Study design and participants
The present study analysed data collected as part of the Otago School Students Lifestyle Survey Two (OSSLS2), a cross-sectional study conducted from February to June 2011 in the Otago province of New Zealand. Data were collected on food choice, psychosocial correlates of diet, eating behaviours, and body composition. Secondary schools in the Otago area were invited to take part in the study. Randomly selected classes from school years 11 to 13 (age 14 to 18 years) from each school were invited to take part, with one class from each school year in smaller schools and up to four classes in larger schools invited. In the week before the projected data collection invited students were given information sheets and consent forms for themselves and their parents. Students were required to sign a consent form in order to participate, while parents were only required to sign a form if they wished their child to opt-out of the study. This study was approved by the University of Otago Human Ethics Committee.
The survey was administered online during class time in the schools’ computer rooms. Students supplied information on their date of birth, age, sex, ethnicity and residential address. Ethnicity was categorised into three groups, in accordance with other national New Zealand surveys : ‘Māori’; ‘Pacific Island’; or ‘New Zealand European or Other’ (NZEO). The NZEO group comprised those who identified as New Zealand European, plus other groups where numbers are too small to allow separate analyses e.g. those who identified as Indian, Asian, European. Home residential address was used to assign a neighbourhood area-based New Zealand Deprivation Index (NZDep06) score for each participant. NZDep06 combines nine variables from the 2006 New Zealand census which reflect eight dimensions of deprivation, including income, owning a house, access to a car . NZDep06 ranges from 1 (least deprived) to 10 (most deprived). This variable was collapsed into 5 categories. NZDep06 was only used to describe characteristics of the participants, as this information was not available for all participants. School decile is determined by the deprivation level, as measured by NZDep06, of students attending the school, with the lowest decile rating reflecting the 10% of schools nationwide with students mostly from high deprivation areas, and the highest decile rating reflecting the upper 10% of schools with students from mostly low deprivation areas. School decile was divided into ‘Middle’ (Deciles 5 to 8) and ‘High’ (Deciles 9 and 10).
Food consumption was assessed, as part of the online survey, using a food frequency questionnaire (FFQ) adapted from the Health Behaviour in School Children study (HBSC) . This questionnaire was pretested in a sample of adolescents from Otago before use in OSSLS2 and showed good repeatability and relative validity . Participants were asked ‘On average, how many times a week do you usually eat or drink any of the following foods?’. The food items assessed were ‘fruit’, ‘vegetables’, ‘sweets’ (e.g. jelly beans, marshmallows, etc.), ‘chocolate confectionery’ (e.g. chocolate, Moro, Crunchie, etc.), ‘standard milk’ (≈ 3.3% total fat), ‘other milk’ (low fat, rice, soy), ‘cheese’, ‘breakfast cereals’, ‘white bread’, ‘brown/wholegrain bread’, ‘potato crisps’, ‘hot chips/fries’, ‘artificially-sweetened ‘soft-drink’ (e.g. Diet Coke, Coke Zero, etc.), ‘regular soft-drink’ (e.g. Coke, Pepsi, etc.), ‘energy drinks’ (e.g. Red Bull), ‘sports drink’ (e.g. Powerade), ‘fruit juice/fruit drink’ and ‘alcohol’. Regular soft-drink, energy drinks and sports drinks were collapsed into one food group for this analysis called sugar-sweetened soft-drinks.
Dieting status was examined using the HBSC question ‘At present, are you on a diet or doing something else to lose weight?’. The response options were ‘No, I am happy with my weight’; ‘No, but I should lose some weight’; ‘No, because I need to put on weight’ and ‘Yes’. This question was dichotomized into 0 (non-dieters) for those participants who answered ‘No, I am happy with my weight’ or 1 for all other responses, “dieters; or those who think they should make changes to their weight”.
Height was measured twice to the nearest millimetre, with a calibrated portable stadiometer (University of Otago, NZ). Fat mass (in kilograms) and fat-free/lean mass (in kilograms) were estimated with BIA using a calibrated segmental machine (BC-418, Tanita Corporation, Japan), which also measured weight (in kilograms). Mid-point waist circumference, between the lower costal border of the last rib and the top of the iliac crest, was measured twice to the nearest millimetre with a body composition non-elastic tape (Seca, Germany). A third measure was taken if the initial two readings differed by more than 0.5cm. All body composition measures were taken by trained research assistants.
BMI was calculated as weight (in kilograms) divided by height squared (in metres). Age-specific BMI z-scores, using the 2007 World Health Organisation method , were determined, as were thinness, normal-weight, overweight and obese categories using the 2012 International Obesity Task Force cut-points . Because of the low prevalence of obese individuals in the current sample, the overweight and obese groups were combined, hereafter referred to as overweight. Likewise, those with a low BMI-for-age (thinness) were combined with the normal-weight category. Fat mass index (FMI) was determined as fat mass (in kilograms) divided by height squared (in metres), fat-free mass index (FFMI) was calculated as lean mass (in kilograms) divided by height squared (in metres). The waist-to-height ratio was calculated as waist circumference (in centimetres) divided by height (in centimetres).
Participants were excluded if they had incomplete demographic information, as this was the first section of the survey. Participants were also excluded if their responses indicated that the survey was not completed properly. Reasons for exclusion included clicking patterns (e.g. selection of extreme left or right answer options for all questions), contradictory responses to similar questions, or giving several unrealistic answer options. Only participants who had complete data for all body composition variables were included in the final analysis. Missing data were imputed only for questions consisting of four or more sub-questions and only for those who had completed at least 75% of the sub-questions. For example the FFQ was considered a question, with each food item considered a sub-question. Missing answers to sub-questions were imputed with the worst-case scenario response. For the current analyses, the only data imputed were for the FFQ question. Of the 681 participants who completed at least 75% of the FFQ, data was imputed for a maximum of two participants for any given sub-question. Dietary pattern scores were derived using Principal Components Analysis (PCA). The PCA was run with all 16-food groups, entered as raw frequencies using varimax orthogonal rotation. Factors were retained on the basis of eigenvalues >1.0, a scree plot of the eigenvalues and interpretability of components . Variables with factor loadings ≥ 0.3 and ≤ −0.3 were considered significant when naming patterns. The patterns produced were transformed to remove skew and converted to z-scores, so factor scores had a standardised mean of zero. All participants received a score for all three patterns, with positive scores indicating high consumption of foods associated with that pattern and negative scores indicating lower consumption, for example the higher the ‘Fruits and Vegetables’ score, the higher an individual’s frequency of consumption for fruit, vegetables, cheese, and brown/multigrain bread.
Body composition values were examined as the dependent variables, while the dietary pattern scores were used as the independent variables. Gaussian family generalized estimating equations, with robust standard errors, were used to account for the complex sample in order to ensure results are representative of the population in the sampled area, with schools as the clustering units. For every unit increase in the dietary pattern score, the beta coefficient reflects the expected change in the body composition outcome. As some of the models were not linear in their associations, some outcome variables (waist circumference, waist-to-height ratio, FMI, and FFMI) were logarithmically transformed. For those models with transformed outcomes, the exponentiated beta coefficient reflects the percent change in the geometric mean of the body composition outcome for every unit increase in the dietary pattern score.
Several models were used to examine associations of interest. Model one was adjusted for age only. As sex , SES  and ethnicity  have been found to be associated with food choice, model two included these factors. Model two was repeated using different individual measures of deprivation, including NZDep06, household crowding and car ownership. As these different markers of SES made no difference to effect sizes, and school decile was the only deprivation variable available for every participant, school decile was used in the final analysis. Model three consisted of model two plus a dieting status by dietary pattern interaction variable, while model four consisted of model two plus the relevant sex by dietary pattern interaction. As differences between dietary patterns and overweight have been found between boys and girls [5, 7], model five was run separately for boys and girls and adjusted for age, school decile, and ethnicity. Model results in tables are presented with no interaction terms. When a significant interaction was found this was indicated with a superscript in the table while the relevant models including the interaction terms were presented visually with figures. A two-sided P-value of <0.05 was considered statistically significant. Statistical analyses were undertaken using Stata statistical software package version 12.0IC (StataCorp, College Station, TX, USA).