Body mass index cut offs to define thinness in children and adolescents: international surveyBMJ 2007; 335 doi: https://doi.org/10.1136/bmj.39238.399444.55 (Published 26 July 2007) Cite this as: BMJ 2007;335:194
All rapid responses
With an effort to develop an international definition of thinness for
children, Cole et al. (1) proposed cut points for underweight (IOTF cut
points). Although the IOTF referent population might represent better
growth patterns of international children compared to the CDC BMI-for-age
growth charts, it is still uncertain if the cut points for underweight are
appropriate to be used in non-representative countries (1).
To determine the consistency between the prevalence of underweight
derived from the CDC and IOTF cut points, we used data from 2820 Chinese
children aged 2–18 y who participated in a representative survey in 9
China provinces (the China Health and Nutrition Survey in 2000; to
represent about 50% of Chinese population) (2). Each child was defined to
be underweight or non-underweight based on the CDC and IOTF cut points (1,
3). Differences in the prevalence (absolute and relative) and Kappa
statistics were estimated to facilitate the comparison. A smaller
difference between corresponding prevalence and a higher Kappa statistic
indicate a higher consistency (4).
The prevalence of underweight defined by the IOTF cut points was
higher compared to those obtained from the CDC BMI-for-age growth charts.
Absolute differences were about 4–7.5% in boys and 6.5–16% in girls.
Compared to the prevalence obtained from the CDC growth charts, those
obtained from the IOTF cut points were about 45–60% larger in boys and
80–190% larger in girls (relative difference). Kappa statistic values
varied from 0.75–0.80 in boys and from 0.45–0.70 in girls (Table 1). These
findings suggested a relatively high inconsistency level between the two
sets of cut points in defining underweight in Chinese children.
Although we might expect some differences in the prevalence of
underweight when the CDC and IOTF cut points were used (they were
developed with the use of different data sets and approaches) (1, 3), it
is a surprise to find a much higher prevalence of underweight derived from
the IOTF cut points. Because American children tend to have a higher
median and mean BMI compared to children from other countries (5), we
expect to observe a lower median, mean, and BMI cut points for underweight
in an international referent population. However, compared to values of
the 5th percentiles for BMI in the CDC growth charts, the IOTF cut points
for underweight are about 0.5 kg/m2 larger in boys and 0.5–1.0 kg/m2
larger in girls (1, 3). The IOTF cut points for underweight are also
larger compared to values of the 5th percentiles for BMI in the earlier
international referent population (6, 7).
A larger sample size in the IOTF referent population would be an
explanation for the elevated BMI cut points. The IOTF cut points were
based on data from almost 200000 children in the US and five other high
and middle income countries (Brazil, Great Britain, Hong Kong, the
Netherlands, and Singapore) (1), which is much larger compared to those of
the CDC BMI-for-age growth charts (about 30000 American children) (1, 3).
The increase in number of children who might share similar growth patterns
would lead to a decrease in values of standard deviation. Because z-score
(which is based on standard deviation) was used as one of the components
in the construction of IOTF cut points (1), the IOTF cut points for
underweight may systematically higher compared to those developed by using
the same method, but with a smaller and / or more heterogeneous sample.
The IOTF referent population should also include samples of healthy
children from other middle and low income countries. Further research is
needed to determine a set of cut points, which are based on the increased
total body fats, current or future disease risk in children.
1. Cole TJ, Flegal KM, Nicholls D, Jackson AA. Body mass index cut
offs to define thinness in children and adolescents: international survey.
2. Popkin BM, Paeratakul S, Zhai F, Ge K. Dietary and environmental
correlates of obesity in a population study in China. Obes Res 1995;3
3. Kuczmarski RJ, Ogden CL, Guo SS. CDC growth charts for the United
States: Methods and development: National Center for Health Statistics,
Vital Health Stat 11(246), 2002.
4. Rosner B. Hypothesis testing: categorical data. Fundamentals of
biostatistics. 6th ed. Belmont, CA: Thomson-Brooks/Cole, 2006:385-463.
5. Popkin BM, Conde W, Hou N, Monteiro C. Is there a lag globally in
overweight trends for children compared with adults? Obesity (Silver
6. Must A, Dallal GE, Dietz WH. Reference data for obesity: 85th and
95th percentiles of body mass index (wt/ht2) and triceps skinfold
thickness. Am J Clin Nutr 1991;53(4):839-846.
7. WHO expert committee. Physical Status: the use and interpretation
of anthropometry. Geneva: WHO, 1995.
Competing interests: No competing interests
It has been pointed out that the paper is unclear, in the Abstract and Table 4,
about the BMI range at age 18 that corresponds to each thinness grade. To
avoid confusion we repeat the grading here, which corresponds to the WHO
BMI cut- off
17 to <_18.5o:p xmlns:_18.5o="urn:x-prefix:_18.5o"/>
16 to <_17o:p xmlns:_17o="urn:x-prefix:_17o"/>
So grade 1 is less
than grade 3.
Competing interests: No competing interests
Dr Connolly goes on to state that ponderal index (PI) is
better than BMI, but it all depends on what is meant by "better". PI
like BMI is an index of weight and height, and so lacks information about
composition. Thus both indices are poor proxies for adiposity, be it thin or
His dimensional argument supporting PI over BMI (i.e.
dividing weight by height3 rather than height2)
when applied across species (e.g. from mouse to elephant) where the
between weight and height is close to 1. But within the species of homo
sapiens the weight-height correlation
only about 0.7, so the best predictor of weight (based on log-log regression)
is 3 x 0.7 ~ 2. This imperfect correlation, which is biologically inevitable,
attenuates the height power and ensures that BMI rather than PI is
In practice the optimal height power to predict weight
varies with age in children, due to heterogeneity in the rate of maturation; it
increases from 2 near birth to 3 in puberty and back to 2 or less in
adults. Thus PI is briefly better
than BMI in puberty.
But this is not the whole story. Obesity itself is
positively related to height (fatter children tend to be taller), so the
optimal index should be correlated with height. Through most of childhood
is positively correlated with height while PI is negatively correlated. Thus
BMI is consistently better than PI.
Dr Connolly argues that the gradients of the BMI cut-
are strikingly attenuated compared to the centiles of normal growth charts.
This is also quite wrong as the cut-offs are centiles, averaged across surveys, and are very similar in shape to
the centiles of the constituent national BMI charts. His argument involves heat
exchange in infancy, but the cut-offs only start at age 2 years, well past
Dr Connolly's final point is that an index incorporating
weight, height and waist should be better than one using just weight and
height. This is almost certainly true, but not for the reasons he gives. Waist
circumference is a fairly direct measure of central fat, and so provides the
information on adiposity that weight-height indices like BMI and PI lack.
TJ, Flegal KM, Nicholls D, Jackson AA. Body mass index cut-offs to define
thinness in children and adolescents: international survey. Brit Med J
TJ, Bellizzi MC, Flegal KM, Dietz WH. Establishing a standard definition for
overweight and obesity: international survey. Brit Med J 2000;320:1240-3.
TJ. Weight-stature indices to measure underweight, overweight and obesity.
Himes JH, editor. Anthropometric assessment of nutritional status. New York:
Wiley-Liss; 1991. p. 83-111.
TJ. Weight/heightp compared to weight/height2
assessing adiposity in childhood: influence of age and bone age on p during
puberty. Ann Hum Biol 1986;13:433-51.
DS, Khan LK, Serdula MK, Dietz WH, Srinivasan SR, Berenson GS.
Inter-relationships among childhood BMI, childhood height, and adult
the Bogalusa Heart Study. Int J Obes Relat Metab Disord 2004;28:10-6.
Competing interests: Dr Connolly suggests that our paper  is fundamentally flawed inviewing body mass index (BMI) as a measure of adiposity. Our earlier relatedpaper  made clear that BMI is only aproxy for adiposity, which to be identified accurately requires a direct measureof body fat. However BMI is popular as it is moderately correlated withadiposity and has the benefit of being widely applicable, since weight andheight measurements are ubiquitous.
Stef van Buuren (email@example.com)
1 TNO Quality of Life, Leiden
2 University of Utrecht
Cole et al. (2007) present new cut off values for three classes of thinness. In their motivation, the authors state that “There are no valid BMI cut offs for assessing underweight or wasting in adolescents or children over 5 years” and “There are no suitable thinness cut offs for this age group”.
I like to point out that some years ago I published criteria for underweight and severe underweight for children 2-18 years (van Buuren, 2004). This work appears to be somewhat unknown, presumably because it appeared in Dutch. I used a similar approach as in Cole et al (2007), and derived cut off values that were anchored at BMI-values of 17 and 18.5. However, having no access to the references of all six studies, I restricted the analysis to the data from the Third Dutch Growth study only (Roede & Van Wieringen, 1985), for which Cole and Roede (1999) had published BMI references.
It is of interest to study the differences between the criteria. Figure 1 plots the cut offs as published by Van Buuren (2004) and Cole et al (2007). Despite the differences in methodology, the values are very similar across the entire age range. The new cut offs for thinness are virtually always located within 0.2 BMI points of the “Dutch cut offs” for (severe) underweight. In clinical work, such differences easily fall within the bounds of measurement error. Differences in prevalence estimates are likely to be minor, and will nearly always be swamped by sampling variation. The differences between both sets of cut offs are considerably smaller than those observed between the centile curves of the six international studies as portrayed in figures 1 and 2 of Cole et al (2007). All in all, it appears that the cut off values proposed by Van Buuren (2004) and Cole et al (2007) are remarkably similar.
This begs the question what cut offs to use. Evidently, the large exposure of BMJ combined with Cole’s authority will be a major factor in the worldwide adoption of the cut offs, which in itself would be a good thing. On the other hand, we should not be blind to the shortcomings of the new thinness cut offs. In the process of combining data from different studies, Cole et al. average centile curves. There is no statistically defensible rationale for this averaging operation, and it results in an intractable statistical distribution. Consequently, as noted by the authors, it is not possible to calculate Z-scores related to the new thinness cut offs. The criteria derived by Van Buuren (2004) do not have this problem. The cut offs correspond to a set of fixed centiles from a well defined reference distribution. The availability of such a distribution allows for the calculation of Z-scores. Thus, in addition to establishing cut offs, it is possible to assess BMI corrected for age and sex in a continuous way. The possibility to calculate Z-scores may well give the edge to the “Dutch” criteria for underweight and severe underweight.
Cole TJ, Flegal KM, Nicholls D, Jackson AA. Body mass index cut offs to define thinness in children and adolescents: international survey. BMJ 2007;335;194.
Cole TJ, Roede MJ. Centiles of body mass index for Dutch children aged 020 years in 1980—a baseline to assess recent trends in obesity. Ann Hum Biol 1999;26:3038.
Roede MJ, van Wieringen JC. Growth diagrams 1980: Netherlands third nation-wide survey. Tijdschr Soc Gezondheidsz 1985;63(suppl):1-34. pdf
Van Buuren S. Afkapwaarden van de 'body-mass index' (BMI) voor ondergewicht van Nederlandse kinderen. [Body-mass index cut-off values for underweight in Dutch children] Ned Tijdschr Geneeskd. 2004;148(40):1967-72. pdf
Competing interests: No competing interests
Dr Cole and his colleagues are to be congratulated on producing
tables that accurately reflect risk of malnutrition as assessed by BMI,
but they might have questioned why their results appear to suggest that
the hazard of thinness is least in infants. In his editorial Dr Cameron
points out that BMI is not a measure of obesity. He suggests it does have
some relationship to body composition, but justifies it by using
correlation coefficient, the incorrect measure of an anticipated
association . One might go further and say that the discussion in the
paper itself is fundamentally flawed in suggesting BMI is a measure of
thinness or indeed obesity. The recognized index which is closest to being
a surrogate for these dimensionless concepts is Ponderal Index
(Weight/Height3). This is a true index in being itself effectively
dimensionless, because with objects of similar density, mass, and
therefore weight, has dimension equivalent to length cubed. Body Mass
Index is a misnomer, as it is not dimensionless, being proportional to
height (Height3/Height2). In simple terms the taller individuals of
similar build, the greater is their BMI. For example a very tall person
(1.9 m) of the same composition as a short person (1.55m) of ideal
proportions (BMI 25) will be assessed as obese (BMI 30.6), or a tall
modern female fashion model (1.85m) of the same degree of thinness as an
undernourished more traditional one (1.65m, BMI 17) will appear to be
adequately nourished (BMI 19.1).
BMI is conventionally used as the measure of risk from obesity and
thinness because it has been shown to perform better than ponderal index
in the populations studied. This implies that obesity is less dangerous or
conversely that thinness is a greater risk in those of short stature
within those populations, but because BMI has a dimension cut –off values
cannot be transferred to other groups.
As might be deduced from the first paragraph, the curves in Dr Cole’s
paper are similar to normal growth charts, but the gradients are
strikingly attenuated. As heat loss per unit mass is inversely
proportional to body size, and so such factors as skin thickness and
calorific reserve become more critical in infants. This explains the
attenuation of the gradient in the graphs in the paper. Had ponderal index
rather than BMI been used the gradients would have been reversed, with a
greater desirable ponderal index in the young rather than older children,
reflecting the greater danger of thinness in infants.
Aside from the potential and recognized errors arising from
differences in the proportion of fat and lean body mass which may be less
than expected , there is a further complication. The relationship
between desirable body mass and both BMI and ponderal index is confounded
because both only consider length along the longest axis. Potential for
heat loss is related to shape, being least when it approximates to
spherical. Therefore probably the best measure to use to determine optimal
mass is the dimensionless index weight/height x waist2.
1. Cameron N (2007) Body mass index cut offs to define thinnessin
children and adolescents. BMJ, 335, 166-167.
2.Bland JM, Altman DG. (1986). Statistical methods for assessing agreement
between two methods of clinical measurement. Lancet, i, 307-310
3.Freedman DS, Wang J, Maynard LM, Thornton JC, Mei Z, Pierson RN, et al.
Relation of BMI to fat and fat-free mass among children and adolescents.
Int J Obes 2005;29:1-8.
Competing interests: No competing interests