Results (primary and secondary outcomes)

Participant numbers through the HeLP study

Figure 2 shows the flow diagram of the progress of schools and children through the phases of the trial. The target of recruiting 32 schools, to ensure that a minimum of 28 schools completed the trial, was met, with all 32 completing the trial. The number of children in each school was greater than we had anticipated before the recruitment of schools started (the prior assumption was a mean of 35 Year 5 children per school). Therefore, the total number of eligible children was greater than expected. All 1371 children in the 32 participating schools were eligible for the trial: 34 were opted out by their parent/carer and 13 left the school before baseline data collection (and randomisation). All of the remaining 1324 children, irrespective of whether or not they had (all) measurements collected, were included in the trial, and so the minimum requirement of 952 children recruited was far exceeded.

FIGURE 2 Trial profile. (Figure not reproduced here; see the original article.)

Up to four visits were made to each school in an attempt to obtain measures for children who had been absent on the original day of data collection, with only a small number of children absent at all subsequent visits. As described in Chapter 2, one class per school was randomly sampled to complete the physical activity data collection. At 24 months, there was a slightly greater loss to follow-up of children in the intervention group than in the control group (6.1% vs. 3.7%). The number of children included in the primary analysis of BMI SDS at 24 months was 1244, 63% greater than the required 762 children from the a priori sample size calculation.

Figure 2 shows the trial profile. ‘N’ refers to the number of schools (clusters) and ‘n’ refers to the number of children (individual participants). Two schools that had been allocated to cohort 2 withdrew while they were waiting to commence the trial and so these were subsequently replaced with two of the four schools on the waiting list before cohort 2 commenced. All schools that started the trial remained in it, and so all of the randomised clusters are present at baseline and at each follow-up point. The percentage given in brackets for the proportion of children with data at both baseline and follow-up is calculated from the total number of recruited children in the schools at baseline. Not all children with a follow-up measure necessarily had a corresponding baseline measure (or vice versa) owing to different children being absent on the day of the main and additional assessments for each of the time points and/or owing to children leaving or moving schools. In all of the analyses, children were analysed in the group (intervention or control) to which the school they were enrolled in at baseline was randomised.

Completeness of outcome measures through the HeLP study

Figure 3 shows the completeness of the outcome measures at each of the time points. A small number of children at each time point did not assent to one or more of the anthropometric measures. In summary, between 93% and 99% of children provided data on BMI, waist circumference, percentage body fat and FIQ. The proportion of children who had usable physical activity data files was 96% at baseline and > 88% at 18 months; similarly, the proportion of children with valid physical activity data was 94% and 84% at baseline and at 18 months, respectively, following the application of the minimum wear requirements (3 weekdays and 1 weekend day, each with a minimum of 10 hours of wear time per day). There was no evidence of differences between allocated groups in terms of the completeness of outcome measures throughout the trial, with the exception of the small difference at 24 months for BMI SDS.

FIGURE 3 Completeness of measures across time points. PA, physical activity. (Figure not reproduced here; see the original article.)

All anthropometric measures were taken blind to group allocation. A further assessment of the success of allocation concealment was made at the 24-month data collection point, when there was a mixt of children from intervention and control primary schools in each secondary school, with the independent (blind) assessors requested to indicate whether or not the child had revealed their allocated group; no child had reported their allocated group.

Figure 3 shows the completeness of the measures across the time points. ‘N’ refers to the number of schools (clusters) and ‘n’ refers to the number of children (individual participants). The percentage given in brackets for the proportion of children with data at baseline and follow-up is of the total number of recruited children in the schools at baseline. Not all children with a follow-up measure necessarily had a corresponding baseline measure (or vice versa) owing to different children being absent on the day of the main and additional assessments for each of the time points and/or owing to children leaving or moving schools. In all of the analyses, children were analysed in the group (intervention or control) to which the school they were enrolled in at baseline was randomised. There were no differences in missing anthropometric data by allocated group at each time point.

Baseline characteristics

School-level characteristics

Table 5 shows how the HeLP study schools compare with schools in Devon and England in terms of the percentage of children eligible for free school meals, the school size, the percentage of children achieving level 4 at Key Stage 2 and the proportion of pupils with English as an additional language or who were non-white British.

Owing to the inclusion criteria for schools for this study (at least one Year 5 class of > 20 pupils and half of schools having ≥ 19% pupils eligible for free school meals), HeLP schools are larger and more deprived than average primary schools in Devon. Although broadly similar to other primary schools in Devon, HeLP schools have a considerably lower proportion of pupils with English as an additional language than do primary schools in England in general.

Table 6 shows the school level-baseline characteristics. In the intervention group; there were an equal number of schools with single and multiple Year 5 classes, while in the control group, 9 out of the 16 schools had single Year 5 classes. In both groups, ≥ 19% of the children were eligible for free school meals in 7 out of 16 schools. The median school-level IMD scores were comparable between the allocated groups.

Individual-level characteristics

Tables 7 and 8 show the comparison of individual baseline characteristics by allocated group. The two allocated groups were similar in terms of the child-level baseline characteristics, including physical activity and food intake scores. The baseline anthropometric measurements, although similar and with considerable overlap, were greater on average in the intervention group, with 12.1% and 14.6% of the intervention group classified as overweight and obese, respectively, compared with 10.7% and 12.6% of the control group.

Primary intention-to-treat analysis of body mass index standard deviation score at 24 months

The unadjusted mean BMI SDS at 24 months was slightly higher in the intervention group, at 0.35, than in the control group, at 0.22, with SDs of 1.25 and 1.22, respectively (Table 9). The effect of the intervention was estimated allowing for clustering within schools, modelled as a school-centred random effect, and other prespecified variables. The results showed no evidence of an intervention effect, with an estimated between-group mean difference (intervention minus control) of –0.02 (95% CI –0.09 to 0.05) (see Table 9), and an estimated ICC of 0.014 (95% CI 0.003 to 0.069). Without adjustment for the baseline BMI scores and gender of each child, cohort and the stratification factors, but after allowing for the school-level clustering, the mean BMI SDS score was 0.11 (95% CI –0.11 to 0.33) greater in the intervention group than in the control group, but this difference was not statistically significant (see Table 9).

Model diagnostics for the primary analysis of body mass index standard deviation score at 24 months

Following the fitting of the full random-effects model, plots of the residuals and best linear unbiased predictors for the random school intercept revealed no systematic deviation from the assumed distribution of the observations, required for linear mixed modelling. Inspection of the observed versus fitted values indicated a good fit within the random variation of the data. Inspection of the dfbeta statistics of individuals and schools did not reveal any extreme of influential values or instability in the estimation of the coefficients, including for the intervention effect.

Secondary analysis of body mass index standard deviation score at 24 months

All of the secondary analyses that were undertaken produced results that were consistent with the primary analysis. An alternative computational approach using generalised estimating equations yielded an adjusted between-group mean difference of –0.02 (95% CI –0.08 to 0.05), allowing for clustering around schools and adjusting for the same variables as in the random-effects model in the primary analysis (see Table 9). The results from the two approaches are, thus, consistent with each other, with no evidence of an intervention effect.

The sensitivity analysis of the fully adjusted random-effects model based on the worst-case scenario for missing observations at 24 months estimated an increase in mean BMI SDS of 0.07 (95% CI –0.02 to 0.16) for the intervention group relative to the control group (see Table 9). Based on the best-case scenario for the same missing observations, the direction and size of the effect of the intervention changed to –0.02 (95% CI –0.09 to 0.05). Therefore, neither scenario provided any tentative evidence of a significant or meaningful intervention effect (see Table 9).

Subgroup analyses of body mass index standard deviation score at 24 months

The results for the prespecified subgroups, defined by gender, school size, free school meal category and cohort, are presented in Table 10. The estimated intervention effects were small and in the same direction within subgroups for gender, cohort and school size, with little or no effect on the BMI SDS at 24 months and no evidence of a significant difference from the null effect. Interaction terms between each factor and the allocated group were added to the random-effects model, and the absence of modifying effects by each factor was confirmed by results from the test of the interaction term. All p-values were > 0.05. Similarly, there was no evidence of a modifying effect on the intervention by the BMI SDS at baseline or from the individual child-level IMD scores. The point estimate for the effect of the intervention in either level of the free school meal factor was not significantly different from the null, but, although this effect was negative in those schools with < 19% of their children eligible for free school meals, there was a small, non-significant, increase in the BMI SDS in those children from schools in which ≥ 19% of pupils were eligible for free school meals.

In summary, there was no evidence of a significant effect of the intervention in any of the prespecified subgroups. As noted in Chapter 2, these subgroup analyses, although prespecified, are considered to be exploratory and the trial was not designed with the aim of being able to detect differences between subgroups.

Longitudinal analysis of body mass index standard deviation score

The analysis was based on fitting a longitudinal model with observations at fixed time points (baseline, and 18 months and 24 months post baseline). The complexity of the between-time-point covariance for each child was reduced to an autoregressive pattern of order one. As the outcome of this longitudinal model comprised all BMI SDS measures, including those at baseline, the effect of the intervention was determined from the interaction between the allocated group and the time point.

According to this model, there was no significant difference between the two groups at baseline (intervention BMI SDS greater than controls by 0.117; p = 0.170). The BMI SDS increased significantly in both groups at each subsequent time point relative to baseline, by an average of 0.04 (95% CI 0.01 to 0.06) at 18 months and by 0.05 (95% CI 0.01 to 0.09) at 24 months, although, relative to the control group, the increase in the intervention group was attenuated, not significantly, by –0.03 (95% CI –0.08 to 0.01) at 18 months and by –0.02 (95% CI –0.07 to 0.04) at 24 months.

Figure 4 shows the predicted marginal BMI SDS by allocated group across the time points of the trial.

FIGURE 4 Predicted marginal BMI SDS with 95% CIs in the intervention group (black with dashed connecting lines) and control group (green with solid connecting lines) across time points. According to the longitudinal model, allowing for hierarchical clustering by child within each school, modelling the within-child covariance between fixed time points as an autoregressive pattern of order one. Reproduced from Lloyd et al.32 This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) license, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: http://creativecommons.org/licenses/by/4.0/. (Figure not reproduced here; see the original article.)

Primary analyses of secondary outcomes at 24 months (anthropometric measures only)

Table 11 shows the summary data of the remaining anthropometric measures at 24 months post baseline. Among the children in the intervention group, slightly more were classified as being overweight or obese, according to the Public Health England definition of categorisation of BMI scores at 24 months, relative to the control group (30.8% vs. 26.6%). However, analysing the weight status categories under the assumptions of an ordinal logistic model with school-level random effects, there was no evidence of an intervention effect in the fully adjusted model (p = 0.472). The odds of being in the higher weight categories were greater among the children in the intervention group than among those in the control group, although this change in odds was not statistically significant [odds ratio (OR) 1.15, 95% CI 0.79 to 1.67]. The intervention effect size was similar in both the unadjusted and adjusted models. Relaxing the assumption of proportional odds, the estimated intervention effect size from the generalised ordinal model, with a robust error variance clustered on schools, was comparable with that from the random-effects ordinal logistic model (results not shown). According to the generalised ordinal model (i.e. the model without assuming proportional odds across the weight status categories), the odds of being classified as overweight or obese relative to healthy or underweight were 1.14 (95% CI 0.77 to 1.68) in the intervention group relative to the control group. The odds were 1.17 (95% CI 0.74 to 1.84) for the risk of being obese relative to the lower weight status categories in the intervention group relative to the control group. The test of proportional odds under the Peterson–Harrell parameterisation indicated that there was no evidence of a change in the odds between binary comparisons of levels of weight status category (p = 0.503), indicating that the assumption of proportional odds for intervention effect had not been violated.

Consistent with the primary outcome measure, the average anthropometric measurements were generally slightly larger for the children in the intervention group at 24 months, as they were at baseline, relative to those in the control group (see Table 11). An exception was found in the standardised body fat measurements before the exclusion of the extreme values (implausible absolute SD scores of ≥ 5). The unadjusted mean differences (intervention minus control) were positive for all anthropometric outcomes except the standardised body fat score that included extreme values. However, there was no evidence of an intervention effect from the unadjusted random-effects model, allowing for clustering on schools. After adjusting for the stratification variables, cohort, gender and the corresponding baseline measurements, all mean differences were negative, indicating lower, albeit non-significant, measurements in the intervention group at 24 months than in the control group.

Primary analyses of outcomes at 18 months

Tables 12–14 show the summarised data of the primary and secondary outcomes at 18 months post baseline. Overall, fewer children were classified as overweight or obese than at 24 months, but more children in the intervention group than in the control group were already overweight or obese at 18 months (28.9% vs. 25.4%). Although not statistically significant, this difference was reflected in both the adjusted and unadjusted analyses when the weight categories were modelled as an ordinal outcome. No violation of the proportional odds assumption was detected, according to the test of the Peterson–Harrell parameters from the generalised ordinal model (results not shown).

The anthropometric outcomes at 18 months displayed a similar pattern to those at 24 months, with larger mean values for children in the intervention group than for children in the control group, and positive mean between-group differences (intervention minus control) from the unadjusted analyses, except for the standardised measurement of percentage body fat with extreme values included (extreme values defined as absolute standard scores of ≥ 5) (see Table 12). After adjustment for all of the prespecified variables, the mean between-group differences in anthropometric measurements were all negative, indicating lower values in the intervention group at 18 months than in the control group. However, there was no evidence of a significant intervention effect for any of the anthropometric measures at 18 months.

There was no significant effect of the intervention on any of the physical activity outcomes (see Table 13). Measures of physical activity, average weekly volume and average time per day in each intensity category were all fractionally larger in the intervention group from a direct comparison of the mean values of the two groups and from both the unadjusted and adjusted analyses. Conversely, the accelerometry data classified as sedentary behaviour (between 6 a.m. and 10 p.m.) were, on average, lower in the intervention group. However, none of the observed differences was statistically significant.

The mean scores of the FIQ outcomes averaged over the week were higher in the intervention group for healthy snacks and positive food markers and lower for the energy-dense snacks and negative food markers (see Table 14). These between-group differences were statistically significant in the unadjusted analysis for the healthy snacks and positive food markers. However, once the analysis was adjusted for baseline score, stratification variables and other potential confounders, these differences were no longer statistically significant, while the between-group differences for energy-dense snacks and negative food marker scores were statistically significant in the adjusted analyses, with the strongest evidence of an intervention effect observed for energy-dense snacks (p = 0.017). According to the adjusted analysis, the mean average weekly scores for the intervention group were 0.37 (95% CI –0.66 to –0.07) lower than for the control group for the energy-dense snacks score and 0.47 (95% CI –0.91 to –0.02) lower for the negative food markers score.

The pattern of differences observed for the weekly averaged FIQ outcomes was repeated for the weekday scores. There was weak evidence of between-group differences from the unadjusted analyses for the energy-dense snacks and negative food marker scores, as well as for the healthy snacks score. After adjustment for the prespecified variables there was evidence of statistically significant differences for the energy-dense snacks (p = 0.013) and negative food markers (p = 0.020). The intervention effect sizes were of a similar magnitude and direction for the two unhealthy food scores, with the mean average weekday energy-dense snacks score being 0.47 (95% CI –0.84 to –0.11) lower in the intervention group than in the control group, and 0.64 (95% CI –1.17 to –0.11) lower for the negative food markers score.

At the weekend, the pattern of the differences between the two groups was the same, with higher mean scores for healthy snacks and positive food markers and lower mean scores for energy-dense snacks and negative food markers for the intervention group than for the control group (see Table 14). The between-group differences for the healthy snacks and positive food markers scores were statistically significant in the unadjusted analyses (p = 0.035 and 0.030, respectively); however, after adjustment for the prespecified variables, neither between-difference was statistically significant (p = 0.086 and 0.145, respectively).

Model diagnostics for the primary analysis of the secondary outcomes

Following the fitting of the full random-effects models, plots of the residuals and best linear unbiased predictors for the random school intercept revealed no systematic deviation from the assumed distribution of the observations, required for linear mixed modelling, for all the secondary outcomes, with the exception of the FIQ scores when the weekday and weekend scores were modelled separately. Given these model violations, the weekday and weekend FIQ scores were reanalysed as ordinal outcomes, after aggregating the sparser larger scores into a single category to leave five levels for each score. The results from the ordinal models are shown in Table 15. In summary, from the adjusted analyses, there remained evidence of statistically significant intervention effects for the weekday energy-dense snacks (OR 0.68, 95% CI 0.51 to 0.90) and negative food markers scores (OR 0.69, 95% CI 0.52 to 0.92), indicating that children in the intervention group were more likely to have lower scores than children in the control group for either of the weekday negative food behaviour categories. The OR for the intervention effect for the weekend healthy snacks score was 1.27 (95% CI 1.00 to 1.61), indicating that children in the intervention group may be more likely than children in the control group to have higher healthy snacks scores at the weekend, although the 95% CI does include the null value of one.

As with the analyses of the weight status categories, the assumption of proportional odds was tested through fitting of the generalised ordinal model, with robust error variances, clustered on schools. While the assumption of proportional odds was violated for some of the adjusted factors, the intervention effects from these models were similar to the presented results from the random-effects models, differing in their coefficients by < ± 0.05. The test of the Peterson–Harrell parameters indicated marginally weak evidence of a violation of the proportional odds assumption for the intervention effect in the model for weekday energy-dense snacks (p = 0.049). This was observed to be due to a different, larger effect in the sparser high energy-dense snack score categories.

Intraclass correlation coefficients

Table 16 shows the ICC and 95% CIs for each of the outcomes from the random-effects models. The estimated ICC for the primary outcome was 0.014 (95% CI 0.003 to 0.069), slightly lower than the estimate of 0.02 used in the sample size calculation. The estimated ICC for weight status when dichotomised (overweight and obese vs. healthy and underweight) was extremely small (estimated as 0.000). This was due to the estimated between-cluster variance for this outcome being particularly small relative to the variance of the standard logistic distribution, assumed to be that of the latent continuous distribution that determines the weight of status of individuals. The wide CI for this ICC is an artefact of the estimated parameter being so close to the boundary of the parameter space, and the logit function,56 which yielded a lower bound very close to zero and an upper bound at one. The estimated ICCs for waist circumference and waist circumference SDS were higher than for the other anthropometric measures.

Complier average causal effect analysis

The original analysis plan included a complier average causal effect (CACE) analysis to estimate the CACE of treatment, as a potentially unbiased estimate of receiving HeLP, with non-compliers defined as children who did not receive at least four sessions of drama and the one-to-one goal-setting session. The TSC subsequently deemed the planned analysis unnecessary, given the small number of children who ‘switched’ between allocated treatment groups: there were only 24 out of 675 children allocated to the intervention group who were categorised as non-compliers (3.6%) and 4 of these 24 did not provide primary outcome data.