A pragmatic evaluation of a family-based intervention for childhood overweight and obesity

Research question 1: Who takes part in MEND?

Children living in disadvantaged SEC are more likely to be overweight than their more advantaged peers. 20 In addition, the proportion of children who are overweight varies by ethnic group. 21 Research also suggests that some families (those in more advantaged SEC, for example) may be more likely to access or benefit from services which support behaviour change. 22

Given the socioeconomic and ethnic inequalities in childhood overweight and potential inequities in the use of services (including weight management services), we examined the question of who takes part in the MEND 7–13 programme. To do this, we used service data from MEND 7–13 and information from routine data sources about the characteristics of children who would be eligible to take part in the programme. MEND 7–13 participants were described by their age, sex, ethnicity, SEC and characteristics of their residential neighbourhood. The MEND 7–13 participants’ sociodemographic profiles were then compared with those of respondents to three nationally representative surveys [the Health Survey for England (HSE), the Millennium Cohort Study (MCS) and the National Child Measurement Programme (NCMP)]. This analysis is described in full in Chapter 2.

Research question 2: Who does MEND work best for?

Interventions might work less well for some groups than others, and so create or exacerbate inequalities through differential effectiveness. 23 Research suggests that some groups (again, those with more favourable SEC) may be more receptive to interventions aimed at promoting healthy patterns of diet and physical activity. 24 MEND 7–13 (like some other weight management interventions) was evaluated in a relatively small (n = 116) RCT15 which (in common with similar intervention studies) was not designed to investigate whether effects varied by sex, ethnicity or SEC. Therefore, questions remain about whether outcomes associated with participation in MEND 7–13, weight management programmes more generally, and indeed public health programmes in general, vary with these characteristics. 23 We examined how changes in outcomes associated with participation in MEND 7–13 varied by the sociodemographic characteristics of participants and the places where they lived.

Interventions need to be understood as part of a complex system, rather than one-off events which can be understood in isolation. 25 The substantial variations in the contexts where the MEND 7–13 service was delivered, combined with the large size of the data set, allowed us to estimate how changes in outcomes varied by characteristics of programmes and contexts. We estimated whether outcomes associated with MEND 7–13 were moderated by obesogenic characteristics of the environment (such as the density of fast food outlets or access to green space in the local neighbourhood). We also estimated whether differences in programmes accounted for differences in outcomes.

Randomised controlled trials are often conducted with highly selected populations in academic settings,10 meaning that conclusions about effectiveness cannot necessarily be generalised for all groups in the population. 26 In the case of MEND 7–13, the RCT was delivered only to children who were obese, although both overweight and obese children are eligible for the MEND 7–13 service following its implementation into practice. Additionally, as discussed above, once scaled up, the service was delivered across a wide range of contexts. There was therefore significant potential for outcomes under service conditions to vary from those observed in the RCT.

Although the RCT of MEND followed children up for 12 months, such long-term follow-up of participants in large-scale service interventions is unusual and would be costly. Our analyses could only assess changes associated with participation in the MEND 7–13 programme from the first to the penultimate sessions (the time of the last measurement), a period of 10 weeks. We compared unpublished RCT data over a comparable period of follow-up with these service data.

These analyses are described in full in Chapters 3 and 4.

Research question 3: What are the economic costs of MEND?

Evaluations of this type are developed in part to inform decision-making by those considering whether or not to commission weight management interventions like MEND 7–13. The economic costs of interventions are clearly an important component of such decisions. We aimed to analyse cost data for MEND 7–13 to estimate the overall cost of delivery and to characterise the types of costs which were involved in running the programme, over and above the costs of the programme itself.

Although our initial intention (as illustrated in Figure 1, research question 3) was to conduct a de novo analysis of MEND financial data, in the course of reviewing the literature we identified work which had recently estimated these costs for MEND 7–13 on a national and local level. These sources were reviewed and used for our analysis rather than repeating work already undertaken. This analysis is described in full in Chapter 4.

Research questions 4 and 5: What is the salience and acceptability of MEND to those who commission and use it, and what types of costs are borne by participants?

Proponents of the evidence and effectiveness agendas27 have long understood the importance of collaborating with citizen, policy, funding and practitioner users if research is to influence practice. 28 Literature on the difficulties of implementation, the tension between programme fidelity and responsiveness to local needs and the importance of context29–31 informed our approach to research questions 4 and 5.

Our qualitative research was designed to address the salience and acceptability of the intervention to users, non-users and providers. Hawe et al. 32 point to the context-level factors which may affect intervention uptake, success, and sustainability and which include primary care agencies and local adaptation of an intervention. In the course of our study, the NHS was undergoing fundamental reorganisation as a result of proposals subsequently set out in the Health and Social Care Act 2012. Commissioners and service delivery partners, never professionals with time on their hands, were experiencing considerable churn in the system, and we were concerned to capture insights from these provider groups as early as possible in the study. They were selected from a database provided by MEND Central. The interviews were designed to explore the ways in which commissioning and recommissioning decisions were made, what worked well or less well, and the tension between programme fidelity and responsiveness to local context (see Appendix 14 for topic guide). Our findings from these interviews are set out in Chapter 5.

Families were selected from the same MEND database as that used in the quantitative analysis, up to 4 years after they had participated in a MEND 7–13 programme. Interviews began with open questions about family make up and memories about MEND, followed by discussions of barriers and levers to participation, costs and benefits within the family, and life after MEND (see Appendix 14 for topic guide). Our findings from these interviews are set out in Chapter 6.

MEND 7–13 service-level data

Service-level data about the MEND 7–13 programme were collected by local MEND programme managers and collated by MEND Central between January 2007 and December 2010. All analyses of service data (n = 21,503) were also restricted to those with no duplicate observations in the data set and those with completely observed information for age, sex and postcode (Figure 2).

FIGURE 2.

Derivation of analysis sample, excluding participants with duplicate records, partially observed data on age, sex or postcode and those outside the age range 6–13 years.

Comparator data – representing the ‘MEND-eligible’ population

The ‘MEND-eligible’ population comprised those children who were living in England during the period when MEND was scaled up (2007–10), who were overweight and within the observed age range for participants on the programme. We estimated the size and sociodemographic composition of the MEND-eligible population using three ’comparator’ data sets: the HSE, the MCS and the NCMP. From these data sets we derived subsets representative of children in the MEND-eligible population.

Subsets of MEND data set used in this chapter

For the analyses in this chapter, the MEND service-level data were divided into three subsets of participants (MEND-HSE, MEND-MCS and MEND-NCMP), each drawn to be comparable with their respective comparator by age and time period. Thus, 18,289 children (all children with observed data for age, sex and postcode) were referred to and contacted MEND (discussed as ‘referrals’ throughout) and were comparable in terms of age (6–13 years) and time period (2007–10) with the HSE data. Of these, 13,998 had BMI measured at the first session and are discussed as ‘starters’, while 8311 starters completed > 75% of sessions and are referred to throughout as ‘completers’. This is the MEND-HSE subsample.

The MEND-MCS subset is made up of 4391 children referred to MEND Central when they were aged 6–8 years inclusive. This subsample was not restricted to the MCS data collection period (mainly in 2008) as this would have severely reduced the numbers, and hence the power for comparison. Of the 4391 referrals, 3360 were starters and 2010 were completers. Although MEND 7–13 was aimed at 7- to 13-year-olds, in practice 317 children (1.6%) who participated were aged 6 years, and we included these participants in our analyses to increase the statistical power to compare MEND participants with MCS respondents.

The MEND-NCMP subsample included 6004 children referred to MEND who would have been in Year 6 in 2007–8, 2008–9 or 2009–10, and who were therefore comparable with the pooled NCMP data. Of these, 4615 were starters and 2711 were completers.

Routine small area-level data

All the individual data sets (MEND, HSE, MCS and NCMP) contained the residential postcodes of respondents. These data were not routinely available for these data sets but were linked to the MEND 7–13 data set by the Institute of Child Health team, to the HSE by the National Centre for Social Research (NatCen), to the MCS by the Centre for Longitudinal Studies (based at the Institute of Education, University of London) and to the NCMP by the NOO. This was then used to describe which of 34,378 lower super output areas (LSOAs) across England the respondents and their families lived in. LSOAs are statistically designed to have a minimum of 1000 residents but average 1500 residents. Using the LSOA identifiers we linked in the IDACI 2007 deciles41 and an indicator of urban/rural status to the MEND and comparator data sets. For analyses reported in Chapters 3 and 4 we also linked in measures of the food and built environment (described below) to the MEND data sets. We decided not to use composite indicators of multiple deprivation such as the Index of Multiple Deprivation 200741 as we wished to assess which specific components were independently associated with our outcomes.

The IDACI 2007 measures the percentage of income-deprived households with children aged 0–15 years in the LSOA. Households are defined as income-deprived if they are in receipt of means-tested benefits such as Jobseeker‘s Allowance or child tax credits. 41 The majority of underlying data sources were collected in 2005, although some are based on 2001 census data. Therefore, this index slightly predates the MEND and comparator data sets. However, it is unlikely that the decile ranking of LSOAs would vary in the 2-year period between measurement of most of the IDACI data and the measurement of the individual data sets. Deciles of the index were calculated relative to the deprivation ranking of all English LSOAs. Thus, MEND participants and survey respondents living in a LSOA classified as decile 10 live in an area which is among the 10% most deprived areas in England (as opposed to within the data set) for children.

In addition to IDACI 2007, we used an index developed for the Department for Environment, Food and Rural Affairs. 42 The index divides England into grid squares, classifying those with more than 10,000 people as ‘urban’. Rural categories are assigned using geospatial statistics to categorise communities as ‘town’ and ‘fringe’ (described throughout this report as ‘suburban’), or as ‘villages’, ‘isolated hamlets’ or ‘dispersed communities’ (described here as ‘rural’). This index was available at the LSOA level and was linked into our data sets.

We also described the environmental characteristics of participants’ residential LSOAs, which might constrain physical activity or be associated with unhealthy or healthy diets, and so moderate the effect of MEND 7–13. We examined physical land use characteristics, specifically the percentage of green space and the percentage of roads in each LSOA. Land use data were linked from the Generalised Land Use Database constructed in 2005. 43 These land use components were inversely correlated (i.e. as the percentage of green space increases, the percentage of roads in an area decreases) and therefore would violate assumptions of collinearity in regression models. Therefore, we combined them into a single ‘built environment’ variable using factor analysis, with a low value indicating a LSOA with a high percentage of green space and low road density. The Local Index of Child Well-being44 includes a relevant measure of access to sports facilities as well as the percentage of green space as two of five underlying environmental measures. However, we decided not to use this composite indicator as we wished to assess which specific components were independently associated with our outcomes.

To examine the characteristics of the local food environment in the MEND data we linked in information from the national Ordnance Survey Points of Interest data set,45 which lists the postcode locations of all food retail outlets in England on an annual basis. Following previous work examining the relationship between food availability and overweight and obesity in 9- to 10-year-old children,46 supermarkets and fruit and vegetable stores were classified as ‘healthy’ and takeout/fast-food outlets and convenience stores as ‘unhealthy’. Population data downloaded at the LSOA level from the mid-year population estimates for England47 were used in conjunction with the Points of Interest data to derive two density measures of the number of healthy and unhealthy food outlets per 1000 population in the LSOA in 2007.

Percentage of MEND-eligible population referred to, starting and completing MEND 7–13

We estimate the percentages of the MEND-eligible population who were referred to, started and completed MEND 7–13 as the number of children referred to, starting and completing the programme divided by the size of the MEND-eligible population. We also calculated the percentage of the MEND-eligible population who started or completed the programme who were obese (as opposed to overweight or obese) when they started the programme.

The MEND-eligible population (the denominator) was estimated three times, using the HSE, MCS and NCMP. The HSE allowed us to estimate the prevalence of children aged 6–13 years who were overweight or obese between 2007 and 2010. We applied this prevalence to the Office for National Statistics (ONS) mid-year population estimates of 6- to 13-year-olds between the same years48 to estimate the MEND-eligible population aged 6–13 years.

Sweep four of the MCS was collected in 2007 and 2008, providing information about children who were born in England between August 2000 and September 2001. These children were aged 6–8 years when they were surveyed. We used the estimates of the MCS population size (537,000 children born between August 2000 and September 2001) as reported by Plewis et al. 49 As with the HSE, we applied the sweep four prevalences of overweight and obesity, which were adjusted for sampling design including attrition.

The NCMP measures the BMI of children in Year 6, predominantly aged 10–11 years. This population cannot be readily represented using ONS population estimates of 10- or 11-year-olds as the school year spans the period of interest for mid-year estimates. Instead, we used the observed population counts from the NCMP as our denominator.

Missingness and multiple imputation

Data were missing for sociodemographic and outcome variables to varying degrees (Table 3). This was attributable to systematic reasons for some variables; for example, parental employment status was only collected in 2009–10, and self-esteem was measured by the Harter personality profile instead of the Rosenberg self-esteem scale until mid-2008.

This number of missing data were considered to have the potential to introduce systematic bias to observed associations, and to reduce both the statistical power of models to detect associations and the precision of estimates. Therefore, we decided to impute missing values using multiple imputation. This is a general approach to producing valid inferences when analysing partially observed epidemiological data. 50 Multiple imputation assumes that data are missing at random (MAR), and that differences between the missing values and the observed values can be explained by differences in the observed data. Therefore, a multiple imputation approach will not, in theory, account for missing data arising from factors not included in the multiple imputation model [where data are missing not at random (MNAR)] or those where the participants with missing data do not differ from those with complete data [where data are missing completely at random (MCAR)]. However, it is important to note that whereas multiple imputation explicitly makes this assumption, other techniques, such as discarding cases with only partially observed data (complete case analysis), implicitly make similar assumptions (MCAR in the case of complete case analysis51). Furthermore, although it may be difficult to verify why data are missing (i.e. how far the MAR assumption is correct) in any given analysis, studies which simulate different missing data patterns have demonstrated that multiple imputation tends to be more robust to departures from these assumptions than available alternatives such as the complete case approach. 52

Multiple imputation for our data set was complicated by two characteristics of the missing data. Firstly, because data were multilevel, participants were not randomly distributed in the population but grouped through attendance at programmes. This meant that the reasons for the data to be missing might vary between participants but also between programmes (when, for example, staff did not input any responses for a particular programme). To take account of this we estimated multiple imputation models using the REALCOM-IMPUTE software (REALCOM, Centre for Multilevel Modelling, University of Bristol, Bristol, UK). 53 This software imputed missing data while taking account of the multilevel nature of the missingness. The second complicating feature was that we had 15 variables which were only partially observed (i.e. had missing data) and these were a mixture of categorical and continuous variables. The imputation of the missing values needed to be conducted in a single model to ensure that the relationships between the variables were estimated as precisely as possible. Again, the REALCOM software allowed these joint models to be fitted, while taking account of the multilevel structure described above.

In summary, we assumed that the reasons for missing data could be explained by the variables included in the multiple imputation model (i.e. all outcomes, sociodemographic, attendance, neighbourhood and programme-level factors, height and weight rounding, and random variation between programmes).

Differences between sociodemographic characteristics of MEND 7–13 participants and the MEND-eligible population

Percentages for the MEND subsamples and comparators were calculated, taking into account survey design where necessary (see Comparator data – representing the ‘MEND-eligible’ population). Differences in percentages were calculated for each MEND subsample (MEND-HSE, MEND-MCS and MEND-NCMP) by subtracting from them the corresponding comparator percentage (HSE, MCS or NCMP). The statistical significance of the differences at the 5% level was then tested using tests of the equality of percentages for two independent samples.

Calculation of differences in percentages for the imputed variables was a three-stage process. First, 10 data sets were imputed from the master data set by the multiple imputation model described above (stage 1). Following this, 10 tests of the equality of percentages were conducted, producing 10 sets of differences (and their variances) in percentages (stage 2). Finally, these parameters were combined into final estimates using Rubin’s rules,54 which take into account the variation introduced by the imputation procedure (stage 3).

These calculations were repeated for all the imputed socioeconomic variables (i.e. ethnicity, family structure, housing tenure and employment status). The process was repeated for the MEND-MCS and MEND-NCMP data sets, which were also imputed and compared with the percentages in the MCS and the NCMP.

Sensitivity analyses were conducted with the MEND-HSE data set to examine whether the sociodemographic composition of MEND participants varied by year. We repeated comparisons of the MEND-HSE data and the MEND-eligible population for starters and completers using complete case data, imputed data for overweight children excluding obese children, and imputed data for obese children.

Differences in neighbourhood income deprivation gradient between MEND 7–13 participants and comparators

We tested for differences in the social gradients of participants in the MEND programme by examining IDACI 2007 deciles graphically. The statistical differences in slopes were tested using regression models where the outcome was the percentage of participants by deprivation in each MEND subsample (MEND-HSE, MEND-MCS and MEND-NCMP) and comparators (HSE, MCS and NCMP) (for further details see Appendix 7). This analysis was repeated, where data existed, for those referred to MEND 7–13, those who attended < 75% of sessions (low to medium attendance) and those who attended > 75% of sessions (high attendance).

Determinants of completion of MEND 7–13

A multilevel poisson regression model was used to estimate the relative risk ratios of children completing the programme. A multilevel model was needed because participants were nested within programmes and so were not statistically independent from each other (a violation of the assumptions of single-level regression models). Relative risks ratios were calculated with 95% confidence intervals (CIs). We examined unadjusted and adjusted associations between completion and BMI at baseline (BMI at the first session), SDQ score at baseline, age, sex, ethnicity, family structure, housing tenure, employment status of the primary earner, IDACI 2007 decile, built environment, urban/rural status, group size and the number of previous programmes per programme manager.

We fitted the final multilevel model in MLwiN Version 2.25 (MLwiN, Centre for Multilevel Modelling, University of Bristol, Bristol, UK),55 fitting a model for each of the 10 imputed data sets in turn before combining the coefficients and standard errors using Rubin’s rules54 as implemented by the Stata 11.2 multiple imputation commands.

The model was estimated using multiply-imputed data for 13,998 participants. As with the calculation of differences in percentages described above, the relative risk ratios and standard errors derived from analysis of all ten data sets were calculated using Rubin’s rules as implemented by the Stata 11.2 missing data (mi est) commands.

Age was grand-mean centred, subtracting the mean across the whole sample from each observation. This means that relative risk ratios from the model could be interpreted as ‘for each year of age greater than the average the risk of completion increases/decreases by . . .’

We categorised all continuous covariates apart from age as this facilitated interpretation. We categorised the SDQ as normal (0–13 symptoms reported), borderline (14–17 symptoms reported) and abnormal (≥ 17 symptoms reported), following the wording and guidance of the score’s developers. 36

What percentage of the MEND-eligible population are referred to, start and complete MEND 7–13?

Estimates of the size of the MEND-eligible population were calculated using prevalence and population estimates data presented in Table 4. The prevalences of overweight and obesity were smallest in the youngest children (in the MCS) and highest in the NCMP data (for Year 6 children).

Table 5 shows the population estimates by the age range of the comparator surveys. Across the entire age range of the MEND 7–13 programme (6- to 13-year-olds), 4.4 million children would have been eligible for the programme and just under half (2.0 million) of those eligible would have been obese.

Of the 21,132 children who self-referred or were referred to MEND, 18,289 had complete age, sex and postcode data. These children constituted 0.42% of the MEND-eligible population (Table 6). A total of 13,998 started a MEND programme, constituting 0.32% of the MEND-eligible population; the 8311 who completed the programme constituted 0.19%. The percentages of those who were referred, started and completed were higher for 6- to 8-year-olds. The percentages who were referred, started and completed were estimated to be higher for 10- to 11-year-olds (Year 6 children), reflecting the high percentage of MEND 7–13 children in that age range (mean age of MEND children at baseline was 10.7 years). Percentages reduced from referral through starters and completers in the same way for the age ranges 6–13, 6–8 and 10–11 years.

The percentages of obese children referred to the programme could not be calculated because anthropometry was not measured until participants started the programme. The percentages of obese children who started and completed MEND 7–13 were higher than when overweight (including obese) children were considered (see Table 6). Starters aged 6–13 years who were obese accounted for 0.58% and completers accounted for 0.35% of the MEND-eligible population who were obese.

Differences between the sociodemographic characteristics of MEND 7–13 participants and those of the MEND-eligible population

We investigated how MEND referrals, starters and completers compared with the MEND-eligible population. Characteristics of all data sets compared in the analyses are available for reference (see Tables 28 and 29). Differences between the MEND referrals and children in the MEND-eligible population are reported only in the text (only sex- and area-level data were available for referrals) and, for starters and completers, sociodemographic differences and the p-values of these differences are reported in Table 7. A negative value for the difference in the proportions indicates that proportionally fewer children of that group were in the MEND participant subsample than might be expected given the proportion of children in that group in the comparator data set (and, by extension, in the MEND-eligible population). Figure 3 presents graphs showing the distribution of MEND subsets and comparator data sets by deciles of deprivation. The results are discussed below.

FIGURE 3.

Percentages of MEND participants and comparator respondents by decile of IDACI 2007, where decile 10 is the most deprived. The dashed black line is the socioeconomic gradient, estimated by linear regression of the MEND-eligible population (represented by HSE, MCS and NCMP). The solid grey line is the socioeconomic gradient of the MEND participants (MEND-HSE, MEND-MCS and MEND-NCMP respectively). Black and grey spots are comparator- and MEND-observed proportions respectively. (a) Social gradient by IDACI decile of MEND referrals vs. comparators; (b) social gradient by IDACI decile of MEND starters vs. comparators; and (c) social gradient by IDACI decile of MEND completers (> 75% of sessions attended) vs. comparators.

Sensitivity analyses

The analyses described above were conducted for all MEND 7–13 referrals, starters and completers who were overweight. However, the majority of MEND participants were obese. To test whether the distribution of MEND 7–13 was steeper by IDACI 2007 decile because of the differential adiposity of the MEND 7–13 and comparator samples, we repeated analyses for obese children only (see Appendix 6, Table 30) and overweight children excluding those who were obese (see Appendix 6, Table 31). Results were very similar.

Some variables in the service data were only partially observed, so that the complete case data set was significantly smaller than the number of participants in MEND. Basing the analysis on this would have reduced the power to detect genuine variations in the outcome, and could have introduced bias. We addressed this limitation using multilevel multiple imputation, which relies on the assumption that the data were MAR. By making comparisons based on imputed data we were able to maximise power and minimise bias. While the MAR assumption could not be directly tested, we ran sensitivity analyses of the final multivariable models with complete case data (as recommended for this type of analysis50). Results were similar between the complete case and the multiple imputation models (see Appendix 6, Table 32), implying that bias introduced by missingness due to MAR was minimal. However, the multiple imputation findings are presented because the variation in the sample is more accurately estimated by these models. Sensitivity analyses suggested that the sociodemographic characteristics of the MEND starters and completers did not vary systematically by year of participation (data not reported).

Determinants of completion of MEND 7–13

The relative risk of completion (recorded as attending > 75% of sessions) varied by sociodemographic, neighbourhood and programme-level characteristics (Table 8). The second column of Table 8 presents the results from models of the unadjusted relative risks of completion. Covariates which were significantly associated with the relative risk of completing from these univariable models were carried forward to a multivariable model (presented in the final column). The unadjusted results show that SDQ score at baseline, sex, family structure, housing tenure, employment status of the primary earner, IDACI 2007 decile, programme group size and the number of previous programmes per programme manager were associated with the relative risk of completing. Conversely, BMI at baseline, age, ethnicity and urban/rural status were not. These variables were not carried forward to the multivariable model.

The multivariable model showed that, on average, children with higher than expected symptoms of psychological distress (‘abnormal’) at baseline were less likely to complete than those with ‘normal’ symptoms (see adjusted relative risk ratios in Table 8). Boys were less likely to complete the programme than girls.

All four markers of SEC were independently associated with completion (see the final column of Table 8). Children living with a lone parent were less likely to complete compared with those living with two parents. Children living in social or private rented housing were less likely to complete than those living in owner-occupied housing. Children living in households where the primary earner was unemployed were less likely to complete than those in households where the primary earner was employed. Finally, children living in the most deprived neighbourhoods were less likely to complete the programme.

Families who started MEND 7–13 programmes in groups with > 10 people were less likely to complete than those who started in groups of one to five children. However, there was no statistical difference in completion between those in groups of six to nine children and those in groups of one to five children. Children attending programmes where managers had run ≥ 10 previous programmes were less likely to complete than those where managers had run < 10 programmes.

As noted previously (see MEND 7–13 service-level data), there is no agreed definition of a completer. 56 We tested relationships between completion and covariates using a lower threshold of attendance at 60% of sessions to define completion (see Appendix 8 for results). The direction and statistical significance of relationships were similar.

Percentage of the MEND-eligible population referred to, starting and completing MEND 7–13

Our analysis shows that only a small percentage of the MEND-eligible population were referred to, started and completed MEND 7–13 programmes. As MEND 7–13 has been reported as the largest intervention of its type in England in a recent mapping study of such schemes,57 this serves to illustrate the challenge for scaling up family-based lifestyle interventions to the population level.

Those who started and completed the programme were more likely to be obese compared with the MEND-eligible population, indicative perhaps that children and families principally self-refer or are referred when levels of adiposity are perceived by families or professionals as already particularly high. The percentage of MEND-eligible children referred to the scheme was slightly higher for 10- to 11-year-olds, which might reflect the role of the NCMP as a health promotion intervention which raises awareness of childhood overweight.

Differences between sociodemographic characteristics of MEND 7–13 participants and those of the MEND-eligible population

Our results showed that in comparison with the MEND-eligible population, proportionally more children who started or completed MEND lived in less favourable SEC (indicated by employment status of the primary earner, family structure and housing tenure). Relative to the MEND-eligible population, proportionally more of those who started or completed a MEND 7–13 programme were girls, Asian and urban-dwelling, and proportionally fewer were of white or other ethnicities. Finally, relative to the MEND-eligible population, proportionally more MEND 7–13 starters and completers were obese, rather than overweight but not obese.

We also found that there were socioeconomic gradients in referral to, starting or completing MEND 7–13 at the neighbourhood level. In the MEND-HSE and MEND-MCS comparisons, the proportions of MEND 7–13 participants were higher for those living in the most deprived areas relative to the MEND-eligible population. In the MEND-NCMP comparison, proportions of MEND 7–13 participants by deprivation were similar to the proportions in the MEND-eligible population. In none of the comparisons was provision of MEND 7–13 in the most deprived neighbourhoods less than expected. This is an interesting finding as it suggests that referral to MEND 7–13 was equitable, in the sense that children from more deprived communities were more likely to be referred or to self-refer to the scheme than those living in more advantaged circumstances.

Finally, MEND children were more likely to live in urban areas and less likely to live in suburban and rural areas than children in the MEND-eligible population. This applied to those who were referred to MEND and also to starters and completers.

Many of the differences between the MEND-eligible population and the MEND referrals, starters and completers were statistically significant but not large in magnitude. Therefore, we conclude that the provision and/or uptake of MEND 7–13 did not appear to compromise and, if anything, promoted participation of those from more disadvantaged circumstances and from ethnic minority groups. This suggests that participation in MEND 7–13, when it was implemented at scale across England, had the potential to make a contribution to tackling health inequalities.

The reasons for the differences between the MEND-eligible population and MEND referrals, starters and completers might be owing to factors related to the differential supply of the service (programmes may be more likely to be commissioned in deprived localities) and/or the differential uptake of the service (families in more disadvantaged SEC may be more likely to take up the programme). We were not able to examine these factors.

Determinants of completion of MEND 7–13

Completers differed from starters. Completers were less likely to be psychologically distressed at baseline than starters. This is consistent with the findings of previous work. 58 Specifically, girls were more likely to complete than boys. This finding was inconsistent with the results of previous research,59 which found no sex difference in dropout rates (dropout rather than completion was measured in previous work). Completion of MEND 7–13 was not associated with ethnicity in unadjusted models. This was in contrast to previous research suggesting that children from BME groups were more likely to drop out of paediatric weight management programmes. 58,59 The difference in findings might reflect the fact that the previous studies were based in different countries (the Netherlands59 and the USA58) with different ethnic, cultural and other contexts.

Family structure, housing tenure, parental employment and neighbourhood deprivation were all independently associated with completion of the programme. All associations were in the same direction, with more disadvantaged families less likely to complete. The same associations were not observed in the study by de Niet et al. ,59 where levels of parental education, maternal employment status and family structure were not associated with dropout between 0 and 3 months. These differences might be explained because the intervention had a lower level of parental involvement (parents only attended three sessions) than that of MEND 7–13.

Programme group size and the number of programmes delivered by the programme manager were both associated with rates of completion. We have not identified other studies which tested whether these types of variables were associated with attrition or completion. However, further analysis of these types of process factors at the programme level, and their relationship with completion of child weight management programmes delivered under service conditions, would be valuable.

The model therefore gives some insight into the predictors of completion of weight management programmes conducted in the English context. It is notable that although there were associations between sex and SEC and the likelihood of completion, these relative risks of completion did not alter the direction of MEND versus MEND-eligible differences for starters and completers, although differences were reduced in magnitude and sometimes attenuated to non-significance (see Who completed MEND?).

Strengths and weaknesses of the study

Data availability prevented us from making identical comparisons of the MEND subsample with each routine data set. For example, family structure was not established in exactly the same way in any of the three data sets (MEND 7–13, HSE and MCS). However, results were consistent across two comparisons (MEND 7–13 with MCS and MEND 7–13 with HSE) which gave us confidence that results were not overly sensitive to differences in variables or measurement. Moreover, we drew on three family socioeconomic variables from two studies (the MCS and HSE). The results from all three variables in both data sets were consistent. We also linked neighbourhood deprivation data into all three samples. Although this is measured at the area level, and so is less appropriate for making inferences at the individual level, it was also consistent with the results from the family SEC data in the sense that participation in MEND 7–13 was proportionate for each decile of neighbourhood deprivation. Taken together, the proportion of MEND 7–13 participants who were living in socioeconomically deprived circumstances was similar or slightly higher than expected given the proportion of children in these circumstances in the MEND-eligible population.

The analysis of attendance may be biased owing to the ways in which the service managers reported attendance. Specifically, we could not differentiate between a missing response (where the data were not entered) and a child who was reported by the programme manager as not attending. For this reason, the number of completers reported in the MEND subsets may have been less than the actual number. In addition, this may have introduced systematic bias into the results if those who were missing as opposed to non-attendees differed in terms of the socioeconomic variables of interest. We attempted to mitigate this last possibility by multiply imputing the attendance and SEC variables, to minimise bias due to differences between participants with complete data and those with partially observed data.

The MEND 7–13 programme was predominantly delivered to obese rather than overweight children. Within a population which is generally overweight (including obese), socioeconomic gradients may be steeper for those who are obese. However, sensitivity analyses suggested that the differences in adiposity between the MEND participants and those in the MEND-eligible population did not explain the findings.

Wider context

Most current work on scaled-up interventions refers to those delivered in low- to middle-income countries, and studies which have evaluated scale-up of public health interventions in higher-income countries are relatively rare. 60 It is unusual to have comprehensive, population-level information about public health interventions and the degree to which they reach populations at particular risk. 61 This study therefore adds to this literature by estimating how far a weight management intervention, when scaled up, was available to overweight and obese children in the population.

To summarise, the provision and/or uptake of MEND 7–13 did not appear to compromise and, if anything, promoted participation of those from more disadvantaged circumstances and from ethnic minority groups. This suggests that participation in MEND 7–13, when it was implemented at scale across England, had the potential to make a contribution to tackling health inequalities. However, this potential was mitigated to some extent because, having started a programme, completion was relatively less likely for those participants living in less favourable SEC and for boys.

Outcomes

The primary outcome was change in BMI after participation in MEND. BMI was calculated at baseline and follow-up [using (weight/height)²]. BMI measures weight after taking account of an individual’s height, and where it exceeds the 91st centile of the UK 1990 growth charts14 it can be used as a measure of overweight, a proxy for having excess fat. A reduction in BMI can therefore be interpreted as a reduction in excess fat. Where this is an average value across a group such as the MEND 7–13 participants, it will likely reflect that the group predominantly reduced levels of excess fat (one of the aims of the intervention), but that a minority either maintained or even increased their levels. Change in BMI was calculated by subtracting baseline BMI from follow-up BMI, so that 0 on the measure indicates no change, a positive value indicates an increase in BMI and a negative value indicates a reduction in BMI. We took an average reduction in BMI to be evidence of the programme assisting families in child weight management at a population level.

To compare the changes with those commonly reported in reviews10 of childhood obesity interventions, we also measured change in BMI after standardising baseline and follow-up BMI for age and sex. This was done using the UK 1990 BMI growth reference,14 created using Cole’s lambda-mu-sigma (LMS) method,62 which standardises BMI for age and sex. Standardised BMI is termed zBMI. As with BMI, change in zBMI was calculated by subtracting baseline zBMI from follow-up zBMI. Also as with BMI, a fall in zBMI implies a reduction in excess fat, whereas a rise indicates increased excess fat.

In our analysis, the two outcome measures, change in BMI and change in zBMI, lead to broadly similar conclusions. Change in BMI is more comprehensible than change in zBMI, and this is why it is the main focus of the report.

Covariates

Participant-level variables included BMI at baseline, sex, age and ethnicity; characteristics of families included housing tenure, parental unemployment, lone parent or couple status; and characteristics of LSOAs included IDACI 2007 deciles, a measure of built environment, food outlet density and urban/rural status. Programme-level variables included group size, number of previous programmes per programme manager and rounding of height and weight measurements. These variables were described in Chapter 2 (see MEND 7–13 service-level data).

In this section we also examine whether participant-reported global self-esteem is associated with change in BMI. We measure global self-esteem using an adapted form of the widely validated Rosenberg self-esteem scale. 35 The scale measures global self-esteem by asking MEND participants to state to what extent they agree with 10 statements about themselves (listed in Box 1). The Rosenberg scale was originally designed for use with adolescents35 and few studies have used it with children within the age range targeted by MEND 7–13. However, the response format was judged by staff at MEND Central to be too confusing for younger children, and so was modified to more fully reflect the cognitive abilities of younger children. In addition, some of the original constructs measured in the scale used American terminology, and these statements were modified by MEND Central to be more accessible to children in the UK (for example, including ‘happy’ in brackets as an alternative to ‘satisfied’ in statement 1 – all 10 MEND-modified statements are listed in Box 1 for reference). Responses to each statement were on a four-point symmetric agree–disagree scale where the coding is given in brackets: ‘a lot like me (0)’, ‘a bit like me (1)’, ‘not like me (2)’ or ‘not at all like me (3)’. Where statements were negative (i.e. statements 2, 5, 6, 8 and 9 in Box 1), these were reverse coded to be positive. The score was then calculated by summing the statements together so that a high value was indicative of high self-esteem.

Missingness

Chapter 2 (see Missingness and multiple imputation) described how we used multiple imputation to address issues of missingness in variables required for the analysis. We used the same approach in the analysis reported in this chapter. We imputed values on covariates where BMI (and therefore zBMI) at baseline and follow-up, age, sex, neighbourhood and programme variables were completely observed (n = 9563).

In addition to exploring sources of variation in change in BMI and variation in change in self-esteem, we wanted to examine how baseline self-esteem related to change in BMI. For this specific model, we estimated models where baseline self-esteem and change in BMI were both completely observed (n = 4962).

Randomised controlled trial data and the randomised controlled trial-comparable subset of the MEND data

We pooled the service data and data from the intervention arm in the MEND RCT. 15 The RCT was not part of the current study. The RCT data (previously unpublished in trial reports) relate to changes in BMI over the course of the programme. This period (10-week follow-up) is the same as that for the service data. The RCT was restricted to 8- to 12-year-old participants who were obese (exceeded the 98th centile of the UK 1990 growth charts14), whereas MEND 7–13 services in the community are available to those who are overweight and aged 7–13 years. To ensure a valid comparison, the pooled sample was restricted to participants in the RCT intervention arm (n = 47) and those in the service data who were obese and of the same age as RCT participants (n = 8054), for whom data were recorded at baseline and follow-up.

Regression to the mean

Data obtained before and after interventions are subject to measurement error, biological variation, intrasubject variability and other influences unrelated to the intervention. As a result, the correlations between serial measurements taken from individuals over time are less than perfect, and this leads to regression to the mean. 63 Regression to the mean is a statistical phenomenon such that, on average, the change in BMI over time is greater in subjects whose BMI at baseline is more extreme; thus, there is an inverse association between BMI at baseline and BMI change even in the absence of any intervention effect. 64 Regression to the mean confounds the interpretation of BMI change associated with an intervention. To address this, estimation of intervention effects needs to include adjustment of the outcome (change in BMI and zBMI in this chapter and change in secondary outcomes presented in Chapter 4) for the baseline value. We undertook this adjustment in all analysis models. The adjustment cannot adjust perfectly for confounding by regression to the mean as it accounts for two separate effects simultaneously: conventional regression to the mean, and unmeasured confounding by variables which are associated with BMI at baseline and change in BMI but are not measured by other covariates in the model.

Multilevel analysis of change in body mass index

In this chapter we used multilevel linear regression models to examine the outcomes of change in BMI and change in zBMI, as these outcomes were continuous. Multilevel regression models make the assumption that residuals at the participant and programme levels are normally distributed. We checked this assumption by graphing the residuals from the unadjusted models and found that residuals at both levels met these assumptions. We calculated the Bayesian information criterion (BIC) for models to assess whether the addition of parameters such as random intercepts and slopes, polynomials and interactions improved the fit of the model parsimoniously. The BIC is a measure of a model’s overall fit, penalised for its parametric complexity. The BIC decreases if parameters added to the model improve the fit over and above the increase in parametric complexity. Following recommendations for the use of Bayesian criteria for model selection, we considered a decrease in BIC of 1–2 as only weak evidence of improved fit while decreases ≥ 3 were considered positive to strong. 65

Our preliminary analysis involved a five-stage series of multilevel models. Complete case data (n = 2150) were used to evaluate whether complex parameters (random intercepts and slopes, polynomials and interaction terms) were required in the multivariate model. Complete case data were used as the BIC cannot be computed for multiply-imputed data.

We hypothesised that relationships between variables might vary with age, sex, ethnic group and whether the children lived with lone parents or parents in a couple. We tested these a priori interactions for change in BMI and change in zBMI (see the description of the stages of the analysis, below, for further details).

A series of models were fitted to estimate associations with change in BMI. To ensure that associations were not ruled out because of a lack of power, we used data that were completely observed for each pairwise comparison (i.e. for the age models, data were completely observed for change in BMI, baseline BMI and age).

The first five stages, using complete case data sets, are summarised below.

1. Compare single-level and multilevel models to test whether a multilevel model is necessary (complete case data to compute change in BIC).

2. Assess whether the relationship between BMI at baseline and change in BMI was linear or polynomial to determine how to adjust for baseline BMI in subsequent models (complete case data to compute change in BIC).

3. Estimate ‘baseline-adjusted’ associations between covariates and change in BMI. These multilevel models were adjusted for BMI at baseline. Models were based on fully observed data for change in BMI, BMI at baseline and each covariate in turn to maximise power.

4. Assess whether a random slope for each coefficient improves the fit compared with stage 3. The random slopes allow the coefficients to vary between programmes (complete case data to compute change in BIC).

5. Assess whether prespecified interaction terms improve the fit of models over and above those including just the main effects specified in the interaction (i.e. the BIC of a model including an age-by-sex interaction was compared with the BIC of a model including age and sex but no interaction).

The sixth, seventh and eighth stages of the analysis used multiply-imputed data. As with the multilevel modelling described in Chapter 2 (see Methods, Determinants of completion of MEND 7–13), we fitted models in MLwiN 2.25 and combined analyses across all 10 data sets using Rubin’s rules54 as implemented by the Stata 11.2 multiple imputation commands. These stages are described below.

1. 6. Fit multivariable model including all significant covariates, and interactions or random slopes which improved fit with imputed data.

2. 7. Repeat multivariable model of change in BMI, replacing outcome with change in zBMI and replacing BMI at baseline with zBMI at baseline.

3. 8. We also explored whether self-esteem at baseline was associated with change in BMI, using the Rosenberg self-esteem scale data collected in 2009/10.

We report the intercept of the model, which describes change in BMI for a given ‘reference group’. We characterised this reference group as girls, and for other categorical variables (e.g. ethnicity or family structure) we chose the largest category as the reference category (e.g. white, parents living as a couple, etc.). For continuous variables (e.g. age), the reference category was the average value (e.g. 10.7 years).

Comparisons with randomised controlled trial results

The RCT and service data were also compared using a multilevel regression model. Again, the outcome was change in BMI. For this analysis, the principal covariate of interest was a binary variable indicating whether participants were included in the RCT or the service data. Two multivariable models were estimated, the first estimating differences in change in BMI between the RCT and service data after adjusting for any differences in baseline BMI, age and sex. The second also included ethnicity and housing tenure in addition to the first model, in order to adjust for differences in samples by ethnic group and SEC.

Description of MEND participants

On average, BMI declined by 0.7 kg/m² among MEND participants, which equated to a reduction of 0.2 units of zBMI. The sample was approximately normally distributed in terms of BMI change (Figure 4). The histogram shows that although BMI was maintained or reduced for the majority of participants (n = 7661, 80.1%), it also increased for some participants (n = 1902, 19.9%). A larger majority maintained or decreased their zBMI (n = 8441, 88.3%) with a corresponding smaller minority increasing their zBMI (n = 1122, 11.7%).

FIGURE 4.

Histogram of change in BMI (n = 9563).

Of 8368 children who were obese at baseline, 8.8% were overweight at follow-up and 0.07% (n = 6) were normal weight. Of 1195 children who were overweight at baseline, 0.4% (n = 38) were obese at follow-up, while 2.8% (n = 264) were normal weight. Demographically, MEND participants were normally distributed by age, centred on a mean of approximately 10 years (Table 9). A disproportionate number of MEND 7–13 participants were girls and from white ethnic groups. Socioeconomically, most MEND participants lived with a parent/carer who considered themselves to be in a couple, reported that they owned their residence and were employed. Higher proportions lived in neighbourhoods which were income deprived relative to England as a whole, and which were urban and built up.

Using the definitions of ‘completion’ chosen for this study (non-completers attended < 25%, partial completers attended 25–75% and completers attended > 75% of sessions), only a very small minority of the sample were non-completers. This is not surprising because to be part of the change in BMI sample (or any other change outcome sample), the children needed fully observed BMI data at baseline and follow-up. A quarter of the sample partially completed the programme whereas the majority completed the programme.

In terms of data quality, most programmes rounded height measurements to 0.5 or 1 cm. A smaller majority rounded weight measurements.

Examination of the overlap in CIs suggests that there are few differences between the imputed and complete case samples in regard to age, sex, ethnicity, housing tenure, residential neighbourhood deprivation, built environment or urban/rural status and programme group size. The imputed proportions were lower for completers and conversely higher for partial completers.

Modelling to inform multilevel models of change in body mass index

The first stage of the analysis compared an unadjusted single-level model of change in BMI with an unadjusted multilevel model of change in BMI. Estimating the between-programme intercept improved the fit of the model (single-level model BIC = 7434, multilevel model BIC = 7354, change in BIC = –79.9). Between-programme variation in change in BMI explained 16% of the overall variation.

The second stage of the analysis tested whether change in BMI was associated with BMI at baseline on a linear or quadratic scale. The multilevel model showed a significant quadratic effect of BMI at baseline. However, the BICs were similar (linear model BIC = 7304, quadratic model BIC = 7303). 65 Thus, subsequent models in the analysis adjusted for BMI at baseline as a linear term only. On average, children with higher BMI at baseline declined slightly more in BMI than those with lower BMI at baseline. Figure 5 illustrates this relationship.

FIGURE 5.

Relationship between BMI at baseline and change in BMI (n = 9563).

The third stage of the analysis used baseline-adjusted models to explore which variables to use in the final multivariate model. After adjustment for BMI at baseline, most variables were significantly associated with change in BMI (Table 10). Food environment variables for both healthy and unhealthy environments were non-significant and were omitted from later models, as were variables measuring rounding of height and weight.

Results were broadly similar for change in zBMI and change in BMI. However, sex and number of previous programmes per programme manager were not associated with the change in zBMI and so were not carried forward to the multivariable model.

A fourth set of models investigated whether the relationships between each covariate and the change in BMI varied randomly between programmes (Table 11). Adding a random slope increased BIC in every case so no random slopes were included. Similarly, the a priori-specified two-way interaction terms all increased BIC, so they too were omitted from the multivariable model.

Multivariable models of change in BMI and zBMI

The multivariable models of change in BMI, based on the imputed and complete case data sets, appear in Table 12. Model 1 was based on 10 imputed data sets (n = 9563) where the missing covariates (including BMI at baseline) and missing change in BMI were imputed. The intercept shows that the average change in BMI associated with the intervention was −0.76 kg/m² for those children belonging to the reference group. The characteristics of the reference group in interpreting this finding are therefore important. The group comprised white girls, of average BMI for MEND participants at baseline, living in favourable SEC (having parents who are in a couple, owner occupiers and employed) and who had completed at least 75% of the MEND programme. This group lived in urban neighbourhoods with average levels of deprivation and built environment and attended MEND programmes with around nine attendees at baseline where the programme manager had run around six previous programmes.

Body mass index fell more in children with higher baseline BMI. Demographically, BMI fell more in younger rather than older children, boys rather than girls and white rather than Asian or black ethnic groups. Socioeconomically, BMI fell more in children whose parents were employed, those living in areas with low proportions of income-deprived families (i.e. low IDACI 2007 scores) and those who attended groups that were larger at baseline. BMI fell less in partial completers and non-completers compared with completers.

Model 2 used complete data (n = 2150) (see the final column of Table 12). The results for models 1 and 2 were similar in direction, though coefficients in the latter model were non-significant for age, sex, black groups, IDACI 2007 decile, group size and non-completers, reflecting the smaller sample size. One coefficient, the number of previous programmes per programme manager, was only significant in model 2.

Results for change in zBMI (Table 13) were broadly similar to the results for change in BMI, except that the baseline BMI/zBMI coefficient was negative for BMI and positive for zBMI, and both were highly significant.

Magnitude of changes in BMI and zBMI

In order to interpret how the results from the multivariable models related to changes in BMI and zBMI we calculated the amount of change which would be expected for key subgroups after adjustment for covariates. We selected subgroups on the basis that they were significantly different from the intercept in the model, and using extreme but meaningful categories (for example, 7 and 13 years were chosen as the extremes of the intended age range for MEND 7–13).

The findings (Table 14) show that all subgroups reduced in BMI and zBMI on average. The top 10% of the sample in terms of BMI showed the greatest BMI reduction overall (0.90 kg/m²). Non-completers reduced the least but still reduced their BMI by 0.55 kg/m² and their zBMI by 0.15 kg/m². Other groups reduced in BMI by 0.61 kg/m² (Asian children), 0.70 kg/m² (children with unemployed parents) and 0.73 kg/m² (children living in the most deprived neighbourhoods). The range of the reduction in BMI for group sizes ranging from two to 18 participants was −0.68 kg/m² to −0.82 kg/m².

Self-esteem at baseline and change in body mass index

Table 15 explores the relationship between change in BMI and self-esteem, using the imputed self-esteem data set (n = 4962) (models 1 and 2). Model 1 omits baseline self-esteem while model 2 includes it. Model 2 shows that self-esteem at baseline is weakly associated with change in BMI; for each extra unit of self-esteem (measured on a 0–30 scale), BMI falls by 0.0048 kg/m² less. This means that, after adjustment for covariates, participants whose level of self-esteem at baseline was average (16.78) reduced in BMI by −0.79 kg/m². For those whose self-esteem was one standard deviation (6.89) above the mean at baseline, BMI reduced less (by −0.76 kg/m²), and for those whose self-esteem was one standard deviation below, BMI reduced more (by −0.83 kg/m²). Model 3 in Table 15 shows that in the complete case model baseline self-esteem was not significantly related to change in BMI.

Results were broadly similar to those shown in Table 12 (change in BMI in a larger sample size of 9563, with no baseline self-esteem included). However, whereas associations between change in BMI and IDACI 2007 decile, non-completers versus completers and group size were non-significant in Table 15, they were significant in Table 12.

Sensitivity analyses of modelling of change in body mass index

We considered it essential to adjust for BMI at baseline before selecting the other covariates. However, we anticipate that, for comparison with previous work, some readers may be interested in the univariable relationships between change in BMI and the covariates without adjustment for BMI at baseline. Associations from multilevel models with the imputed data but without adjustment for BMI at baseline are presented in Appendix 8 (see Table 34).

Comparison of change in body mass index observed in service and randomised controlled trial settings

The mean change in BMI in service-level data for 8- to 12-year-old obese children was −0.71 kg/m² after adjustment for BMI at baseline, age and sex (Table 16). The coefficient for the RCT shows that the BMI of those in the RCT reduced by 0.17 kg/m² more, but this difference was not statistically significant. After adjustment for ethnicity and housing tenure the change in BMI in service-level data was −0.79 kg/m², and for those in the RCT it reduced by 0.25 kg/m² more. Again, this difference was not statistically significant.

Changes in body mass index associated with MEND 7–13 delivered under service conditions

Our principal objective in this chapter was to investigate how outcomes associated with MEND 7–13 varied according to participant, family, neighbourhood and programme factors. However, before the results relating to this objective are discussed, we discuss whether or not MEND 7–13, when delivered under service conditions at the population level, was associated with any reduction in BMI.

Our findings show that MEND 7–13, as delivered across the country in the service setting, was associated with a mean reduction in BMI of 0.76 kg/m². Previous work suggests that there are unlikely to be large increases in height over a 10-week follow-up period,66 which means that a reduction in BMI associated with MEND suggests that the programme may be associated with weight maintenance or weight loss. An outcome of weight maintenance/loss is in line with current national guidance on paediatric weight management interventions from the National Institute for Health and Care Excellence (NICE). 9

Beyond knowing that average BMI is reduced in the programme, it may also be instructive to consider how much of a reduction in BMI might be important from a clinical perspective. However, there is little consensus as to what constitutes a clinically significant effect. Meta-analyses reported by Cochrane reviewers10 of four lifestyle treatments for childhood overweight and obesity suggest that treatment interventions for children aged < 12 years lead to changes in zBMI of −0.06 at 6-month follow-up. This reduction in zBMI was described by the authors as both ‘statistically significant and clinically relevant’ (p. 16). Therefore, the magnitude of the reduction of 0.18 in zBMI observed in the MEND 7–13 service might also be considered ‘clinically relevant’ but it is important to note that the findings reported here relate to a shorter follow-up (10 weeks) than the Cochrane review findings (6-month follow-up). As a targeted intervention, the magnitude of the reductions in BMI and zBMI noted here might be important for the health of overweight and obese children at the population level. However, as with the debate on whether the findings are clinically relevant, there are few sources available which suggest by how much BMI might need to reduce following a scaled-up intervention for it to be important at the population level. Changes observed in the service data were not significantly different from the change in BMI observed in the intervention arm of the MEND RCT (Sacher et al. 15). This suggests that BMI fell by a similar amount under research and service conditions during the programme (and that changes were of a similar order of magnitude), although these results should be treated cautiously owing to the small sample size of the RCT intervention arm. As we discuss above, it would be interesting to consider whether this difference was important from a clinical or population health perspective. However, in the absence of a clear consensus view on the size of a clinically relevant difference between groups, or what constitutes an important reduction at the population level, we cannot comment further.

In the MEND RCT, the reduction in BMI seen by the end of the intervention was maintained at 12 months. 15 We did not have service data at a comparable follow-up to assess whether changes were maintained after MEND 7–13 was delivered in service settings (discussed in more depth below; see Strengths and limitations of the study).

We note that our findings do not reflect an assessment of the effectiveness of MEND 7–13, or a comparative analysis of MEND 7–13 versus alternative weight management or health promotion schemes.

Variations in outcomes associated with MEND 7–13

The extent of the reduction in BMI varies with BMI at baseline, demographic group, family SEC and neighbourhood deprivation, group size and levels of attendance at the programme. Similar variations were also observed for models using change in zBMI as an outcome and where complete data were used. Our findings suggest that reductions in excess fat associated with this weight management intervention were smaller for those living in less favourable SEC and in black and Asian groups. However, statistically significant variation between groups does not necessarily equate to variation which is clinically relevant. As discussed above, a reduction in zBMI of 0.06 can be considered a ‘clinically relevant’ threshold. Even the groups which experienced statistically significantly smaller reductions in average BMI (for example, older children, girls or Asian children) still reduced their zBMI by more than 0.06 on average (see Table 14).

Higher baseline zBMI was associated with a smaller reduction in zBMI, whereas higher baseline BMI was associated with a larger reduction in BMI. We investigated the underlying reasons for this, finding that these different patterns might be expected given the ways that age and sex standardisation impact differentially on the variability of baseline BMI/zBMI and on the relationships between age, sex and change in BMI/zBMI. These patterns are discussed in more detail in Appendix 10.

We have found no studies which have systematically tested how the impact of weight management interventions varies according to demographic factors. Review articles have suggested that this is a key area for future research. A recent review of reviews23 examining the differential effectiveness of interventions by SEC found no studies which examined this question for change in BMI or change in zBMI (although diet and physical activity interventions were found; see Chapter 4, Change in lifestyle behaviours associated with MEND 7–13).

Our study investigated whether or not self-esteem at baseline was associated with change in BMI after adjustment for BMI at baseline and other covariates. We found that higher levels of self-esteem were associated with slightly smaller reductions in BMI; conversely, children with lower self-esteem reduced in BMI more. This may be because children who have lower self-esteem before the programme have more potential to benefit from MEND than children whose self-esteem is already high. Although several intervention studies have examined self-esteem as a secondary outcome in weight management intervention studies (see Griffiths et al. 67 for a recent review), we have found none which examined self-esteem as a moderator of weight management interventions.

Baseline-adjusted models showed that change in BMI associated with MEND was not associated with variations in local food environments. Thus, participants living in areas with a high density of fast food and other unhealthy food outlets did not reduce in BMI less than those living in areas with a lower density. Conversely, participants living in areas with high proportions of healthy outlets did not reduce in BMI more than those living in areas with less healthy outlets. This suggests that any changes associated with the programme were, in the short term, apparently resistant to differential exposure to food environments that either promote or constrain a healthy diet. This is perhaps not surprising in the UK context; despite the popular characterisation of ‘food deserts’ (areas where healthy food is less readily available), the evidence that variations in the availability of healthy food in the local environment are associated with levels of overweight is relatively weak. 68,69

Change in BMI was not independently associated with characteristics of the built environment or with urban/rural status. Although participants living in more built-up areas did reduce less in BMI, and those in rural and suburban areas reduced more, when these variables were included in a model which also adjusted for neighbourhood deprivation (IDACI 2007 decile) and family socioeconomic status, these relationships were attenuated to non-significance. This suggests that the unadjusted differences may have been due to differences between urban and rural areas in terms of families’ SEC.

Change in BMI was associated with levels of attendance at the programme; completers showed greater reductions in BMI than partial completers and non-completers. This supports the assertion that the variations in the outcome are attributable to participation in the intervention. This is important because our study did not have access to measurements for controls who did not receive the intervention. This means that we cannot make causal assertions about whether or not changes in BMI observed among MEND participants were attributable to the intervention itself (as opposed to another temporal change).

Strengths and limitations of the study

This analysis was based on a service-level data set collected under service (or ‘real world’) conditions across all regions of England. The data set was large enough to have substantial statistical power to identify associations between change in BMI and factors at the individual, family, neighbourhood and programme levels when modelled in a multivariable regression. The availability of such data to examine how interventions operate is a major strength as it allows us to quantify changes associated with such programmes delivered under a wide range of circumstances across the country.

We have noted challenges associated with using the MEND service data in terms of missingness and our methods of mitigating these issues using multiple imputation (see Chapter 2, Strengths and weaknesses of the study). The same discussion also applies to these analyses of change in BMI and change in zBMI.

The statistical power to relate sociodemographic and programme factors to change in BMI was a strength of the study. However, as with most secondary data sets, there were other variables which we would have liked to examine which were not measured at the family level, such as parental BMI and parental lifestyle factors which may have moderated the programme’s relationship with participants’ BMI. In addition, there was substantial unexplained random variation between programmes accounting for approximately 16% of the variation in change in BMI. This may well be explained by features which were unmeasured in the service data, such as variations in facilitator qualifications and experience, or features of the venues where sessions were delivered. Again, it would have been useful to explore how these factors related to changes in outcome, and service organisations might consider how to routinely collect such information.

MEND Central recommends anthropometry equipment and trains delivery partners in measurement. However, programmes did not record what equipment was used and so we could not check how far programmes adopted this guidance. In addition, when we analysed the height data, many measurements appeared to have been rounded. As this study was based on service data, the staff delivering programmes were also responsible for the data collection. Therefore, it is possible that observer bias may have influenced the measurement of results. This is not uncommon in routine data collection and it has been demonstrated to impact on estimates of prevalence. 70 Our analysis considered height and weight rounding at the programme level, and these were not found to be associated with change in BMI in baseline-adjusted models; however, it is possible that observer bias may remain.

The analysis allowed us to estimate variations in change in BMI as the primary outcome. In the absence of controls we cannot attribute changes in BMI to the programme itself, as children may have been participating in other interventions elsewhere. However, it is unlikely that secular changes would account for changes in such a short time period. In addition, although there were issues with data quality in terms of height and weight rounding, it is unlikely that these could systematically have biased the results. That said, as with all studies, and especially those utilising data not collected under research conditions, there may have been other, unmeasured sources of error which may have introduced bias and we cannot assess how this may have impacted on our findings.

Most interventions measured under research conditions report change at periods of follow-up, typically 6 or 12 months. In this study, data were not available beyond the end of the programme. This limitation is probably common if using service-level data. Although the estimates derived here cannot be used to comment on the sustainability of the intervention effect over a longer time period, the intervention’s effectiveness was assessed under research conditions at 6 months and 1 year in a RCT. 15 The change in BMI and zBMI at the end of the programme was similar in both the RCT and our analysis.

Secondary outcomes

The four secondary outcomes were change in self-esteem, change in SDQ score, change in physical activity and change in diet. These are defined as follows.

Change in self-esteem was assessed using the Rosenberg scale35 (described in Chapter 3, Covariates; a higher score indicates higher self-esteem). Change in self-esteem ranges theoretically from −30 to 30, and was derived by subtracting baseline self-esteem from self-esteem at follow-up.

Change in psychological distress was assessed using change in SDQ score. A positive value on change in SDQ score means that the participant’s level of psychological distress increased. The SDQ score ranges theoretically from −50 to 50, and was derived by subtracting baseline SDQ score from SDQ score at follow-up.

Change in physical activity was measured by the difference in the number of hours spent walking or cycling in the previous week, as reported by parents in the first and last sessions of the intervention. We excluded values at baseline or follow-up which exceeded 20 hours of exercise (affecting between 1.2 and 3.4% of participants) as unrealistic. Change in physical activity could therefore range theoretically from −20 to 20 hours, and was derived by subtracting physical activity at baseline from physical activity at follow-up.

The diet score was developed for use in the MEND 7–13 programme. The score measures ‘MEND-friendliness’, a construct to assess a combination of diet and food behaviours considered to reflect a healthy balanced diet. The score is derived from parent reports on 14 items that assess:

• the types of foods the child eats

• how much the child eats

• how the food is prepared and eaten

• the adult’s perspective on the child’s food.

Responses are scored 0 (‘MEND unfriendly’), 1 (‘MEND neutral’) or 2 (‘MEND friendly’), and summed over the items to create a score out of 28. A higher score indicates that the child’s diet and food behaviour is ‘MEND friendly’. The coding matrix (provided by MEND Central) is reproduced in Appendix 12.

Covariates

The set of covariates at the participant, neighbourhood and programme levels used in models of change in secondary outcomes was the same as that used for change in BMI (see Chapter 3, Covariates). The only difference was that instead of adjusting for BMI at baseline we adjusted for each secondary outcome at baseline (for example, we adjusted for self-esteem at baseline for the model of change in self-esteem).

Missingness

As with the analyses of change in BMI, data were missing for ethnicity, family structure, housing tenure, employment status and session attendance. Data were handled using multilevel multiple imputation models (see Chapter 3, Missingness).

Multilevel models of secondary outcomes

The analytic sequence for changes in self-esteem, SDQ score, physical activity and diet was the same as that used for change in BMI (see Chapter 3, Multilevel analysis of change in body mass index).

Bivariate multilevel model

We also examined a bivariate multivariable model of change in self-esteem and change in BMI. This allowed us to model changes in self-esteem and BMI simultaneously (using the imputed data for change in self-esteem and change in BMI described in Chapter 2, Missingness and multiple imputation). This type of model directly estimates the covariance between changes in self-esteem and BMI (although inferences about the direction of causality cannot be made). It also shows whether characteristics associated with change in BMI were associated with change in self-esteem in the same way.

Cost of providing MEND 7–13

To identify the cost of providing MEND 7–13 we originally planned to undertake a de novo analysis of selected currently and previously funded MEND 7–13 programmes. However, we identified two recently conducted studies that evaluated the cost of providing MEND 7–13 and reported results in sufficient detail for our analysis. 71,72 These studies were undertaken separately and yielded similar results so we felt that further independent analysis was unnecessary. Based on these studies we present figures for the cost of providing a MEND 7–13 programme. We also calculated the mean cost of providing MEND 7–13 per participant, using published figures and our own calculations based on the number of children who started each MEND 7–13 programme. We focus on the number of starters because it is relatively straightforward to define, and because this is likely to be the basis on which local programme running costs (e.g. staffing and venue hire) are calculated. Costs were calculated from the perspective of the NHS and PSS and reported in UK£2010/11.

To identify the factors likely to affect the cost of providing MEND 7–13 we undertook interviews with staff at MEND Central, and examined the main drivers of the cost of providing MEND 7–13 from the two studies described above. Group size was identified by MEND staff as a factor likely to affect the cost of providing MEND 7–13. Data on group size were taken from the same sources used in Chapter 3.

To identify the relationship between variation in the cost of providing MEND 7–13 and the outcomes of participants we used the analysis for research question 2 (see Chapter 3) to identify, where possible, whether or not the factors likely to affect the cost of providing MEND 7–13 affected the outcomes of participating in MEND 7–13.

Description of MEND 7–13 participants

The complete case and imputed data sets for change in self-esteem and change in SDQ score are described in Table 17. Those for change in diet and change in physical activity are shown in Table 18. Changes in the secondary outcomes showed that self-esteem increased by 3.2 units, SDQ score (i.e. psychological distress) decreased by 2.9, parent-reported walking and cycling increased by 0.8 hours a week and the MEND-friendliness of children’s diets increased by 6.1 units.

As with the data on change in BMI (see Table 9), the imputed and complete case data were very similar in terms of the estimated proportions in the data. The sociodemographic distribution of MEND 7–13 participants in the change in BMI data set was described previously (see Chapter 3, Description of MEND participants and Table 9). The sociodemographic composition of the participants in the secondary outcome data sets (see Tables 17 and 18) was similar to that of the data set looking at change in BMI.

Modelling to inform multilevel models of change in secondary outcomes

The fit of all models was improved by multilevel modelling (changes in BIC in comparisons of single and multilevel models were −18 for self-esteem, −6 for SDQ score, −5 for physical activity and −21 for diet). The percentages of variation explained by between-programme variation differed by outcome as follows: 8.2% for change in self-esteem; 6.2% for change in SDQ score; 7.6% for change in physical activity; and 11.0% for change in diet. These were lower than the 16.0% observed for change in BMI (see Chapter 3).

The set of variables included in the final model was identified independently for each outcome by fitting a series of preliminary multilevel models, one for each covariate after adjustment for the outcome at baseline. As with the analysis of change in BMI, the covariates significant at the 5% level were carried forward. The density of healthy or unhealthy food outlets and changes in self-esteem and SDQ score were not tested as they were not considered applicable.

Self-esteem at baseline, ethnicity, parental employment status, IDACI 2007 decile, the number of programmes run by the programme leader up to that point, and attendance were all carried forward to the change in self-esteem model. The remaining covariates were not associated with change in self-esteem (Table 19, model 1).

Baseline SDQ score, sex, ethnicity, IDACI 2007 decile, measure of built environment, the number of programmes run by the programme leader up to that point, attendance and the amount of height rounding at the programme level were all carried forward to the change in SDQ score model. Other covariates were not associated with change in SDQ score (see Table 19, model 2).

Baseline hours of physical activity, housing tenure, parental employment status and IDACI 2007 decile were carried forward to the change in physical activity model. Other covariates were not associated with change in physical activity (see Table 19, model 3).

Baseline diet score, sex, ethnicity, measure of built environment, urban/rural status, the number of programmes run by the programme leader up to that point, group size, attendance, height and weight rounding were carried forward to the change in diet model. Other covariates were not associated with change in diet (see Table 19, model 4).

The next stage of the analysis tested random slopes for each outcome. None of the modelled random slopes improved the fit of the change in self-esteem model so none were included (Table 20). In contrast, the between-programme intercept in change in SDQ score varied by employment status (change in BIC = –16). For this reason, this random effect was included in the multivariable model of change in SDQ score. No other random slopes improved the fit of the model (see Table 20). Inclusion of random slopes for the change in physical activity and change in diet models did not improve fit so the final models for these outcomes excluded these parameters (see Table 20).

The a priori two-way interactions described in Chapter 3 (see Multilevel analysis of change in body mass index) were tested for each secondary outcome, but none were included in any of the multivariable models (Table 21).

Multivariable models of changes in secondary outcomes

As each outcome is measured on a different scale, each set of results is presented with the range, mean and standard deviation of the score at baseline to aid interpretation. Increases in self-esteem and physical activity and improved diet can all be interpreted as ‘beneficial’ changes, whereas an increase in SDQ score represents an increase in psychological distress and would therefore be interpreted as an ‘adverse’ change.

Relationship between change in self-esteem and change in body mass index

Change in self-esteem and change in BMI negatively covary at the programme and individual levels (see between-programme and between-participant covariance terms in Table 23). This means that when BMI falls, self-esteem rises. However, the direction of causality cannot be inferred.

The bivariate multivariable model (Table 23) extends the model of change in BMI moderated by self-esteem (see Table 15, model 1) by modelling change in self-esteem and change in BMI simultaneously. The associations from the bivariate multilevel model of change in self-esteem and change in BMI did not differ substantively from the univariate models of change in BMI and change in self-esteem.

The bivariate multivariable model shows that BMI and self-esteem at baseline are independently associated with change in BMI (higher BMI at baseline is associated with a larger fall in BMI; higher self-esteem is associated with a smaller fall in BMI). Higher baseline BMI and higher baseline self-esteem were both associated with a smaller increase in self-esteem (Table 22).

Cost of providing MEND 7–13

As noted, cost data were derived from two recent evaluations of the MEND 7–13 programme. 71,72

Techakehaki and Hutton71 calculated the costs of providing MEND 7–13 to all eligible children across England from an NHS and PSS perspective in UK£2010/11. Their analysis was based on activity data collected from the MEND 7–13 programme based on four cost components: programme set up; participant recruitment; programme activities; and evaluation and monitoring. For each of these items, fixed costs such as office rental and variable costs such as staffing, equipment and materials were included. Costs were not itemised by cost component. They were calculated by contract size and weighted by the number of children attending programmes with different contract sizes. Mean costs per participant were calculated by dividing the total programme costs by the number of children in the programme (it is not stated if these are numbers starting or completing the programmes, or some other measure of participation). Using this approach, Techakehaki and Hutton71 calculated that the mean cost of the programme per participant was £416.

Steeds72 calculated the total costs of MEND 7–13 programmes in South Gloucestershire, England between May 2008 and November 2009. The analysis was based on a detailed study of the costs of providing seven MEND 7–13 programmes within the region. Costs were disaggregated into local programme costs, comprising start-up costs and running costs, and costs incurred from MEND 7–13 Central to organise and run local programmes. Total costs were reported for seven programmes, and the base year for costs was assumed to be 2008/9. The costs are summarised in Table 24; they have been converted into mean costs per programme and inflated to UK£2010/11 values using the NHS Hospital and Community Health Services Pay and Prices Index. 73 The cost per participant, based on a reported 10 children per programme (figures for a subset of four programmes suggested that 46 children started these programmes and 36 completed them), was £399.

Based on these figures, the cost per programme is around £4000, and the mean cost of providing MEND 7–13 per participant is £399–416.

Factors likely to affect the cost of providing MEND 7–13

Broadly, the factors affecting the mean cost of providing MEND 7–13 per participant are those that affect local programme costs and MEND Central costs (as these are the two components of the cost per MEND 7–13 programme, the numerator in the calculation of the cost per participant), and the number of participants in each programme, i.e. the programme group size (the denominator in the calculation of the cost per participant).

More specifically in terms of local programme costs, as shown in Table 24, the main cost components are staff costs (programme co-ordinator, theory leader and exercise leader), accounting for around 60% of the cost per programme, and venue hire, accounting for around 20%. Assuming these figures are largely representative of all MEND 7–13 programmes, they are the main factors likely to affect local programme costs. Staff costs will vary depending on the specialism, qualifications and level of experience of staff. For example, as shown by Curtis73 (see table 17.2), the salary of qualified allied health professionals (AHPs) at Agenda for Change Pay Band 7 is more than twice that of unqualified AHPs at Pay Band 3 (mean annual basic salary per full-time equivalent £37,600 vs. £17,500). In terms of both staff costs and venue hire, local programme costs will be lower when these can be provided by in-kind support at zero or negligible financial or opportunity cost; for example, via the free use of unoccupied rooms.

MEND Central costs are likely to depend on the size of the contract between MEND Central and those commissioning MEND 7–13 programmes locally. For example, commissioners drawing up larger contracts with MEND Central are likely to be able to negotiate lower costs per participant.

Steeds72 reported a programme group size of 10 children per programme. Using the data analysed in Multivariable models of changes in secondary outcomes and Chapter 3, Multivariable models of change in BMI and zBMI, we calculated the programme group sizes at the first session of 1940 MEND 7–13 programmes (see Table 2). The mean, median, mode, interdecile range and overall range were 8.6, 8, 7, 5–13 and 1–19 respectively. If we applied the mean value in these data to the costs per programme calculated by Steeds,72 then the mean cost of providing MEND 7–13 per starting participant would increase to around £463. If we applied the programme group size values based on the limits of the interdecile range, then the costs would vary between £307 and £797.

The costs presented above are based on mean programme sizes, using data on the mean number of children starting MEND programmes. Programme size could instead be based on the mean number of children completing programmes, and, assuming the completion rate of programmes is less than 100%, this would increase the cost of providing MEND per participant. Using our definition of completers as those attending > 75% of sessions, the completion rate based on imputed data for participants who started the programme was 59.9%. If we applied this to the mean 8.6 children who started each MEND programme, then 5.2 children completed each programme on average and the cost per completer was £773. If we applied the completion rate to the limits of the interdecile range for starters, then the costs would vary between £511 and £1329.

Relationship between the variation in the cost of providing MEND 7–13 and the outcomes of participants

As noted, the factors likely to affect costs are staff costs (which will vary depending on the specialism, qualifications and level of experience of staff), venue hire costs, contract size and programme group size. What evidence is there that these factors affect the outcomes of participating in MEND 7–13? Using the analysis we undertook to address changes in outcomes (see Multivariable models of changes in secondary outcomes and Chapter 3, Multivariable models of change in BMI and zBMI), we were unable to obtain data on the specialism, qualifications or level of experience of local staff and so could not assess the impact of these factors on outcomes. In terms of venue hire costs, we did not know if the venue was provided at zero or negligible cost. We were also unable to obtain other possible proxies such as the type of facility where the programme was delivered. We could not find out the size of the contract between MEND Central and those commissioning the programmes; however, we did include funding source [whether primary care trust (PCT), Big Lottery Fund (BLF) or another organisation], and this was found to have a non-significant impact on outcomes, except in one instance (Table 25).

Programme group size slightly moderated change in BMI but no other outcomes (see Multivariable models of changes in secondary outcomes and Chapter 3, Multivariable models of change in BMI and zBMI).

Programme group size was found to have a statistically significant impact on outcomes (see Multivariable models of changes in secondary outcomes and Chapter 3, Multivariable models of change in BMI and zBMI), with larger group size being associated with smaller BMI reductions. Hence, larger programme group size is associated with lower costs per participant, but worse outcomes.

Therefore, there is limited evidence that the factors affecting the cost of providing MEND 7–13 are related to the outcomes of participating in MEND 7–13. We found evidence of an effect for programme group size but not for other factors, as we did not have sufficient data for the analyses of specialism, qualifications and level of experience of staff, venue hire and contract size.

Changes in psychological health associated with MEND 7–13

On average, MEND 7–13 was associated with improvements in self-reported global self-esteem and reductions in parent-reported psychological distress (change in SDQ score). Conservatively, this can be interpreted as showing that MEND 7–13 does not adversely impact on children’s psychological health (i.e. it does no harm). More liberally, it suggests that participation in weight management interventions such as MEND may have at least short-term benefits for psychological health when delivered under real-world conditions. This is consistent with evidence from childhood weight management programmes delivered under research conditions. 67,74

Variations in psychological outcomes associated with MEND 7–13

Higher baseline self-esteem and Asian compared with white ethnicity were associated with lower levels of improvement in self-esteem. Otherwise, change in self-esteem did not vary by sociodemographic, neighbourhood or programme characteristics.

Benefits associated with the programme, in terms of reduced parent-reported participant psychological distress (change in SDQ score), also varied by participant characteristics. Psychological distress reduced less in children with lower baseline psychological distress (consistent with regression to the mean); boys compared with girls; white compared with black ethnic groups; children living in more deprived neighbourhoods; those who attended programmes where the manager had delivered a higher number of previous programmes; those who completed < 75% of sessions compared with ‘completers’; and those who attended programmes where height rounding was common. Although change in psychological distress was not associated with family structure, housing tenure or parental employment status, between-programme variation was higher for participants with unemployed parents.

Thus, change in psychological distress was associated with more covariates than change in self-esteem. Self-esteem and psychological distress are conceptually and operationally distinct, and therefore liable to vary in different ways with the same covariates. Furthermore, SDQ is reported by parents whereas self-esteem is reported by participants. SDQ scores may partially reflect children’s and parents’ distress, and might therefore be affected by factors which influence both children’s and parents’ mental health. The same is not true for self-esteem, which is self-reported. Therefore, it might be expected (as was found here) that the parent-reported measure (psychological distress) would be associated with a wider range of factors than the participant-reported measure (self-esteem).

Covariation in changes in self-esteem and change in body mass index

The bivariate model of changes in self-esteem and BMI allowed us to show that self-esteem at baseline and BMI at baseline were both independently associated with changes in self-esteem and BMI. This is important when considering the longstanding debate about the direction of the relationship between changes in self-esteem and BMI, as it shows that all possibilities are supported by the data. This is perhaps consistent with a model whereby self-esteem and BMI genuinely covary, rather than one causing variations in the other. To our knowledge, no other studies have investigated BMI and mental health outcomes in a bivariate model to assess this possibility.

Change in lifestyle behaviours associated with MEND 7–13

Parent reports suggested that their children’s physical activity (walking and cycling) increased after participation in MEND 7–13. Higher levels of physical activity at the start of the intervention were associated with smaller increases over the course of the intervention. No individual-, family-, neighbourhood- or programme-level covariates were associated with change in physical activity, apart from housing tenure; families living in social rented accommodation increased their physical activity more than owner occupiers. This suggests that the intervention is associated in a similar way for most participants and their families across the service, once baseline physical activity is taken into account.

The finding that SEC did not moderate physical activity outcomes (apart from slightly larger increases in physical activity among children living in social rented accommodation) is broadly consistent with a recent review which draws together the limited evidence base to suggest that physical activity interventions are equally effective for low and high socioeconomic groups. 23 However, as physical activity did not vary with any other covariates, the lack of a statistically significant difference associated with SEC should be interpreted with caution because of the known errors in reported physical activity. 75

The results show that the MEND-friendliness of participants’ reported diets increased over the course of the MEND 7–13 programme. Those with more MEND-friendly diets at baseline reported a smaller increase, as did boys compared with girls, Asian and black children when compared with white children (there was no significant difference between other ethnic groups and white children), children in programmes where the number of previous programmes per manager was higher, non-completers and partial completers compared with those completing the programme, and those in programmes where height and weight measures were often rounded.

Cost of providing MEND

In summary, we met the aims of our analysis as follows:

1. The cost of providing MEND is around £4000 per programme. The mean cost per starting participant is £463 and the mean cost per completer is £773.

2. The cost per participant varies according to:

   1. local programme costs, which are affected by, among other things, the number and type of staff employed, administration costs and venue hire

   2. the costs incurred from MEND Central, which are affected by the size of the contract between the purchaser/deliverer and MEND

   3. the number of participants on each programme.

3. There is limited evidence that the factors affecting the cost of providing MEND are related to the variation in outcomes. We found evidence that larger programme group size is associated with lower costs per participant but smaller BMI reductions per participant. In other cases there is no evidence because we do not have sufficient data to undertake appropriate analyses (e.g. with regards to the specialism, qualifications and level of experience of staff, venue hire and contract size).

Strengths and limitations

All the secondary outcomes were reported by participants in the intervention or their parents. As such, all are subject to sources of bias which can affect the internal validity of measures. These might include social desirability bias; for example, respondents might be less likely to report feelings of low self-esteem or psychological distress because of stigma surrounding these issues. Positive outcomes might be inflated because the questionnaires were distributed by the same people who managed the intervention.

The psychological measures used in the study have been validated against several different populations. However, the majority of the research literature uses multicomponent measures of self-esteem67 (as opposed to the adapted Rosenberg measure used here) and no studies were found which had used the SDQ in this context.

A recent Cochrane review of childhood obesity interventions10 suggests that there are considerable challenges in measuring lifestyle behaviours such as physical activity and diet, and recommends that research uses measures that have been validated to ensure that the data reliably measure the behaviours of interest. Our analyses of dietary patterns and physical activity were based on scales which have not been validated in previous research studies. Accordingly, the findings from these scales should be treated with caution and the results cannot be compared directly with those of other studies.

Psychosocial and lifestyle outcomes associated with MEND

Psychosocial (self-reported self-esteem and parent-reported psychological distress) and lifestyle (parent-reported physical activity and diet) outcomes improved overall. These varied by group:

• Improvements in self-esteem were smaller where baseline self-esteem was higher and for Asian compared with white children, but otherwise did not vary.

• Reduction in psychological distress was smaller for children who were less distressed at baseline; boys; black compared with white children; those living in more deprived neighbourhoods; children attending programmes delivered by managers who had delivered larger numbers of previous programmes; those who did not complete the programme; and those who attended programmes where height data were rounded.

• Increases in physical activity were smaller for those who were more physically active at baseline and those living in owner-occupied compared with social rented accommodation, but otherwise did not vary sociodemographically or by completion status.

• Dietary improvements were smaller for those who had healthier diets at baseline; boys; Asian and black children compared with white children; those attending programmes delivered by managers who had delivered larger numbers of previous programmes; children who did not complete the programme; and children who attended programmes where height or weight data were rounded.

Costs of providing MEND

The cost of providing MEND is around £4000 per programme. The mean cost per starting participant is £463 and the mean cost per completer is £773. The cost per participant varies according to:

• local programme costs, which are affected by, among other things, the number and type of staff employed, administration costs and venue hire

• the costs incurred from MEND Central, which are affected by the size of the contract between the purchaser/deliverer and MEND

• the number of participants on each programme.

There is limited evidence that the factors affecting the cost of providing MEND are related to the variation in outcomes.

Recruitment

From a full list of 151 contract holders, shortlists were developed to meet inclusion criteria for a maximum variation sample. There was a geographical spread of interviews, with commissioners and delivery partners located in London, the south-west and the north-east of England.

We selected commissioners according to the sociodemographics of the local population (index of multiple deprivation and area ethnic composition), the type of organisation (PCT or not) and contracting details (number of programmes contracted and current contracting status). The sampling approach is fully described in Appendix 13.

Although there were few refusals, a considerable number of commissioners could not be contacted during this period of rapid change in the NHS.

Twenty commissioners and nine delivery partners were recruited into the study. Some delivery partners had been delivering MEND for some time, whereas other commissioners were new in post.

The recruitment of interviewees is described in Figure 6.

FIGURE 6.

Recruitment process for commissioner interviews. Please note that 54 individuals were contacted but the final sample increased to 66 because of onward referral. ICH, Institute of Child Health.

Interviews

Our proposal described recruiting up to 30 commissioners in the expectation that we were unlikely to generate significant new data after that point. In the event, 29 interviewees took part in 26 interviews (see Figure 6) in 2011. They included those who had commissioned MEND and those who had managed it or contributed to the programme delivery. Interviewees included health and well-being development officers, strategic leads for obesity, medically qualified public health consultants and MEND operations managers or co-ordinators. For some commissioners, MEND had been commissioned or was running by the time they had taken up post. Not all had observed the delivery of the MEND programme first hand but those who had not had access to reports from colleagues.

Interviews with commissioners were largely conducted by telephone, with five face-to-face meetings. Consent forms and information sheets were sent in advance and the interview started only once written/e-mailed consent had been received. Some interviewees asked to see the topic guide before being interviewed and this was provided, with the caveat that this was a guide and not a questionnaire. All interviewees agreed to the interview being audio recorded. Interviewees took a highly professional approach. Despite our interviewing them at a time of considerable professional uncertainty for many, we found them generous with their time, thoughtful in expressing their views and experience, and well prepared. Some had gathered data or papers in advance or offered to send them subsequently. Some expressed concerns related to confidentiality, particularly where they had criticisms of the scheme. Although the information sheets emphasised that the research was independent of MEND, several respondents were aware of the relationship between MEND and UCL Institute of Child Health (see Chapter 1, Management of competing interests). The interview material reported here has all identifiers removed.

The topic guide (see Appendix 14) for the interviews was developed on the basis of our proposal to the NIHR and was adapted following initial interviews with commissioners on our advisory group. The wording was further adapted according to whether we were interviewing delivery-partner contract holders or commissioners. Interviews typically lasted around 45 minutes. Encrypted data were sent to a secure drop box facility and then uploaded using a secure file transfer system for transcription. Verbatim transcriptions were returned to the qualitative research team after removal of identifying features such as names.

For ease, the term ‘commissioner’ will be used to refer to both delivery partners and commissioners, unless otherwise stated. Individuals are identified using unique interview numbers (e.g. C6).

Analysis

Before commencing analysis, the qualitative researchers held a team workshop, also attended by JF and our colleague Hannah Lewis (who had dealt with recruitment), to discuss the strengths and weaknesses of various approaches to analysing our three data sets (family, individual and commissioners). The constant comparative approach, which compares new emerging concepts from the data, guided our thinking and we decided that a framework analysis, enabling us to draw on this, was the best way to analyse data from both the commissioners and the families.

The interviews were analysed using framework analysis, drawing on work by Ritchie et al. 78 This involves familiarisation with the data and drawing up a thematic framework on the basis of research questions and interview data. Summaries of data are indexed, mapped and tabulated by case and theme within this framework to allow analysis and interpretation and reconsideration of themes.

The context for commissioning MEND

Engaging with families for children

MEND’s family-based approach was largely endorsed by interviewees. They valued the active involvement of parents and carers, but acknowledged that this brought challenges for recruitment and retention. They also emphasised the need for skilled staff who could engage with families, including families who might have multiple health or social needs.

Delivering MEND for local populations

The sociodemographic and ethnic diversity of families was considered important, along with the content and delivery of MEND programmes and local adaptations in enabling better engagement with families on the programme.

Interviews

Data were gathered through individual and family-based group interviews. A topic guide was developed by the qualitative team, informed by the literature, discussions with the wider project management group, the advisory group and the commissioner interviews. Transcripts from the first family interviews were reviewed by the research team and the topic guide was refined. The guide was a tool, not a rule, and was used to map rather than constrain our discussions. In order to make sense of MEND from the participants’ perspectives, we needed to elicit their narratives. Interviews began with open questions about family make up and memories of MEND, after which families were encouraged to take the lead with prompts towards the following broad topics:

• the opportunities and trade-offs presented by participation in a MEND programme

• costs and benefits within the family

• access to schemes

• positive and/or negative experiences of participation.

Researchers were alert to whom within the group had opportunities to speak, and managed conversations in order that everyone could contribute.

Ethics, informed consent and data protection

The work was approved by the UCL Ethics Committee (2842/001). Amendments were submitted to allow further advertising for participants in the under-represented categories. Age-appropriate participant information and consent forms were used, including versions for children with learning disabilities. Families were provided with information several weeks prior to participation, and the research was explained in person before interviews began. All parents and children were asked for written consent, except for one very young child who gave verbal consent. Translation/interpretation was offered to families for whom English was an additional language. Researchers ensured that all members of the family (including children) were given the opportunity to opt out of the research, and some children did choose to leave interviews before they ended. Researcher safety for interviewers was addressed by ensuring that another member of the research team had details of interview locations and mobile telephone contact details.

Interviews were recorded and encrypted sound files were sent to a secure drop box facility and then for transcription. Verbatim transcriptions, with identifying features removed, were returned to the qualitative team. Hard copies of consent forms were securely archived.

Analysis

As with the commissioner interviews, data were analysed using framework analysis. 78 The qualitative researchers collaborated to develop the initial framework for data management. All contributed to tabulating data within this, including both individual and family group data. One team member drew up a list of higher-order themes and interpretations, extended after discussion with the whole team. We took into account issues raised by the NCB’s young people’s PEAR group (see Chapter 1, Public involvement), consulted prior to the research proposal being submitted, including the accessibility of MEND, how it fitted into parents’ busy lives, the potential for positive peer groups and the potential for attending MEND to be stigmatising.

As an aim of individual interviews had been to hear from those with ‘alternative perspectives’, a further analysis was carried out to explore tensions between participants’ accounts within group interviews, or between group and one-to-one interviews within the same family. A framework approach was used where participants’ accounts were at odds, or where participants challenged or contradicted one another, and a typology was developed to describe these.

Recruited sample

Of 68 families contacted, 10 declined to take part (15%), 18 did not respond to calls or voicemails (26%), 14 had invalid phone numbers and did not respond to letters (21%) and 26 (38%) agreed to be interviewed. Of these, 22 went on to be interviewed and 4 declined on further contact. One further family was recruited through snowballing. Appendix 15 provides further details on recruitment.

We interviewed 23 families comprising 64 individuals, including 22 mothers (one a foster mother), six fathers, two grandmothers, two aunts, 12 male and 12 female child MEND participants, five brothers of child participants, eight sisters of child participants and one family friend. Our sample included adult participants who worked in health or social care, including two who had at some time been responsible for commissioning decisions about MEND.

Although our final sample captures variation across most of the dimensions of interest (Table 26), there was a larger proportion of owner occupiers and MEND completers than planned. Most had attended earlier programmes (most families were referred in 2008). This afforded an opportunity to explore participants’ reflections on change over time. In order to ensure anonymity, individuals are identified using a unique family number and a brief descriptor (e.g. ‘father, F4’).

Some families initially found it difficult to remember details of MEND, but talking jogged their memories and most participants had clear views on what they had found useful, enjoyable and difficult about MEND, both at the time and looking back. We were aware that these interviews were sometimes difficult for families; unhappiness with weight was expressed by adults and children, and family members disagreed, contradicting each other’s memories and interpretations of events. Occasionally, children expressed unhappiness that their parents had not made changes that they wanted; some parents felt that their children were resistant to change.

Where mothers and fathers were both present in interviews, mothers had the most to say but fathers too made important contributions. In four cases, men (one grandfather and three fathers) had chosen not to take part in interviews, but were in fact present. Two of the fathers who were ‘listening in’ in this way asked to join the conversation and were consented.

Our findings fall into four main sections mapping on to our research questions about the salience, acceptability and costs of the programme to families, and the perceived costs of maintaining a healthy lifestyle afterwards. The first section explores reports of why families attended and who made the decision to do so. The second explores experiences of attendance, including positive experiences, barriers to attending and perceived costs associated with attending. Finally, we consider long-term change as understood by families, beginning by setting out perceptions of long-term change, and perceptions about costs and constraints associated with trying to maintain behaviours promoted by MEND after the programme had finished.

Referral to MEND and decision to attend

Most families self-referred, having seen leaflets or heard about MEND from friends or family who knew that parents were concerned about diet or weight. Four families were referred through some part of the health service (e.g. a nutritionist) and in one case a community worker recruited a group of local families to MEND and arranged shared transport for them. Schools were an important source of information, including leaflets received with NCMP letters. The impetus for contacting/attending was not always clear in discussions, but reasons offered fell into four areas: experience of bullying, transition to secondary school, concerns about diet or weight, and concerns about other health problems.

The most common reason given was that children had experienced bullying, and parents and children felt that losing weight might reduce this: ‘I used to get bullied a lot about it. I just decided to drop the weight and stuff’ (boy, F18).

This chimes with children’s focus on the social impact of overweight observed in other studies,79,80 and in our study this was often associated with the transition to secondary school. This seemed to be an important moment, because of fear or experience of bullying at the ‘teenage school’ (F61) and because it was a time of transition for the children. There was a sense that they could either capitalise on changes for the better, or slip back into familiar habits.

A small number of parents were worried about their child having a poor diet and felt that MEND might be useful in correcting this. Many families (parents and children) said that they were worried about excess weight, although parents seldom explicitly used the term: ‘I heard about it from my friend [name] because her lad was like [son’s name], you know, a big lad’ (mother, F61).

Several parents referred to their own weight. One father (F4) said of himself, ‘I’m not the smallest person in Britain’. Overweight was sometimes identified as a family problem: ‘we knew there was an issue in our family and we didn’t want our children to end up like our parents or us either’ (mother, F70).

Some parents described weight gain, and attending MEND in the context of another problem for the child such as learning disabilities, cystic fibrosis or endocrine problems:

[I]t’s kind of a genetic hormone imbalance, he can be prone to weight gain so he was starting to put on weight at a certain stage and I just happened to see a leaflet advertising the MEND programme and I thought, ‘Oh, that’s something . . .’ I couldn’t afford to go and hire a private dietician or anything, so it was something I was able to go to[.]

Father, F57

Despite some shifts in the domestic division of labour nationally and internationally,81 it remains largely mothers who protect and promote the health of children. In our study, responsibility for undertaking MEND lay largely with women. In almost all families, the mother was behind the decision to attend. Most often it was she who heard about the scheme, decided the child should attend and persuaded him or her to sign up:

I think it was basically both of us. Mum said to me, ‘Oh, there’s this going on, would you like to maybe try it for a week or so, see how it goes, don’t like it, don’t have to go, if we do, stay there.’

Girl, F26

In some cases this was not negotiated: ‘you just told us one day we’re going’ (boy, F50).

Only in one case – that of a father parenting alone – was a mother not the route by which the family entered MEND. Where children lived with both of their parents (n = 11), the mother accompanied the child to MEND in all but one family. This was often described as being because mothers worked part-time or had more flexible working arrangements than their partners. Only five fathers reported ever attending any MEND sessions; two regularly accompanied their sons and three went to some sessions when asked or when their partner could not attend. Other family members also helped out; grandmothers, aunts and older sisters attended when parents could not. It was very common for siblings to attend alongside the child participating in MEND, usually because there was no other child care available. In two families more than one child in the family was referred to MEND.

In the interviews, non-attending adults told us that it simply was not possible for them to attend because of their work, but they felt they played a role in supporting changes at home and that they did what they could: ‘You had me running around the bloody field. What do you mean I didn’t get involved?’ (father, F27).

Two families felt that the wrong decision had been made about which adult should attend MEND. In family 67, the mother attended with her daughter, but felt in retrospect that as her partner did most of the food shopping, changes in diet would have been more successful if he had attended instead. In family 4, the father and his son attended, but the mother felt excluded from MEND and believed that the lack of change in their diet was because she did not know what she was supposed to do: ‘Because that’s the problem with not going, I was relying on two boys to bring the information back’ (mother, F4). Her partner agreed that lack of communication had impeded family change following MEND, but her son felt that they had left the information for her, and it was her responsibility to read it.

Family experiences of attending MEND

Perceived costs of MEND participation

Parents most often told us that their participation was constrained by the time cost, rather than the financial cost. One mother of three children, one with special needs, whose husband often works away from home, said: ‘I just felt like I’d invested a lot of time, forget about the money, but a lot of time and effort’ (F67).

The main message from parents was that they would find a way to pay for the things their children need: ‘I don’t mind paying for something if he likes doing it. I mean, my parents will help out, you know, because they know he likes doing it, they don’t mind’ (mother, F61).

The same mother also told us that money had been ‘tight’ lately, and that ‘there isn’t a lot of money to spend on healthiest food: I mean I’m a single mum, you know, I can’t go out and buy all this decent low fat food, all the yogurt’. She and others discussed the fact that low-cost ranges and low-cost supermarkets do not include the healthier alternatives they had been encouraged to buy, such as lean mince or turkey breasts. As for the more exotic foods they had tried with MEND: ‘at the end it was all healthy stuff, but it was more expensive’ (girl, F26).

Others said that some food costs had gone down because they ate less meat, or there was less waste because they were making things with fruit and vegetables, but also that it need not be expensive to eat healthily: ‘Nowadays they’re [fruit] a pound a bowl . . . And they’re everywhere. So there’s no excuse saying it’s expensive’ (mother, F13).

Any change in costs to families was related to prior buying habits. A study of food prices has suggested that although a change from low-cost food purchased at budget supermarkets does carry a cost, families may be able to keep costs to a minimum if they have the time and resources to seek out the keenest prices. 83

Many of the barriers to attending MEND described by parents would probably have resulted in financial costs, but were not identified as such by families. Time off work, child care and, in particular, driving to the venue were likely to have carried costs. Several families said it was easy to attend because the venue was just a short drive away, but did not identify this as costly because they already owned and ran a car: ‘The only cost would be travelling there, but you drive anyway‘ (foster mother, F28).

Similarly, several parents arranged shifts or child care differently in order to attend MEND, and this was not perceived as a new cost, but part of the ordinary running of family arrangements.

Alongside time, the greatest perceived burden on families was the emotional demand. Parents spoke about this, and of people crying in sessions. They were aware that the children sometimes found the sessions emotionally demanding. An exchange with one father (F57) describes his observations:

. . . I think the children are crying out for help, but they don’t know what’s going wrong and why they are the way they are. And it’s kind of the parents’ fault really and it just makes them aware of that. I think some of them feel quite upset and bad about it.

So do people get upset in some of the classes?

Yeah, a couple of the kids did when they kind of get up to introduce themselves and why they’ve come to the MEND programme.

This could be perceived as intrusive and inappropriate:

And then it was almost like they were trying to look for some psychological reason . . . and I should imagine everyone has a psychological reason, but it was almost like they were looking for something deep and meaningful. I thought to myself, ‘I haven’t come here to be psychoanalysed’ which is what I felt that the leaders were trying to do.

Mother, F67

Although on the one hand children described enjoying the opportunity to exercise with others with weight problems, there was an emotional cost to attending a weight intervention programme.

The top, I didn’t like the top . . . I got out of the car, launched it on, and then launched a coat over it. It was just really embarrassing . . .

Why didn’t you want people to see you?

Just if they laughed or something, called me names.

Did people know what MEND was about?

I didn’t tell anyone. I just said I had to go to the leisure centre after school.

Study limitations

An important limitation of the study was our failure to recruit people referred to the scheme but who did not attend. This was due to limitations in the selection procedure applied to the contact data set. Although it is not uncommon for those who do not engage with a service to be under-represented in data sets,89 our experience in recruiting such samples in the past had been less problematic, and in particular we had found that people who decline or drop out of programmes are not reluctant to say why.

Some features of research governance may have made it more difficult for us to recruit to this study. Our recruitment process was lengthy. At the request of the ethics committee, the letter of invitation to take part in this study was followed by a 21-day ‘cooling off’ period before researchers could make contact. Families were not reluctant to opt out once telephoned, but none did so during the 21 days, suggesting that the time lag may not have been helpful to participants or researchers. After the delayed contact, many had no recollection of having received a letter, or said they had thrown it away or had read it but forgotten what it had said. For those who agreed to an interview, there was often a further delay in order to find a time when family members could meet the researcher. This long process was evidently unattractive to people for whom the intervention was a distant memory, lower on their agendas as participants than on ours as researchers. We believe that for low-risk studies such as this one, direct contact from researchers and a shorter opt-out period is likely to reduce barriers without compromising participants’ opportunities to opt out.

Finally, we may not have provided sufficient incentive for families to consider taking part. In recent work with an advisory group of parents for another study, parents suggested three changes to information materials that are needed to make them attractive. Firstly, written information should be as parsimonious as possible, as too much written detail is off-putting; such detail is better described in person by the researcher once initial interest has been shown. Secondly, it would be appropriate to emphasise the ‘thank you’ payment offered for those taking part in research. Thirdly (and this is perhaps the most difficult to achieve), the parents wanted reassurance that their contribution would make a difference to policy or practice. The careful wording and long information sheets that have evolved from more complete research governance may put families off taking part, and researchers may want to consider in what circumstances written information may not be the most appropriate strategy.

The sample of families recruited here does not claim to be representative. Rather, it offers a range of views and experiences which help to understand service delivery and service acceptability. However, ‘refusers’ are missing voices, and as those we interviewed are ‘compliers’ both in terms of MEND attendance and allowing us to interview them, it is likely that our work presents a less rounded picture than would have been the case had we been able to learn from low or non-attenders. Young people who may have wanted to attend but could not as there was no family member to accompany them, or parents unable to persuade their overweight child to attend, are unrepresented in this study.

In addition to the lack of low and non-attenders, we recruited fewer individuals than planned, but believe that this did not have a significant impact on findings. Families sufficiently motivated to attend MEND together, and to then agree to a research interview as a group, may be so familiar with the give and take of family discourse that the silenced and the overtalkative were uncommon. Siblings did not seem to be constrained in their teasing or squabbles because an interviewer was present and observing family exchanges allowed researchers to better understand the dynamics of the household. 90 The additional analyses we undertook identified convergence or divergence of narratives, and demonstrated occasions when children were able to use the interview to challenge their parents.

Family findings which relate to the salience and acceptability of MEND

There were no interviews in which no one in the family felt they had benefitted, but few felt that they had managed long-term change.

What types of costs are borne by families when participating in MEND, and in sustaining a healthy lifestyle afterwards?

• Costs are mostly associated with a move to higher-quality food.

• ‘Treats‘: ‘bad’ food was still thought of as a treat.

• ‘We find a way’: parents prioritise spending on things to benefit their children.

• Time costs have to be negotiated with work and child-care arrangements (a hidden financial cost).

• Transport costs: probably underestimated by parents (e.g. ‘It was only a 10-minute drive away’).

• Social/emotional cost of making/maintaining behaviour changes generally unsupported by the wider environment.

Challenges of using routine data for research

Data collected through service provision are normally intended for monitoring or other service delivery purposes, rather than in anticipation of research. For example, the MEND 7–13 data did not distinguish whether attendance data were missing owing to reporting error or actual non-attendance. This meant that there was a degree of uncertainty about the construct validity of attendance variables. Although we retrospectively addressed these biases to some extent using data cleaning and multiple imputation, these types of issues are, like missingness, best addressed at the data collection stage.

Where possible, we have conducted analyses to test the sensitivity of our findings to analytic decisions. For example, in Chapter 2 we estimated differential completion, where completers were defined as those attending > 75% of sessions. However, there is a lack of consensus in the literature about what constitutes completion, and therefore we also estimated the model with completers defined at a 60% threshold. Results were similar, giving us confidence that the findings were not threshold dependent.

Design issues

Our analysis explored internal variations in a specific subgroup of the population: children who were overweight or obese. In study design this is, in essence, a large case series. We did not have a comparator group who had not gone through the intervention, and therefore we cannot make causal inferences about whether or not the programme achieved changes in outcomes beyond what might be expected over a similar period for children who were overweight but did not receive this intervention. This means that we cannot make assertions about whether or not the intervention was effective or differentially effective. We were, however, able to explore whether or not the outcomes associated with the intervention varied for subgroups of children participating in MEND under service conditions.

Perspectives from quantitative and qualitative work

The mixed-methods approach used here allowed us to reflect on the MEND intervention from three distinct angles: the quantifiable outcomes, the perceptions of families and the perceptions of commissioners. The commissioner/service delivery and family studies were designed to derive expert views from those receiving, commissioning or delivering the intervention within the context of a sampling strategy designed to achieve maximum variation. The quantitative work aimed to summarise the average change in weight management and psychosocial outcomes associated with participation in the MEND 7–13 programme.

Our quantitative analysis was based on short-term prospective data, whereas the qualitative analysis was based on retrospective recall of the programme over a longer time period. Both had distinct advantages and disadvantages. The quantitative analysis showed that, on average, there were short-term benefits in terms of weight management and psychosocial well-being, but our data could not be used to comment on longer-term changes. In contrast, the qualitative work suggested that although families perceived short-term benefits (there were none who felt that nothing had been gained from the programme), these were felt to be relatively modest and difficult to sustain in the longer term in an obesogenic environment. Whereas, in the short term, the commissioners commented favourably on the evidence-informed ‘solution in a box’, concern was expressed about the longer-term cost implications once Big Lottery funding was no longer available.

Gaining information from families who did not engage with MEND

People referred to a programme or intervention who do not take it up, or who withdraw, are of particular interest to researchers and those charged with commissioning services. This is especially important when the per capita cost of the programme, relative to the population health gain, is at least partly conditional on achieved recruitment and retention rates.

Families who were referred to MEND 7–13 but who did not attend, or attended few sessions, were under-represented in the qualitative component of the study and had less data available in the quantitative component. As discussed in Chapter 2 (see Findings, Determinants of completion of MEND 7–13), those in less well-off circumstances who started the programme were less likely to complete than those who were better off. However, when we compared starters (including completers) and completers with the MEND-eligible population, the differences suggested that participation in MEND 7–13 was socioeconomically and ethnically equitable even after differential completion was taken into account.

We did not have sufficient information to identify those referred to the programme accurately, and this led to a lack of information in the qualitative study relating to the perceived acceptability of the MEND 7–13 prior to attendance, and reasons for dropping out of the programme early. This was important because some of the commissioners interviewed expressed concerns about retention rates in the programmes. It is a limitation of our study that as a result of data problems at the sampling stage described above, our findings do not include the perspectives of those who were referred to the programme but did not attend, or dropped out early on. Participants in the qualitative component were eloquent in telling us what had made attending and engaging hard. We can only surmise that these or other difficulties were even less surmountable for the non-attendees or those who dropped out early.

Intervention-generated inequalities

MEND 7–13 has been scaled up to include many families across the country, so it is important to consider what our findings mean in terms of the potential impact of the scaled-up programme on population-level inequalities in health. The evidence base for interventions delivered in a community setting is relatively small and equivocal,23,93 with a lack of information on how outcomes associated with family-based weight management interventions for children vary by ethnicity and SEC. This is an important gap given the variation in overweight and obesity for these groups. Inequity might arise at several points in the delivery of a programme, from differential access to differential outcomes. These pathways are discussed below with respect to the findings from our project before general conclusions for population-level health inequalities are drawn out.

Implications of findings for translational research in public health

In order to generalise from a study of this kind, findings must be understood as part of a framework linking intervention development, public health decision-making and population health. One such framework was put forward by Ogilvie et al. 93 and is reproduced in Figure 7. We consider where our study contributes to this framework. Findings from evidence syntheses directly and indirectly influence policy responses and commissioning decisions (‘Public realm’ box in Figure 7; see Chapter 5, Findings). The ability of organisations to scale up delivery of an intervention in practice will determine how available the intervention is, and how far it is taken up by all groups in the population (‘Professional practice’ box in Figure 7; see Chapter 2, Findings). The salience and acceptability (‘Professional practice’ box; see Chapter 6, Findings) of the programme may partially determine how far the programme is taken up and subsequently engaged with by families (and by which types of families). The intervention may then determine changes in individual and family behaviours which may translate to small changes in population health behaviours (‘Health-related behaviour’ box in Figure 7; see Chapter 4, Findings) and population-level health (‘Heath and well-being’ box in Figure 7; see Chapters 3 and 4).

FIGURE 7.

Translational framework for public health research (adapted from Ogilvie et al. , 200993).

Currently, and to our knowledge, no studies have assessed how family-based childhood weight management interventions operate in the real world, although a study to examine the scale up of an adult weight intervention has been outlined. 95 Our study evaluated one family-based intervention and it is difficult for us to say that our findings are generalisable to other weight management interventions for children, or to other community-level interventions. However, the model of study we have undertaken might be useful in evaluating other interventions shown to be effective in RCTs when rolled out at scale. Research designed to explore the real-world impact of other interventions should continue to be developed.

The sensitivity of MEND 7–13 to the obesogenic environment

Obesogenic environments are generally discussed with reference to their role in the development of overweight and obesity at the population level96,97 and, following from that, as environmental determinants to be modified by area-based intervention. 98 Obesogenic environments have not, to our knowledge, been empirically tested with respect to their interaction with family-based weight management interventions, partially because individual-level intervention studies are rarely statistically powered to test whether or not outcomes are moderated by factors such as the availability of healthy food and the nature of the built environment. In this study we had a large enough sample to examine whether families’ food and/or built environments might lead to variations in the outcomes associated with participating in MEND 7–13; in particular, whether or not children and families living in environments which were more obesogenic may have been less likely to benefit from participation. Families interviewed were receptive to healthy food messages but some reported not always being able to afford these foods, or that they were not available in low-cost supermarkets (see Chapter 6, Perceived costs of MEND participation). Other families found that healthier food was not more expensive. When investigated using service-level data, we found no differences in outcomes for those living in areas which were very dense in terms of fast food and other unhealthy outlets compared with areas with high proportions of healthy outlets. This suggests that any changes associated with the programme are, in the short term, apparently resistant to differential exposure to food environments which either promote or constrain a healthy diet. This is perhaps not surprising in the UK context; despite the popular characterisation of ‘food deserts’ (areas where healthy food is less readily available), the evidence that variations in the availability of healthy food in the local environment are associated with levels of overweight is relatively weak. 68,69 Although the quantitative analysis did take into account the local retail environment in terms of fast food shops and access to healthy food, it is important to note that qualitative work highlighted that exposure to highly calorific foods may go beyond this environment, and that festivals and celebrations may also have a role. In addition, food environments measured in the quantitative analyses, using broad exposures such as fast food restaurant density, may not be precise enough to estimate relationships with short-term change in outcomes. Finally, although the quantitative analysis took into account broad sociodemographic characteristics of families, it could not examine detailed characteristics of families’ lives suggested by the qualitative work, such as school meals or the influence of grandparents, which might help the family to keep unhealthy foods out of children’s diets.

In addition to energy intake, energy expenditure is a central component of overweight and obesity, and evidence suggests that interventions may be able to influence this by increasing physical activity. However, the degree of physical activity that children (in particular) can pursue is subject to the constraints and affordances of local built environments, such as the availability of green space, and family resources including the time of parents and other carers. In this study we found unadjusted associations between the built environment and change in BMI. These were attenuated to non-significance in multivariable models (see Chapter 3, Modelling to inform multilevel models of change in BMI and Multivariable models of change in BMI and zBMI) but the unadjusted associations were consistent with reports by participants in qualitative family interviews that when they tried to change their physical activity behaviours they felt constrained by their local environment.

In the qualitative work, participants valued the opportunities afforded by the programme to exercise in a non-competitive environment with others of similar build. The lack of such opportunities following the programme was considered by some to be a problem.

MEND as an event in the life course

The children who are referred to and participate in MEND 7–13 are, by definition, overweight or obese. The natural history of this group is one associated with adiposity in early life which is likely to track into adolescence and adulthood. Moreover, children who are overweight at this point may well live in families where parents struggle to maintain healthy lifestyles. In the context of this natural history, MEND 7–13 and other short-term interventions tend to be one-off events, of relatively short duration in the life course. Many families struggled to remember the programme in detail until prompted in family interviews (see Chapter 6, Recruited sample).

That said, most families did report having gained something positive from the experience of participation in MEND, with some also reporting changing aspects of their lifestyles (see Chapter 6, Family experiences of attending MEND). However, several things made it difficult to sustain changes associated with the intervention in the longer term. These sometimes related to the constraints imposed by the wider environment (as discussed in the section above), or could be due to the changeability of family life or simply the ‘pull’ of normal ways of living, not accompanied by the time, financial and emotional costs associated with participation in the programme. Although MEND Central runs a website giving MEND graduates a forum to stay in touch, feedback from family interviews suggests that this was perceived as somewhat ironic, and of course counter to advice to be less sedentary (see Chapter 6, Perceived costs of MEND participation). Families’ concerns about sustainability were also shared to some extent by commissioners, who felt that MEND 7–13 was supportive while it was running but needed to continue to support families in the longer term.

Our quantitative findings showed that younger children reduced in BMI more than older children. This suggests that interventions such as MEND 7–13 might be more effective when families have more control over children’s lifestyles. On the other hand, qualitative data suggested that, in some cases, parental and family influence actually limited children’s engagement with physical activity or new foods. Selective uptake, whereby more proactive parents, whose children may benefit more from weight management interventions, take part in programmes such as MEND at an earlier point in their child’s growth, might also account for these differences.

Interventions in the real world, a reflection on complexity

Tony Delamothe, writing in the BMJ on complexity theory,99 refers to the work of Shiell et al. ,100 saying:

. . . evaluating complex interventions is a doddle beside the challenges of evaluating the effects of interventions on complex systems. In complex systems ‘everything is interconnected, changes in one part of the system feed through to other parts of the system and feedback on themselves.’ [Quote from Shiell et al.] Attributing causality is difficult.

Family-based weight management schemes are only one way that overweight and obese children can be encouraged to adopt healthier lifestyles. For this reason, we have situated our findings in the light of the social model of health and with reference to health inequalities, obesogenic environments, life course theory and frameworks of translational research.

Financial competing interests

Three members of the project management group (Sacher, Chadwick and Radley) are employees of MEND Central Ltd. Sacher and Chadwick are also shareholders. They therefore have an interest in MEND Central’s success. Results from the project might support, threaten or have no effect on MEND Central’s success.

The MEND programme was aimed at 7- to 13-year-olds, based on research undertaken at UCL Institute of Child Health. Under a contractual relationship, UCL Institute of Child Health receives a small proportion of the revenues generated from this programme. These sums have been less than £10,000 p.a. for the last 3 years. UCL Institute of Child Health uses these funds to support nutrition research and none of the project management group has access to these funds, nor works in the department where they are held.

Other competing interests

Researchers in the project management group may predict particular results which may influence discussion and interpretation, or may have known intellectual positions on which their reputations may rest. In addition, some results may be easier to publish than others. These competing interests are common in health research, though are not often articulated.

To date

Discussions to manage competing interests took place prior to submission of the grant application in April 2010 and again after reviewers’ comments on that proposal had been received (May 2010). A number of measures were proposed to manage competing interests in the grant application and these are the basis of the proposals in this paper. In addition, in response to reviewers’ comments, Sacher and Chadwick shifted from being project co-applicants to project partners. This move was accepted by the funding body in June 2010. Sacher and Chadwick remain on the project management group as project partners. In addition, a third member of MEND Central’s staff, Duncan Radley (MEND Central Research Manager) has also joined the project management group to directly support the involvement of MEND Central in the project.

To assess whether further action was required, CL discussed the issues with the following:

• Dr Peter Davison, Director of the Health Technology Assessment programme, NIHR National Evaluations, Trials and Studies Coordinating Centre.

• Dr Vic Larcher, Consultant Paediatrician and Ethicist, Great Ormond Street Hospital NHS Trust.

• Professor Sir Michael Rawlins, Chairman of National Institute for Health and Clinical Excellence (Helen Roberts also attended this meeting).

Two common themes were transparency and proportionality (i.e. taking steps in proportion to the magnitude of the competing interest and the risk of this affecting the research and being open about both what competing interests might exist and what these steps are).

Going forward

In our grant application we proposed a range of measures to address competing interests. Having received approval of these measures from the PAG, we now intend to follow these measures. The text below has minor adaptations of language but not meaning from that submitted in our grant application (for example, we originally referred to PAG as a Study Steering Group in our application).

We propose a Programme Advisory Group (PAG), to be chaired by an independent academic or commissioner. We propose that the composition of members (including the chair) will be two commissioners, two academics, two service users (parents) and the principal investigator of this application. Co-investigators and project partners will be invited to be observers. The PAG will oversee the conduct of the study and will be asked particularly to ensure that processes to address potential competing interests are adhered to. These include:

Project partners from MEND will not take part in data analysis (quantitative or qualitative), though they will be involved in planning the analysis and in discussions relating to interpretation.

The results of the data analysis (and if they request it, the primary data) will be presented to the PAG prior to report or paper writing being commenced.

The reports of the results will be scrutinised by the PAG and checked against the presentation of the results made earlier.

We have further developed the protocol to mitigate risks related to competing interests.

Qualitative component

Risks of selection bias have been mitigated by separating the tasks carried out by researchers in sample selection from the tasks of MEND Central, who supply information only. Selection of both commissioners and families who participated in MEND is made by Institute of Child Health researchers on the project management group according to pre-specified criteria which aim to achieve maximum variation in the samples. MEND supply contact information on the selected sample. Follow-up of commissioners and families approached is also made by researchers from the project management group.

Input to the topic guides for commissioners and families: the topics to be covered were set out in our application and ethics approval documents. Inevitably, in qualitative research which asks open questions, respondents may themselves raise issues not in the topic guide. All members of the project management group, including those employed by MEND, have had an opportunity to comment on the application. Final decisions on questions asked, and the ways they are asked will rest with the qualitative researchers. While we do not foresee difficulties, any which arise will be reported to the advisory group chair.

Quantitative component

Data have been transferred from MEND Central to the Institute of Child Health. We will note MEND Central’s data preparation protocols but will carry out all data preparation (checking for outliers, assessing missing data, etc.) within the Institute of Child Health research team.

Catherine Law and Helen Roberts

16th June 2011

Ola

• Her family are not encouraging healthy lifestyles. She needs to reduce the amount of time she spends doing sedentary activity.

• School dinners may be a healthier option for her (depending on what she has in her lunchbox).

• She might go to MEND, but her parents might not want to as they are overweight themselves and they might not have enough time.

Alize

• Rather a sad person and may be comfort eating.

• Probably does little exercise because of social isolation.

• Might not go to MEND because of language and transport issues. On the other hand this might be a good way to make friends.

Jon

• Lazy and following the family pattern.

• May not know about healthy lifestyles. Also, his mother is very busy.

• Might not go to MEND as he is lazy and his mother might be too busy to go with him. Can he go by himself?

Naufil

• Eats well and is privileged but doesn’t use what is on offer enough.

• Needs a confidence boost – this would help him increase his physical activity at school.

• Also needs to think about active transport.

• Probably would go to MEND because he might like the peer support and it might increase his confidence.

General themes

• Accessibility and fitting into parents’ (busy) lifestyles.

• Many group members felt parents would be unwilling to go to MEND because of the time commitment. This was more commonly expressed than children and young people being unwilling.

• Attending MEND may be seen as a stigma and a stimulus for bullying.

• There was lots about the social aspects of MEND and the positive peer group support it might provide.

Introduction

This section reports on the NCB Young Research Advisors consultation session about the PLEM project on 14 July 2012. The aim of the session was to find out how young people interpreted the PLEM results, and to establish whether their interpretations provided additional insights to be considered by the researchers.

The session was introduced by presenting a short overview of MEND 7–13 as a weight management intervention for children and their families and an overview of the methodologies used in the population-level evaluation.

The young people were split into three groups and then each group was presented with a page of findings from the qualitative and quantitative work (see Findings sheets), and a worksheet with relevant questions (see Feedback about the information sheets). Some questions were specific to each group to allow the children and young people to actively engage with the work, while two questions were given to all the groups and related to what they thought the findings might mean for MEND and whether they were useful. Each group then had 10 minutes to report back their findings and discussion to the rest of the advisors and then there was a short discussion before the session ended.

Individual group responses

Group 1 thought that:

• Families will have to devote time to MEND, and change their schedules so that it does not clash with other priorities. Families might need to arrange extra support in order to attend MEND, such as child care.

• If one family member is unemployed, he or she will have more time to look after the children. Unemployed parents might not be able to afford healthy foods.

• Religion might be a bigger priority than attending MEND.

• Some ethnic groups might eat more fatty foods and therefore do worse on MEND than other ethnic groups.

MEND could improve its intervention by:

• surveying parents of 7- to 13-year-olds about the timings of MEND sessions

• changing session times and frequency as required (sessions should not necessarily have set times)

• providing or signposting exercise classes and other support after MEND finishes

• promoting MEND within schools; however, this should be done discreetly so that children are not embarrassed if they want to go to MEND

• focusing sessions on cheaper, healthy food/budgeting/lifestyle for unemployed parents

• highlighting healthier options of cultural foods

• making the parents aware of the benefits of MEND

• having supportive parents.

These findings are useful because:

• they can now give support and extra help to the children who need it

• child obesity is a big problem

• statistics allow the intervention to have more impact.

Group 2 thought that:

• Losing weight makes you feel more confident about your appearance/body, and feeling like you are doing something about your weight may improve your self-esteem.

• Meeting other people ‘like me’ will make you feel like you are not alone in being overweight.

• Older children are more conscious of their weight. When you are younger you don’t understand why people look at you, but as you get older you are exposed to different people’s opinions which may influence what you think.

• When you start secondary school you meet new people.

• When you are younger it is easier to change and to make it a long-term change.

• Overall – you will be more successful if you want to make the changes.

MEND might work because:

• the ‘commitment’, having to physically go to something, is better than doing it alone

• it is a sustainable way for children to manage their weight because it is a lifestyle change

• it presents scientific evidence, which means children are more likely to attend

• support and advice is given, which means children are more likely to want to go

• children are given help to go through it, rather than simply being told what they need to do

• it brings back people’s confidence and provides mental support

• it is sociable; other people also attend.

These findings are useful because:

• It is important to get the children’s views; there is a need to find out what users think in order to make MEND better, as this will lead to an increase in the number of users.

• There is currently no evaluation, and therefore no way of knowing whether or not it is effective.

Group 3 thought that:

• It is important that both parents and children attend MEND as this provides support and motivation, and parents will remember things from the sessions. Parents are also in charge of shopping and so should attend MEND sessions.

• However, perhaps it is not essential that parents attend the sessions with their children, as children should take responsibility for themselves.

• The number of fast food outlets in a child’s neighbourhood did not have any effect on the reduction in BMI because fast food might be too expensive and children might not like it.

Children who attended more sessions might do better because:

• they get to do more exercise

• they are more motivated

• they set more aims and goals

• they have diet plans.

MEND could improve its intervention by:

• offering incentives (e.g. gym passes, free t-shirts and pens, grocery vouchers)

• doing more advertising, e.g. within schools

• providing more statistics as evidence of how well it works

• linking with Change4Life

• educating people about what is contained in food

• promoting the ‘traffic light’ system (of food labelling) because it is good

• providing alternative exercise to sport, as some people do not like sports

• promoting sports

• giving advice on supplements.

Summary of young research advisors’ thoughts

• Family: the role of the child’s family is important as it will affect attendance at MEND and the success of MEND; however, as children get older they should rely less on their families and take more responsibility for their own weight.

• Self-esteem: MEND may improve self-esteem, not just because of improved appearance but because of the process of taking control over one‘s weight and having the social support of other overweight children.

• Age: as children get older they are more aware of how people view them and are more exposed to people’s opinions. The younger children are, the easier it is to make changes.

• The young people thought that weight management interventions could be optimised by:

  • providing flexible access

  • providing post-intervention support

  • being accessed discreetly within schools

  • being practical and conscious of food costs

  • being tailored to different cultures

  • offering incentives for attendance and weight loss

  • linking with public health campaigns or big businesses

  • using statistics in their promotion

  • providing a range of physical activities

  • educating people on what is contained in food.

• The young people felt that this research is valuable because:

  • obesity is a serious and large problem

  • the results can be used to improve weight management interventions

  • children’s views were taken into account in the qualitative component

  • there is currently no other population-level evaluation of a weight-loss intervention.

(PDF download)

Stage 1: maximum variation purposive sample

We were interested in five dimensions of variation. This was analogous to drawing a sample with six binary dimensions, leaving 25 (32) possible profiles of participants. These profiles were identified in the data using Stata version 11. Two commissioners were selected from each profile. Where there were multiple participants per profile, the commissioners within that profile were ranked by multiple deprivation score and selected according to the highest (for those in the most deprived profiles) or lowest (for the least deprived profiles) deprivation scores.

We aimed to recruit 30 commissioners and anticipated that 50% might opt out of the study when we contacted them. We oversampled to allow for this attrition. The first pass over the data left a sample too small for our criterion of at least 60 respondents; a second pass using the same method produced an overall sample of 68 participants.

Detailed breakdown of dimensions of variation

Stage 2: critical-case sampling

A small number of organisations (n = 13) commissioned between 41 and 119 programmes. The project team believed that these organisations might have a different perspective on commissioning, and these were included in the sampling strategy.

Family structure, ethnicity and housing tenure

MEND programme managers collate questionnaires from parents which report family structure, participant ethnicity and housing tenure information. There were considerable numbers of missing data on all the MEND programme-reported information (42% of families did not report family structure status, 41% did not report housing tenure and 38% did not report ethnicity). Although participants with missing data may be distinct from those with complete data in interesting ways, we took a decision to exclude families with missing data. Missingness on one or more variables would have prevented us from selecting families across all dimensions of variation, and as one of our aims in the study was to look at inequalities, we felt that this was the better of the choices open to us. The sample size of our data set with complete data on all variables was 825.

The question on family structure relates to single-parent families in comparison with couples. Family structure was sampled so as to try to pick up 50% of single parents and 50% of couples across the whole sample. This was disproportionate to the overall MEND data set, where single parents made up only 36% (see Table 26).

Housing tenure was sampled to pick up 50% of owner-occupied tenures and 50% of council and private rental tenures, a situation which broadly reflects the overall composition of the MEND participants.

Our sample was drawn to comprise 50% from black and Asian groups and 50% from white groups. This decision was made because young people (aged 0–15 years) from these groups are strongly represented in the south-east (Asian and black) and south-west (Asian) sample areas (Table 38). Overall, MEND participants in our subsample were predominantly white (69%), but Asian (21%) and black (5%) groups made up significant minorities, whereas other groups (including mixed-ethnicity families) were only minimally represented (see Table 38).

Multiple deprivation

Multiple deprivation at the LSOA level (mean population size across England approximately 1500 residents) was described using the English Index of Multiple Deprivation 2007. 41 We linked this to our subsample using residential postcode. This variable was complete for all 825 participants. To maximise variation in multiple deprivation, we sampled those who lived in areas which were in the 30% of most deprived areas in England (top tertile) and the 30% least deprived areas (bottom tertile).

Completion status

The project team aimed to identify four categories of completion status. The first group (‘referrals’) includes participants who were referred to MEND, registered with the company but did not subsequently attend the programme (had no BMI data measured). This group made up almost one-third (28.5%) of MEND’s records. Three other groups were identified based on the percentage of sessions that they attended, corresponding loosely to non-completers (< 25%), partial completers (25–75%) and completers (> 75%). Overall, 33.6% of MEND participants completed the programme (i.e. attended more than 75% of sessions), 16.8% partially completed the programme and 21.2% only attended 25% of the sessions. MEND records were used to purposively select to approximately reflect this balance in our sample, so that our sample of 68 included proportionate numbers of those whose MEND records suggested they had not confirmed, or had attended < 25%, 25–75% or > 75% of all sessions.