Cdc Growth Chart Sas Program
INTRODUCTIONSince the early 1970s, the prevalence of obesity in the United States has increased by more than a factor of 3 in 6- to 19-y-olds and by 70% in 2- to 5-y-olds ( 1, 2). In response to these secular trends, much data on the weight and height of children and adolescents have been collected in schools ( 3– 7) and clinics ( 8– 11). Because the quality of these measured weight and height data can vary substantially, investigators must decide how to address extreme values that could represent errors. These decisions would influence estimates of the prevalence of obesity ( 12), trends over time, and associations with other characteristics.Cutoffs that are based on z scores of weight, height, and BMI (in kg/m 2) that are considered to be biologically implausible values (BIVs) 7 are included in the SAS program for the 2000 CDC growth charts ( 13, 14). These cutoffs, which were based on recommendations made by a WHO expert committee in the mid-1990s ( 15), have been widely used ( 3, 4, 10, 11, 16– 18), and extreme values are flagged by the program. However, other than stating that the proposed cutoffs are “likely to be errors and may be treated as missing” 15(p217), no justification was given for the specific cutoffs that were chosen.
Furthermore, it is difficult to know how the exclusion of BIVs would influence the accuracy of the prevalence of obesity.It is possible that the exclusion of these BIVs resembles other techniques that have been used to exclude data that are considered to be incorrect or confounded but have eventually been shown to result in various biases ( 19– 22). The application of these BIV cutoffs to data from the United States and other countries that have a high prevalence of obesity could exclude subjects who have very high, but correct, BMIs and, therefore, underestimate the prevalence of obesity. Estimates of the prevalence of obesity in the NHANES ( 1, 2) do not exclude BIVs.Our objectives were to evaluate the distribution of weight, height, and BMI in 2- to 19-y-olds in NHANES 1999–2000 to 2011–2012 and to assess the prevalence of BIVs on the basis of WHO cutoffs with the use of CDC growth charts. We also assessed other body-size measures to determine whether the BIVs are consistent with information on waist and arm circumferences and leg length. METHODS SampleThe NHANES uses a multistage, stratified, cluster-sampling design to select a representative sample of the US civilian, noninstitutionalized population. The surveys were approved by the Ethics Review Board of the National Center for Health Statistics; parental permission was obtained for minors z scores in CDC growth chartsBMI was calculated as weight divided by the square of height. Weight, height, and BMI z scores and percentiles were calculated with the use of CDC 2000 growth charts and expressed a subject’s values relative to those in children and adolescents of the same sex and age from 1963–1965 to either 1976–1980 (6- to 19-y-olds) or 1988–1994 (2- to 5-y-olds) ( 14).
Cdc Growth Chart Sas
Obesity was defined as BMI greater than or equal to the 95th percentile of these growth charts ( 24).Although z scores calculated from the L (power transformed for skewness), M (mean), and S (CV) variables ( 25) in the CDC growth charts capture the distribution of height, weight, and BMI of most subjects, projections into the upper tails are problematic ( 26, 27). This problem may be exacerbated in the CDC growth charts because the LMS variables were derived from smoothed percentiles rather than from the underlying distributions ( 28).Because the upper z score limit at most ages in the CDC growth charts is 3.0–3.5 ( 27), a modified z score ( z score M) approach was used to identify extremely high values. One-half of the difference between a z score of 0 and 2 was extended outwards to estimate a subject’s modified z score; this modification allowed for z scores that were much higher than 3.5. A somewhat similar approach was used in the WHO growth standards, but the extrapolation was based on the distance between 2 and 3 z scores ( 29, 30).These scores are referred to as a z score M or a modified weight z score (weight MZ), modified height z score (height MZ), and modified BMI z score (BMI MZ). The following ranges are currently flagged as BIVs by the SAS program for the 2000 CDC growth charts ( 13): weight MZ less than −5 or greater than +5; height MZ less than −5 or greater than +3; and BMI MZ less than −4 or greater than +5.We considered body-size measures below these cutoffs to be low BIVs and values above these cutoffs to be high BIVs. These cutoffs for weight and height were based on those proposed by the WHO in 1995, whereas the cutoffs for BMI correspond to those used for weight-for-height ( 15).Of the 146 subjects with a high BMI BIV, 7 subjects were missing information for both arm and waist circumferences.
Of the sixty-six 8- to 19-y-olds who had a high height BMI, 3 subjects did not have a leg-length measurement. Subjects with missing circumferences or leg lengths were excluded only from analyses that made use of these measurements. Statistical analysesAll analyses were performed in the R program ( 31), and with the exception of the quantile-quantile plots, all estimates used the sample weights and accounted for the complex survey design ( 32). Because many of the estimated proportions for the BIVs were close to zero, 95% CIs were calculated with the use of the modification proposed by Korn and Graubard ( 33).After describing the prevalence of BIVs for weight, height, and BMI, we examined the distributions of body-size measures with the use of quantile-quantile plots ( In these figures, the observed z scores M were plotted ( y axis) against their expected quantiles from a normal distribution ( x axis). Additional analyses ( 34) indicated that the distribution of positive z scores M for weight and BMI approximated an exponential distribution with rate parameters of 1.11 and 1.07, respectively. For BMI z scores M with this distribution, the probability of observing a value that was at least as extreme would have decreased by 66% (1 − exp(−1.07)) for each 1-unit increase. Although 2.5% of values in a normal distribution would be expected to have a z score M ≥1.96, the comparable percentage would be 12.3% for this exponential distribution.We examined the circumferences (waist and arm) and leg lengths of the 2- to 19-y-olds.
Both of the circumferences were highly correlated with BMI and, along with sex and age, accounted for ( R 2) 92–93% of the variability in BMI. The R 2 of a similar model that predicted height from leg length, sex, and age was lower (0.88); leg length was measured only in 8- to 19-y-olds. Because percentiles of waist circumference, arm circumference, and leg length were not estimated in the CDC growth charts, we used quantile regression ( 35– 37) to estimate these smoothed, weighted percentiles in the current study.
Rather than estimating the mean response by minimizing the sum of the squared errors as in ordinary least-squares regression, quantile regression minimized the absolute error around various percentiles of the outcome variable.Within each sex, we modeled the 85th, 95th, and 99th percentiles of the circumferences and leg length as a function of age with the use of restricted cubic splines with 5 knots ( 38). The circumferences of subjects who had a high BMI BIV were overlaid on various percentiles of the circumferences, and the leg lengths of subjects with a high height BIV were overlaid on the corresponding percentiles of leg length. If BIVs were data errors, it would be expected that the circumferences and leg lengths of subjects who had a BIV would be scattered throughout the distribution.We also examined inconsistent values of BMI and waist circumference with the use of Studentized (leave-one-out) residuals. These residuals, which indicate possible outliers, have a SD of 1.0 and are calculated by ignoring the current observation.
These residuals were derived from a regression model that predicted BMI from sex, age, waist circumference, and various interactions. BMI and waist circumference were log transformed in this model to improve normality.Although we also assessed the skinfold thicknesses of children and adolescents with BIVs, 50% of subjects with a high weight or BMI BIV were missing data on either the triceps or subscapular skinfold thickness. Of the 61 subjects who had a high BMI BIV and a non-missing subscapular skinfold thickness, 56 subjects had a skinfold thickness greater than the 95th percentile. RESULTSDescriptive characteristics of the children and adolescents are shown in Table 1. Overall, the prevalence of obesity was 16%, but it varied by sex, race-ethnicity, and age. Approximately 0.9% ( n = 277) of subjects had a BIV for weight, height, or BMI, and almost all instances ( n = 275) were due to extremely high values.
One child (a 4-y-old girl) had a high height BIV (135 cm) along with a low BMI BIV (8.0). Of subjects who had a high BIV, 138 participants had high weight, 92 participants had high height, and 146 participants had high BMI; 89 subjects had a high BIV for both weight and BMI. Twelve of the 92 subjects who had a high height BIV also had a high weight BIV.The prevalence of BIVs was consistently greater in boys than in girls and most BIV prevalences differed by race-ethnicity; estimates for height and BMI BIVs differed by age group. The exclusion of the 277 subjects who had a BIV from the analysis reduced the overall prevalence of obesity from 16.0% to 15.5%, but the prevalence of obesity in non-Hispanic blacks decreased by greater than one percentage point (from 20.3% to 19.0%). In non-Hispanic black boys, the prevalence of obesity decreased from 18.3% to 16.6%.Figure 1 shows the distribution of weight MZ, height MZ, and BMI MZ.
In contrast with the 275 children and adolescents who had a high BIV, there were only 3 subjects who had a low BIV (open triangles) as follows: a 15-y-old with had a height of 120 cm, a 2-y-old with BMI of 11.5, and a 4-y-old with BMI of 8.0. The distributions of both weight MZ and BMI MZ were skewed with very heavy right tails, and the distribution of positive values for both weight MZ and BMI MZ were consistent with an exponential distribution with rate parameters of 1.11 (weight) and 1.07 (BMI). With the possible exception of the very-highest value of height MZ, the distribution of extremely high height MZ values appeared to be fairly continuous with little evidence that the distance between the highest values were larger than might be expected. (This result was in contrast to that observed for 2 of the low BIVs.) Although levels of height MZ were approximately normally distributed, there were 92 subjects who had a high height BIV ( z score M ≥ 3.0).We examined the waist circumferences of children and adolescents who had BMIs that were considered to be implausibly high ( Figure 2, top panels).
Lines represent the 85th, 95th, and 99th percentiles of waist circumference in all 2- to 19-y-olds, and circles represent waist circumferences of the 113 subjects who had a BMI MZ between 5 and 6.9; the solid triangles represent circumferences of the 26 subjects with a BMI MZ ≥ 7.0. As noted in Table 1, the prevalence of BMI BIVs varied substantially by age ( Table 1).With only one exception, all of the children and adolescents with a high BMI BIV also had a waist circumference that was greater than the 95th percentile (middle smoothed lines); 79% of children had a waist circumference greater than the 99th percentile.
Furthermore, the very highest BMI MZ values (solid triangles) were generally associated with the highest waist circumferences. Fairly similar results were seen in analyses that focused on the arm circumferences of subjects with a high BMI BIV ( Figure 2, middle panels). None of the subjects who had a BMI BIV had an arm circumference less than the 95th percentile, and 74% of subjects had an arm circumference greater than the 99th percentile. Similar results were seen in subjects who had a high weight BIV with almost all of them having waist and arm circumferences greater than the 95th percentile (data not shown).Comparable analyses of the leg length of the 63 children and adolescents who had a high height BIV ( Figure 2, bottom panels) indicated that 47 subjects had a leg length that was greater than or equal to the 95th percentile, and 30 subjects had a leg length that was greater than or equal to the 99th percentile. (Leg length was measured only in 8- to 19-y-olds. However, there were 4 subjects who had a leg length less than the 85th percentile, and 1 subject who had a leg length that was slightly less than the 50th percentile.)We also assessed the Studentized (leave-one-out) residuals of a regression model that predicted BMI from sex, age, and waist circumference (data not shown). The 3 subjects with the largest absolute residuals had relatively high BMI (81st to 98th percentiles) together with a waist circumference less than the 10th percentile.
Of 100 subjects with the largest residuals, only 5 subjects had a high BMI BIV; 2 subjects had a waist circumference greater than the 99th percentile, 2 subjects had a waist circumference between the 98th and 99th percentiles, and 1 subject had a waist circumference between the 95th and 96th percentiles.Table 2 shows levels of several of the body-size measures for children and adolescents who had 1) a high BMI or weight BIV and one or both circumferences less than the 97th percentile (top 8 rows), 2) a BMI MZ ≥ 8 and a weight MZ ≥ 8 (middle 6 rows), or 3) a height MZ ≥3.75 (bottom 8 rows). (These cutoffs were used only to illustrate possible inconsistencies between BIVs and levels of circumferences and leg length.) Levels of BMI MZ in subjects who had at least one circumference less than the 97th percentile ranged from 3.3 to 5.6, whereas all waist and arm circumferences of subjects with a BMI MZ 8 and a weight MZ 8 were greater than or equal to the 99th percentile. All subjects who had a height MZ ≥3.75 also had a leg length greater than or equal to the 97th percentile with the exception of an 11-y-old boy who was at the 87th percentile.Table 3 shows the minimum and maximum values of the body-size measures by year of age, with age-specific sample sizes that varied from 1280 to 1774.
With the exception of 3-and 4-y-olds, maximum BMI MZ values ranged from 5.8 (10 y of age) to 9.2 (7 y of age) across ages; however, maximum values in 3- and 4-y-olds were ≥10.6. A fairly similar pattern across ages was also seen for the maximum values of weight MZ, and the highest height MZ (5.8) was also seen in 4-y-olds. Although the minimum and maximum values of weight, height, and BMI generally increased with age, there were some exceptions. For example, the lowest minimum BMI was seen in 4-y-olds ( 8), whereas a 16-y-old had the highest maximum weight (215 kg). Note that in contrast to the CDC/WHO cutoff for low BMI BIVs (−5 z scores M), the minimum BMI MZ values in 5- to 19-y-olds were approximately −3.0 (data not shown).There were substantial sex differences in the maximum values of weight, height, and BMI after 10 y of age (data not shown).
For example, in 19-y-olds, the maximum values in boys were 23 kg higher (weight), 16 cm higher (height), and 3.5 higher (BMI) than those in girls. DISCUSSIONAlmost 1% ( n = 277) of the children and adolescents examined in the NHANES from 1999 to 2012 had a body-size measure that was flagged as being biologically implausible on the basis of the 1995 WHO cutoffs ( 15) and with the use of the CDC growth charts. Of subjects who had a BIV for weight or BMI, most had arm and waist circumferences that were greater than the 99th percentile, and only one subject had a circumference that was less than the 95th percentile. The prevalence of high leg lengths in subjects with a height BIV was lower, but 75% of subjects had a leg length that was greater than the 95th percentile. Although the NHANES data collected from 1999 to 2012 are likely to have very few errors for extreme body-size values because of extensive quality control, our analyses provide additional support for the validity of the extreme values and indicate that many high BIVs are very unlikely to be errors.
If these flagged values had been incorrectly treated as errors, the overall prevalence of obesity would have been reduced by 0.5% (from 16.0% to 15.5%) in all subjects and by 1.7% in non-Hispanic black boys.The use of BIVs in anthropometric measures was addressed in a 1995 WHO report ( 15) which recommended that z scores M 3 (height) or 5 (weight and weight-for-height) should be treated as improbable and set to missing. Although little justification was provided for these specific cutoffs, it was stated that they were the “criteria for anthropometric values that are most likely to represent errors.” These cutoffs may have been selected in part on the basis of the very-low probability that values this extreme occur in a normal N(0,1) distribution; a z score M 5 would be expected only once in a sample of 4 million observations. However, the distributions of BMI MZ and weight MZ in the current study had very heavy right tails, which resulted in positive z scores M being approximately exponentially distributed. The prevalence high BMI BIV values (0.37%), e.g., was 12,000 times higher than that expected for a normally distributed variable. Although height MZ values were closer to being normally distributed, 0.4% of subjects also had height that was considered to be biologically implausible.The CDC program to calculate BMI z scores and percentiles ( 13), which also flags values considered to be biologically implausible, has been widely used ( 3, 4, 10, 11, 16, 17). Previous studies have also noted that the use of the WHO cutoffs can result in a large number of extremely high values, which are unlikely to be errors, being identified as biologically implausible. For example, Lo et al.
( 39) used several data-cleaning methods and showed that, after the exclusion of 30 children who had a height BIVand 3 children who had extremely low values (e.g., weight 30%. Estimates of obesity did not exclude values considered to be BIVs. P values indicate whether there was a significant difference in levels of each column variables across sex, race-ethnicity, or age groups. These differences were assessed in ANOVA models that included race, sex, and age group as predictor variables.P 5 or a BMI MZ 5 and at least one circumference that was less than the 97th percentile, for 6 subjects (middle rows) who had a BMI MZ ≥ 8 and a weight MZ 8, and for 8 subjects (bottom rows) who had a height MZ 3.75. Within each section, subjects are listed in ascending order of the modified z score. BIV, biologically implausible value; BMI MZ, modified BMI z score; height MZ, modified height z score; weight MZ, modified weight z score.