Empirical dietary patterns

This review focused on studies that utilised empirical methods to identify patterns in total food intake. Studies employing diet indices or diet quality scores, e.g. Healthy Eating Index⁽ Reference Guenther, Casavale and Reedy ^{9 )} were not included, as diet indices are not empirically derived and they usually focus on selected foods or nutrients rather than the total diet. Diet indices require many arbitrary decisions regarding cut-offs for scoring, weighting item contributions to the total score, and the scaling of scores. Some, for example, the Mediterranean diet score⁽ Reference Panagiotakos, Pitsavos and Stefanadis ^{10 )}, are strongly culturally influenced and may not be applicable in different populations.

To date, the most popular empirical dietary pattern methods have included principal components analysis (PCA), factor analysis (FA) and cluster analysis (CA). In brief, these are a posteriori data reduction techniques that identify underlying constructs or patterns in food intake. PCA and FA produce a number of dietary patterns, each consisting of a set of factor loadings (one for each food) that characterises a particular food and nutrient profile at the population level^{( 11 , 12 )}. Individuals are then given a z-score quantifying the degree to which their reported dietary intake reflects the identified dietary patterns, relative to others in the same study population. These z-scores can be used as predictor variables to model disease risk. CA is a less frequently applied dietary pattern method that attempts to classify individuals into mutually exclusive groups or clusters with similar food intakes⁽ Reference Devlin, McNulty and Nugent ^{13 )}. Unlike PCA and FA, CA does not produce continuous dietary pattern z-scores but rather, individuals are assigned membership to one of the identified clusters, which precludes examination of dose–response relationships.

PCA, FA and CA aim to explain the maximum variation in food intakes, but they do not necessarily explain the maximum variation in nutrient intakes relevant to the health outcome of interest⁽ Reference Hoffmann, Schulze and Schienkiewitz ^{14 )}. It is after all nutrients, the product of food intake, that interact at the mechanistic level to influence disease risk. Furthermore, these methods are entirely exploratory, and their dependence on the data covariance can limit their reproducibility.

Reduced rank regression (RRR) is both an a priori and a posteriori empirical method⁽ Reference Hoffmann, Schulze and Schienkiewitz ^{14 )}. RRR can be used to identify combinations of food intake, or dietary patterns, that explain the maximum variation in a predefined set of intermediate (or response) variables. These intermediate variables are hypothesised to be on the causal pathway between food intake and a specific health outcome of interest. The intermediate variables may be nutrient intakes or biomarkers, but should be susceptible to food intake and not be proxies for the outcome. Similar to PCA and FA, RRR dietary patterns are characterised by a set of factor loadings at the population level, and individuals are scored for the dietary patterns using a z-score. By including a priori information in the form of intermediate variables, RRR can be a powerful method for examining specific diet–disease pathways that builds on prior knowledge. The incorporation of a priori information in RRR dietary patterns may also make them better at predicting disease risk than dietary patterns identified using exploratory methods such as PCA and FA⁽ Reference Hoffmann, Schulze and Schienkiewitz ^{14 ,} Reference Hoffmann, Boeing and Boffetta ^{15 )}.

Identifying the literature

Studies eligible for this review were (i) of prospective design, (ii) employed an empirical method (i.e. PCA, FA or RRR) to identify dietary patterns in childhood or adolescence, and (iii) reported associations between these dietary patterns and a prospectively measured obesity-related outcome. Eligible studies were identified in the literature published between 1970 and June 2013 using electronic searches of PubMed and Web of Science (English language only) and the following search terms: dietary patterns; PCA; FA; RRR; children; adolescents; obesity; BMI; overweight; weight. Reference lists in eligible articles and relevant reviews⁽ Reference Smithers, Golley and Brazionis ^{16 ,} Reference Togo, Osler and Sorensen ^{17 )} were also hand searched. Studies on growth confined to infancy were excluded. Studies using CA were excluded because their dichotomous scores were not considered comparable with continuous dietary pattern z-scores produced by PCA, FA and RRR.

Study quality

At present, there is no universally accepted tool for scoring the methodological quality of observational studies⁽ Reference Sanderson, Tatt and Higgins ^{18 )}. In this review, a scoring tool applied by te Velde et al.⁽ Reference te Velde, van Nassau and Uijtdewilligen ^{19 )} (in a review of dietary and physical activity behaviours and obesity in preschool children) was used to assess the methodological quality of each eligible study. The tool assessed study quality according to information, validity and precision related to: (i) study population and participation, (ii) study attrition, (iii) data collection methods and (iv) data analysis methods (Table 1). The scoring tool consisted of twelve criteria. Each could be assigned one of three possible scores: a positive score (+) if the study met the quality criterion, a negative score (−) if it did not meet the quality criterion, and a ‘?’ if a decision was not possible because insufficient information was provided. Citations to earlier papers were searched if information relevant to the criteria could not be found in the primary paper. A total quality score for each study was based on the number of positive scores for criteria reflecting validity or precision⁽ Reference te Velde, van Nassau and Uijtdewilligen ^{19 )} (Table 1). Data from each study were extracted into a summary table and scored for methodological quality by two independent reviewers (G. L. Ambrosini and D. Johns).

Table 1.

Criteria for assessing methodological quality of studies (adapted from⁽ Reference te Velde, van Nassau and Uijtdewilligen ^{19 )})

Criteria related to: I, information; V, validity; P, precision.

^a Sufficient to enable study replication.

Summarising the evidence

Meta-analysis can be used to combine statistical evidence from more than one study to estimate a pooled effect. However, this requires a sufficient number of comparable studies with a common outcome and exposure measurement, and limited between-study heterogeneity. Where appropriate, a meta-analysis was undertaken using a random effect model in STATA^{( 20 )} to combine the log-transformed results of eligible studies⁽ Reference Egger, Smith and Phillips ^{21 )}. The estimate of heterogeneity was used to gauge the appropriateness of combining studies for a pooled effect.

As the number of eligible studies was not expected to be large, a harvest plot was also employed to summarise the evidence. Harvest plots are an alternative to meta-analyses when large heterogeneity between studies precludes their statistical combination⁽ Reference Ogilvie, Fayter and Petticrew ^{22 )}. A harvest plot was created by plotting each eligible study according to its quality score and grouping studies by size and finding (null, negative or positive)⁽ Reference Ogilvie, Fayter and Petticrew ^{22 )}.

Eligible studies

The literature searches identified 166 papers. Of these, nine papers met the eligibility criteria, including two that analysed the same cohort at different follow-up ages⁽ Reference Johnson, Mander and Jones ^{23 ,} Reference Ambrosini, Emmett and Northstone ^{24 )} and one study published as a conference abstract⁽ Reference Ambrosini, Mori and Appannah ^{25 )}. The most common reason for exclusion at the screening stage was cross-sectional study design. After full screening, two of the nine eligible papers were deemed invalid and excluded for issues we have previously raised regarding the use of RRR⁽ Reference Ambrosini, Johns and Jebb ^{26 )}. At the end of the screening process, seven papers were eligible for review (Fig. 1).

Fig. 1.

(colour online). Flow diagram of the study identification process. RRR, reduced rank regression.

The seven eligible studies were conducted in the UK⁽ Reference Johnson, Mander and Jones ^{23 ,} Reference Ambrosini, Emmett and Northstone ^{24 )}, USA⁽ Reference Cutler, Flood and Hannan ^{27 )}, Australia⁽ Reference Ambrosini, Mori and Appannah ^{25 )}, Norway⁽ Reference Oellingrath, Svendsen and Brantstaeter ^{28 )}, Finland⁽ Reference Mikkilä, Räsänen and Raitakari ^{29 )} and Colombia⁽ Reference Shroff, Perng and Baylin ^{30 )} (Table 2). The median duration of follow up was 4·5 years. Follow-up ages ranged from 3 to 39 years, however, most studies included follow up until the mid-to-late teenage years. With the exception of two studies, sample sizes approximated 1000 or more subjects. All studies included appropriate adjustments for confounders, including age, gender, physical activity, socioeconomic status and maternal education. Two studies adjusted for total energy intake⁽ Reference Mikkilä, Räsänen and Raitakari ^{29 ,} Reference Shroff, Perng and Baylin ^{30 )}.

Table 2.

Studies of common empirical dietary patterns* in childhood or adolescence and later obesity risk

PCA, principal components analysis; RRR, reduced rank regression; ALSPAC, Avon Longitudinal Study of Parents and Children; Project EAT, Project Eating Among Teens.

* Dietary pattern characterised by high intakes of energy-dense, high-fat and low-fibre foods.

† Reported associations between dietary pattern and prospective measure of weight status.

‡ A total of 2516 were included from both age groups but the number in each age group was not reported.

Five studies applied PCA or FA and two applied RRR to identify dietary patterns. A comparable dietary pattern was identified in all the seven studies that consisted of a high consumption of energy-dense, high-fat and low-fibre foods. These were labelled ‘snacking’, ‘junk/convenient’, ‘sweet and salty snack foods’ ‘western’, ‘traditional’ and ‘energy-dense, high-fat, low-fibre’ (Table 3). All of these patterns were characterised by high intakes of confectionery and chocolate. High intakes of refined grains (low-fibre bread and breakfast cereals, rice and pasta), cakes and biscuits, savoury snacks and crisps, and processed meat (sausages, bacon, ham, salami, frankfurter/hot dogs) featured in most (83 %) of the patterns. High intakes of sugar-sweetened beverages, full fat milk, fried potato (chips), fats and oils, and takeaway foods (burgers, pizza, kebabs, fried/crumbed chicken or fish) featured in at least half of the patterns. Five of the seven studies reported that these dietary patterns had either negative factor loadings or were associated with lower intakes of fruit, vegetables and fibre (indicated by fibre, wholegrain, high-fibre breakfast cereal or high-fibre bread intake) and higher intakes of total fat and saturated fat⁽ Reference Johnson, Mander and Jones ^{23 ,} Reference Ambrosini, Emmett and Northstone ^{24 ,} Reference Cutler, Flood and Hannan ^{27 ,} Reference Mikkilä, Räsänen and Raitakari ^{29 ,} Reference Ambrosini, Oddy and Robinson ^{31 )}. In the remaining two studies⁽ Reference Oellingrath, Svendsen and Brantstaeter ^{28 ,} Reference Shroff, Perng and Baylin ^{30 )} it was not possible to determine the foods or nutrients negatively associated with their dietary pattern, because neither negative factor loadings nor nutrient profiles were reported⁽ Reference Oellingrath, Svendsen and Brantsaeter ^{32 ,} Reference Isanaka, Mora-Plazas and Lopez-Arana ^{33 )}.

Table 3.

Mapping of common dietary patterns and their key components

ALSPAC, Avon Longitudinal Study of Parents and Children (UK); Bogota, Bogota School Children Cohort (Colombia); Project EAT, Project Eating Among Teens (USA); Raine, Western Australian Pregnancy (Raine) Cohort (Australia); Telemark, Telemark County (Norway); Young Finns, Cardiovascular Risk in Young Finns Study (Finland). +, indicates that the food group/nutrient had a positive factor loading or a positive association with the dietary pattern; −, indicates either a negative factor loading or a negative association with the dietary pattern.

* Both ALSPAC papers identified the same dietary pattern⁽ Reference Johnson, Mander and Jones ^{23 ,} Reference Ambrosini, Emmett and Northstone ^{24 )}.

† Sausages, bacon, ham, salami, frankfurter/hot dogs.

‡ Low-fibre breakfast cereals and bread, rice, pasta.

§ Burgers, pizza, kebabs, fried/crumbed chicken or fish.

Quality of eligible studies

There was 93 % agreement between the two reviewers and differences in study quality scores were discussed to reach a consensus. The final study quality scores ranged from 43 to 71 % (Supplementary Table 1). The quality criteria met least often related to participation rates and selective attrition. All studies scored negatively for achieving a participation rate at baseline (beginning of analysed follow up) of at least 80 % of the original cohort. Only two studies reported investigating selective attrition during follow up⁽ Reference Johnson, Mander and Jones ^{23 ,} Reference Cutler, Flood and Hannan ^{34 )} and three studies did not achieve at least 70 % follow up of the baseline sample 12 months later⁽ Reference Johnson, Mander and Jones ^{23 ,} Reference Cutler, Flood and Hannan ^{27 ,} Reference Oellingrath, Svendsen and Brantstaeter ^{28 )}. Other factors impacting study quality included reliance on a very small number of overweight or obese children⁽ Reference Oellingrath, Svendsen and Brantstaeter ^{28 )} and the use of self-reported height and weight⁽ Reference Cutler, Flood and Hannan ^{27 )}.

Summary of evidence relating to childhood empirical dietary patterns and later obesity

Various obesity-related outcome measures were reported in the seven eligible studies (Table 2). One study analysed dietary pattern z-score as the outcome⁽ Reference Oellingrath, Svendsen and Brantstaeter ^{28 )}. Only three studies: Avon Longitudinal Study of Parents and Children (UK)⁽ Reference Ambrosini, Emmett and Northstone ^{24 )}; Western Australian Pregnancy (Raine) Cohort (Australia)⁽ Reference Ambrosini, Mori and Appannah ^{25 )}; Project Eating Among Teens (USA)⁽ Reference Cutler, Flood and Hannan ^{27 )} had a comparable outcome measurement (OR) and dietary pattern measure (continuous z-score) suitable for meta-analysis. Of these, Project Eating Among Teens⁽ Reference Cutler, Flood and Hannan ^{27 )} reported OR for four age- and gender-specific subgroups. Therefore, a meta-analysis combining the six OR from these three studies was undertaken. This produced a large estimate of between-study heterogeneity (78 %, P<0·0001), indicating that statistically combining these studies was inappropriate, and any estimated pooled effect misleading. There were an insufficient number of studies to conduct meta-analyses on other outcomes.

The harvest plot in Fig. 2 summarises the findings from all seven eligible studies. It can be seen that all four studies reporting a positive association between the dietary pattern and later obesity were of consistently higher quality than studies reporting null findings. No studies reported inverse associations between the dietary pattern and later obesity. The balance of evidence therefore supported a positive relationship between dietary patterns consisting of energy-dense, high-fat, low-fibre foods in childhood or adolescence and later obesity risk.

Fig. 2.

Harvest plot summarising prospective studies examining a dietary pattern in childhood or adolescence characterised by high intakes of energy-dense, high-fat, low-fibre foods and later obesity risk. ALSPAC, Avon Longitudinal Study of Parents and Children (UK); Bogota, Bogota School Children Cohort (Colombia); Project EAT, Project Eating Among Teens (USA); Raine, Western Australian Pregnancy (Raine) Cohort (Australia); Telemark, Telemark County (Norway); Young Finns, Cardiovascular Risk in Young Finns Study (Finland).