Body mass index and mortality: a meta-analysis based on person-level data from twenty-six observational studies

Purpose

For this report, we examined the relationships between the conditions of being overweight and obese and mortality from all causes, heart disease, cardiovascular disease, and cancer.

Methods

We defined the categories of body weight according to level of body mass index, $\mathrm{BMI}=\frac{\mathit{wt}\left(\mathit{kg}\right)}{\mathit{ht}(m{)}^2},$ using classifications suggested by the National Institutes of Health and the World Health Organization. These classifications are as follows: “normal weight” is defined as BMI ≥ 18.5, but less than 25; “overweight” equals BMI ≥ 25, but less than 30; and “obese” individuals have BMIs ≥ 30. Our investigation is based on person-level data from 26 observational studies that include both genders, several racial and ethnic groups, and samples from the US and other countries. The database consists of 74 analytic cohorts, arranged according to natural strata including gender, race, and area of residence. It includes 388,622 individuals, with 60,374 deaths during follow-up. We use proportional hazards models to examine the relationships between the BMI categories and mortality, controlling for age and smoking status. We use random-effects models to assess summary relative risks associated with the overweight and obesity conditions across cohorts.

Results

The relative risks among the heaviest individuals for overall death, death caused by coronary heart disease (CHD), and death caused by cardiovascular disease (CVD) are 1.22, 1.57, and 1.48, respectively, when compared with the those within the lowest BMI category. The summary relative risk among the heaviest participants for death from cancer is 1.07.

Conclusions

We document once again, excess mortality associated with obesity. Our results do, however, question whether the current classification of individuals as “overweight” is optimal in the sense, since there is little evidence of increased risk of mortality in this group.

Introduction

The estimation and interpretation of the relationship between body mass index [$\mathrm{BMI}=\frac{\mathit{weight}\left(\mathit{kg}\right)}{\mathit{height}(m{)}^2}$(1)] and mortality remains a topic of considerable controversy. The inability to reach consensus on this question is largely a result of contradictory evidence. Epidemiological studies have reported the association between BMI and mortality to be positive 2., 3., 4., 5., J-shaped 6., 7., 8., 9., 10., inverse J-shaped (11), U-shaped 12., 13., 14., 15., 16., 17., nonexistent or even inverse 18., 19., 20.. In this report, we focus on the narrow issue of mortality among those individuals falling into three general categories, including normal weight, overweight, and obesity. These categories correspond to those suggested by both the National Institutes of Health and the World Health Organization, which define “normal weight” to be BMI ≥ 18.5, but less than 25, “overweight” to equal BMI ≥ 25, but less than 30, and “obese” individuals to have BMIs ≥ 30 (21). Participants with BMI < 18.5 are not included in our analyses. We present data from 26 observational studies to examine the relationships between these categories and mortality from all causes, coronary heart disease (CHD), all cardiovascular disease (CVD), and cancer.

The data included in this study comprise the Diverse Populations Collaboration (DPC). The Diverse Populations Collaboration examines the variation in the results of epidemiological investigations in population samples from several countries and cultures. Participation in the collaboration has continued to grow since its inception in 1996, and for this report, we include data from 26 studies. A brief description of the studies with a full list of investigators is included in the Appendix. For this investigation, we use person-level data from 26 studies that include samples from both genders, several racial and ethnic groups, and the US and other countries, offering wide geographic diversity. The analysis includes 388,622 individuals with 60,374 deaths. The list of studies included in our analysis is provided in Table 1 and references for these studies are provided in the Appendix.

Underlying cause of death was determined from death certificates according to ICD-9 in most of the studies (codes 410–414 and 429.2 as CHD), in a few studies death certificates and ICD-8 was used (codes 410–414 as CHD), and for a small number of studies cause of death was assigned by a panel of physicians.

As described above, we use categories of BMI levels suggested by The National Institutes of Health and the World Health Organization for our analyses (21).

• “Normal weight”: 18.5 ≥ BMI < 25.

• “Overweight”: 25 ≥ BMI < 30.

• “Obese”: BMI ≥ 30.

Many of the studies in our database contain natural strata specifying race, sex, treatment group, sample selection, and area of residence. When we include these strata, we have 74 analytic cohorts available, and we conduct separate analyses for these cohorts. For example, Table 1 notes that ARIC contained strata according to both gender and race. So there are potentially 4 strata to be constructed from the ARIC study. (All of the potential strata listed have two levels.) We fit proportional hazards models to the survival-time data for each of the 74 groups, incorporating BMI categories, age, and smoking status into the models. We use these models to examine four mortality groupings: all-causes of death, deaths due to CHD, deaths due to CVD, and deaths due to cancer.

We use the proportional hazards model to relate BMI categories to mortality (22). The model contains two indicator variables to designate the BMI groups to which participants belong and the normal range category used as the reference. In addition to the indicator variables, we include age and cigarette smoking status (yes/no) as covariates. The relative risks of mortality for overweight and obese individuals are calculated using a standard method by exponentiating the estimated proportional hazards coefficients. For each cause of death, we fit models to those groups that have at least 30 deaths from the cause. For some studies, however, particular causes of death were not available, and these restrictions result in 74 cohorts for all causes of death, 65 cohorts for CHD death, 64 cohorts for CVD death, and 58 cohorts for cancer death.

To examine heterogeneity of results and to summarize results across analytic groups, we used methods suggested by DerSimonian and Laird (23). We found significant heterogeneity for the four groupings of causes of death and, therefore, present summary estimates based on a random effects model (23).

After completing the analytic procedures described above, we conducted sensitivity analyses to determine whether the results could have been influenced excessively by a single cohort. This methodology consisted of reanalyzing all the meta-analyses, while omitting each of the cohorts, one by one. We concluded that the summary statistics were not significantly impacted by any particular cohort.

We also conducted separate examinations for the interactions between BMI and age and BMI and smoking for each of the four causes of death for each cohort. Results show only a few of the cohorts to exhibit significant interactions between BMI and the two covariates. When testing a large number of interactions, we would expect a few to be significant. Thus, we decided not to present the results of analyses using models containing these interaction terms.

All of our analyses were accomplished using the statistical package Stata (24).