The results in Table show a substantial amount of misclassification of the BMI based on self-reported height and weight (self-reported BMI) compared with the BMI based on measured height and weight (measured BMI). More than 43% of respondents classified as "underweight" and 16% of respondents classified as "overweight" based on their measured BMI were classified as "normal weight" using the self-reported BMI. In addition, 19% of respondents classified as "obese" using measured BMI were misclassified as "overweight" using self-reported BMI. The general trend is for classification errors to be larger in the extreme (over- or underweight) categories. Sensitivity values (the proportions of overweight or obese persons according to physical measurement, who are classified as overweight or obese according to their self-reported measures) are 91.4% for overweight or more and 83.3% for the obesity classification. The corresponding positive predictive values (the proportion of self-reported 'overweight' or 'obese' persons who actually are overweight or obese based on measured height and weight) are 95.8% and 93.9%, respectively. However, a closer look at the misclassifications reveals that the majority of the misclassified cases have BMI values within an interval of just one unit from the category boundary in question. For example, while 43.5% of adults classified as "underweight" based on their measured BMI were classified as "normal weight" based on self report, about three-fourths of these individuals (32.1%) had self-reported BMI values at the lower end of "normal" category --- with a self-reported BMI value between 18.5 and 19.5. Similarly, while 16.0% of overweight adults were misclassified as "normal weight" using self-reported data, nearly two-thirds of these overweight adults (10%) had a self-reported BMI values between 24 and 25; likewise, 19% of obese persons were misclassified as "overweight" using self-reported BMI, but more than half of these adults (10.6%) had a self-reported BMI value between 29 and 30. Finally, among extremely obese individuals whose self-reported BMI fell below 40, 9.7% actually had a self-reported BMI between 39 and 40. Generally, deviations of BMI values based on self-reported height and weight from BMI values based on measured height and weight are moderate: an estimated 56% have self-reported BMI values within a one-unit interval of their measured BMI, and 81.5% have self-reported BMI values within two units of their measured BMI (not shown). More extreme deviations tend to occur at the very top (underestimates) and bottom (overestimates) of the BMI distribution.
Cross-Classification of Measured BMI and Self-Reported BMI for Standard BMI Categories*
Table displays arithmetic means for self-reported (SR) and measured (M) heights, weights and BMIs, as well as the discrepancies (DIS) between the self-reported and measured variables. In addition to information about the standard errors, standardized effect sizes for the discrepancy scores are added. As the rightmost column shows, the overall population mean for self-reported heights represents a modest overestimate of one centimeter, while self-reported weight represents an underestimate of measured weight by 3/4 of a kilogram (1.6 pounds). The net result is that the mean population estimate of the BMI, based on the self-reported heights and weights, is lower than the mean estimate of the measured BMI by 0.59 units. The effect sizes of the height and weight discrepancies for the total population suggest that height overestimates contribute more to the total population BMI discrepancies than weight underestimates. However, the discrepancies in each BMI class hint at linear trends such that overestimates of self-reported height become larger with larger measured BMI categories. The weight discrepancies indicate overestimates of self-reported weight in lower BMI categories and underestimates of self-reported weight in the higher BMI categories. The net result is a linear trend towards declining self-reported BMI values relative to measured BMI values. A linear regression model predicting the discrepancy between self-reported and measured BMI values (BMIDIS) based on the measured BMI (BMIM) values leads to the following estimates: BMIDIS = 2.283 - 0.102 BMIM, with standard errors of 0.134 for the intercept and 0.005 for the regression coefficients. Using this equation, self-reported BMI values would be unbiased estimates (i.e., BMIDIS = 0) of actual (measured) BMI for persons with a BMI of 22.4 (2.283/0.102), but would overstate actual BMI at lower BMI values and understate it at higher BMI values. Furthermore, the effect sizes for the height and weight discrepancies in the underweight category suggest that weight overestimates play a large role in the BMI overestimates in this group. By comparison, the effect sizes among the obese indicate that both height overestimates and weight underestimates contribute in equal measure to BMI underestimates.
Means of self-reported, measured, and discrepancy scores of height, weight and BMI values by BMI categories based on measured height and weight
Three multivariate linear regressions were conducted to predict variations in the discrepancy variables, i.e., the over- or underestimates of height (R2 = 0.061, p ≤ 0.001), weight (R2 = 0.070, p ≤ 0.001) and BMI (R2 = 0.045, p ≤ 0.001). All three models included the same set of predictors: gender, age, race/ethnicity (non-Hispanic white, Mexican American, other Hispanic, non-Hispanic black, other), marital status (married, widowed, divorced, separated, single/never married, living with a partner), education (less than high school, finished high school/GED, some college or more), pregnancy status and household income (less than $20,000 vs. $20,000 or more). The first model showed that overestimation of height was greater among non-Hispanic blacks (+0.15, p < 0.05) and Hispanics other than Mexican Americans, (+0.43, p < 0.03), than among all other racial/ethnic groups. In the second model, overestimation of self-reported weight occurred among widowed respondents (+.40, p < 0.02) compared with married individuals, weight was underestimated among college educated individuals (-0.51, p < 0.01) compared with individuals with less than a high school education, while pregnant women underestimated their weight, on average, more than 5 kg (-5.43, p < 0.01). The net effect of race/ethnicity, marital status, education, household income and pregnancy status on over- or underestimating the BMI remained small: only college-educated individuals and pregnant women had self-reported BMI values that were significant underestimates (-.13, p < 0.01 and -2.07, p < 0.01 respectively) of their measured BMIs.
Figures , , are based on the same regression models and illustrate how age and gender are associated with the discrepancies between measured and self-reported height, weight and BMI. Both men and women overstate their height, particularly at older ages, although the extent of over-reporting is greater for men than for women, p < 0.01 (Figure ). Men also tend to overstate their weight, although by relatively small amounts (on average, less than 1 kg) (Figure ). In contrast, women understate their measured weight with the greatest understatement (on average, more than 3 kg) found among young women (Figure ). The net effect on the BMI of these patterns in self-reporting of height and weight is depicted in Figure . Although self-reported BMI understates measured BMI for both men and women across all age groups, the discrepancy is significantly greater (p < 0.001) for women except at the very oldest ages. For both genders, the most accurate BMI values based on self-reported height and weight are obtained between the ages 42 and 55, with larger underestimates of measured BMI among the youngest and oldest individuals.
Associations of Gender and Age with Discrepancies between Self-reported and Measured Height: Fitted Values.
Associations of Gender and Age with Discrepancies between Self-reported and Measured Weight: Fitted Values.
Associations of Gender and Age with Discrepancies between Self-reported and Measured BMI: Fitted Values.
Given the pattern that self-reported BMI values underestimate "true" BMI values, except at very low levels of BMI, and given that the discrepancies between self-reported and measured BMI vary systematically with height, weight, age, gender, pregnancy status, marital status, income, and, to a lesser extent, with race and ethnicity, we used these variables to predict measured BMI scores (Table ). The results from this regression model show that self-reported height and weight, in conjunction with a few demographic characteristics, account for more than 92% of the variation in measured BMI scores. (Note: We used a polynomial regression approach for height and weight and age, eliminating higher power terms, if they showed no significant effect on the dependent variable [24
].) The predicted BMI scores from this model represent "adjusted" BMI scores, which take account of all the predictor variables in the equation. While the mean discrepancy between the adjusted and measured BMI is close to zero, since it is a residual score, these discrepancies continue to show a systematic, though smaller, bias in relation to the measured BMI. The simple linear regression model using the measured BMI values as predictors of the adjusted discrepancy scores yields the following results: BMI-(Adjusted) Discrepancy = 2.204 - 0.078 BMI, with standard errors of 0.109 for the intercept and 0.004 for the coefficient. Based on this equation, the actual BMI value at which the adjusted self-reported BMI scores are unbiased is 28.3 (2.204/0.078). Adjusted self-reported BMI values in the range below 28.3 are overstated and are understated in the range above 28.3. However, the biases are not large; at an actual BMI of 18.5, the average overestimate is 0.78 BMI units, while the average underestimate at a BMI of 40 is 0.91 units.
Regression of BMI, based on Measured Height and Weight, on Self-reported Height, Weight & Demographic Predictors (NHANES 2001-2006)
Classification of individuals as overweight or obese is improved by employing the adjusted BMI score, predicted from the regression model shown in Table . The adjusted BMI score improves the sensitivity of being classified as overweight or more (BMI 25+) to 94.6% (from 91.4%) and the sensitivity of being classified as obese (BMI of 30+) to 91.5% (from 83.3%). Population estimates are also improved, as shown in Table . The three NHANES estimates of the percentages of overweight and obese individuals in the population show that the adjusted self-reported measure mirrors the BMI categories based on measured height and weight more closely than the unadjusted self-reported measure. The NHIS, which exclusively relies on interview data, yields a very low estimate of the percentage of obese U.S. residents, even when compared to the interview data from the NHANES. However, after applying the NHANES prediction model to the NHIS to adjust for systematic biases in self-reporting height and weight, population estimates of overweight and obesity among U.S. adults using data from the NHIS more closely approximate the NHANES (measured) estimates, although significant differences remain for the obesity category.
Estimates of U.S. Civilian Resident Population by BMI Categories using Physical Measures, Self-Reports, and Adjusted Self-Reports
Table shows selected results from five logistic regression models (three based on the 2001-2006 NHANES and two based on the 2001-2006 NHIS) used to predict the odds of having ever been diagnosed with diabetes, based on the BMI measures. All five multivariate logistic regression models include an identical set of covariates: gender, age, education, race/ethnicity and household income. From the odds ratios associated with the BMI measures (both in linear and quadratic form), it is apparent that the adjusted self-reported measures deliver risk estimates that are close to those obtained on the basis of the measured BMI scores. In particular, the estimates obtained from the adjusted NHIS data are almost identical in size to those for the BMI scores based on measured height and weight in the NHANES.
Odds Ratios predicting prevalence of Diabetes on the basis of measured, self-reported and adjusted self-reported (SR) BMI scores with a standard set of covariatesa
Finally, in order to provide independent evidence of the applicability of the adjustments of BMI scores based on self-rated height and weight, we used the regression weights from Table and created an adjusted BMI score using the 1999-2000 NHANES sample of adult respondents (N = 5448). As with the 2001-2006 NHANES data, the adjusted BMI score from the prediction model improves the sensitivity of being classified as overweight or more (BMI 25+) to 94.2% (from 91.2%) and the sensitivity of being classified as obese (BMI of 30+) to 90.9% (from 83.9%).