Subjects and study design
The present data were obtained from a study of primary school pupils in Telemark County, Norway. Data collection took place in the spring of 2007 and spring of 2010, when the children were in primary school grades 4 (9 to 10 year old) and 7 (12 to 13 years old) respectively. The detailed methods for the 4th grade data collection have been described previously [3]. An identical procedure was used for data collection in the 7th grade. In brief, all primary schools in Telemark County were invited to participate in the study at both time points. Of 110 invited schools 70 agreed to participate in the 4th grade data collection and 53 of 104 participated in the 7th grade data collection. In total, written parental consent to inclusion in the study was received for 1,045 out of 1,477 invited children in the 4th grade and 1,095 out of 1,503 invited children in the 7th grade. This represented about half of the county's 4th and 7th grade pupils at the respective time points.
Weight and height measurements were obtained for 955 (4th grade) and 865 (7th grade) children, while complete weight/height and dietary data were obtained for 924 and 691 pupils, respectively. In total, 427 children provided complete weight, height and dietary data at both time points (Figure 1).
The research protocol was approved by the Regional Committee for Ethics in Medical Research and the Norwegian Data Inspectorate. Informed written consent was obtained from the parents of all participating children in both 2007 and 2010.
Dietary information
The children's food and drink intake was reported by their parents using a retrospective food frequency questionnaire (FFQ), which asked about habitual daily consumption of 40 food items, 11 types of drink, 13 types of snacks (between meals) and five main meals (breakfast, lunch, afternoon meal, dinner, supper) during the last six months. Identical FFQs were used at both time points. This questionnaire was based on a short FFQ developed for use among fourth- and eighth-grade children in Norway, but was modified to include more dietary questions. The modified FFQ is appropriate for exploring dietary patterns based on frequencies, but has not been validated for estimating total intakes of energy or nutrients. The alternative frequencies for food and drink items were: "rarely/never", "1-3 times a month", "1-3 times a week", "4-6 times a week", "once a day", "twice a day", and "3 or more times per day". Meal patterns were registered as the daily frequencies of five main meals (breakfast, lunch, afternoon meal, dinner, supper), with 8 response alternatives ranging from "never/rarely" to "daily". The questions about snacking between meals had three answer categories: "never/rarely", "sometimes" and "often/always". As we used meal and snacking events in addition to food consumption frequencies as input variables in the PCA, the components were denoted as 'eating patterns' rather than 'dietary patterns'.
Other variables
In addition to providing dietary information, the parents answered questions about their own weight, height, educational level and work situation, and family income. They also provided their subjective opinion regarding their child's physical activity level compared with that of other children of the same age, and of time spent on screen-based activities and other sedentary activities outside school (e.g. reading or homework).
Parental educational level was divided into three categories: "primary and lower secondary education" (10 years or less), "upper secondary education" (three to four years of secondary education), and "university or university college".
Family income was divided into three categories: "both parents < Norwegian kroner (NOK) 300,000 (EUR 33,909)", "one parent ≥ NOK 300,000", and "both parents ≥ NOK 300,000".
A variable categorising leisure physical activity by reference to other children was used as an indicator of the children's physical activity level. Parents indicated on a scale from 1-5 whether the child was "less physically active than other children of the same age" or "more physically active compared to other children of same age". The question was taken from a battery of validated questions used in a study of children's activity and inactivity in the Netherlands [26], and translated into Norwegian for use in the present study.
Inactivity was defined as time spent on screen-based activities and other sedentary activities outside school. These activities were combined and divided into two categories: "less than 4 hours per day", and "4 hours or more per day".
BMI categories
The weight and height of the children were measured by public health nurses at each school at both time points. The children were weighed wearing light clothing (i.e. trousers, T-shirt, socks), using calibrated, electronic scales measuring in 100 g increments. BMI (kg/m2) of each child was calculated on the basis of these measurements. Child BMI categories were calculated using International Obesity Task Force (IOTF) cut-off points (underweight, normal weight, overweight, obese), based on growth curves and BMIs of 17, 25 and 30 kg/m2 at age 18 years [27, 28]. The respective cut-off points for 9.5-year-old and 12.5-year-old boys and girls were used. Due to small numbers we included underweight children in the normal weight group and obese children in the overweight group.
Changes in BMI categories between the 4th grade and the 7th grade were divided into four categories: "unchanged normal weight", "overweight to normal weight", "normal weight to overweight", and "unchanged overweight". Parent BMI categories were calculated on the basis of self-reported height and weight and the IOTF cut-off points for adults (overweight at BMI ≥ 25 kg/m2).
Statistical analyses
At each of the two time points, we used PCA with varimax rotation to identify eating patterns from the reported dietary responses [29]. Food and drink frequencies were assigned values from 1 for "never/rarely" to 7 for "3 or more times daily", while meal frequencies were assigned values from 1 for "rarely/never" to 8 for "daily" and snacking between meals were assigned 1 for "never/rarely", 2 for "sometimes" and 3 for "often/always". Missing values for a given variable were replaced by rarely/never. Respondents were excluded from the analysis if answers were missing for 16 (23%) or more of the questions about food and drink items or if answers were missing for more than two questions (40%) about meals (n = 31 and n = 68 for the 4th grade and the 7th grade, respectively).
PCA constructs new linear factors by grouping together correlated variables. The coefficients defining the factors are called factors loadings and represent the correlations of each input variable with the factors. The number of components chosen from the factor analysis was based on the scree plot, eigenvalues and the interpretability of the components [29]. The criteria for choosing the components were identical at both time points. Variables with factor loadings > 0.25 or < -0.25 were considered important for interpretability of the components. The components (called eating patterns) were named after the nature of the foods, beverages and meals with the highest factor loadings within each pattern. The four eating patterns previously identified for the children in the 4th grade [3] were: a "snacking" pattern, characterised by snack items and sugar-sweetened drinks, low intake of water, vegetables and brown bread and a low frequency of eating breakfast and dinner; a pattern labelled "junk/convenient", characterised by high-fat and high-sugar processed fast foods; a "varied Norwegian" pattern, characterised by food items typical of a traditional Norwegian diet, close to what is recommended by the health authorities; and, finally, a "dieting" pattern, containing foods and drinks often associated with dieting and weight control.
Individuals were given factor scores for each of the patterns. Factor scores were standardised to a mean of zero. Positive factor scores indicate higher consumption of foods, drinks, snacks and meals in that pattern, while negative factor scores indicate low consumption. The factor scores for each eating pattern were used in the further analysis as continuous variables, or ranked into categorical variables (tertiles).
The factor scores were approximately normally distributed. Therefore, Pearson's correlation coefficients were used to evaluate the agreement between the factor scores for similar and different eating patterns at the two time points. Additionally, Cohen's weighted kappa (κ) [30] was used to compare individual factor scores as categorical variables (tertiles) across time. Cohen's κ for being in the same weight group at both time points was also calculated. In accordance with the scale of Landis and Koch [31], we interpreted the κ-values to represent the following agreement between time points: 0.01 to 0.20: "slight"; 0.21 to 0.4: "fair"; 0.41 to 0.60: "moderate"; 0.61 to 0.80: "substantial"; and 0.81 to 1.00: "almost perfect". BMI was not normally distributed, and Spearman's rho was used for correlation analysis.
Analysis of variance (ANOVA) was used to examine differences in pattern scores between groups. We used multiple logistic regression to calculate adjusted odds ratios (OR) and 95% confidence intervals (CI) for being overweight in the 7th grade and for staying overweight from the 4th to the 7th grade. Categorised pattern scores (low, medium and high, based on tertiles) and a dichotomous variable denoting increase/decrease in pattern scores were used as categorical independent variables in the models, respectively.
We used multiple linear regression and independent samples T-test to examine the changes in eating pattern scores in relation to changes in BMI categories. Changes in eating pattern scores over time were calculated as the difference in pattern scores from the 4th to the 7th grade. The difference within each pattern was examined as the dependent variable, while changes in BMI categories were used as independent variables, with "unchanged normal weight" as the reference category and "overweight to normal weight", "normal weight to overweight", and "unchanged overweight" as independent dichotomous variables.
In order to examine adherence to eating patterns over time, we categorised the changes in eating pattern scores from the 4th to the 7th grade into dichotomous variables denoting either unchanged/increased adherence to the pattern (no change or positive change) or reduced adherence to the pattern (negative change).
The potential confounding variables in the multiple regression models were: maternal and paternal overweight, maternal and paternal education, family income, child physical activity, child sedentary activity, and child gender. We applied a forward conditional selection and included variables significantly associated with overweight in each model. Adjusting for all of the variables had little additional impact on the effect estimates, and led to no changes in the main conclusions in this article.
For all tests, P < 0.05 was considered significant. The questionnaires were scanned by Eyes and Hands (Readsoft Forms, Helsingborg, Sweden), and all the statistical analyses were carried out using SPSS version 15.