Ethics

As this project involved only secondary analysis of anonymised publically available data ethical approval was not required. Ethical approval for the original data collection was granted by the London Multi-Centre Research Ethics Committee.

Participants

Data from the HSE 2007 was utilised. Information on under 16 year olds was obtained from two components of the survey. First, data on children living with adults were gathered as part of the stratified random 'core sample' of 7,200 households in England. Second, a 'child boost' component to the survey obtained information on children from a stratified random sample of 26,100 selected addresses [32]. In both cases, where more than two children resided at the address two children were randomly selected for interview. Consequently a total of 6,882 adults and 7,504 children were interviewed, with 1,727 children from the core sample and 5,777 from the boost. Those aged 13-16 were interviewed directly about health and lifestyle issues whilst this information was obtained via parents for younger participants. The full methodology of the HSE 2007 is detailed in the survey technical documentation and reports. In terms of sociodemographic characteristics the samples were representative at both a regional and national level [32]. For the purposes of this analysis only data from children aged 5-16 years was utilised; this is the age range for which the Strengths and Difficulties Questionnaire (SDQ) has been validated.

Measures

Interviewers measured the weight and heights of children. These were first converted to BMIs (kg/m²) then to standardised BMI z-scores that were adjusted for age and gender using data obtained from the 1990 growth reference dataset [33]. Children were then classified as overweight or obese according to the International Obesity Task Force (IOTF) recommended cut-offs for standardised BMI [34].

Socioeconomic status was evaluated according to equivalised household income (total household income adjusted for the number of people dwelling there). Ethnicity was reported to interviewers and grouped into White/Black/Asian/Mixed and 'Chinese or other' ethnicities. Estimated time spent engaged in physical activity over the preceding week was also reported to the interviewer. Where reported activity levels were less than 30-59 minutes of moderate to vigorous physical activity per day over the last seven days the child was categorised as having activity levels likely to be significantly below the current Department of Health for England recommendations [30].

The parentally completed version of the Strengths and Difficulties Questionnaire (SDQ) was used to evaluate child psychological wellbeing [35]. The SDQ is traditionally divided into five subscales (Conduct Problems, Emotional Symptoms, Hyperactivity, Peer Problems and Prosocial Behaviour) according to the originally proposed factor structure. An overall estimate of psychological adjustment is derived from the summed scores of the former four of these five subscales (the total difficulties score). The SDQ has been validated against semi-structured diagnostic interviews in terms of the instruments ability to detect clinically significant behavioural or emotional disturbance. The parental version of the instrument has 62.1% sensitivity to detect any psychiatric disorder, 73.5% sensitivity to detect clinically significant conduct problems and 69.2% sensitivity to detect depression in children aged 5-10 years. For children aged 11-15 years these values are 59.4%, 77.3% and 61.1% respectively [36]. Thus, as might be expected, parental reports using the questionnaire are better at detecting behavioural rather than emotional problems. Despite this, it should be noted that the parental SDQ is better at detecting depression in children and adolescents than the self-report version of the instrument. A recent reanalysis of a large community-based sample of SDQ respondents suggests that in non-clinical (i.e. low-risk) populations a scoring system based on a three factor structure (internalising, externalising and prosocial behaviour) may be more appropriate [37]. This, more parsimonious, structure was reported to show the clearest and most consistent evidence of convergent and discriminant validity across informants and reliability with respect to the diagnosis of clinical disorder. Thus, using the broader internalising and externalising dimensions may therefore be more appropriate as predictor or dependent variables for epidemiological studies. For this reason, when evaluating emotional and behavioural symptoms, factor scores were utilised as the estimates for the internalising (emotional) and externalising (behavioural) latent variables respectively. Factor (rather than summed) scores were utilised in this case as in the present sample factor loadings were found not to be tau-equivalent (i.e. factor loadings significantly varied across items). However, normative data on this alternative SDQ structure is not yet available. Therefore for mental health screening purposes the recommended cut-off score of five or more for both Conduct Problems and Emotional Symptoms subscales of the SDQ was utilised [36]. Screening also usually utilises the SDQ 'impact score'. This reports whether the parent considers the child's functioning has been affected by any reported symptoms. As the impact supplement was not included in interview schedule for the HSE 2007 screening thresholds were defined on the basis of subscale total scores only, computed on the basis of the algorithm provided by the questionnaire authors on the SDQ website [38].

Statistical Analysis

As clustering occurred due to second stage sampling procedures a multilevel approach to model evaluation was utilised to allow for the non-independence of observations from children nested within the same home. Thus, a random intercept with covariates model was used to explore the relationship between the dependent (reported psychological adjustment) and predictor variables. Sampling weights can potentially be employed in the multilevel analysis of complex survey data but both cluster and individual level weights must be rescaled [39]. As cluster level probability sampling weights were not available for children in the child boost sample this strategy could not be used. When investigating potential cross-level effects, random coefficients for the regression slopes between obesity and internalising/externalising factor scores were also introduced. Household income was therefore treated as a level two variable whilst other observations were on the child level (level one). Dummy variables were created for categorical items used in regression-based analyses. Continuous explanatory variables were mean-centred. In order to examine the likelihood of a child exceeding the SDQ screening threshold score for a potentially clinically significant emotional or behavioural disorder a multilevel logistic regression was performed. Thirty quadrature points were specified to ensure accurate estimates.

All analyses were performed using Stata SE version 11 [40], with the exception of the investigation of cross-level interaction and derivation of factor scores which utilised Mplus version 6 [41]. Factor scores were derived via a Confirmatory Factor Analysis (CFA) performed using Robust Weighted Least Squares as the estimation method to allow for the ordinal nature of the SDQ ratings.

Sociodemographic characteristics

For those participants aged 5-16 years with a valid BMI and completed SDQ the mean age was 10.1 years (sd 3.1) and 51% (2,017) were male. Average equivalised household income was £25,644/year and the mean daily physical activity levels reported were 89 minutes/day (sd 88 minutes). In terms of ethnicity 3,392 (85.8%) of the sample were classified as White, 258 (6.5%) as Asian, 137 (3.5%) as Black, 135 (3.4%) as Mixed and 31 (.8%) as Chinese/Other. Ethnicity was not reported in three cases.

Univariate Analysis

A univariate analysis was performed to explore the relationship between parentally reported psychological adjustment and obesity and also to identify any potential confounding/mediating variables. Both mean total SDQ score (as a marker of overall psychological adjustment) and the internalising (emotional) and externalising (behavioural) symptoms factor scores were significantly higher in children classified as obese but not overweight, according to the IOTF recommended cut-offs (see Table 1). In order to explore the crude association between mental wellbeing and weight, total SDQ core was regressed on age and gender adjusted BMI. A random intercept term was introduced to allow for the non-independence of children within the same families. As adjusted BMI was in the form of a Z score, a constant was added so that all values were positive, allowing the addition of quadratic terms to the model. Indeed, the addition of quadratic and cubic terms, though not higher polynomials, increased the fit of the modelled association between adjusted BMI and SDQ total score, reflecting a curvilinear relationship between weight and psychological wellbeing. This modelled relationship is depicted in Figure 1 for the SDQ total scores. However, the overall amount of variance in the SDQ total scores explained by BMI was small at 1.9% (R² for within family effects=.005, between effects=.024, overall R²=.019).

IOTF Classification Status	Mean SDQ Total (sd)	Mean Internalising Factor Score (sd)	Mean Externalising Factor Score (sd)	N (%) Emotional Disorder Screen Positive	N (%) Conduct Disorder Screen Positive
Obese (n = 305)	10.18 (6.1)*	.17 (.4)*	.21 (.6)*	53 (17%)*	48 (16%)**
Overweight (n = 760)	8.07 (5.4)	.06 (.4)	.02 (.6)	84 (11%)	84 (11%)
Normal weight (n = 2,704)	7.57 (5.4)	-.01(.4)	.00 (.6)	226 (8%)	275 (11%)
Underweight (n = 192)	7.59 (5.1)	.01 (.4)	.00 (.6)	16 (8%)	14 (7%)

Note: *Difference between "Obese" and other groups significant at p < .001 level; **Difference between "Obese" and other groups significant at p < .01 level

Increasing child age was significantly associated with increasing total SDQ internalising score and a significant trend to increased adjusted BMI. No gender difference in internalising factor scores were observed. Equivalised household income was associated with both increased BMI and SDQ internalising factor scores. In terms of ethnicity, those reporting Asian ethnicity had higher internalising symptom scores but lower BMIs and household incomes, on average, when compared to non-Asian participants. When treated as a continuous variable reported weekly physical activity levels were observed to have a quadratic relationship with internalising symptoms scores. When physical activity was dichotomised as below/above recommended levels for England low activity status was associated with higher internalising symptom scores compared to those who reported exceeding the recommended levels of physical activity. Thus, low physical activity levels, income, age and BMI/obesity status were entered into the multilevel multiple regression model predicting internalising symptoms factor score as potential confounding/mediating variables.

In terms of externalising symptoms: obesity was associated with higher scores and a similar curvilinear relationship with adjusted BMI was observed (not shown); no associations with ethnicity were observed. There was no association between low physical activity status and externalising factor scores. Girls had lower mean externalising scores than boys and slightly lower adjusted BMIs. Increasing income was associated with both lower externalising behaviour scores and adjusted BMI. Increasing age was correlated with higher BMI but lower externalising scores. Consequently, only income and gender were entered into the multivariate regression model exploring the association between reported externalising behaviours and obesity.

Multilevel modelling

Using adjusted BMI as a continuous measure, the cubic relationship with internalising symptoms factor scores was reduced but remained statistically significant (p=.02) once the effects of age, low physical activity levels, equivalised household income and non-independence of observations from children nested in the same household were adjusted for. Likewise the cubic association between adjusted BMI and externalising factor scores was slightly reduced in magnitude but remained statistically significant (p=.009) once the effects of gender and household income were controlled for (full results not shown).

A random slope model was used to investigate cross-level interaction; in this case whether household income modified the relationship between obesity and reported emotional or behavioural symptoms. There was no evidence of a moderating effect of household income on the relationship between obesity and either internalising or externalising symptom factor scores (β=.01, p = 0.4 and β=.00, p=.99 respectively).

Residual diagnostics were performed for the multilevel multivariate models used in the analysis via plots of residual values for both the fixed and random effects. These indicated that the residuals were normally distributed. In order to check for endogeneity a Hausman test was conducted, which did not indicate significant model misspecification via endogenous within household effects (p=.5).

Comparison with Previous Findings

This sample of children had, on average, higher BMIs than those used to derive normative values in 1990 [33] reflecting the overall trend for increased obesity rates over the last two decades. As the IOTF recommended cut-offs for overweight and obesity were employed the rates presently reported will be lower than those already described in the HSE 2007 report, which utilised normative data from the UK only [31]. Our observation of higher rates of obesity in girls compared to boys under 10 years is a trend that has been observed in health survey data since the mid 1990s [44].

Our finding of an independent association between obesity and internalising (emotional) difficulties is echoed by findings from a smaller, mainly non-White multiethnic sample of 11-14 year olds from East London. In the survey by Viner and colleagues, 17% of those of White British ethnicity (N = 267) who were classified as obese scored above screening threshold for self-reported SDQ total difficulties compared to 9% of ideal weight children of the same ethnic group [19]. Overall differences in SDQ total difficulties scores remained significant even after controlling for gender, age and socioeconomic status. A significant, independent association with depression and chronic obesity was observed in boys (but not girls) in an all-white sample of 9-16 year olds (N = 991) drawn from the US-based Great Smoky Mountains study. The authors reported that boys with depression were 1.7 times more likely to be chronically obese than non-depressed boys after controlling for SES and age [21].

However, the above findings stand in contrast to those reported by several previous studies; one Dutch survey of 614 children aged 13-14 reported a statistically significant relationship between obesity and only the peer problems/prosocial behaviour subscale scores of the self-report version of the SDQ, once age, gender and educational status had been adjusted for [45]. A separate survey of 4,320 London-based school students age 11-12 years utilised the self-report SDQ and reported only a small ( < 1 point on the SDQ) though statistically significant (p=.01) trend for the SDQ Emotional Symptoms subscale score to be raised in obese and overweight children compared to ideal weight peers [19]. The authors attempted to control for the effect of potential confounding variables by sub-group analysis according to ethnicity, socioeconomic group (based on Townsend scores) and gender. As in our study, the authors concluded that there was no evidence that socioeconomic status was a moderating variable, although a sub-group analysis may have lacked power to detect a difference, should it have existed. Ethnicity and gender were highlighted as potential moderating factors with the closest association between obesity status SDQ total scores being observed in the subgroup of girls of white ethnicity (mean score of 12.1 [obese] vs 13.4 [ideal weight]). The lack of association between overweight, as opposed to obesity, and poor mental health observed in our cohort of British children echo the findings from a community-based survey of 2,341 French children aged 6-11 years [46]. This latter study found no association with Conduct Problems or Emotional Symptoms SDQ scores and weight exceeding the 85^th centile once sociodemographic and lifestyle (including physical activity levels) were adjusted for. These findings, along with the curvilinear relationship between adjusted BMI and emotional symptoms reported by the present study, strongly suggest the presence of a threshold effect of childhood BMI on psychological wellbeing. Thus, we would hypothesise that the risk of significant emotional problems would rapidly increase in children with BMI z-scores exceeding approximately 2.0 (i.e. exceeding the 97^th centile). As with existing studies, BMI explained only around 2% of the variance in SDQ scores. Nevertheless, taking a categorical approach, obesity would appear to be associated with a clinically significant risk of poor psychological adjustment, at least in terms of emotional difficulties due to the potential threshold effects outlined above. In addition, it must be noted that the SDQ was developed as a screen for mental health problems in young people and the instrument may be less useful as a metric of wellbeing. However, the variation in published findings are unlikely to be wholly explained by the different measures employed. Rather, there may be genuine differences in the relationship between childhood obesity and wellbeing as a result of both cultural and cohort effects which require further exploration. The choice of potential mediating/confounding variables may also shape the final results.

This is not the first study to observe some relationship between BMI and externalising problems in children. Indeed, findings from both a British cohort reported higher rates of externalising problems in obese boys aged 3-5 years [22]. Moreover, a study of a North American cohort of children of both sexes reported that children with externalising behaviour problems at 2 years old had significantly higher BMIs when followed-up at age 12 years [47]. However, overall, the association of behavioural problems with obesity seems less consistent than that with emotional difficulties, as echoed by the present findings.

In the present study we did not observe a difference in internalising factor scores according to gender. Given the previously documented excess of depression and anxiety in adolescent females this was initially surprising. However, in the present study the average age of the study sample was only about 10 years and the gender difference in emotional problems may only become apparent in later teenage years. For example, depression is twice as common in adolescent girls compared to boys but this difference is only observed by the age of 15 years [48]. Moreover, higher rates of comorbidity between internalising and externalising difficulties have been reported in pre-pubescent boys [49] and this also may have contributed to a lack of an observed gender difference.

Study Strengths and Limitations

This was a relatively complete and representative national sample of children where the effects of a number of key sociodemographic variables were able to be controlled for. Moreover, the use of multilevel modelling appropriately adjusted the standard errors of the estimates for the non-independence of observations from children nested within households. However, although there were a very large number of clusters the average number of children nested within families was small at 1.4. Indeed, given this average cluster size and the intraclass correlations for observations nested within families the design effects were relatively small, and the curve in Figure 1 would not appear very different were these not controlled for by the introduction of a random intercept to the model. Nevertheless, given the clearly hierarchical nature of the data and the risk of dependency amongst residuals from observations within each cluster we felt the use of multilevel, rather than single level, modelling was justified. Moreover, this approach provided an opportunity to explore, albeit tentatively, within family effects and cross-level interaction. However, when considering the power of multilevel modelling studies both cluster number and size, as well as the parameters being estimated must be taken into account. When estimating parameters associated with fixed-effects (e.g. the effect of obesity status externalising factor scores) the number of clusters are of prime importance- where fewer than 50 clusters exist parameter estimates may be biased downwards [50]. Therefore it can be assumed that any fixed effects were estimated accurately. However, in this analysis we also introduced a random slope parameter in order to investigate the possibility of cross-level interaction. Again, cluster size is of secondary importance to the number of clusters with a recommendation of at least 100 groups with around 10 individuals in each group [51]. However, in our study average cluster size was considerably lower than this, although the number of clusters was very large. Therefore the parameters associated with potential cross-level interactions may be relatively poorly estimated and we may not have detected a significant effect where one existed. This is a potential limitation of the present study. Nevertheless, our findings were in keeping with that of Drukker et al.[45] who also reported that SES did not appear to be a moderating factor. However, neither the present or these latter findings can be taken as definitive evidence of this as both studies may be subject to low power.

Ideally, more detailed biometrics would have been utilised to derive obesity status. However, the IOTF recommended cut-offs correlate to a moderate to high degree with more sophisticated methods to estimate adiposity [52]. Whilst valid BMIs were obtained, self-reported physical activity levels may be less reliable than more objective based estimates, such as those based on accelerometry or heart rate, although moderate levels of correlation are generally reported [53]. No information on pubertal status was available in this sample. The relatively small numbers of non-white ethnic groups within this survey, whilst reflecting the general population from which the sample was drawn, makes it difficult to draw firm conclusions about ethnic differences. Probably the most significant limitation in this survey was that data on psychological wellbeing was restricted to the parentally reported SDQ, in the absence of the SDQ impact supplement. The use of SDQ internalising and externalising factor scores as the main outcome measure may have been more appropriate than using SDQ subscale scores consisting of only five items each. Moreover, parentally reported SDQ scores may be more sensitive to emotional disturbance than the self-report version of this instrument in 11-15 year olds [36]. Indeed, the use of totalled subscale scores in previous studies could partly explain the failure to report firm associations between obesity and emotional problems in young people. The SDQ is widely used and well validated, but the addition of self-report versions for those children over ten years would have resulted in increased sensitivity for the screening for potentially clinically significant disorders [35, 36]. The exclusion of the impact supplement from the survey pack may have reduced the reliability of the screening thresholds for conduct and emotional disorders as defined by the respective SDQ subscales. Despite this, the relative risks may have remained relatively unchanged as the decreased accuracy would apply to both obese and non-obese children. In addition, we did not have any detailed information of family environment available, although we felt it was important to include family level economic status as this is known to be a risk factor for both childhood obesity [54] and certain psychological problems [55].

Directions for Future Research

The conflicting findings from previously published research suggest that further datasets containing relevant measures of wellbeing and biometrics should be utilised in replicating the present analyses. However, in order to model hypothesised underlying mechanisms driving the association further longitudinal data are required. A number of ongoing studies of health and development are potential sources of such information, though it may be that new studies based in mixed qualitative/quantitative methodologies would be more effective in exploring this area and contextualising classes of observed trajectories. There are some indications that in adults poor mental health (and in particular, depression) may precede obesity [16]. There is little longitudinal research published regarding under 18s but the available evidence suggests this predominant direction of causality may also apply to children and adolescents. One US based longitudinal study involving 9,374 adolescents reported no association between obesity and depression at initial assessment. In contrast, at one year follow-up, depression significantly predicted onset of obesity (OR 2.05; 95% CI 1.04 to 4.06) independent of self-esteem ratings, conduct problems, socioeconomic status, gender and parental obesity [56]. A separate cohort study also suggested that childhood depression was a risk factor for obesity in adulthood, at least for women [44]. 'Temperamental Difficulties' were also noted to predict weight gain in a cohort of 138 North American children aged between 4 and 9 [57]. From these scant studies a tentative model could be proposed whereby temperament (largely hereditary in nature), interacting with early environment gives rise to a tendency to dysphoric mood and low self-esteem that increases the risk of over-eating. The reasons for the non-linearity of the relationship between BMI and psychological adjustment require further exploration. It may be that socio-cultural factors are the predominant influence, with children who obviously exceed the normative range of adiposity being at an exponentially increasing risk of adverse experiences, such as peer rejection.