Design and Procedure
Ethical approval was obtained from the School of Education Research Ethics Committee at Queens University Belfast (reference numbers 100314, 210216 and 111217). Data were gathered in three waves (2014, 2016 and 2018) as part of the Wellbeing in Schools (WiSe) study, which was a longitudinal survey of the health and wellbeing of 11-16-year-old school children recruited from a representative sample of ninety-nine post-primary schools in Northern Ireland (NI). A list of all post-primary schools in NI was obtained from the Department of Education NI (DENI) website. A letter detailing the aims, objectives and procedures of the study, along with an invitation to participate, was posted to the head teacher in all schools. At time one, of the 203 schools contacted, 104 replied with 99 agreeing to participate. At subsequent time points, of the 99 schools who participated at time one, 89 agreed their continued participation at time two, and 79 continued to participate at time three.
At time one, each school randomly selected one class from year 8 to participate. Informed consent was obtained in writing from parents and pupils. At times two and three, all pupils were required to provide full written consent prior to participation and parents were provided with an opportunity of opt-out consent for their child’s continued participation.
The resultant sample at time one (11-12 years) comprised N=1912 of whom 49.5% were male and 50.5% female. At time two (aged 13-14 years) there was 85.77% continued participation, this resulted in a total of 1640 adolescents who responded at time two, of whom 51% were male and 49% female. At time three (aged 15-16 years) at total of 1566 adolescents continued to participate of whom 49.5% were male and 50.5% female. Data were collected from pupils in clusters (each comprising one class) within each school. Data collection took place in the schools with responses entered onto either a study iPad or a school computer onto the LimeSurvey on-line platform. The researcher was present during data collection to address any issues/questions arising.
Measures
Kidscreen10
Kidscreen10 [32] is a ten-item measure of HRQoL designed for use with children aged between 8 and 18 years. Kidscreen10 has been used previously in cross-sectional studies of HRQoL and diet. Recent systematic review [6] located four previous studies that employed the Kidscreen10. Kidscreen10 has also been used in longitudinal research into HRQoL in young people [17, 19], and HRQoL along with breakfast consumption [28], adherence to the Mediterranean diet [29] and physical fitness [33]. Together this implies the suitability of Kidscreen10 for research of this type.
The tool asked: ‘thinking about the last week, have you’: ‘felt fit and well’; ‘felt full of energy’; ‘got on well at school’; ‘been able to pay attention’ and for which responses were on a five-point scale: ‘not at all; slightly; moderately; very; extremely. The tool also asked: ‘thinking about the last week, have you’ ‘‘felt sad’; ‘felt lonely’; ‘been able to do the things you want to do in free time’; ‘had enough time for yourself’; ‘had fun with friends’; ‘parents treated you fairly’ and for which responses were on a five-point scale: never; seldom; quite often; very often; always. The item responses for Kidscreen10 scale were coded so that higher values indicate better wellbeing. Reliability (Cronbach’ alpha) for Kidscreen10 was good at time one (α=0.79) and time two (α=0.84), but only moderate at time three (α=0.64).
Food Frequency Questionnaire (FFQ)
Food choices were assessed using a 17-item FFQ previously employed in the Young Persons Behaviour and Attitudes Survey (YPBAS) [34]. Responses were on a five-point scale: more than once a day; once a day; most days; once or twice a week; less often or never. In the analyses a higher value represents a more frequent consumption pattern. Items related to the frequency of consumption of: sweets/chocolate/biscuits; buns/cakes/pastries; fizzy/sugary drinks; diet drinks; crisps; chips/fried potatoes; boiled/baked potatoes; fried foods (sausage eggs, bacon); meat products; meat/meat dishes’; fish (not fried); beans/pulses; fruit; vegetables/salads (except potatoes); bread; rice/pasta; milk (to drink; on cereal; puddings) cheese/yoghurt. Reliability for FFQ was good at all three time points, time one (α=0.760), time two (α=0.770) and time three (α=0.773).
Data Analysis
The original full dataset consisted of N=2143. Data which were missing (n=79) or stated as ‘other’ than female or male (n=3) were excluded from the analysis. An additional 481 cases had data missing on all variables used in the analysis and were also excluded. The analysis was conducted on the remaining sample (N=1912). The structure of all available data, on each of the three occasions, is shown in Supplementary File 1. The table also provides details relating to the model estimated means and variances for the items used in the analysis.
Missing data across points in time were assumed to be missing at random since the reason for the missingness was a result of the study design [35-36]. A planned missing strategy was employed enable analysis at three points in time. On time-point one (2014, aged 11-12 years) food frequency questionnaire data were available for only 285 individuals. Individuals with structured missingness on occasion one and were more likely to be missing completely at random (MCAR) [35] and as such were included within the full analysis, based on data from all occasions. The Kidscreen10 measure at time-point one was available for approximately 50% (n=1089) of the full data set. Since full data for the Kidscreen10 and the food frequency questionnaire were available for the two later time-points (2016, aged 13-14 years; 2018, aged 15-16 years), data available from the first occasion were included within the overall analysis. This was achieved using a missing at random strategy for handling missing data [37] where full information maximum likelihood (FIML) was used to estimate the model.
The dimensional structure of responses to the 17 FFQ items was determined by means of exploratory factor analysis using a geomin (oblique) solution with chi-square testing of model fit. Because responses were on a five-point Likert scale, they were treated as ordinal. Results indicated a five-factor solution: 1 ‘Junk Food’; 2 ‘Meat’; 3 ‘Protein’; 4 ‘Fruit and Vegetables’; 5 ‘Bread/Dairy’ [see 20]. Factor loadings that were statistically significant at the 0.05 level in the exploratory factor model were then included within a five-factor solution modelled within a confirmatory framework. This two-step approach was taken in an attempt to maximise the factorial invariance of the measures (keeping the measure the same across occasions and groups).
Given we are using a cross-lagged model, variables are required to be restricted to be equal across certain parameters otherwise results can be biased. Cross-wave equality constraints were therefore placed on structural coefficients [38-39]. All modelling then assumed that the measures were invariant across both groups (females and males) unless otherwise stated. Correlated residuals across time periods were introduced into models based on the fit statistics and the modification indices. For the purpose of factor mean comparisons female scores on the latent variables were set at zero.
Single-Factor Model (Kidscreen10)
A one-factor model was constructed for a multi-group (females and males) factor model in order to describe the Kidscreen10. This resulted in a replication of the same factor within each group, across the three occasions, with correlated residuals between time periods. Factor loadings were constrained to be equivalent for females and males across the three occasions. Where intercepts differed between females and males these were relaxed following an examination of the modification indices.
Cross-Lagged Panel Model (CLPM)
The fit model for the Kidscreen10 was then combined in turn with each of the five food choice factors in a CLPM [40-41] (Figure 1). In obtaining a fit statistic for these new models the results from the Kidscreen10 were estimated as for the single factor model (step 1 above). No adjustments were made to the Kidscreen10 result when placed alongside the other measures in the CLPM. Model adjustments were based on the modification indices in the context of substantive judgements.
The food choice factors and the Kidscreen10 measure, on the three occasions, were auto-regressively related across time (Figure 1). This was to establish the temporal stability of the rank order of individuals through time [42]. In this model the parameter estimates for females and males were constrained to be equivalent.
In a series of five separate analyses, each of the food factors on each of the three occasions were regressed onto the measure on the previous occasions. These regression effects were constrained to be equal, as a test of stability.
Given this is a cross lagged analysis the time points are the variables. The independent variable is baseline (timepoint 1) and the dependant variables are the subsequent timepoints two and three. A series of cross-lagged regression coefficients were introduced between the food and Kidscreen10 factors, for example, each food factor at time two was regressed onto the Kidscreen10 factor at time one. Likewise, the Kidscreen10 factor at time two was regressed onto the food factor at time one. This pattern of going forward in time was then repeated on the third occasion. Parallel restrictions were placed on the cross-lagged regression coefficients where the same measure was present going forward in time, as a test of stability [43].
This pattern of cross-lagged regression coefficients was maintained for both females and males, in an invariant manner. All measures were corrected for measurement error and again these properties were constrained to be equal for females and males where this was appropriate in terms of model fit. The residual errors, in the structural part of the model, moving through time were left unequal.
Maximum likelihood estimation was used with robust correction for the standard errors and fit statistics using the TYPE=COMPLEX strategy in Mplus V8.8 [44].
Insert figure 1 here