Articles were identified by searches of PubMed and EMBASE through September 30, 2012. Details of search strategies appear in eTable 1 at http://www.jama.com . No language restrictions were applied. All articles were reviewed for inclusion by 1 reviewer (K.M.F.). An independent review of all articles was conducted by a second set of reviewers (B.K.K., H.O., and B.I.G.). The articles were reviewed to identify those that used standard BMI categories in prospective, observational cohort studies of all-cause mortality among adults with BMI measured or reported at baseline. Studies that addressed these relationships only in adolescents, only in institutional settings, or only among those with specific medical conditions or undergoing specific medical procedures were excluded. We included multiple articles from a given data set only when there was little overlap between articles by sex, age group, or some other factor.
In some cases, authors used standard BMI categories for overweight and obesity but had used a slightly broader reference BMI category of less than 25 or a slightly narrower reference BMI category of 20 to less than 25 rather than the standard normal BMI category of 18.5 to less than 25. We included these articles but have noted the cases in which the reference BMI category was less than 25 or 20 to less than 25. We classified studies that included a mix of self-reported and measured weight and height according to the preponderant type.
Abstracted items included sample size, number of deaths, age at baseline, length of follow-up, HRs and 95% confidence intervals, sex, age, type of weight and height data (measured or self-reported), country or region, source of study sample, adjustment factors, exclusion and inclusion criteria, and sensitivity analyses. Authors of screened articles were queried for additional information when necessary. In studies that only presented results stratified by smoking or health condition, we selected results for nonsmokers or never smokers or for those without the health condition. We selected the most complex model available for the full sample and used a variety of sensitivity analyses to address issues of possible over-adjustment or underadjustment.
We categorized HRs into 2 age groupings either as limited solely to people aged 65 years or older or as a mixed-age category (eg, aged 25–64 years or 40–80 years). We classified articles as adequately adjusted, possibly overadjusted, or possibly underadjusted. We categorized HRs by adjustment level, by whether the data were measured or self-reported, by whether the analysis was performed separately for men and women or for both sexes combined, and by region (North America, Europe, and other).
We used a random-effects model5 (link) to summarize the results overall and within subgroups and based statistically significant heterogeneity on a P value of less than .05. We calculated the quantity I2 to describe the degree of heterogeneity with values of 25%, 50%, and 75% considered low, moderate, and high, respectively.6 (link) We also used a sequential approach similar to that described by Patsopoulos et al7 (link) to assess consistency of findings when heterogeneity was reduced. All analyses were performed with SAS version 9.3 (SAS Institute Inc).
For sensitivity analyses, we examined the effects on HRs of incorporating results from a recent large pooled study for overweight.8 (link) For comparative purposes, we also constructed approximate HRs relative to normal weight from several recent large studies9 (link)–14 (link) that had used finer BMI groupings and thus did not meet our inclusion criteria. To do this, we averaged HRs from the finer BMI groupings over groups corresponding to the standard BMI categories, weighting the HRs by the number of deaths.
In some cases, authors used standard BMI categories for overweight and obesity but had used a slightly broader reference BMI category of less than 25 or a slightly narrower reference BMI category of 20 to less than 25 rather than the standard normal BMI category of 18.5 to less than 25. We included these articles but have noted the cases in which the reference BMI category was less than 25 or 20 to less than 25. We classified studies that included a mix of self-reported and measured weight and height according to the preponderant type.
Abstracted items included sample size, number of deaths, age at baseline, length of follow-up, HRs and 95% confidence intervals, sex, age, type of weight and height data (measured or self-reported), country or region, source of study sample, adjustment factors, exclusion and inclusion criteria, and sensitivity analyses. Authors of screened articles were queried for additional information when necessary. In studies that only presented results stratified by smoking or health condition, we selected results for nonsmokers or never smokers or for those without the health condition. We selected the most complex model available for the full sample and used a variety of sensitivity analyses to address issues of possible over-adjustment or underadjustment.
We categorized HRs into 2 age groupings either as limited solely to people aged 65 years or older or as a mixed-age category (eg, aged 25–64 years or 40–80 years). We classified articles as adequately adjusted, possibly overadjusted, or possibly underadjusted. We categorized HRs by adjustment level, by whether the data were measured or self-reported, by whether the analysis was performed separately for men and women or for both sexes combined, and by region (North America, Europe, and other).
We used a random-effects model5 (link) to summarize the results overall and within subgroups and based statistically significant heterogeneity on a P value of less than .05. We calculated the quantity I2 to describe the degree of heterogeneity with values of 25%, 50%, and 75% considered low, moderate, and high, respectively.6 (link) We also used a sequential approach similar to that described by Patsopoulos et al7 (link) to assess consistency of findings when heterogeneity was reduced. All analyses were performed with SAS version 9.3 (SAS Institute Inc).
For sensitivity analyses, we examined the effects on HRs of incorporating results from a recent large pooled study for overweight.8 (link) For comparative purposes, we also constructed approximate HRs relative to normal weight from several recent large studies9 (link)–14 (link) that had used finer BMI groupings and thus did not meet our inclusion criteria. To do this, we averaged HRs from the finer BMI groupings over groups corresponding to the standard BMI categories, weighting the HRs by the number of deaths.
