We used BMI as our basic indicator of obesity. Flegal and Graubard have shown that the proportion of deaths attributable to obesity does not vary significantly with the indicator chosen.10
In our baseline analysis, we assumed that the relative mortality risks in various BMI categories by age and sex that were recorded in a study by the Prospective Studies Collaboration (PSC) are applicable to all countries considered.5
The PSC study is the largest and most detailed of several large compilations of data on obesity and mortality.11
The synthesis includes data on 895000 participants from 57 prospective studies, of whom 63% were from Europe and Israel, 29% were from the United States and Australia, and 8% were from Japan. Results of the PSC investigation have been presented by sex, age group (35–59, 60–69, 70–79, and 80–89 years), and detailed BMI categories (2.5-unit intervals within the range 15.00–34.99 and a single interval for 35.00–49.99).5
We estimated population distributions of BMI from nationally representative survey data. Height and weight data for estimating an individual’s BMI are based on self-reports obtained through in-person interviews, except in Canada and England, where we used data on measured height and weight. For the United States, we used both self-reported and measured values.
We obtained data for European countries, excluding England, from the Survey of Health, Ageing and Retirement in Europe, including individuals interviewed in wave 1 (2004) as well as a refresher sample from wave 2 (2006–2007). We obtained data for England from wave 2 (2004–2005) of the English Longitudinal Study of Ageing. US data came from the National Health and Nutrition Examination Survey (NHANES) cycles 2003–2004, 2005–2006, and 2007–2008. Previous research has found no significant national trend in adult obesity for either sex during this period in the United States.12
Data for Canada were taken from cycle 3.1 (2005) of the Canadian Community Health Survey. We constructed period life tables by country, age (in single-year age intervals), and sex using data from the Human Mortality Database (HMD)13
on deaths and population in 2006.
To identify the proportion of deaths in a particular country, age, and sex category that were attributable to obesity, we hypothetically redistributed the population above the optimal BMI category (i.e., the lowest-mortality category) in that group to the optimal category and calculated the proportional reduction in mortality that would occur under this redistribution. This is known as the population attributable fraction (PAF). We constructed estimates of BMI prevalence in the same age-sex-BMI groupings as used by the PSC except that we applied the PSC mortality values for ages 35 through 59 years to ages 50 through 59 years. In the PSC, the lowest-risk BMI category is 22.50 to 24.99, except for males aged 80 through 89, for whom it is 20.00 to 22.49, and women aged 70 through 79, for whom it is 25.00 to 27.49. We used the term obesity to refer to all weight categories above the optimal, including overweight (BMI= 25.00–29.99). We did not change the proportion of persons below the optimal BMI category because our interest was in the effect of obesity on mortality. Throughout our analysis, we assumed the mortality risk from obesity to be zero after age 90. We estimated the PAF for population i (where i is an indicator for each country, age, and sex combination) as
=proportion of population i
in BMI category j
=death rate in BMI category j in the standard drawn from PSC data, and C*ij
=proportion of population i
in BMI category j
if all individuals above the optimal BMI were redistributed to the optimal category.
would give the same value of the PAF if the death rates were in the form of relative risks (e.g., if numerator and denominator were divided by the death rate in the optimal category).
We applied the country-, age-, and sex-specific PAFs to death rates in the HMD in single-year age intervals to estimate what these rates would be if no one were obese. We then calculated life expectancy at age 50 years using the modified death rates. Conventional methods of calculating life tables were used.14
We then compared hypothetical life expectancies obtained in this manner with the actual values, also computed from the HMD, by country and sex. To identify the extent to which the US shortfall in life expectancy is attributable to obesity, we compared differences in actual life expectancy between the United States and each country with the differences that would be expected in the absence of obesity. When Canada and England were compared with the United States, we used measured rather than self-reported heights and weights.
We conducted analysis of uncertainty for PAFs and life expectancy estimates using a bootstrapping procedure.15
We combined uncertainty estimates from 2 sources: uncertainty in the BMI data resulting from sampling variability and uncertainty in estimation of the relative risks. For each country, age, and sex combination, we sampled BMI values randomly with replacement as many times as there were nonmissing observations on BMI in that country, age, and sex category. To incorporate uncertainty from the relative risks, vectors of the underlying effect parameters of relative risks of length corresponding to the number of BMI intervals were drawn from independent normal distributions with age- and sex-specific standard errors provided to us by the PSC. We applied the resulting vectors of risks to the simulated BMI distribution data to obtain country-, age-, and sex-specific PAFs. We repeated these steps to obtain 500 estimates of each country-, age-, and sex-specific attributable fraction from which we extracted the 2.5 and 97.5 percentile values as 95% confidence intervals.
We explored the sensitivity of results to the assumed set of risks associated with obesity and to misreporting of height and weight. Flegal et al. have suggested that the relative risks of death associated with obesity have declined in the United States.16
To investigate the effect of a possible reduction in obesity risks on international comparisons, we introduced an alternative set of risk factors adapted from Adams et al. that applies to a more recent period.17
These were derived from a large study of 527000 enrollees in the National Institutes of Health–American Association for Retired Persons Diet and Health Study,17
which was conducted in 6 US states and 2 cities. Enrollees were followed from enrollment in 1995 and 1996 through the end of 2005. As in the PSC results, relative risks were adjusted for smoking. In contrast to PSC procedure, relative risks in Adams et al. were also adjusted for social status and physical activity.
We used the published results of Adams et al. to estimate relative risks in the age categories that were used in the baseline analysis using data from the PSC. To do so, we fit a linear age trend using weighted least squares to risks that were originally reported in 4 age intervals (50–65, 56–70, 61–75, and 66–81 years). From primary data, we recalculated the proportions in various BMI intervals in each country to align with the categories used by Adams et al. We approximated standard errors for uncertainty estimation because of the smoothing procedure we employed to obtain risks for the relevant ages.
Analysis of NHANES data shows that American women tend to underestimate their weight, whereas both men and women tend to overestimate height at older ages.18
To explore whether our results were sensitive to error in self-reports of height and weight, we replicated all analyses after correcting self-reported height and weight for misreporting, using an approach similar to one applied elsewhere.19
Using data on adults aged 50 years and older from NHANES 2003–2008, for each sex, we estimated linear regression models of measured height (weight) vs self-reported height (weight), age, and the square of age.
We conducted analyses using Stata 10.1 (Stata Corp, College Station, TX) and R 2.11.1 (R Foundation for Statistical Computing, Vienna, Austria).