The use of measures of obesity in childhood for predicting obesity and the development of obesity-related diseases in adulthood: a systematic review and meta-analysis

Simple measures of obesity

There are a range of simple, anthropometric, indirect measures of adiposity. 25,26 The simplest measure is weight; however, on its own this provides little useful information, as confounding variables such as a person’s height and body composition are not taken into consideration.

Body mass index is the most commonly used simple measure of obesity, and is also the only measure recommended for use in children in the UK. It is calculated by dividing weight in kilograms by height in metres squared (kg/m²). BMI increases sharply in infancy, peaking at around 9 months, and then falls to its lowest at around 6 years. 27 BMI (adjusted for age and sex) is recommended by the National Institute for Health and Care Excellence (NICE) (CG43) as a practical estimate of overweight in children and young people, but the guidance warns that the result needs to be interpreted with caution because BMI is not a direct measure of adiposity. 28 The main limitation of BMI is that it measures excess weight, rather than excess fat, which is what determines whether someone is obese or not. Therefore, those with strong bones and/or well-developed muscularity, but little fat, will have a high BMI and could be categorised as obese, as bone and muscle are more dense than fat. 29 Additionally, BMI is not consistent across the normal height range, with shorter heights producing higher BMI values. 30 BMI also does not give any indication as to the distribution of fat in the body; in adults, central adiposity is more closely associated with health risks than general adiposity,31,32 and the use of BMI and waist circumference (WC) is recommended in adults in the UK.

Simple measures other than BMI include:

• WC A measure which gives an indication as to the distribution of excess body fat. However, in isolation it provides insufficient information regarding overall adiposity. Based on a UK reference population, 95th centile cut-off points range from 57.0 cm at age 5 years to 85.2 cm at age 16 years in boys, and from 57.2 cm at age 5 years to 75.1 cm at age 16 years in girls. 33

• Neck circumference (NC) The use of this measure has been investigated in children, and it is thought to be reliable for identifying children with high adiposity. 34–36 Based on a US population, thresholds for obesity are suggested as 29 cm in prepubertal boys, 28 cm in prepubertal girls, 32.5 cm in pubertal boys and 31 cm in pubertal girls. 36

• Rohrer’s Ponderal Index (Rohrer’s Index, Ponderal Index or Corpulence Index) Similar to BMI, but calculated by dividing weight in kilograms by height in metres cubed (kg/m³), rather than height in metres squared. 29 This has been compared with BMI in respect of its ability to predict percentage body fat in children and adolescents, and its long-term associations with adult obesity. 37,38

• Benn’s Index Calculated by dividing weight in kilograms by height^p (kg/m^p); p is a power index derived from the weight-to-height ratio (p = bH0/W0), making the index independent of height. 29,39 Benn’s Index is rarely used as p is neither constant nor necessarily a whole number, which means the calculations are complicated. 29 A study in boys from four countries showed that the power of p required to produce a correlation of zero between the index and height varied with age and ethnicity. For US, Japanese and Singaporean boys, the p value was < 2.8 at age 6 years, increased to < 3.5 at age 9–10 years and decreased to < 2.0 by age 16 years. For boys in the UK, p started at < 2.3, increased to < 2.6 then decreased to < 2.0. 40

• Waist-to-height ratio (WHtR) As with the waist-to-hip ratio (WHR), WHtR is a measure of fat distribution, and primarily identifies those with abdominal obesity. A person is typically considered obese if their WC is over half their height (threshold of 0.5 across multiple countries and ethnicities). 41,42

• WHR A measure of regional fat distribution used as a marker for intra-abdominal fat in adults. A WHR of ≥ 0.80 in females and ≥ 1.0 in males is considered to be indicative of a high risk of health problems. 43 One of the limitations with ratios using waist measurement is the potential for measurement error, as there is inconsistency in the actual site of the waist measurement.

• Body adiposity index (BAI) This was suggested by Bergman et al. (2011)44 as reflecting the percentage body fat in adults regardless of sex or ethnicity without numerical correction, and it was therefore considered an improvement on BMI. BAI is calculated by the equation [hip circumference/(height)^x] − 18, with x being a unitless power term. The correlation between BAI and percentage adiposity was highest when x was between 1.47 and 1.5, and therefore the use of 1.5 as the power term in Mexican American adults was suggested. 44 The measure has not been validated in children, and a power term for children has not been specified.

• Skinfold thickness (SFT) A direct anthropometric measure of adiposity. Skinfold measurements are considered as good indicators as they are a direct measure of the fat layer, but the measurements are site- and sex-specific. 27 The most common sites are the subscapular and the triceps;25 other potential sites include the chest, axilla, abdomen, suprailium and thigh. 26 Scores can be presented adjusted for age and sex. There are also a large number of equations available that can be used to obtain an estimate of percentage subcutaneous fat from SFT measurements,25,26 although these may introduce biases and have not been standardised for children over the age of 6 years. 45 In adults, women with 30% or higher, and men with 25% or higher, body fat are classified as obese. 46 One of the limitations of SFT is that visceral fat (fat in the abdominal cavity) is not measured. As with waist measurements, measurement error is a particular limitation of this method. There can be considerable variability across practitioners, leading to the requirement for specific training.

• Bioelectrical impedance analysis (BIA) This measures the opposition to the flow of an electric current applied to extremities of the body (usually the wrists or ankles; in children, foot-to-foot currents are used as this only requires the child to stand barefoot on scales), using a low-level (below the threshold of human perception; usually a single frequency of 50 kHz) alternating electrical current passing through the body. This is used to estimate total body water. As fat contains little water compared with other body tissues, impedance can be used as a proxy measure for fat-free body mass. Body fat can then be calculated from the difference between fat-free body mass and overall weight. BIA has several limitations. First, the body is a number of cylinders of different length and diameter (legs, torso, arms), rather than a single cylinder, and given that resistance decreases as the cross-section of a cylinder increases, the arms and legs contribute more to resistance than the torso. The limbs represent 50% of body weight but 90% of the body’s impedance; this means that impedance is more closely related to changes in the muscle mass of the limbs. In addition, as there is an assumption that fat-free mass is 73% water,26 factors such as dehydration, exercise, diuretics or a full bladder will affect the results. 47 Furthermore, equations based on height and weight are used to determine overall adiposity and these vary across manufacturers. The equations used by some instruments may be unknown, and results may vary across different instruments and populations. 48 Given these limitations, BIA is currently not recommended for use in the UK.

• Fat mass index (FMI) Calculated from fat mass as determined by BIA (kg) divided by height squared. FMI plus the fat-free mass index (fat-free mass/height squared) = BMI. 49,50

• Near-infrared interactance (NIR) A beam of infrared light is transmitted into the biceps; the light is reflected by underlying muscle and absorbed by fat, and therefore the proportion of reflected light will indicate the proportion of fat. The NIR light penetrates the tissues to a depth of 4 cm and is reflected back to the detector, which measures the optical density at wavelengths of 940 nm (optical density 1) and 950 nm (optical density 2). The underlying principle is that optical densities are linearly and inversely related to percentage body fat, and thus, the smaller the optical density, the greater the absorption of NIR light and the higher the fat composition. 26

More complex measures of obesity

Apart from SFT, the anthropometric measures listed above are indirect measures of adiposity. There are a number of more direct measures of adiposity available; these are not routinely available and need to be conducted by those who have been specifically trained, and therefore they are not useful as population measures. Such measures include:

• Densitometry [hydrostatic weighing; underwater (hydrostatic) weighting] This method measures underwater weight. It distinguishes fat mass and fat-free mass, assuming specific densities of these two tissues, and therefore requires measurement of total body density (body mass/body volume). 51 Whereas the density of fat is relatively constant, that of fat-free mass varies according to its composition. This variability is partly explained by the process of chemical maturation that occurs before adulthood, but interindividual variability is also significant, even in healthy children.

• Densitometry [air displacement plethysmography (ADP)] A new system that uses the displacement of air rather than water. This is thought to have better precision than hydrodensitometry in children, and is acceptable in children as young as 4 years. 52 An infant ADP has become available, allowing the measurement of body volume during the first 6 months of life. In general, densitometry is unsuitable for application as a two-component technique in patients in whom the composition of lean mass may be abnormal, such as those with excess fluid retention and undermineralisation, as these decrease the density of lean mass and lead to an overestimation of fatness. 52,53

• Multicomponent models These models combine the results of three, four or five different measures. The three-component model divides body weight into fat, water and fat-free dry tissue, and requires measurements of body weight, body water by hydrometry and body volume by densitometry. The four-component (4-C) model further divides fat-free dry tissue into protein and mineral, and requires the measurement of bone mineral by dual-energy X-ray absorptiometry (DEXA). Its advantage is that it is the most accurate approach, with all the component measurements being acceptable. It has the disadvantage of being expensive and thus generally limited to specialist research. 51 Body volume index is a new multicomponent measure that uses BMI, WC and WHR along with other volumetric and body composition analyses. The method requires a three-dimensional full-body scanner. 54

• DEXA DEXA51 is a three-component model, as it measures bone mineral, bone-free lean mass and fat mass. For accurate estimation of body fat, the hydration of lean body mass needs to be established; often, a constant for hydration of lean body mass (0.73 ml/g) is assumed, which is based on non-elderly, healthy, adult measurements. The use of this assumption can lead to an error in the amount of lean tissue, particularly in children, the elderly and the sick; DEXA is likely to be more reliable in healthy adults with a constant tissue hydration. 55,56 DEXA may therefore have difficulty accurately assessing body mass in infants, with increasing accuracy as body size increases. DEXA also has difficulty distinguishing between bone and non-bone lean mass in high bony areas (thorax, forearm). 55 In a review comparing DEXA with 4-C models, DEXA tended to underestimate body fat in most studies, with this underestimation usually being greater in leaner people. 57 Some studies did find DEXA overestimated body fat, and some no difference; there seemed to be differences in accuracy with different DEXA technology. 57

• Deuterium dilution method (D₂O) Deuterated water (²H₂O) is currently the most accurate way of measuring total body water. The difference in secreted markers (urine, saliva) between baseline and post ingestion is used in an algorithm to calculate total body water and fat-free mass. The difference between total body mass and fat-free mass gives fat mass, allowing body composition to be worked out, but the method is most effective for total body water when used in a multi-(four-)component model with body density [underwater (hydrostatic) weighting], BIA and DEXA. The method is safe and has been validated in a wide range of populations but does have disadvantages; it is expensive, time-consuming (requiring sample collections and analyses) and rather invasive because it requires participants to ingest a dose of deuterium-enriched water. 58

Aim of the literature search

Three separate literature searches were undertaken to identify studies to answer the following questions:

1. How accurately do measures of childhood obesity predict CVD, diabetes and/or cancer in adulthood?

2. Which simple measure of obesity in children most accurately predicts the tracking of obesity into adolescence and adulthood?

3. How accurately do simple measures of obesity reflect actual adiposity in children (actual adiposity in terms of ability to correctly classify a child as normal weight, overweight or obese)?

The three search strategies were devised using a combination of indexed subject heading terms and free-text search terms appearing in the title and/or abstract of database records. Search terms were identified through discussion within the project team, by scanning background literature and ‘key articles’ already known to the project team and by browsing database thesauri. The search strategies were peer reviewed for accuracy by other information specialists based at CRD (Lisa Stirk and Melissa Harden). Once the MEDLINE search strategies were agreed and peer reviewed, they were adapted so that they could be used in the other databases. Bibliographic records were managed using EndNote XI bibliographic management software (Thomson Reuters, CA, USA).

Full details of the search strategies, dates of searches and results from all the databases and resources searched are given in Appendix 1.

Concepts of the search strategy

All three search strategies included a set of search terms for the following three concepts: ‘obesity’, ‘children’ and a collection of ‘simple anthropometric measures’ (index tests). The ‘obesity’ concept included search terms for ‘adiposity’. As BMI is the most widely used and recognised simple anthropometric measure for obesity, using this as a search term inevitably increased the sensitivity of the search strategy and the volume of literature retrieved.

The database searches were supplemented with reference checking and citation searching, as well as additional targeted searches for any gaps in the literature identified after screening of the initial set of literature search results.

The search strategy for the question on the prediction of adult morbidities was structured using the following concepts:

 (Obesity OR adiposity)

 AND

 (Children OR adolescents)

 AND

 (CVD OR diabetes OR cancer)

An initial search to identify only systematic reviews was conducted, and the searches were limited to the Database of Abstracts of Reviews of Effects (DARE), the Cochrane Database of Systematic Reviews (CDSR), the Database of Promoting Health Effectiveness Reviews (DoPHER) and MEDLINE (with a methodological search filter designed to retrieve systematic reviews).

Having assessed the systematic reviews of measures for predicting morbidity in adulthood, the team was able to identify gaps in the literature and decided on the limits for the searches for primary studies. Targeted searches were conducted for the combinations of measures and morbidities shown in Table 1.

For the review question on the tracking of obesity, the search strategy was structured using the following concepts:

 (Obesity OR adiposity)

 AND

 IT

 AND

 (Tracking OR Cohort studies OR longitudinal studies OR follow-up studies)

 AND

 (Children OR adolescents)

 AND

 Adults

 AND

 Date limit (2007–2013)

During the initial scoping searches, a systematic review about ‘tracking’ of childhood obesity into adulthood was identified [Singh et al. (2008)],16 and it was therefore decided to limit the searches to run from the date when the searches for this review were completed (2007). The final agreed search strategy was unable to identify one of the Singh review included studies, owing to the absence of any terms for ‘tracking’ or study design in the title or abstract and inadequate subject heading indexing.

The diagnostic accuracy literature searches were designed to identify studies that compared index tests with reference standards, so both concepts were combined alongside terms for ‘obesity’ and ‘children’. The diagnostic accuracy search strategy was structured using the following concepts:

 (Obesity OR adiposity)

 AND

 (Children OR adolescents)

 AND

 Index tests

 AND

 Reference standards

 NOT

 Animal studies

Initially, the search for studies on the acceptability of simple measures was to be conducted for those tests previously determined to be the most promising measures in terms of predictive accuracy. However, given the lack of non-BMI studies, searches were designed to retrieve qualitative studies about the accessibility and ease of use of the following simple measures of obesity: BMI, SFT, WHR and WHtR. The searches were limited to the following databases: MEDLINE, EMBASE, PsycINFO and Cumulative Index to Nursing and Allied Health Literature (CINAHL). They were also limited by date range (2008–13).

Resources searched

The literature searches involved searching a wide range of databases. The following databases and resources were searched:

• MEDLINE (OvidSP)

• MEDLINE In-Process & Other Non-Indexed Citations (OvidSP)

• PubMed (National Library of Medicine)

• EMBASE (OvidSP)

• PsycINFO (OvidSP)

• CINAHL (EBSCO)

• CDSR (Wiley Online Library)

• Cochrane Central Register of Controlled Trials (CENTRAL) (Wiley Online Library)

• DARE (CRD interface)

• Health Technology Assessment (HTA) Database (CRD interface)

• NHS Economic Evaluation Database (NHS EED) (CRD interface)

• Science Citation Index (SCI) [ISI Web of Science (WoS)]

• Conference Proceedings Citation Index – Science (CPCI-S) (ISI WoS)

• Health Management Information Consortium (OvidSP)

• Trials Register of Promoting Health Interventions [Evidence for Policy and Practice Information and Co-ordinating Centre (EPPI-Centre) interface]

• DoPHER (EPPI-Centre interface)

• Obesity and Sedentary Behaviour Database (EPPI-Centre interface)

• OAIster (http://oaister.worldcat.org/)

• OpenGrey (www.opengrey.eu/).

The following obesity-related organisation websites were searched:

• Public Health England – Obesity Knowledge and Intelligence (previously National Obesity Observatory) (www.noo.org.uk/)

• Association for the Study of Obesity (www.aso.org.uk/)

• Obesity Learning Centre (www.obesitylearningcentre.org.uk/)

• National Obesity Forum (www.nationalobesityforum.org.uk/)

• British Dietetic Association (www.bda.uk.com/index.html)

• Centre for Diet and Activity Research (CEDAR) (www.cedar.iph.cam.ac.uk/)

• The Nutrition Society (www.nutritionsociety.org/)

• International Association for the Study of Obesity (www.iaso.org/)

• European Association for the Study of Obesity (EASO) (www.easoobesity.org/)

• European Congress on Obesity (www.easo.org/eco2013)

• European Childhood Obesity Group (ECOG) (www.ecog-obesity.eu/)

• Centers for Disease Control and Prevention (CDC) Division of Nutrition, Physical Activity, and Obesity (USA) (www.cdc.gov/nccdphp/dnpao/index.html)

• Weight-control Information Network (USA) (http://win.niddk.nih.gov/)

• The Obesity Society (USA) (www.obesity.org/)

• myhealthywaist.org (International Chair on Cardiometabolic Risk of Université Laval – QC, Canada) (www.myhealthywaist.org/).

Citation searches were conducted in SCI and Google Scholar. All sources proposed in the protocol are listed above.

Review of prediction of adult morbidities

• Outcomes The study had to report RRs, odds ratios (ORs), hazard ratios (HRs) or summary estimates of predictive accuracy, or sufficient data from which these could be derived, for the association between childhood obesity and adult CVD, type 2 diabetes or cancer. For the purposes of this review, CVD incorporated major cardiovascular events such as cardiovascular death, myocardial infarction, stroke, heart failure, hypertension, hypercholesterolaemia and metabolic syndrome.

• Interventions Data for the following simple measures were included: BMI, NC, WC, WHR, WHtR, BAI, Rohrer’s Ponderal Index, Benn’s Index, FMI, SFT, BIA and NIR. Studies were included for each of these measures, regardless of how the measurement was conducted (i.e. any level at which WC was measured was eligible).

• Study design Prospective, longitudinal studies that were sufficiently powered (1000 participants) and that evaluated any one of the interventions of interest used in childhood for the prediction of a morbidity of interest in adolescence or adulthood were eligible; case–control studies and retrospective studies were excluded.

There are many other types of morbidity where evidence of an association with obesity exists, including mental health problems, respiratory conditions and musculoskeletal conditions, among others. This review considered only cardiovascular conditions, diabetes, cancer and metabolic syndrome, in order to focus on those conditions most likely to lead to mortality, where the evidence of association with obesity was strongest, and where there was most likely to be long-term evidence in cohort studies of their association with childhood obesity.

Tracking of childhood obesity into adolescence/adulthood or adolescent obesity into adulthood

• Population Studies recruiting children and/or adolescents (up to the age of 18 years) were eligible for inclusion. Studies recruiting a mixture of adults and children/adolescents were included if the results for children/adolescents were reported separately. Studies had to recruit either population-based samples of children or overweight/obese children; studies conducted only in children who were not overweight or obese were excluded. We accepted the definition of obesity/adiposity used in the study.

• Interventions Data for the following simple measures were sought: BMI, NC, WC, WHR, WHtR, BAI, Ponderal Index, Benn’s Index, FMI, SFT, BIA and NIR. Studies were included for each of these measures, regardless of how the measurement was conducted (i.e. any level at which WC was measured was eligible).

• Outcomes The study had to report estimates of test accuracy, or sufficient data from which these could be derived, for the association between the weight status in childhood and/or adolescence and the incidence of obesity/overweight in adulthood. This was a change from the protocol, required as data on the association between weight status in childhood and adulthood would be insufficient to assess the predictive accuracy of childhood obesity/overweight. Studies that reported only correlations between the childhood and adult measures were also excluded.

• Study design Prospective, longitudinal studies that evaluate any one of the interventions of interest were eligible; case–control studies and retrospective studies were excluded. Inclusion was initially restricted to studies that recruited at least 100 children; this was increased to 1000 children in studies of BMI given the volume of evidence available and so as to include only well-powered studies. As inclusion was restricted to those studies where predictive accuracy of the anthropometric measure could be established, only studies that used an acceptable reference standard in adulthood [i.e. a multicomponent model, D₂O, underwater (hydrostatic) weighting, ADP or DEXA] were included. BMI was added to the list of reference standards as it is generally accepted as a suitable estimate in adults and we expected it to be the most commonly used adult measure (see Appendix 2).

The diagnostic accuracy of childhood measures of obesity

• Population Studies recruiting children and/or adolescents (up to the age of 18 years as defined in the NICE CG43 obesity guidelines) were eligible for inclusion. Studies recruiting a mixture of adults and children/adolescents were included if the results for children/adolescents were reported separately. Studies had to recruit either a population-based sample of children or overweight/obese children; studies conducted only in children who were not overweight or obese were excluded.

• Interventions The following simple measures were evaluated: BMI, NC, WC, WHR, WHtR, BAI, Ponderal Index, Benn’s Index, FMI, SFT, BIA and NIR.

• Reference standard A multicomponent model that measures four or five components was considered the gold standard for assessing the accuracy of simple anthropometric measures of adiposity in children, because the precision of such a model is considered to be higher than other complex measures. The other complex measures, such as DEXA, D₂O and densitometry [underwater (hydrostatic) weighting or ADP], were accepted as reference standards as they are more commonly used in research studies; these were considered imperfect reference standards.

• Outcomes The study had to report either summary estimates of diagnostic accuracy or sufficient data from which these could be derived.

• Study design Prospective single-gate (diagnostic cohort) studies that evaluated any one of the interventions of interest in comparison with any one of the reference standards were eligible for inclusion. For measures for which these are not available, prospective two-gate (diagnostic case–control) studies were included; these had to match cases and controls on at least age and sex, or provide estimates of sensitivity and specificity that had been adjusted for these variables.

The acceptability and ease of use of childhood measures of obesity

The original intention was to identify primary studies that undertook a robust evaluation (such as the use of questionnaires or interviews) of the acceptability and ease of the most promising measure(s) in terms of diagnostic and predictive accuracy. Given that the majority of the data identified were for BMI, this was amended to a search for studies of acceptability and ease of use for the four measures most commonly used in clinical practice: BMI, SFT, WHR and WHtR. Studies could be from the perspective of the child, parent or health professional. Studies that discussed acceptability or ease of implementation with no direct measurement were not included. Given the recent changes in distribution across the population, the emphasis placed on education and intervention for obesity and general attitudes towards obesity, the inclusion of primary studies was initially restricted to those conducted within the previous 5 years (search start date 2008). It was expected that the number of studies addressing this issue would be small. Only one study was identified that met the inclusion criteria. 73 Two further studies were identified from the searches for questions 1, 2 and 3 that were published in 200674 and 2007;75 given the paucity of evidence, the results of these are also presented.

To supplement the systematic review, a simple elicitation exercise (survey) was performed to obtain some indication of the attitudes of children, school nurses and parents to the measures being evaluated. Ethical approval from the Carnegie Faculty Committee, Leeds Metropolitan University, UK, was obtained. Overweight and obese children attending a weight management summer camp, parents of the children attending the weight management camp and school nurses were asked to complete a structured questionnaire developed specifically for the project that was suitable for the age group concerned. Opinions were elicited for the same four measures being evaluated in the systematic review of acceptability and ease of use: BMI, SFT, WHR and WHtR.

Review of prediction of adult morbidities

The review of the association between childhood obesity and adult obesity-related morbidity, and the prediction of adult morbidities from childhood obesity status, was based on the identified cohort studies. Most cohorts were reported in multiple publications. Different publications concerning the same cohort were all included, provided they reported on different morbidities or reported results at different childhood ages.

Although studies reported many morbidity outcomes, only the following were considered in the meta-analyses:

• diabetes (adult-onset, type 2)

• CHD

• stroke

• hypertension

• breast cancer

• all other cancers, combined.

These were the only protocol-specified morbidities reported in at least two cohorts. Cancers were combined because a range of different cancers were reported within and across cohorts. In the narrative review we also considered hypercholesterolaemia and metabolic syndrome.

As results were reported at different ages, we grouped ages as:

• under 7 years

• 7–11 years

• 12–18 years.

The first group represents young children who may experience adiposity rebound, where BMI declines up to around age 7 years. The other two groups represent typical UK division between primary- and secondary-level education.

Studies reported results in a variety of ways, including numbers of morbidities in different BMI categories, ORs comparing different categories of BMI, ORs between different centiles of BMI, ORs per standard deviation (SD) increase in BMI (or, equivalently, per BMI z-score unit) and across different reference populations. Given the diversity in reporting, a direct meta-analysis of results was not feasible. In order to perform a meta-analysis, all results were converted into ORs per SD of BMI (with 95% CIs), and into estimates of predictive performance in terms of sensitivity at specified obesity thresholds. The approach used for this conversion followed the methods of Morris and Wald70 for converting reported ORs into estimates of predictive performance. It requires some assumptions about the distributions of obesity in the childhood population, specifically:

• that BMI follows a normal distribution

• that the SD of BMI is the same in people with and without the morbidity

• that the morbidity is rare.

It is acknowledged that the first assumption is invalid, as BMI is not normally distributed and has a positive skew, but this assumption is required for a meta-analysis to be performed, and it is applied equally to all studies, so any error applies uniformly across the analysis. The other assumptions are more reasonable in this context. Tests on simulated data suggest that, for positively skewed distributions of BMI, these assumptions may slightly overestimate ORs, but the overestimation is unlikely to exceed an absolute increase of 0.1. Using these assumptions will therefore also lead to overestimates of predictive accuracy by increasing estimates of sensitivity. The effect of the assumptions is uncertain, but tests on simulated data suggest that the area under the summary receiver operating characteristic (ROC) curve could be overestimated by up to 20%. As data in publications were presented as ORs without data on the distribution of BMI, it was not possible to transform these data to better represent a skewed distribution.

For each morbidity outcome and within each age group, a DerSimonian–Laird random-effects meta-analysis was performed to pool ORs per SD of BMI across cohorts. This produced a summary estimate of the association between childhood obesity and adult morbidity. As the SDs are specific to each cohort, these results should be interpreted as the OR per SD in a ‘typical’ population. To aid interpretation, some ORs were converted into odds per kg/m² of BMI, based on the SD of BMI in adolescents in the UK (about 2.5 kg/m²). 81

Meta-analyses of the predictive accuracy of childhood obesity for predicting future morbidity were also performed. For each study, the sensitivities for predicting the morbidity among overweight (85th centile of BMI) and obese (95th centile) children were calculated using the methods of Morris and Wald70 described above. These sensitivities and their 95% CIs were then pooled across studies in standard DerSimonian–Laird random-effects meta-analyses to summarise the predictive sensitivity of childhood obesity to predict adult morbidity. These pooled estimates were converted into summary ROC curves. The assumption that the disease is rare means that specificity cannot be calculated in these analyses; only sensitivity at specified thresholds of BMI were presented, rather than true ROC curves. All analyses were performed using the R version 3.1 software (The R Foundation for Statistical Computing, Vienna, Austria).

No analyses of potential for publication bias were performed given both the limited number of studies in each meta-analysis and the uncertainty over the validity of such tests when reviewing predictive or prognostic outcomes.

All cohort studies in this review reported single measurements of BMI rather than repeated measurements over time. The use of single measurements may lead to regression to the mean, because children with an extreme measurement at a particular time are more likely to have a less extreme result at some later time, leading to underestimation of the association. Although regression to the mean is an acknowledged problem in measurements that change rapidly over time, such as blood pressure or blood lipids,82 it is unlikely to be a problem for BMI, which by its nature can vary little over short periods of time.

The aim of this review was to assess the ability of childhood obesity to predict adult morbidity; for this reason, no adjustment for adult morbidity was considered.

Review of tracking of obesity into adulthood

The analysis of the studies tracking obesity from childhood to adulthood was based on the diagnostic accuracy data extracted from these studies. Data were reported for tracking from childhood obesity (defined as BMI above the 95th centile) or overweight (above the 85th centile) to adult obesity (defined as BMI above the 95th centile or over 30 kg/m²) or adult overweight (above the 85th centile or over 25 kg/m²).

This allowed for four separate tracking analyses:

• childhood obesity to later obesity

• childhood obesity to later overweight or obesity

• childhood overweight or obesity to later obesity

• childhood overweight or obesity to later overweight or obesity.

Some studies reported data on only one of these comparisons, some on two and some on all four, so certain studies contributed to only some of the analyses.

Ages were split into four categories:

• childhood (ages 7–11 years)

• adolescence (ages 12–18 years)

• adulthood (age 20 years and over)

• aged over 30 years (to examine longer-term obesity tracking).

Studies reporting on obesity below age 7 years were excluded to avoid including children with adiposity rebound.

Given these ages, analyses were conducted to examine tracking:

• from childhood to adolescence

• from childhood to adulthood

• from adolescence to adulthood

• from adolescence to aged over 30 years.

Analyses were not stratified by sex as these data were too limited. Only BMI was considered as only two studies reported data on any other obesity measure (WC).

To assess the association between childhood obesity and adult obesity, the risk of being obese in adulthood given obesity status in childhood was calculated for each study from the diagnostic data. The RR of adult obesity, with its 95% CI, when comparing obese and non-obese children was calculated. These RRs were pooled in a DerSimonian–Laird random-effects meta-analysis, stratified by age group at time of obesity measurement.

The diagnostic accuracy data were synthesised in meta-analyses to estimate summary sensitivity, specificity and positive predictive values (PPVs) across the studies. As sensitivity and specificity are known to be correlated, the bivariate model was used to jointly analyse sensitivity and specificity. 83 Summary ROC curves were produced using the hierarchical summary ROC (HSROC) model,84 which is mathematically equivalent to the bivariate model. PPVs were similarly jointly modelled with negative predictive values (NPVs) using an adaptation of the bivariate model. These analyses did not correct for differences in obesity prevalence, which will lead to heterogeneity in PPVs and NPVs. The bivariate model requires full diagnostic accuracy data (i.e. numbers of true and false positives, and true and false negatives); therefore, studies that only presented summary sensitivity and specificity data were excluded. All results were presented as summary estimates with 95% CIs and HSROC curves, plotted in ROC space. The more formal 95% joint confidence regions for sensitivity and specificity were not presented for reasons of clarity.

No analyses of potential for publication bias were performed given uncertainty over the validity of such tests when reviewing predictive or prognostic outcomes.

All analyses were performed using the R software, using in-house code designed to be equivalent to the metandi library available for Stata version 13 (StataCorp LP, College Station, TX, USA).

Review of diagnostic accuracy

The review of studies of diagnostic accuracy was primarily a narrative review owing to the substantial clinical heterogeneity across the studies. The assessment of diagnostic accuracy was conducted as a supportive question to the primary reviews of predictive accuracy. As a result, data extraction was planned for those thresholds used by studies in the reviews of predictive accuracy. Given that the majority of studies in the predictive accuracy reviews evaluated BMI, using a range of commonly used cut-off points (85th centile, 95th centile, one and two SDs), these data were extracted from the diagnostic accuracy studies evaluating BMI. The other index tests of interest were rarely evaluated in the studies of predictive accuracy, and the thresholds for these tests are less well established. As a result, data were extracted for a more diverse range of cut-off points. Where a study did not report the threshold used, these data were extracted. Study details and results were tabulated.

For the analysis of BMI, inclusion was restricted to those studies that used a cut-off point of 85th centile for overweight and 95th centile for obesity and were conducted in a population representative of the UK child population. This gave us the most homogeneous data set that could be derived from the BMI studies, and we considered these to be the most informative data and those that would be the most reflective of the accuracy of BMI when used in a setting such as a school, where screening often takes place. To produce summary estimates of accuracy for BMI, HSROC curves were produced and the bivariate model fitted as described above for the tracking review using Stata version 13.

For the studies evaluating tests other than BMI, a narrative synthesis was presented, as there were fewer studies that were too heterogeneous to pool. We considered the most robust data to be from those studies that conducted direct comparisons of more than one test (two or more tests conducted in the same patients, using the same reference standard in all patients). We therefore used these studies to gain an insight into the comparative accuracy of the tests being evaluated. The results of these studies were tabulated and a narrative synthesis provided; limited numbers of studies and the range of obesity measures considered precluded any meta-analysis of these data. To aid interpretation of these data, estimates of sensitivity and specificity from studies which compared the diagnostic accuracy of BMI with some other simple obesity measure were plotted in ROC space, and diagnostic ORs (DORs) were calculated. This could only be undertaken for those studies from which 2 × 2 data could be extracted or derived; therefore, there are some test results in the tables of sensitivity and specificity that do not appear in the diagrams.

To help understand why one test may have a better predictive accuracy than another, we planned to compare the sensitivities and specificities for predictive and diagnostic accuracy of those tests evaluated in the predictive accuracy reviews. There were insufficient predictive accuracy data for non-BMI measures for this to be completed.

Review of the acceptability and ease of use of simple childhood measures of obesity

Study details and results have been summarised in tables and a narrative provided. The results of the elicitation exercise have been summarised in a narrative, with the summary results presented in tabulated form.

Summary of the existing systematic reviews of the prediction of adult morbidities

Of the 13 reviews with potential for being updated that were identified,18–22,209–217 the reviews by Park et al. (2012),18 Reilly et al. (2011),22 Owen et al. (2009),21 Lloyd et al. (2010)20 and Lloyd et al. (2012)19 were the most recent and relevant to our review of adult morbidities. This subsection presents a brief critical summary of these five reviews.

Reilly et al. (2011)22 did not restrict their searches to any particular childhood measure of obesity, but only found studies on BMI and self-reported/recalled ‘weight status’. All other reviews included only BMI as a simple childhood measure of obesity. Most studies included in the reviews used cohort designs, although a smaller number of case–control studies were also eligible. Population eligibility criteria varied across the studies; Reilly et al. (2011)22 and Lloyd et al. (2010)20 included children from birth to age 18 years, but Park et al. (2012)18 and Lloyd et al. (2012)19 excluded children aged under 2 years. The review by Owen et al. (2009)21 reported on a significantly wider age range as it included individuals from birth up to age 30 years. The large majority of the studies were conducted in general populations from Europe (particularly Scandinavian countries and the UK) and the USA. None of the studies appeared to include specifically selected populations of children (such as children with cancer or children undergoing a weight reduction intervention). None of the reviews reported a restriction on sample size, and therefore the size of the studies varied widely, from under 100 to over 1.1 million participants.

Eligible outcomes varied across the reviews. The reviews by Park et al. (2012)18 and Reilly et al. (2011)22 had the broadest scope as they focused on both adult mortality and morbidity. Lloyd et al. (2010)20 reported outcomes relating to CVD events and CVD risk, and Owen et al. (2009)21 focused on CHD events. The review by Lloyd et al. (2012)19 was the only one to focus exclusively on metabolic syndrome and its individual components [type 2 diabetes, total cholesterol, triglycerides, high-density lipoprotein (HDL-)cholesterol, low-density lipoprotein (LDL-)cholesterol, fasting insulin, insulin resistance]. The number of individual studies included in the reviews ranged from 11 [Lloyd et al. (2012)19] to 39 [Park et al. (2012)18].

The quality of included studies was evaluated in three reviews. 18–20 However, only Lloyd et al. (201020 and 201219) reported the results of their quality assessments, which were based on the Newcastle Ottawa Scale. The key methodological limitations that they identified related to variation in type and number of adjustments made in the analyses of the primary studies, and the relatively young age of adults at follow-up in some studies. Quality relating to participant selection, comparability of study groups and measurement of outcomes was generally considered acceptable.

Only Owen et al. (2009)21 combined studies in a meta-analysis, stratified by baseline age group. All others reported the results of their included studies narratively by outcome. Lloyd et al. (201020 and 201219), Park et al. (2012)18 and Owen et al. (2009)21 also grouped studies based on whether or not they adjusted for adult BMI in their analyses.

In all reviews, a majority of studies reported a positive and statistically significant association between childhood BMI and adult morbidity. However, results varied by outcome. Both Park et al. (2012)18 and Reilly et al. (2011)22 found consistent evidence of a statistically significant positive association between childhood BMI and diabetes in adulthood. Park et al. (2012)18 reported that ORs for a one-unit increase in BMI SD score ranged from 1.22 (95% CI 1.10 to 1.36) at school entry (n = 5793) to 2.04 (95% CI 1.7 to 2.4) at 16 years of age (n = 10,683). Only 1 small study out of 10 found no association between the two variables. However, all three studies that adjusted for adult BMI found that the association was no longer significant after adjustment. In the review by Reilly et al. (2011),22 all three studies reporting diabetes outcomes found a significant increase in risk of developing the morbidity in those who were overweight or obese in childhood. Effect sizes were reported for only one study, showing significant association between overweight and obesity at age 5 years and self-reported diabetes at age 21 years (OR 2.6, 95% CI 1.3 to 5.2; n = 2639). Lloyd et al. (2012)19 found little evidence that childhood obesity is an independent risk factor for diabetes mortality, although their result was based on a single study.

Evidence of an association between childhood BMI and adult CHD was somewhat more mixed. The pooled analyses from Owen et al. (2009)21 showed that a 1-SD increase in BMI in early childhood [2–6 years (three studies)] was associated with reduced CHD risk (RR 0.94, 95% CI 0.82 to 1.07), although the association was not statistically significant. A weak positive association was found between later CHD risk and a 1-SD increase in BMI in later childhood [7 to < 18 years; RR 1.09, 95% CI 1.00 to 1.20 (seven studies)]. Both analyses showed high levels of heterogeneity. In the review by Park et al. (2012),18 10 of the 15 studies that explored the relationship between BMI in childhood and objective measures of CHD events reported that increased BMI or overweight at ages 2–25 years was associated with increased risk of CHD in later life. HRs ranged from 1.53 (95% CI not reported) for CHD mortality associated with high BMI at age 11 years to 5.43 (95% CI 2.77 to 10.62) for CHD associated with high BMI at age 17 years. One study (n = 181) found that higher BMI at ages 13–18 years was associated with increased risk of CHD morbidity and mortality in men but not in women. The remaining studies reported no association between BMI at ages 2–22 years and CHD. Reilly et al. (2011)22 and Lloyd et al (2010)20 both found that two out of three studies which examined the association between childhood BMI and CHD events reported a positive association.

Evidence of an association between childhood BMI and stroke in adulthood was mixed. Park et al. (2012)18 found that, out of nine relevant studies, only four found positive association between BMI and stroke in adulthood. Four studies (including three British studies) reported no association and one showed an inverse relationship between higher BMI at age 7 years and future stroke. Reilly et al. (2011)22 reported that two out of three studies which reported on stroke found a significant association between childhood BMI and morbidity in adulthood. Lloyd et al. (2010)20 found a pooled study from three historical cohorts and a single study which reported no significantly increased risk of CHD and stroke incidence in adults who were overweight or obese as children.

Evidence of an association between childhood BMI and the risk of developing metabolic syndrome in adulthood was limited. Lloyd et al. (2012)19 found only three studies which considered the link between the two variables. Two studies found that higher childhood BMI was associated with increased odds of developing metabolic syndrome in adulthood. The third study was the only one to adjust for adult BMI, and found that higher childhood BMI was associated with lower risk of metabolic syndrome in adult life when accounting for adult BMI.

Both Park et al. (2012)18 and Reilly et al. (2011)22 found consistent evidence of a statistically significant positive association between overweight or increased BMI in childhood and hypertension in adulthood. Park et al. (2012)18 reported ORs for overweight or increased BMI in childhood that ranged from 1.35 (95% CI 1.13 to 1.64) at age 7 years to 3.75 (95% CI 3.45 to 4.07) at ages 16–19 years in five studies. In the four studies identified by Reilly et al. (2011),22 ORs ranged from 1.46 (95% CI 1.01 to 2.12) to 5.1 (95% CI 1.4 to 18.1). On the other hand, Lloyd et al. (2010)20 found less consistent evidence that childhood obesity was an independent risk factor for hypertension.

None of the reviews reported studies focusing on hypercholesterolaemia as an outcome in adulthood. Lloyd et al. (2012)19 reported on studies focusing on LDL- and HDL-cholesterol levels. The review found little evidence to support the view that childhood obesity is an independent risk factor for higher LDL- and HDL-cholesterol levels.

Significantly less evidence was found on cancer. Park et al. (2012)18 found limited evidence that childhood BMI is associated with increased risk of colorectal and kidney cancers. Reilly et al. (2011)22 found a single study on the association between BMI and cancer, which reported a significant 30% increase (95% CI 10% to 54%) in smoking-related cancers per 1-SD increase in childhood BMI (n = 2997).

Reviews differed in their interpretation of the strength and clinical significance of the associations found. They also differed in the emphasis and importance placed on adjustments made for adult BMI. This, in addition to the fact that there was relatively little overlap across the reviews in terms of included studies, may explain why the conclusions of the reviews differed.

Reilly et al. (2011)22 concluded that there is a relatively large and highly consistent body of evidence which shows that overweight and obesity in childhood and adolescence have substantial and adverse long-term consequences for physical health. However, we feel that these conclusions were too strong; there was some variation in the size and direction of the effect estimates reported. Additionally, there were concerns about several aspects of the review methods and limited reporting of the study results. Therefore, the conclusions of Reilly et al. (2011)22 appear insufficiently cautious and may not be reliable.

The conclusions of Park et al. (2012)18 were more nuanced and more accurately reflected the evidence found. They stated that there is a consistent body of evidence for associations between childhood overweight, unadjusted for adult BMI, and cardiovascular outcomes and mortality in adulthood. They also concluded that evidence for stroke outcomes is mixed, and that there is limited evidence that childhood BMI is associated with increased risk of colorectal and kidney cancers. They also remarked that few studies have assessed the independent effects of childhood overweight on adult disease, and the results of these have been inconclusive.

Owen et al. (2009)21 concluded that higher levels of BMI may be related to increased levels of CHD risk from early childhood onwards. They stated that the control of obesity from childhood onwards may be an important priority for long-term CHD prevention, although they noted that the independent contribution of early obesity to CHD risk still needs to be established. The review by Owen et al. was the only one that conducted a meta-analysis, and the choice of methods appeared appropriate in view of the evidence found, although evidence of significant heterogeneity was reported. However, it is unclear why the authors concluded that higher BMI was associated with an increased risk of morbidity from early childhood even though they reported a weak inverse relationship with adult CHD risk in children under age 7 years. Additionally, the small magnitude of the pooled estimates in later childhood suggests that their clinical relevance may be limited in the context of adult CHD risk prediction. Therefore, the conclusions of the review may not be sufficiently cautious.

The conclusions from Lloyd et al. (2010)20 placed greater emphasis on the influence of adult BMI in the association between childhood BMI and adult morbidity than those of Owen et al. (2009)21 or Park et al. (2012). 18 The authors stated that they found little evidence to suggest that childhood obesity is an independent risk factor for CVD risk. They stated that, instead, relationships observed are dependent on the tracking of BMI from childhood to adulthood. Although the evidence found by Lloyd et al. (2010)20 appears insufficient to support the notion that childhood obesity is an independent risk factor for CVD risk, the review did not discuss the risks associated with standard adjustment for adult BMI, which may lead to overadjustment bias. 218 Additionally, the evidence presented by the review was insufficient to suggest that the relationships observed were dependent on the tracking of BMI from childhood to adulthood. Therefore, the conclusions of Lloyd et al. (2010)20 may not be reliable.

Lloyd et al. (2012)19 also concluded that there is little evidence to support the view that childhood obesity is an independent risk factor for adult blood lipid status, insulin levels, metabolic syndrome or type 2 diabetes. The conclusions reflected the evidence found, although the volume of the evidence on diabetes was significantly smaller than that reported by Park et al. (2012)18 or Reilly et al. (2011). 22

Among all five reviews, it appeared that the review by Park et al. (2012)18 was better conducted and yielded the most reliable conclusions. However, with the exception of the review by Owen et al. (2009),21 none of the reviews conducted a meta-analysis, which limited the strength of their results. Additionally, none included studies on childhood measurements other than BMI, which limits their relevance to our review questions. Some reviews addressed the issue of BMI tracking into adulthood as a potential confounder in the association between childhood BMI and adult morbidity. However, none formally evaluated the extent to which childhood measures of obesity accurately predict the tracking of obesity into adulthood. Finally, all reviews reported on the association between BMI and adult morbidities (generally expressed as ORs, RRs or HRs), but none in terms of predictive accuracy, and none used formal methods to calculate the predictive performance of childhood obesity for predicting future morbidity. This overview of recent systematic reviews highlights important gaps in the evidence for the predictive power of childhood obesity as a predictor of adult morbidity.

Quality and assessment of bias of the included primary studies for the prediction of adult morbidities

Thirty-seven studies (22 cohorts) met the inclusion criteria for the review of predictive accuracy. 117,118,123,129,140–172 The assessment of risk of bias is presented in Table 2 and the full quality assessment in Appendix 4. The most common methodological weaknesses were attrition bias and gaps in reporting about attrition rates, followed by outcomes bias (due to self-reporting of morbidities). Biases associated with selection, childhood measurements, use of confounders, reporting and analysis were low in most studies.

Seven studies used self-reported outcome measures and were considered at high risk of outcome bias. 118,123,129,148,155,160,168 One study used drug medication use as a proxy for hypertension and risk of outcome bias was unclear. 143 Bias associated with childhood weight status measurement was high in one study, where BMI was calculated based on adolescents’ self-reporting of height and weight. 129 One study did not report sufficient information on the site of WC measurements,168 and the only study that used SFT as a childhood measurement did not report how the sum of SFTs was calculated. 171 Thirty-two studies adjusted for age and sex in their analyses, and were therefore considered at low risk of confounding bias. Two studies did not adjust for sex in their analyses;118,169 one adjusted for sex but not age;146 and one did not adjust for either sex or age. 168 One study had a high risk of reporting bias due to partial reporting of results. 158

Characteristics of studies included in the review of prediction of adult morbidities

Findings from the review of adult morbidity

Meta-analysis

The meta-analysis was based on 26 published articles representing 13 cohort studies,117,129,140–142,145,146,148–151,153–155,158–166,168,169,172 which provided data that could be converted into an OR per 1-SD increase in BMI for each morbidity. There were 32 different classifications of morbidity reported. This analysis considered only those morbidities described in Chapter 2.

Analysis of the association between childhood obesity/overweight and adult morbidity

A series of plots of the OR per SD increase in BMI for having a morbidity (with 95% CIs) are presented for each cohort and each morbidity, according to the age at which BMI was measured. Figures 4–7 present results for diabetes, CHD and cancers, as examples. A plot showing the results for all morbidities is provided in Appendix 5.

FIGURE 4.

Childhood BMI and incidence of adult diabetes. NGHS, National Growth and Health study; NLSAH, National Longitudinal Study of Adolescent Health; PFS, Princeton follow-up study; SPEC, Staff Periodic Examination Center.

FIGURE 5.

Childhood BMI and incidence of adult CHD. SPEC, Staff Periodic Examination Center.

FIGURE 6.

Childhood BMI and incidence of adult breast cancer. MRC NSHD, Medical Research Council National Survey of Health and Development.

FIGURE 7.

Childhood BMI and incidence of other cancers.

Figure 4 shows the ORs for having adult diabetes, according to age. This shows that most cohorts showed a statistically significant association between childhood BMI and adult diabetes, with ORs ranging from 1.2 to 2.5 per SD BMI. Figure 5 shows the association with CHD. Results here were more varied, with two cohorts (Helsinki,150 Aberdeen159) showing no association and the others showing a modest increase in odds of CHD with increased BMI. Figure 6 shows the association of BMI with breast cancer. In this case there were no statistically significant results in any cohort. Figure 7 shows the results for other cancers. The cohort studies reported on a large number of cancers, but the general trend appears to be that higher BMI is associated with higher odds of most cancers.

The meta-analyses of ORs per SD BMI of these cohort studies are presented in forest plots, categorised by age at BMI measurement (Figures 8–13).

FIGURE 8.

Forest plot of the association between childhood BMI and diabetes. NGHS, National Growth and Health study; NLSAH, National Longitudinal Study of Adolescent Health; PFS, Princeton follow-up study; SPEC, Staff Periodic Examination Center.

FIGURE 9.

Forest plot of the association between childhood BMI and CHD. SPEC, Staff Periodic Examination Center.

FIGURE 10.

Forest plot of the association between childhood BMI and stroke.

FIGURE 11.

Forest plot of the association between childhood BMI and hypertension. BCAMSS, Beijing Child and Adolescent Metabolic Syndrome study; NLSAH, National Longitudinal Study of Adolescent Health.

FIGURE 12.

Forest plot of the association between childhood BMI and breast cancer. MRC NSHD, Medical Research Council National Survey of Health and Development.

FIGURE 13.

Forest plot of the association between childhood BMI and cancers.

Figure 8 shows the results of the meta-analysis of diabetes. There is a positive association, with an OR of 1.7 for every SD increase in BMI. That is approximately equivalent to a 24% increase in odds per kg/m² increase in BMI for an adolescent. There is no evidence that this is different in the under-twelves and over-twelves, although data for under-twelves are limited.

Figure 9 shows that there is a positive association between childhood BMI and CHD, but it is weak, with an OR per SD BMI of 1.2 in those aged over 12 years (approximately equivalent to an 8% increase in odds per kg/m² increase in BMI for an adolescent). Again, there is no evidence that the association differs between children aged under and over 12 years, although an increasing trend with age was observed in two cohorts (see Figure 5). 142,150

Figure 10 shows the association with stroke. There is little evidence of any association between childhood BMI and stroke; the statistically significant result in those aged over 12 years is based on a single-cohort study. 142 Figure 11 shows the association with hypertension. There is some evidence that childhood BMI is positively associated with odds of adult hypertension, but the results are based on only three cohorts. 117,129,165

Figure 12 shows the relationship between childhood BMI and breast cancer. The results are not statistically significant, but the trend is negative. Figure 13 shows the association between childhood BMI and other cancers. No pooled result is presented because of the diversity of types of cancer and because more than one cancer is reported in some cohorts, but the overall trend is for a modest positive association between BMI and the odds of cancer. The results are comparatively homogeneous, with an increase in odds of around 1.2 per SD increase in BMI.

It should be noted that, because of the incorrect assumption that BMI is normally distributed, all the summary ORs presented in Figures 8–13 are likely to be overestimates.

Prediction of future morbidity from childhood body mass index

The analyses so far have examined the association between childhood BMI and adult morbidity. This section considers how well childhood BMI predicts future morbidity.

Given the normality assumptions made about the distribution of BMI described in Chapter 2, the association between BMI and morbidity can be used to estimate the predictive performance of childhood obesity for predicting future morbidity. This is because there is a simple relationship between OR per SD and the sensitivity at a specified threshold of BMI (i.e. the proportion of adults with the morbidity who were above the specified threshold of BMI in childhood). As noted in Chapter 2, this assumption is likely to result in overestimation of these sensitivities.

Results for sensitivity at the 85th centile of childhood BMI and above are presented, that is what proportion of adult morbidity occurs in children who might be considered overweight or obese, based on their BMI. Figure 14 presents the results for diabetes, and Figure 15 those for CHD. Plots for other morbidities are given in Appendix 5.

FIGURE 14.

Predicting future diabetes from childhood BMI. NGHS, National Growth and Health study; NLSAH, National Longitudinal Study of Adolescent Health; PFS, Princeton follow-up study; SPEC, Staff Periodic Examination Center.

FIGURE 15.

Predicting future CHD from childhood BMI. SPEC, Staff Periodic Examination Center.

The pattern of these two figures is identical to that seen in Figures 4 and 5, although the scales differ. The pattern of results is likewise the same for all other morbidities (see Appendix 5). These two examples show that, despite the association between obesity and later morbidity, childhood obesity is a poor predictor of morbidity. At best, around 40% of adults with diabetes were overweight or obese (above the 85th centile of BMI) as children. Only around 20% of adult CHD morbidities occurred in the most overweight 15% of children. Based on these findings, it appears that childhood BMI is not a useful predictor of future heart disease.

Summary receiver operating characteristic curves

To summarise the ability of childhood BMI to predict future morbidity, summary ROC curves are presented, based, again, on the assumptions of normally distributed BMI, and using the results from the analyses presented in the forest plots above (see Figures 4–13). As noted in Chapter 2, these summary curves probably overestimate predictive accuracy.

There are two figures: Figure 16 shows summary ROC curves for predicting morbidities for ages 7–11 years, and Figure 17 shows the ROC curves for children aged 12 years and over (these data are too sparse to present ROCs for under-sevens). Breast cancer and stroke are not included, as the analyses above found no evidence of an association between BMI and these morbidities.

FIGURE 16.

Receiver operating characteristic curves for predicting future morbidity from BMI at ages 7–11 years.

FIGURE 17.

Receiver operating characteristic curves for predicting future morbidity from BMI at ages 12 years and over.

The figures show that the power of BMI measurement to predict adult morbidities is poor. BMI cannot adequately predict future CHD or cancer (the cancer ROC curve is hidden under the CHD curve in Figure 17). Only around 20% of future CHD and cancer occurs in overweight or obese children. If we consider only obese children (above the 95th centile), then obesity cannot predict CHD or cancer; the results are no better than chance. Even for diabetes, which has the strongest association with obesity, only around 35% of adult diabetics were obese or overweight as adolescents. The ROC curve for hypertension is presented, but may not be reliable as it is based on very limited data (two cohorts). It should also be noted that, because of the assumption that BMI is normally distributed, even these poor levels of predictive accuracy may be overestimates.

Review of morbidities: summary of findings and conclusions

This review identified 37 cohort studies on the association between childhood weight status and adult morbidities. Of these, 26 were included in the meta-analyses. All the studies were large or very large cohort studies, most being from large national or community data sets. However, large proportions of participants were lost at final follow-up in many studies, making attrition the most common source of bias. Although most studies adjusted for key confounders such as age and sex, the observational design of the studies means that confounding bias cannot be ruled out. Studies were very diverse, with weight status measured at different ages and in different populations. Most studies tended to focus on older cohorts. Follow-up duration and age at final follow-up also varied significantly between the studies. In addition, relatively few studies were found for some outcomes. Given the range of morbidity-related outcomes that could be reported (particularly the large number of types of cancer), there is a possibility of selective reporting, where cohorts reported only outcomes showing a positive association between BMI and the outcome. These limitations should be considered when interpreting the results.

Overall, most studies reported a positive, statistically significant association between measures of childhood weight status and adult morbidity, but the findings varied by morbidity. There was consistent evidence of a statistically significant but modest association between childhood BMI and adult diabetes; there was insufficient evidence to show whether or not the association is stronger among older children. Results for CHD were more mixed, with some cohorts showing a negative association (in early childhood) or no association, and others showing a modest increase in odds of CHD with increased BMI. There is insufficient evidence to show whether or not the association differs significantly between under- and over-twelves. However, evidence from some cohorts suggests an increasing trend with age. There is little evidence of any association between obesity and stroke; the statistically significant result in the over-twelves is based on a single cohort study.

There was some consistent evidence that childhood obesity and overweight are positively associated with metabolic syndrome in adulthood. There was some very limited evidence suggesting that the strength of the associations may vary according to age and measurement tool. There was also some consistent evidence that childhood obesity and overweight are positively associated with odds of adult hypertension, although the results are based on a relatively limited number of cohorts. There was some evidence of a statistically significant positive association between BMI in adolescence and hypercholesterolaemia in early adulthood, although the evidence is based on a single cohort with limited follow-up.

The cohorts reported on a large number of cancers (although few cohorts were found for each site-specific cancer), and the general trend appears to be that obesity leads to a modest increase in risk for most cancers for which evidence was available (except for breast cancer and rectal cancer).

Despite the association between obesity and later morbidity, childhood BMI is a poor predictor of adult morbidity. At best, around 40% of adults with diabetes were overweight or obese (above the 85th centile of BMI) as children. There was some evidence from a single cohort that BMI and WC had equally high specificity and poor sensitivity when used to predict future diabetes. At most, 20% of future CHD and cancers occur in overweight or obese children. If we consider only obese children (above the 95th centile), then obesity cannot predict CHD or cancer; the results are no better than chance. Therefore, childhood BMI is not a good predictor of future heart disease or cancer.

There was some evidence that BMI has good specificity but poor sensitivity when used to predict metabolic syndrome in adults, although this finding is based on very few and small cohorts. This means that most adults who do not develop metabolic syndrome as adults were also not obese as children; however, this tells us nothing about obese children. There is some evidence from a relatively small cohort that BMI values at ages 11 and 16 years were weakly predictive of hypertension in adulthood. No evidence was found on the predictive value or association between childhood weight status and hypercholesterolaemia in adulthood.

The evidence for the predictive accuracy of simple measures other than BMI was limited, and could not be included in the meta-analyses. Only three relatively small studies used WC, and only one used WHR and measurements of SFT. Based on this evidence, the extent to which these simple measures of childhood weight status can be compared is very limited. Only one study reported predictive accuracy data on WC and BMI, and found that both had equally high specificity and low sensitivity when predicting future adult morbidity.

Overall, there was consistent evidence of positive associations between childhood measurements of weight status and morbidities in adulthood. Although higher childhood BMI tends to be associated with a higher risk of morbidity in adulthood, overall it is a poor predictor of adult morbidity. This is because the association between obesity and morbidity would need to be strong (i.e. a large RR) in order for obesity to be an effective predictor of later morbidity. 225 There is also no evidence favouring the accuracy of any of the other simple measures of childhood weight status for predicting future adult morbidity.

Summary of the existing systematic reviews by Singh et al. (2008) and Brisbois et al. (2012)

The systematic review by Singh et al. (2008)16 examined the evidence of the persistence of childhood overweight. The review included prospective or retrospective longitudinal studies that made at least one anthropometric measurement during youth (age ≤ 18 years) and one during adulthood (age ≥ 19 years), using BMI, SFT or WC. Eighteen studies in 25 articles were included in the review. The quality of the included studies was assessed in duplicate by the two reviewers and 52% of the articles were rated as being of high methodological quality.

All the included studies reported that overweight or obese youths were at increased risk of becoming overweight or obese adults. Four studies that stratified for different levels of youth overweight suggested that persistence of overweight was greater with increased level of overweight. It was unclear if persistence varied with sex. When only high-quality studies were included in the analysis, the risk of overweight children becoming overweight adults was at least double that for normal-weight children. Overall, the percentage of obese adolescents becoming overweight or obese adults was higher than that of obese children (24–90% based on three high-quality studies). The review concluded that there is strong evidence of moderate persistence of overweight in both overweight and obese children. A significant limitation of the review was that it was unclear how comprehensive the searches had been, and whether or not relevant studies may have been missed. Furthermore, it presented only a narrative synthesis of the results from the included studies and focused on the association between childhood and adult obesity rather than the predictive accuracy of childhood obesity. In addition, as the review’s searches had been conducted in February 2007 the list of included studies was potentially out of date. Therefore, this review was used only as a source of relevant studies for the current review.

The systematic review by Brisbois et al. (2012)17 evaluated a number of factors in children aged 5 years or under which are the most significant predictors of the development of obesity in adulthood. Studies were eligible if they included healthy individuals, performed child and adult measurements for each individual and reported quantitative results. The searching appeared adequate but the lack of overlap with Singh et al. (2008)16 raised questions over how comprehensive it had been.

The review included 24 studies, and almost all reported a significant association between childhood overweight or obesity and adult obesity. The review concluded that early childhood obesity was a probable marker of adult obesity.

The main limitations of this review were that only early childhood was considered and it was possible that studies had been missed. Furthermore, many of the studies included reported only correlation coefficients, and only a narrative synthesis of study results was conducted. Therefore, as with the review by Singh et al. (2008),16 this review was used as a source of relevant studies for the current review.

Quality and assessment of bias of the included primary studies of tracking of obesity into adulthood

In total, 23 studies were identified for the review of the tracking of childhood obesity into adulthood. These are listed in Table 8.

The results of the quality and risk of bias assessment are summarised in Table 8 and given in full in Appendix 4. All the studies were cohort studies, and so they are already subject to greater bias than a randomised controlled trial would be. However, all the studies were large, most being from national or community data sets. The main potential source of bias arises from the lack of control in retaining participants in the cohorts; thus, although the initial cohorts were very large, many thousands of participants were not followed up at the final assessment. In itself this might not be a source of bias, but the majority of study reports did not confirm that the final cohort followed up was representative of the complete initial sample. Confounding was another significant potential source of bias in these studies. Although some presented results by age and sex, most did not, and the wide age ranges included in some studies increased the potential for age-associated effects that could not be accounted for.

Characteristics of studies included the review of tracking of obesity into adulthood

There was some overlap between a number of the studies, as they were based on the same initial cohorts. Three were based on the Bogalusa cohort in the Southern States of the USA, and two each were based on the National Child Development Study (1958 British birth cohort) (NCDS), the National Longitudinal Study of Adolescent Health (NLSAH) cohort, the Young Finns cohort, Tokushima Prefecture and the Mater-University Study of Pregnancy (MUSP) cohort in Australia (Table 9). Each study included in the review provided some unique data in terms of age of childhood assessment, adult assessment or analysis by age, sex or race. Only unique data are included in the meta-analysis and narrative synthesis.

Findings of the review of tracking of obesity into adulthood

Data extracted from the included studies, and calculated sensitivity, specificity and PPV by study, are presented in Appendix 3.

Overview of the studies of tracking of obesity into adulthood

To give an overview of these data, we present three figures showing sensitivity, specificity and PPV according to childhood age of BMI measurement and sex for each study. These plots show the diagnostic performance of using obesity in children (> 95th centile), according to age at measurement, to predict obesity in adults (aged ≥ 20 years).

Figure 18 shows that sensitivity is generally low – less than 40% – with the exception of one study [Thompson et al. (2007)137]. Most obese adults were therefore not obese in childhood. There is no obvious trend in sensitivity with age or sex. There appears to be considerable heterogeneity across studies, with sensitivities ranging from 10% to over 60% at age 16 years. Figure 19 shows that specificity is much higher and more consistent, ranging from 93% to nearly 100%. This means that nearly all non-obese adults were also not obese as children; however, this tells us nothing about the obese adults and children who are of concern. The PPV tells us what proportion of obese children go on to become obese adults. This is shown in Figure 20. The results are very heterogeneous, ranging from 40% to over 90%. The PPV depends on the prevalence of obesity, so some of this heterogeneity may be a result of varying obesity prevalence across studies.

FIGURE 18.

Sensitivity for predicting adult obesity from childhood obesity.

FIGURE 19.

Specificity for predicting adult obesity from childhood obesity.

FIGURE 20.

Positive predictive value for predicting adult obesity from childhood obesity.

The results do suggest that obesity is moderately persistent, at least between adolescence and adulthood.

Meta-analysis of tracking of obesity into adulthood

All tracking studies that reported full diagnostic data (i.e. numbers of true and false positives, and true and false negatives) were included in the meta-analyses, except for the study by Liddle et al. (2012),127,128 which was of children aged 5 years only and therefore did not fall into the lowest age category of the meta-analysis (7–11 years); the results of this study are presented in the previous section. Different studies reported data on different comparisons between childhood and adult obesity or overweight, and at different ages. Table 12 gives a summary of which studies were included in which analyses; not all publications provided data for the meta-analysis. Studies reported tracking in a variety of ways, with some, for example, examining tracking of childhood obesity to adult obesity, some adolescent overweight to adult obesity and some childhood overweight to adolescent overweight, and therefore most meta-analyses include few studies.

TABLE 12 - Inclusion of tracking studies in the various meta-analyses

Note: four studies did not provide unique data for the meta-analysis: Freedman (2001),120 Juhola (2011),125 Mamun (2005)128 and Merten (2010). 129

Study	Obesity categories				Age categories
	Childhood obesity to:		Childhood overweight or obesity to:		Child to adolescent	Child to adult	Adolescent to adult	Adolescent to over 30 years
	Adult obesity	Adult obesity or overweight	Adult obesity	Adult obesity or overweight
Cheng (2011)117	✗	✗	✗	✗	✗
Power (1997)133	✗	✗				✗	✗	✗
Reilly (2011)134	✗	✗	✗	✗	✗
Starc (2011)135	✗	✗	✗	✗	✗
Venn (2007)138	✗	✗	✗	✗		✗
Freedman (2005)122	✗		✗			✗	✗
Gordon-Larsen (2004)124	✗		✗				✗
Juonala (2006)126	✗		✗				✗	✗
Thompson (2007)137	✗					✗	✗
Wang (2008)139	✗		✗				✗	✗
Freedman (2005)121		✗		✗		✗
Patton (2011)132		✗		✗			✗
Engeland (2004)119			✗	✗			✗	✗
Nakano (2010)130				✗	✗
Nakano (2010)131				✗	✗
Stovitz (2010)136				✗	✗

Association between childhood weight status and adult obesity

Figure 21 shows the forest plot for the meta-analysis of RR of being obese as an adult among children who were obese. The association is strong, with obese children being more than five times more likely to be obese as adults than non-obese children. There is no apparent difference in this RR between younger and older age groups.

FIGURE 21.

Forest plot of the association between childhood obesity and adult obesity.

Tracking from childhood weight status to adult obesity

Childhood obesity to later obesity

Figure 22 shows the results of the joint meta-analyses of sensitivity and specificity when predicting adult obesity (≥ 95th centile or > 30 kg/m²) from childhood obesity (≥ 95th centile of BMI). Four results are shown: child to adolescent (three studies), child to adult (four studies), adolescent to adult (six studies) and adolescent to over 30 years (three studies). The points show the summary sensitivities and specificities with their 95% CIs; the curves are the summary HSROC curves.

FIGURE 22.

Tracking of childhood obesity to adult obesity.

Figure 22 shows that obesity tracking is strongest over short time periods. Childhood obesity does track reasonably well to adolescent obesity; 62% of obese adolescents were obese in earlier childhood. A high specificity of 95% means that nearly all non-obese adolescents were non-obese children. For tracking into adulthood, the specificities of around 98% mean that nearly all non-obese adults were not obese in childhood or adolescence. However, by contrast, the low sensitivities mean that most obese adults were not obese as children. About 70% of obese adults were not obese in childhood or adolescence and 80% of over-thirties were not obese in adolescence.

Childhood overweight or obesity to later obesity

Figure 23 shows the results for whether or not childhood obesity or overweight (≥ 85th centile of BMI) tracks to later obesity (≥ 95th centile or ≥ 30 kg/m²). Four results are shown: child to adolescent (three studies), child to adult (two studies), adolescent to adult (five studies) and adolescent to age over 30 years (three studies). The pattern of results is similar to that for childhood obesity, with low sensitivity to detect adult obesity; around 60% of obese adults were neither overweight nor obese as adolescents. Childhood overweight is, however, a good predictor of adolescent obesity; 90% of obese adolescents were overweight or obese in earlier childhood.

FIGURE 23.

Tracking of childhood overweight or obesity to adult obesity.

Tracking from childhood to adult overweight or obesity

Childhood overweight or obesity to later overweight or obesity

Figure 24 shows the results of the joint meta-analyses examining whether or not childhood overweight or obesity (≥ 85th centile of BMI) tracks to later overweight or obesity (≥ 85th centile or > 25 kg/m²). Three results are shown: child to adolescent (six studies), child to adult (two studies) and adolescent to adult (two studies). There was only one study with data from adolescence to age over 30 years, so no analysis could be performed. We note that there were few data here for tracking into adulthood.

FIGURE 24.

Tracking of childhood overweight or obesity to adult overweight or obesity.

The results for childhood to adolescence suggest that 65% of obese or overweight adolescents were obese or overweight at younger ages. The limited data for tracking to adulthood suggest very low sensitivity; about three-quarters of overweight adults were not overweight in younger childhood.

Childhood obesity to later overweight or obesity

Figure 25 shows how childhood obesity (excluding overweight) tracks to adult obesity or overweight. Three results are shown: child to adolescent (three studies), child to adult (three studies) and adolescent to adult (two studies). The results are broadly similar to those obtained for using childhood overweight to predict adult overweight. The results for childhood to adolescence suggest that 35% of obese or overweight adolescents were obese at younger ages. The substantial majority of overweight adults (around 90%) were not obese as children.

FIGURE 25.

Tracking of childhood obesity to adult overweight or obesity.

Positive predictive values

Meta-analyses of the PPV, which estimates the proportion of obese children who will go on to become obese adults, are presented in Figures 26 and 27. The analyses are similar to those presented above, with the same studies in each analysis as for the equivalent analyses given above.

FIGURE 26.

Positive predictive value for childhood obesity when tracking to adult obesity.

FIGURE 27.

Positive predictive value for childhood overweight and obesity when tracking to adult obesity.

Figure 26 shows the PPV and NPV for predicting adult obesity from childhood obesity, classified according to age as before. This shows that obesity is moderately persistent into adulthood; just under 80% of obese adolescents were still obese as adults, although that declines to 70% when adult ages are restricted to over 30 years. Childhood obesity does not track so strongly into adolescence, however, with just over half of all children still obese as adolescents, although data were limited and therefore CIs are wide. The NPV results show that non-obesity does track into adulthood; more than 85% of non-obese children will still be non-obese in adulthood. However, this does mean that a small proportion (10–15%) of non-obese children will be obese by early adulthood.

Figure 27 shows the equivalent PPV and NPV for tracking of childhood overweight or obesity to later obesity. The pattern of results is similar, but the PPVs are smaller as many overweight children will not become obese adults. Around 66% of obese or overweight adolescents will be obese in early adulthood, dropping to around 55% by age 30 years. Only about 30% of obese or overweight children will go on to be obese in adolescence.

As both PPV and NPV depend on the prevalence of obesity, the results of these analyses should be interpreted as predictive values in a population with average obesity prevalence.

Conclusions of the tracking review

This review identified 23 studies117–139 which tracked childhood obesity or overweight into adolescence or adulthood. Of these, 20 were included in the meta-analyses. These studies were very diverse, including studies from Europe, the USA and Asia, with widely differing sample sizes. The age at which childhood assessment was carried out ranged from 7 to 18 years. Most studies included in the meta-analysis were at an unclear risk of attrition bias and confounding was considered only in terms of age and sex, if at all. BMI was measured at various ages in childhood, adolescence and adulthood, and in various populations. As a result of this diversity, very few studies were included in each specific meta-analysis, leading to considerable uncertainty in the results and wide CIs. This uncertainty should be considered when interpreting the results.

The analysis found a strong association between childhood obesity and adult obesity, with obese children being about five times more likely to be obese as adults than similar, non-obese children. This strong association is reflected in the high PPVs: just under 80% of obese adolescents go on to be obese adults. This drops to 70% when only adults over age 30 years are considered, so consequently around 30% of obese adolescents will not be obese 20 years later. Tracking of obesity from childhood to adolescence was generally good, with around 60% of obese adolescents having been obese as children, and around half of obese children still being obese in adolescence.

By contrast, most obese adults were not obese as children or adolescents: just 30% of obese adults were obese in adolescence, and around 60% of obese adults were of healthy weight during childhood and adolescence. Using childhood obesity to predict adult obesity will therefore fail to identify most adult obesity.

No data were available for tracking beyond age 40 years, but given the available data it seems that many obese children will not be obese by their forties or later (when obesity-related morbidities are most common), and most people who are obese at later ages will not have been obese in childhood.

Quality of the evidence for diagnostic accuracy of childhood measures of obesity

Thirty-four studies met the inclusion criteria for the review of diagnostic accuracy. 173–206 The full results of the quality assessment, along with the guidance for scoring each criterion, are given in Appendix 4. There were some duplicate papers for some cohorts, where more than one paper provided unique data. Three papers reported data from the NHANES study;101,103,180 data were extracted from the one that provided 5 × 5 data in the largest population (n = 7356) for both BMI and SFT using DEXA as the reference standard. 180 The other two NHANES papers were not extracted; one reported only PPV for BMI,103 and the other 3 × 3 tables using cut-off points < 5th, 5th–85th and > 85th internally derived centiles, so providing 2 × 2 data for overweight only, for a small subgroup of children (n = 58) with cerebral palsy and ‘reference children’ (n = 6754). 101 Three papers were identified from the Stockholm Weight Development Study. 102,192,193 One of these was a summary of the results presented in the other two papers, and was therefore excluded. 102 Data were extracted from both of the other studies; one provided data for BMI, WC and hip circumference (HC),193 and the other for a further three classification systems of BMI. 192 Two studies reported results from the Pediatric Rosetta Body Composition study. 179,190 Both studies were included in the review as one reported sensitivity and specificity for BMI of the entire population, along with SFT for boys and girls separately and for the entire population,190 and the other reported a contingency table for boys and girls separately for BMI only, from which summary measures were calculated for those subgroups. 179

All 34 studies avoided differential verification bias and incorporation bias; only one study was subject to partial verification bias, with the results of DEXA being imputed for some children. 180 Only two studies used the gold standard, a multicomponent model;175,182 the other studies used an imperfect reference standard. Of the 32 using an imperfect reference standard, three used D₂O,173,204,205 one used ADP193 and four used underwater (hydrostatic) weighting,186,187,189,197 but most used the least reliable of the eligible reference standards, DEXA. 174,176–181,183–185,188,190–192,194–196,198–203,206 Brief study details are provided in Table 15.

Eighteen studies recruited either consecutive or randomly selected children from populations representative of the UK population who would be eligible for the test in clinical practice. 173,177–181,186,189–195,201,202,205,206 One study recruited a population representative of the country in which the study was conducted. 188 Thirteen studies did not recruit a representative population, for either the country in which the study was conducted or the UK;174–176,184,185,187,196–200,203,204 these studies generally focused on specific subgroups of children. The representativeness of the population was unclear in two studies. 182,183

Most of the studies were single-gate diagnostic cohort studies undertaken specifically to determine the accuracy of a simple anthropomorphic measure (the index test). 173,175–179,181–184,186–189,193,194,196–205 Although the time lag between the index test and reference standard test was not stated, it is unlikely that this was sufficiently long for progression bias to be present. Five studies were cross-sectional, with aims beyond the determination of the diagnostic accuracy of an index test, and had either a prospective185,190,191,206 or retrospective/post-hoc180 analysis; these were also considered to have a low likelihood of progression bias. Three studies were longitudinal, and the timing of the index and reference standard tests was not specifically reported; therefore, the potential for progression bias was unclear. 174,192,195

The description of the index tests was adequate in most studies; the details provided for the reference standards were generally more poorly reported, making reproducibility of the reference standard tests more difficult. Most of the studies apparently had no uninterpretable results and no withdrawals. The number and expertise of the assessors for both the index tests and reference standards were unclear in most studies.

Owing to the poorer accuracy of DEXA compared with the other reference standards, none of the studies using DEXA was considered to be of high quality. Of the 10 studies that used a reference standard other than DEXA, eight were considered to be high quality,173,186,187,189,193,197,204,205 one did not explain withdrawals175 and another excluded withdrawals from the study. 182

Diagnostic accuracy of body mass index

A total of 30 diagnostic accuracy studies evaluated BMI. The study and population characteristics varied across these studies. In clinical practice, weight status is commonly defined using the cut-off points of 85th centile for overweight and 95th centile for obesity. In addition, the measures tend to be used in a diverse population of children. Several of the studies did not report the thresholds used, used other thresholds to define overweight and obesity or conducted the study in a population that was not representative of the UK child population. This section therefore concentrates on the 11 diagnostic accuracy studies of BMI that used the 85th centile for diagnosing overweight and the 95th centile for diagnosing obesity (regardless of the reference population used to standardise the measure for age) in unselected boys, girls or children of both sexes reflective of the UK child population (see Table 16). Results for the other studies of BMI are presented in Appendix 3, which includes the findings of all the included diagnostic accuracy studies.

The sensitivities and specificities of the 11 BMI studies are presented in Table 16. Of these, only one study was considered to be of high quality (results shown in italics in Table 16). 186 Sensitivity varied considerably across these studies, ranging from 23% to 96% for diagnosing obesity and 19% to 94% for diagnosing overweight. Specificity was less variable, ranging from 89.4% to 100% for diagnosing obesity and 82% to 100% for diagnosing overweight. Data were available as 2 × 2 contingency tables for eight studies diagnosing overweight177,179,180,188,191,192,197,203 and six studies diagnosing obesity;177,179,180,186,188,195 2 × 2 tables were not available for two studies. 181,190

Using the 2 × 2 data, the pooled sensitivity, specificity, positive likelihood ratio (LR+) and negative likelihood ratio (LR−) were calculated using the HSROC model (Figure 28); analyses were conducted for boys and girls separately and combined (Table 17). Two analyses were conducted for boys and girls combined; one used the separate boy and girl subgroups,177,179,180,186,188,191,192,195,197 adding data for the one study that did not report for girls and boys separately,203 and in the other, data from the within-study boy and girl subgroups were combined to create ‘all children’ populations prior to analysis. 177,179,180,186,188,191,192,203 The former included the one study that reported results for boys only;197 this study was excluded from the latter analysis.

FIGURE 28.

Hierarchical summary ROC curves for BMI for (a) unselected boys; (b) unselected girls; (c) all children (boy and girl subgroups kept separate); and (d) all children (within-study boy and girl subgroups combined prior to analysis).

The results of these analyses (see Table 17) show that BMI has a relatively low sensitivity, but a high specificity, across the different subgroups for diagnosing overweight/obesity. With such high specificities, there are few false-positive results; therefore, BMI is better at ruling obesity in than ruling it out, hence the moderately high LR+ but poorer LR−. These summary estimates should be treated with caution, as they are based on a subset of the BMI studies, and heterogeneity was substantial (see Figure 28); there were insufficient studies to investigate the impact of this heterogeneity. The analyses also combined data from the same population using different thresholds from some studies. As a result, the uncertainty surrounding the pooled estimates (95% prediction region) is large.

Two studies provided insufficient data to derive 2 × 2 tables. Of these, one reported sensitivity of 92.5% and specificity of 89.4% for diagnosing obesity (95th centile, CDC 2000219),190 and the other reported an AUC of 0.983 (95% CI 0.968 to 0.999) for boys and 0.981 (95% CI 0.961 to 1.00) for girls for diagnosing obesity (95th centile, reference population unclear). 181

Diagnostic accuracy of skinfold thickness

Of the 34 included studies, 10175,180,182,184,186,187,189,190,196,197 evaluated SFT. Where reported, the most commonly used cut-off point was the 85th centile. Most of the studies used the measure to diagnose obesity; data for diagnosing overweight were more limited. Four studies reported the sensitivity and specificity of triceps SFT, or data to calculate these outcomes, for the diagnosis of overweight or obesity (Table 18);186,187,189,190 of these, three were considered to be high quality (results shown in italics in Table 18). 186,187,189 Sensitivity ranged from 23% to 98.5%, and specificity from 78% to 100%, across the populations and thresholds.

Two studies evaluated subscapular SFT. 186,190 Sensitivity ranged from 30% to 98.5%, and specificity from 79.1% to 99% across the two studies (Table 19). Six studies evaluated sum SFT (Table 20), where more than one site was measured and an overall percentage body fat was determined. 175,180,184,186,189,197 As with the single-site SFT measurements, sensitivity varied across the six studies, but was, on the whole, higher than that for triceps and subscapular SFT individually, ranging from 57% to 98% across populations and thresholds. Most had a sensitivity of 80% or over. Specificity was more variable for the sum SFT than for the individual SFT measures, ranging from 47% to 96%, but was still reasonably high for most population and threshold combinations (see Table 20).

Diagnostic accuracy of waist circumference

Four studies reported sensitivity and specificity for WC for diagnosing obesity,181,193,195,205 and four studies for diagnosing overweight184,193,197,199 (Table 21); of these, three were considered to be of high quality (results shown in italics in Table 21). 193,197,205 As with BMI, there was substantial variation in sensitivity across the studies, ranging from 34.7% to 100% for diagnosing obesity and 53% to 90.9% for diagnosing overweight. Specificity was less variable, ranging from 81% to 100% for diagnosing obesity and 92% to 100% for diagnosing overweight.