SAS Program for CDC Growth Charts

Overview

This SAS program calculates percentiles and z-scores (standard deviations [SD]) for a child's sex and age for BMI, weight, height, and head circumference from CDC growth charts.1 Weight-for-height z-scores and BMI percentiles are also calculated. The program allows for the identification of outliers, but keep in mind that outliers might be correct values. Consider reviewing outliers, also referred to as "extreme values", so you can decide whether to include or exclude them.

This SAS program calculates BMI z-scores and percentiles for children and adolescents with and without obesity (BMI at or above the 95th percentile). In 2022, BMI z-scores and percentiles calculated for those with obesity changed to use extended BMI z-scores and percentiles. The z-scores and percentiles calculated for children with obesity from this SAS program will differ from those calculated from earlier versions of this SAS program (before 2022).

For infants and children younger than 2 years, the World Health Organization (WHO) growth charts are recommended. Programs on the WHO and CDC websites are based on these WHO growth charts and should be used instead of this SAS program for children under 2.

Instructions for SAS users

Step 1: Download the SAS program (cdc-source-code.sas) and the reference data file (CDCref_d.sas7bdat). Move these files to a folder (directory) that SAS can access. For the following example, the files are in:

Example SAS code corresponding to Steps 2 to 6 below. After downloading the SAS code and the reference data, cut and paste the following 4 lines into your SAS program. You will likely need to change the libname and %include statements to point at the folder (directory) for the downloaded files. You'll also probably need to rename and recode variables, as explained in Steps 2 to 6.

libname refdir 'c:\sas_growth_charts';
data mydata; set whatever-your-original-dataset-is-named;
%include ‘c:\sas_growth_charts\cdc-source-code.sas’;
proc means data=_cdcdata; run;

Step 2: Create a libname statement in your SAS program to point at the folder location of 'CDCref_d.sas7bdat'. An example would be:

Note: the SAS code expects this name to be refdir. Make sure you specify this in the libname statement.

Step 3: Set your existing data that contains height, weight, sex, age, and other variables into a temporary dataset named 'mydata'. Use the information in Table 1 to rename and code the variables.

Table 1. Guidance for SAS users on renaming and coding variables in dataset

Z-scores and percentiles for the anthropometric variables not in mydata (or that are missing) will be coded as missing (.) in the output dataset, _cdcdata. It is unlikely that the SAS code will overwrite variables in your dataset, but you should avoid having variable names that begin with an underscore or with 'mod_'.

Step 4: Copy and paste the following line into your SAS program after the line (or lines) in Step 3.

If necessary, change this statement to point at the folder or directory containing the downloaded cdc-source-code.sas file. The %include will run your data through cdc-source-code.sas and create a dataset named _cdcdata.

Step 5: The output dataset, _cdcdata, contains your original data and z-scores, percentiles, and flags for extreme values shown in Table 2. Additional information on the extreme z-scores is given in the Extreme Values, Implausible Values, and Data Errors section.

Step 6: Examine the new dataset, _cdcdata, to verify that the z-scores and other variables have been created. If z-scores and percentiles for a variable in your dataset are unexpectedly missing, make sure your dataset is named _mydata and your variables are named and coded as shown in Table 1. The program will not modify your original data but adds new variables to your dataset. Table 2 shows the names and descriptions of several variables in _cdcdata.

Table 3. Additional variables in the output dataset, _cdcdata

LMS method

Z-score = ((BMI / M)L - 1) / (L × S)  [equation 1]

The L (transformation for normality), M (median), and S (coefficient of variation) values for the CDC growth charts, which vary by sex and month of age, are in CDCref_d.sas7bdat. These z-scores are then transformed into percentiles with the SAS probnorm function. For example, a z-score of 1.645 is the 95th percentile. For more information on the LMS method, developed by Tim Cole and P.J Greene in the 1990s, see:

• Establishing a standard definition for child overweight and obesity worldwide: international survey
• Smoothing reference centile curves: the LMS method and penalized likelihood

Also see sex- and age-specific L, M, and S values.

The LMS method for BMI results in a curvilinear relation between BMI and BMIz, as shown in the following figure for children aged 5, 12, and 18 years. The range of BMIs (x-axis) corresponds to those observed in NHANES 1999–2000 through 2017–2018. At low BMIs, a small change in BMI results in a sizeable BMIz change. In contrast, at very high BMIs, the same BMI change results in a much smaller BMIz change. This is most evident among males aged 12 years and females aged 18 years. Larger view of chart below.

Further, if a child's BMI is very large relative to the median BMI, then (BMI ÷ M)^L in the LMS equation approaches 0, and the maximum possible BMIz value at that sex/age is (-1) ÷ (L × S). For most ages older than 5 years, the maximum possible BMIz, regardless of the BMI magnitude, is less than 4.0 SDs. Further, among males aged 7 to 15 years and females aged 15 to 19 years, BMIz cannot be greater than 3.3 SDs, limiting the usefulness of these z-scores in characterizing the extremely high BMIs (for example, 40 kg/m² or higher) shown in the figure above.

The 2000 CDC Growth Charts are based on data collected from 1963 to 1980 for most children and adolescents. It was advised that extrapolation beyond the 97th percentile be done cautiously.1 Further, the 2000 CDC Growth Charts' BMI z-scores were not intended for use among children and adolescents with extremely high BMI values.23 Several studies have highlighted the limitations of LMS-calculated BMIz in characterizing very high BMIs.456

Extended BMI percentiles and z-scores

To explore alternative metrics for BMI, the National Center for Health Statistics convened a workshop in 2018 and published a 2022 report7 that evaluated several alternatives to LMS-BMIz. This report recommended using "extended BMIz" and "extended BMI percentiles" to characterize the BMIs of children and adolescents with obesity (BMI ≥ 95^th percentile for a child's sex and age).

These extended metrics were constructed from the BMIs of children and adolescents with obesity in the CDC growth chart reference population and more recent NHANES surveys (through 2015–2016). These BMI data were modeled within each sex and 6-month age group as a half-normal distribution, a truncated normal distribution with only values at or to the right of the peak having a probability density greater than 0.8 Characterizing these distributions' shape parameter, sigma, allows calculating BMI percentiles for children and adolescents with obesity, even those with extremely high BMIs. These percentiles can then be transformed into z-scores.

Updates from December 20227 make it easier to use these extended metrics. The calculated values for BMIz and BMI percentile (bmiz and bmipct) in the SAS program combine the LMS-based values for children and adolescents without obesity with the extended values for children and adolescents with obesity. As a result, the original BMI metrics—constructed using only the L, M, and S parameters—have been renamed as original_bmiz and original_bmipct. Note that bmiz and original_bmiz (and bmipct and original_bmipct) are identical for children and adolescents without obesity.

The following figure shows the relation of BMI to the original and new (extended) values of BMIz. The dashed lines represent the original, LMS-based BMI z-scores from the 2000 CDC growth charts. The solid lines represent the extended bmiz values for BMIs greater than or equal to the 95th percentile (z-score = 1.645 SDs).

Among children and adolescents without obesity, the LMS-based z-scores and the new BMI z-score are identical. At higher BMIs, the relation of BMI to bmiz is fairly linear and does not approach a horizontal asymptote. However, the extended BMIz values are lower than the original values for some BMIs above the 95th percentile, which is most evident for males aged 5 and 18 years in the figure. These lower values arise because children and adolescents with obesity in more recent NHANES surveys have higher BMIs than those in the original CDC reference population.

Although the original LMS-based BMIz was problematic for z-scores greater than 1.88 (97th percentile)1, extended BMIz has a maximum value of slightly less than 8.21 because of the limitations of computer precision. Percentiles very close to 100, such as 99.999999999999993, are indistinguishable from 100 and cannot be converted to a z-score. SAS outputs these z-scores as missing, but the SAS macro converts these z-scores to 8.21. An extended BMIz of 8.20 corresponds to BMIs of about 33 at age 3 years, 55 at age 12 years, and 82 at age 17 years. Actual BMIs this high are very uncommon. These children and adolescents could be monitored using their BMI value.

Severe obesity

Because the original LMS-based z-scores for very high BMIs resulted in percentiles that differ from those estimated from the data3, a BMI ≥ 120% of the CDC 95th percentile has been widely used for classifying severe obesity since 2013.9 This cut-point is approximately equal to the empirical 99th percentile in the growth charts.3 However, among older adolescents, a BMI can be ≥ 35 kg/m² but be less than 120% of the 95^th percentile. Therefore, severe obesity is defined as either a BMI ≥ 120% of the 95^th percentile or a BMI ≥ 35 kg/m². This aligns with guidelines from the American Heart Association9 and the American Academy of Pediatrics.10

The SAS program outputs the variable, bmip95, which expresses a child's BMI as a percentage of the CDC 95^th percentile for the child's age and sex, which can range from below 50 to more than 220. For example, a bmip95 of 140 would indicate that a child has a BMI equal to 1.4 times the 95th percentile. You can also calculate the arithmetic difference between a child's BMI and the CDC 95th percentile. For example, the CDC 95th percentile for a 60-month-old boy is 17.9 kg/m². If this boy had a BMI of 21.3 kg/m², the arithmetic difference would be 3.4 kg/m² (21.3 – 17.9), and bmip95 would be 119% (100 × 21.3/17.9).

Extreme values, implausible values, and data errors

As explained in the modified z-scores documentation, the SAS code also calculates modified z-scores that you can use to identify extreme values that may be errors. These modified z-scores were computed by extrapolating one-half of the distance between 0 and +2 (or between 0 and -2) z-scores to the distribution's tails. Although these z-scores were developed to identify outliers at a single examination, they have been incorporated into algorithms for cleaning longitudinal data.11

The output from the SAS program contains biologically implausible value (BIV) flag variables for weight, height, and BMI that are coded as -1 (modified z-score is very low), +1 (modified z-score is very high), or 0 (modified z-score is between these two cut-points). These BIV flags in the output dataset, _cdcdata, were included in Table 2. It is essential to realize that an extreme value is not necessarily incorrect. The value should be further examined, possibly along with other characteristics of the child.

The upper thresholds for the modified z-score cut-points were initially based on a 1995 WHO publication12 but were changed in 2016. Several papers131415 showed that these cut-points excluded many children whose weight, height, or BMI were very likely to have been recorded correctly. These BIVs can flag potentially problematic data points, but the BIV cut-points are not a gold standard. The cut-points were chosen to balance including extreme values likely to be correct and excluding those likely to be incorrect.1415

These paper results led to an increase in the upper cut-points in 2016 from:

• +5 to +8 for modified z-scores for weight and BMI.
• +3 to +4 for modified z-scores for height.

These new z-score cut-points roughly correspond to the modified z-scores for the maximum values of the body size measures among children aged 2 to 18 years in NHANES at many, but not all, ages. However, please be careful in using these cut-points to exclude data, as different decisions could alter the prevalence of obesity and severe obesity by up to 1%.1415

Other cut-points for the modified z-scores may be more appropriate based on additional information in your data. For example, does a child with an extremely high BMI also have a high skinfold thickness or arm circumference or is very tall? If so, the very high BMI value is more likely to be correct. Similarly, in a longitudinal study or an analysis of electronic health records (EHRs), one could assess whether a child has extreme values of weight and BMI at multiple examinations.

Although +8 SDs is the threshold for a high BMI BIV, two young (less than aged 5 years) boys in NHANES (2005–2006 and 2017–2018) have a modified BMIz greater than 11 SDs. Further, EHR datasets that comprise millions of children indicate that many children consistently have a modified BMIz between 10 and 12 SDs at consecutive examinations. Growthcleanr11, an R package, helps identify errors in longitudinal datasets containing multiple records for each child.

You can use the modified z-scores construct other cut-points for extreme values rather than relying on the BIV flag variables. For example, if you feel using a BMI-for-age cut-point of +8 SDs would exclude many values likely to be correct, then you could use mod_bmiz > 10 as the definition of a high BMI BIV. This could be recoded as:

if -5 <= mod_bmiz <= 10 then _bivbmi=0; *plausible;
else if mod_bmiz > 10 then _bivbmi=1; *high BIV;
else if . < mod_bmiz < -5 then _bivbmi= -1; *low BIV;