This study analysed data from the Behavioral Risk Factor Surveillance System (BRFSS), a cross-sectional telephone survey of non-institutionalized adults in the USA, for the period between 1986 and 2005. The BRFSS has been used for tracking health behaviours over time and study details are documented elsewhere.12
Individuals were classified into weight categories based on their BMI (weight in kilograms divided by the square of height in meters) calculated from self-reported weight and height. In addition to the standard ‘obese’ category, defined as a BMI of greater than or equal to 30 kg/m2
, the main groups of interest were more extreme categories: BMI≥35, BMI≥40 (often referred to as morbid obesity and roughly corresponding to 100 pounds (45 kg) overweight), BMI≥45, and BMI≥50 (sometimes referred to as super obesity). There is a well-known tendency towards under-reporting weight and over-reporting height.13–15
The under-reporting of weight increases with weight and absolute levels of prevalence are therefore lower than if BMI were calculated by independent measurement. The effect on trends is probably minor, but the bias will underestimate the increase among the heaviest groups.
The statistical analysis used individual-level logistic regression with an indicator of a specific weight category as the dependent variable. Time trend was measured as a linear spline (in the log odds) with knots at 1991 and 1996, and 2001 (i.e., linear trends within each 5-year period, but trends can differ between 1986 and 1990, 1991 and 1995, 1996 and 2000, 2001 and 2005). The spline function smoothes estimates compared to year indicators, and was mainly needed because of the small sample sizes in the early years and in the heaviest BMI groups. The results were adjusted for socio-demographic changes to isolate the unique trend in obesity rates. Regressors included: age (in 5-year intervals), educational achievement (less than high school, high school, some college, college degree), racial group (white, black, Hispanic, other), and gender. State indicators were included to control for the changing survey participation by states over time. Tests were based on the regression model and all results were considered statistically significant at P<0.01 unless indicated otherwise. The adjusted results were based on the sociodemographic characteristics in the year 2005 survey.
The study design implies several limitations. First, telephone coverage varies by state and also by subpopulation. Telephone coverage averaged 97.6% for US states as a whole in 2003 (the midpoint of the new analysis), but non-coverage ranges from 1.1% in Connecticut and New Hampshire, to 6.6% in Mississippi.12
The second limitation, inherent in every form of survey data, is non-response. Biases may change over time and noticeable declines in participation or response rates have been reported, especially since the mid-1990s.16,17
A meta-analysis of research papers in epidemiology and public health found that participation rates during 1970–2003 changed between −0.54% and −0.67% per year in similar studies.17
The BRFSS was not exempt from this trend and the cooperation rate (median across states) dropped from 69.8% in 1996 to 53.2% in 2000.12
Since then, aggressive attempts to stem this decline raised the cooperation rate to 58.7% in 2005. Because sociodemographic characteristics of responders differ from the underlying population, post-stratification weights are used to adjust for telephone non-coverage and non-response.12
Despite strong sociodemographic patterns in response rates, biases from non-response were found to be minor in a nutritional survey.18
In another study of non-responders, sedentary lifestyle was found to be higher among initial non-responders.19
To the extent that this bias is relevant, it would suggest that the results here underestimate trends.
A subtler issue is that the social acceptability of obesity may change and alter biases in self-reported weight. This hypothesis can be tested with the National Health and Nutrition Examination Survey (NHANES), which collects both self-report and objectively measured height and weight over time.20
Because the sample size is much smaller (about 5000 versus 300 000 in BRFSS), NHANES cannot be used to estimate changes in morbid obesity rates. However, there is no evidence that people have become more willing to report higher weights. Instead, the gap between self-report and objective measured BMI is increasing over time and was 0.51 units in 1999/2000 and 0.56 units in 2003/2004. This increase in the gap remains statistically significant (P
<0.05) even after controlling for objectively measured BMI. Thus, if anything, the results here underestimate the increases in each weight category compared to data based on objectively measured height and weight.