-Hodgkin lymphoma (9). There are scant data on the effects of BMI at diagnosis among other populations of cancer patients, and few analyses have made comparisons across multiple cancer types or across different stages of cancer diagnosis.

We analyzed the association between baseline BMI and cancer survival using data from 22 clinical trials of common cancer treatments conducted across 14 cancer types and stages within SWOG, formerly known as the Southwest Oncology Group, a member of the National Cancer Institute's National Clinical Trials Network (1UG1CA189974-01).

Previously published SWOG phase II and III clinical trials with ≥5 years of follow-up, BMI at trial enrollment, and survival data were selected for inclusion. Baseline and survival data were collected according to study-specific case-report forms. All trials included in this analysis were approved by local Human Investigations Committees and all patients provided written informed consent.

Overall survival was analyzed by BMI for 14 common treatments across multiple cancers. Underweight patients (BMI < 18.5 kg/m²) were excluded given they comprised a small proportion (2%) of total patients and may have had a qualitatively different health status. Trials of similar histology and stage were combined to increase power to detect potential outcome differences by BMI. Within a disease, analyses were limited to patients on similar treatment regimens. For a given treatment regimen, study arms of similar histology and stage were combined to increase power. The 14 cancer/treatment combinations defining the datasets are shown in Supplementary Table S1, including trials of early stage (n = 4) and advanced (n = 18) disease.

To identify a BMI cutpoint that would best differentiate between groups, we examined the predictive value of BMI for overall survival as a binary indicator variable across a range (22–35 kg/m²) of potential BMI cutpoint values. BMI cutpoints, which defined a subgroup with <10% of overall patients, were excluded due to power considerations. For each treatment analysis, the χ² statistic associated with the effect of the BMI indicator variable on survival in a multivariable regression model was generated. Survival times across the 14 analyses were iteratively truncated in one-year increments from year 1 of follow-up to year 10 of follow-up, to allow a fair comparison across the diverse set of cancer types with different durations of survival. The BMI cutpoint was identified as that level which maximized the average χ² associated with the effect of BMI on outcome, and was used as the base case for the subsequent survival analyses.

Using the identified BMI cutpoint, overall survival was analyzed by BMI status above and below the cutpoint within each cancer/treatment cohort. Overall survival was defined as the time between study registration and date of last contact (censored) or death due to any cause. Estimates of overall survival were derived using Kaplan–Meier estimates (20). Multivariable analyses using Cox regression were performed (21). The base case analyses were based on overall survival truncated at 5 years, which across a diverse set of cancers and prognoses, likely represents the window of follow-up during which treatment has its greatest impact on outcomes. We evaluated whether patients with a higher BMI experienced a different overall survival, and whether patterns of overall survival by BMI status differed by patient sex.

An omnibus test (i.e., P value) of the global trend across studies was based on a one-sample t test of the study-specific (signed) χ² values representing the statistical significance for the BMI indicator variable in multivariable regression models. Secondary analyses included examination of the BMI cutpoint of 30 kg/m². To establish whether there was a durable association between baseline BMI and survival, conditional survival analysis examined survival patterns in the subset of patients who survived at least one year.

Regression models incorporated major clinical prognostic factors for each disease, plus demographic factors age (continuous), race (black vs. other; white vs. other), and sex (where appropriate). Each analysis was stratified by study-specific treatment arm to account for possible temporal differences, potential differences in administration/dose, and differences in patient populations not captured by the other prognostic factors. Two-sided P values are reported.

Analyses of the association of BMI and outcome were conducted for 14 different treatments based on 11,724 patients from 22 unique trials. Dataset sizes ranged from 241 ovarian cancer patients receiving paclitaxel to 3,145 breast cancer patients receiving doxorubicin, cyclophosphamide, and paclitaxel. Supplementary Table S1 provides disease, study, and regression covariate descriptions for each cancer/treatment cohort. Details on the eligibility criteria for each of the trials have been published previously (22).

Table 1 shows the distribution of BMI scores within analyses and averaged over all analyses. On the basis of raw totals, 68% of women and 66% of men were overweight or obese. Using percentages based on study-specific rates, 58% of women and 63% of men were overweight or obese.

Figure 1 shows the χ² statistic for each BMI cutpoint from 22 to 35 kg/m². Results indicate that a BMI of 25 kg/m² maximized the association of BMI and overall survival. The remainder of the primary analysis utilizes a BMI cutpoint of 25 kg/m² as the base case.

Figure 2 shows Kaplan–Meier curves for each of the 14 analyses with survival truncated at 5 years, using a BMI cutpoint of 25 kg/m². The HRs, 95% confidence intervals, and P values for overall survival by disease/treatment combination are shown in the forest plot in Fig. 3, in ascending order of the HR. Patients with elevated BMI had a lower risk of death for those treated with Bacillus Calmette-Guerin for bladder cancer (P = 0.02), treated with imatinib for gastrointestinal stromal tumor (P = 0.006), treated with carboplatin + paclitaxel for non–small cell lung cancer (P = 0.01), and treated with androgen deprivation therapy for prostate cancer (P = 0.01). In no cancer/treatment combination was elevated BMI associated with a statistically significantly higher risk of death. The simple (unweighted) overall mean HR across all cohorts was 0.96 (P = 0.06), indicating a nonsignificant trend of a lower risk of death among patients with elevated BMI in this set of 14 groups of trials. While not statistically significant, the HR for the analyses of breast and ovarian cancer consistently show a detrimental association of elevated BMI with survival. A sensitivity analysis using 3 levels for BMI (BMI < 25 vs. BMI 25–29.9 vs. BMI ≥ 30) as an ordinal categorical variable generated similar results in terms of the magnitude and direction of the association of BMI and overall survival (data not shown).

Figure 4,A illustrates forest plots for the association of BMI and overall survival grouped by sex. For males, elevated BMI tended to be protective for overall survival (mean HR = 0.82; P = 0.003), and for females, elevated BMI tended to be neither protective nor harmful for overall survival (mean HR = 1.04; P = 0.86). Results within male and female subgroups were similar when the analysis was limited only to regimens with dose based on body surface area (BSA; Fig. 4,B) and when sex-specific analyses were excluded (Fig. 4,C).

We examined the survival patterns among patients who survived ≥1 year (Supplementary Fig. S1 A). There was no evidence that BMI ≥ 25 kg/m² at baseline was better or worse among those who survived ≥1 year. The mean HR was 1.01 (P = 0.20). Similar to the analyses in all patients, there appeared to be sex-specific variation in the HR patterns. We further examined the interaction of BMI and sex among those who survived one year (Supplementary Fig. S1B and S1C). Consistent with analyses presented above, elevated BMI was protective on average for males (mean HR = 0.85; P = 0.03) and neither protective nor harmful for females (mean HR = 1.19; P = 0.21).

To understand the associations between being obese at diagnosis and cancer survival, we also examined patterns of survival based on a BMI cutpoint of 30 kg/m². Obese patients had an elevated risk of death among those treated with cytarabine and daunorubicin for acute myelogenous leukemia (HR = 1.18; P = 0.02) and those treated with cyclophosphamide, doxorubicin, and 5-fluorouracil for breast cancer (HR = 1.34; P = 0.03). In no other cancer/treatment combinations was elevated BMI associated with a statistically significantly higher or lower risk of death (Supplementary Fig. S2). The overall mean HR across all cohorts was 1.00, indicating no evidence of a trend toward higher or lower risk of death across these 14 groups of trials (P = 0.71).

We examined whether patterns of survival according to a BMI cutpoint of 30 kg/m² differed by patient sex. In a similar fashion as observed for a BMI cutpoint of 25 kg/m², obesity was associated with better overall survival for males (mean HR = 0.88; P = 0.04) but neither better nor worse overall survival for females (mean HR = 1.05; P = 0.28; Supplementary Fig. S3 A). These results were generally consistent within male and female subgroups when the analysis was limited only to regimens with dose based on BSA (Supplementary Fig. S3 B) and when sex-specific analyses were excluded (Supplementary Fig. S4).

To remove the possibility that cancer stage confounded the association between BMI and survival, we restricted the analyses to patients with advanced stage cancers. In these analyses, the results were consistent with the earlier analyses showing that among men, BMI ≥ 25 kg/m² was protective (mean HR = 0.84; P = 0.007) and among women, there was no association between the BMI cutpoint and overall survival (mean HR = 1.03; P = 0.61; Fig. 5).

In an analysis of 22 clinical trials representing 14 different cancer/treatment combinations, the association between BMI and survival was not consistent across cancer type or stage. There was no statistically significant evidence that BMI at a cutpoint of 25 kg/m² was associated with overall survival. However, there was good evidence that the relationship between BMI and cancer survival differed by sex. On average, elevated BMI was associated with better overall survival in males, but neither better nor worse overall survival in females. This association was maintained when sex-specific cancers were excluded, when treatment regimens were restricted to dose based on BSA, and when early-stage cancers were excluded. The results remained consistent when a BMI cutpoint of 30 kg/m² was examined and when analyses were restricted to individuals who survived at least one year following trial enrollment. The latter analysis emphasizes the potentially durable impact of BMI on long-term survival.

The majority of cancers included in this analysis were of advanced stage. The literature on the association between BMI and survival among individuals with advanced staged cancers is sparse and inconsistent. One reason for the inconsistency may be due to the timing of BMI assessment. A prior meta-analysis shows that timing of BMI assessment affects observed associations between BMI and colorectal cancer survival (23). Our analysis uses BMI data collected at trial enrollment, which is advantageous because it allows for examination of BMI as a prognostic factor assessed prior to treatment.

We observed no association between BMI and cancer survival among breast cancer patients, which is somewhat inconsistent with prior studies that reported an inverse association between BMI at breast cancer diagnosis and survival (3, 5, 11). However, each of the individual breast cancer trials included in this analysis showed a nonsignificant relationship in the same direction as prior studies. It is possible that the use of cutpoints at 25 and 30 kg/m² masked the potential association of very high (class II or III) obesity and worse survival, as suggested by an observational study in the After Breast Cancer Pooling Project, which showed an association of BMI and survival when class II or III obesity was attained (24). It is further possible that our analysis was underpowered and that a meta-analysis would be needed to better define this association.

We observed that men with advanced prostate cancer who had a higher BMI at trial enrollment had better survival. These results are inconsistent with some prior observational studies where higher BMI among prostate cancer patients was associated with poorer survival (6, 7, 25). It should be noted that in our analyses the trial of androgen deprivation therapy and the trial of chemotherapy for hormone-refractory prostate cancer both had similar findings, raising the possibility that the mechanism by which BMI affects survival may be independent of treatment effects. Prior studies have shown that obese patients with prostate cancer may be underdosed (26), which could account for poorer survival. However this is unlikely to be the explanation for our findings as underdosing would be uncommon in the clinical trial setting.

Perhaps our most provocative finding is that baseline BMI had a different association with survival for female and male patients. The finding held after removing potential confounding factors from the analyses, including sex-specific cancers, treatments not dosed on BSA, and early-stage cancers that may have had a better prognosis. Differences between sexes may be related to differences in the biology of the disease, or may be related to chemotherapy dosing, distribution, or tolerance. The role of muscle may be important in explaining sex differences. Prior studies have shown that muscle mass is associated with better cancer prognosis (27, 28). The ratio of muscle to fat is higher for men than women at each BMI category (29, 30), which may explain in part the interaction between sex and BMI with respect to survival outcomes. These differences need to be explored in future investigations.

The strength of this analysis is the systematic examination of the relationship between baseline BMI and survival across multiple cancer types and stages using a consistent analytic approach. As all participants were enrolled in a therapeutic clinical trial, and each cancer–drug combination was analyzed separately, the study design removes course of treatment as a prognostic factor and potential confounder. Prior studies have noted that the choice of BMI cutpoints and categories affects study results and inferences (31, 32). We used an innovative method to identify a meaningful analytic cutpoint to discriminate between BMI groups. The finding with respect to better overall survival for elevated BMI patients in males was remarkably consistent, regardless of which subsets of studies were examined. Despite these strengths, there are also limitations. We do not have data on BMI prior to diagnosis, during treatment, or in the posttreatment phase. These measures collected systematically would allow for a better understanding of whether change in weight following diagnosis is associated with cancer prognosis. The analyses also do not account for lifestyle behaviors or circumstances that could affect the relationship between BMI and survival, and we had no data on comorbidities or body composition. These measures would allow for better understanding of potential targets for modifiable behaviors and risk factors. We do not have data on breast cancer subtypes, which has been associated with prognosis (2, 33). It is possible as well that some of the individual analyses were underpowered to detect differences between BMI groups. Our analysis was based on comparing two groups of patients based on specified single cutpoints. However, it is possible that more complicated relationships between BMI and survival exist (such as a quadratic or U-shaped relationship using multiple cutpoints). Finally, as these are trial patients, compared with nontrial patients they likely have fewer comorbid conditions, especially severe comorbidities and end-organ damage, potentially limiting the generalizability of the findings.

These findings should be considered within the context of findings on BMI and overall mortality risk in the general population. National Health and Nutrition Examination Survey (NHANES) data show a U-shaped association between BMI and overall mortality (32, 34). Overweight status is associated with the lowest risk of mortality, grade 1 obesity (BMI 30–35 kg/m²) is unassociated with mortality and underweight and grades 2 (BMI 35–40 kg/m²) and 3 (BMI ≥40 kg/m²; refs. 32, 34) are associated with higher mortality. Furthermore, the risk associated with obesity was found to be age dependent, with the effect of obesity on mortality being lower among older individuals. There has been a great deal of discussion on how to interpret these findings; one issue is the possibility that illness-related weight loss may bias observational studies (35–37). Analyses of men from the Baltimore Longitudinal Study of Aging find that weight loss prior to death begins 9 years before death and for cancer-related deaths weight loss accelerated significantly 3 years prior to death (38). Therefore, it is possible that prior to enrollment into the trials analyzed here, subjects at highest risk of mortality were already experiencing weight loss. Another issue to be considered is that obesity is associated with comorbidities that may have excluded subjects from entering the trials and that the obese individuals entering the trials may have been systematically healthier for noncancer risks than obese individuals overall.

The American Cancer Society (ACS) and American Institute for Cancer Research (AICR) guidelines for cancer survivors recommend being as lean as possible without being underweight (39, 40). Our findings are observational and therefore do not provide definitive proof that BMI is causally related to cancer survival. Nonetheless, our findings, for men in particular, are inconsistent with this recommendation. In fact, these results suggest that recommendations for body size and weight management following a cancer diagnosis may be disease, stage, and even gender specific. The ACS and AICR guidelines are intended to apply across all cancer types, and are based upon hypothesized carcinogenesis mechanisms related to hormonal, inflammatory, and oxidative stress pathways. Our results suggest that more tailored recommendations may be in order. This could be particularly true in cancers with long survival times, wherein patients may be at risk for other BMI-associated diseases, including cardiovascular and metabolic disorders. Studies are needed to test the effects of weight management strategies in specific high-risk patient populations.

In summary, our findings suggest that the association between BMI and cancer survival is not consistent across all cancer types and stage, and that sex-related factors may interact with BMI and affect cancer survival. Higher BMI at the time of a cancer diagnosis may be protective for men. Among women there was much more variability in the association between BMI and overall survival. These findings suggest that the effect of BMI may be different for different cancer types, and that body size recommendations for cancer survivors that are disease, stage, and gender specific may be warranted.

No potential conflicts of interest were disclosed.

Conception and design: H. Greenlee, J.M. Unger, M. LeBlanc, S. Ramsey, D.L. Hershman

Development of methodology: J.M. Unger, S. Ramsey

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): J.M. Unger, D.L. Hershman

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): H. Greenlee, J.M. Unger, S. Ramsey, D.L. Hershman

Writing, review, and/or revision of the manuscript: H. Greenlee, J.M. Unger, M. LeBlanc, S. Ramsey, D.L. Hershman

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): J.M. Unger

Study supervision: J.M. Unger

We thank all of the patients who participated in the SWOG trials and the clinical research staff who facilitated their participation.

This work was supported by grants from the National Cancer Institute at the NIH (K23 CA141052 to H. Greenlee and 1UG1CA189974-01 to J.M. Unger, M. LeBlanc, S. Ramsey, and D.L. Hershman).