Studies of the overall diet have emerged as an important research field complementary to ‘reductionist’ single component studies( Reference Newby and Tucker 1 – Reference Jacobs and Steffen 4 ). The rationale for this is threefold. First, dietary exposure consists of a multitude of different nutrients and other bio-active constituents. Some components act synergistically, where other components work in opposition. The complex interactions and cumulative effects cannot be captured well by studying the effects of single dietary components( Reference Newby and Tucker 1 , Reference Schulze and Hoffmann 3 ). Second, people do not eat nutrients or constituents. They consume foods; and food consumption often occurs in patterns of meals and in-between meal consumption. Consumption patterns are shaped by income, prices, individual preferences and beliefs, cultural traditions, as well as geographical, environmental, social and economic factors( 5 ). Third, dietary change is usually not restricted to one dietary component because of substitution and compensatory effects of other dietary characteristics( Reference Kant 2 ).
Various approaches to study the overall diet can be distinguished( Reference Moeller, Reedy and Millen 6 ). In hypothesis-driven or a priori approaches, researchers define scores or indices of the overall dietary quality. The scores are usually based on guidelines for a healthy diet or on diets known to be healthy( Reference Waijers, Feskens and Ocke 7 ). In contrast, a posteriori approaches are driven by the underlying dietary data. Statistical methods such as principal component analysis, exploratory factor analysis and cluster analysis are applied to derive dietary patterns that are available in the data( Reference Schulze and Hoffmann 3 ). In principal component and exploratory factor analyses, new dietary pattern variables are obtained based on the underlying interrelationships between the dietary components. Whereas in cluster analysis, subgroups of the study population are created that combine people with similar dietary intakes within a cluster and put people with different dietary intakes in different clusters( Reference Newby and Tucker 1 ). Finally, hybrid approaches were introduced to study the overall diet. Hybrid approaches are driven by a combination of biological pathways and the underlying dietary data. For example, reduced rank regression can be applied to define linear combinations of food intakes that maximally explain predictors of disease( Reference Hoffmann, Schulze and Schienkiewitz 8 ). Methods such as decision tree analysis can be used to identify subgroups of a population whose members share dietary characteristics that influence a disease or disease predictor( Reference Hearty and Gibney 9 ).
This paper aims to present an overview of approaches for assessing the overall diet. Attention is particularly given to newer methodologies. The advantages and limitations of each approach are presented, as well as ways to evaluate the obtained overall diets and their main uses. The description is given from a public health perspective rather than from focus on individuals or on clinical settings.
Dietary quality scores
Principles
Scores or indices of dietary quality express the overall healthiness of the diet. They are usually developed based on a specific dietary pattern that is known to be healthy or based on pre-existing dietary guidelines for the general population or for the prevention of a specific dietary-related disease( Reference Waijers, Feskens and Ocke 7 , Reference Arvaniti and Panagiotakos 10 ). Table 1 shows key aspects to be considered in the development and interpretation of dietary quality scores, as well as commonly chosen implementations for those aspects. A more extensive description of the make-up of dietary quality scores is given in Waijers et al.( Reference Waijers, Feskens and Ocke 7 ).
Key characteristics of dietary quality scores and commonly used implementations
* Indirectly through energy adjustment or standardisation.
In the application and interpretation of a dietary quality score, the chosen make-up should be considered. For example, population-specific median values as cut-off levels for scoring the individual components of a score will limit the usefulness of comparing the quality scores across populations. However, it will make sure that all components of the score contribute to the overall score. In contrast, in case of a score with fixed cut-off levels and a binary scoring system, it might occur that all persons score the same for a specific component. This component does not then contribute to the overall score. However, a score with this system is suitable to compare the dietary quality of various populations( Reference Waijers, Feskens and Ocke 7 ).
Well-known examples
Recent reviews, in 2007( Reference Waijers, Feskens and Ocke 7 ) and 2009( Reference Wirt and Collins 11 ), identified twenty and twenty-five different scores of overall dietary quality, respectively. Dietary quality scores have been widely used in adult populations, whereas in children the use is limited( Reference Lazarou and Newby 12 ). Two well-known examples of dietary quality scores are the Healthy Eating Index( Reference Guenther, Reedy and Krebs-Smith 13 , Reference Kennedy, Ohls and Carlson 14 ) and the Mediterranean Diet Score( Reference Trichopoulou, Kouris-Blazos and Wahlqvist 15 ).
The Healthy Eating Index is a measure of dietary quality according to the United States Department of Agriculture Food Guide Pyramid. It was developed in 1995( Reference Kennedy, Ohls and Carlson 14 ), revised according to the revision of the guidelines in 2005( Reference Guenther, Reedy and Krebs-Smith 13 ), and is again being updated at the moment( Reference Krebs-Smith, Guenther and O'Connell 16 ). Index components in the 2005 Healthy Eating Index were: total fruit, whole fruit, total vegetables, dark green and orange vegetables and legumes, total grains, whole grains (each five points), milk, meat and beans, oils, saturated fat, Na (each ten points), and energy from solid fats, alcoholic beverages and added sugars (twenty points). The included food groups were expressed in servings per 4184 kJ (1000 kcal); the included nutrients and energy from solid fats, alcoholic beverages and added sugars are also expressed using an energy density approach. The scoring is proportional to the extent to which the dietary guideline is met. The overall score can range from zero (poor diet) to 100 (excellent diet)( Reference Guenther, Reedy and Krebs-Smith 13 ). Adapted versions were developed for other countries (e.g. Canada( Reference Glanville and McIntyre 17 )) and for specific population groups (e.g. children( Reference Feskanich, Rockett and Colditz 18 )).
Advantages of the Mediterranean diet were described already in the 1950s by Keys( Reference Keys and Grande 19 ) and by many others afterwards( Reference Mila-Villarroel, Bach-Faig and Puig 20 , Reference Bach, Serra-Majem and Carrasco 21 ). The first scoring system to express adherence to the Mediterranean diet, i.e. the Mediterranean Diet Score, was developed by Trichopoulou et al.( Reference Trichopoulou, Kouris-Blazos and Vassilakou 22 ). The original Mediterranean Diet Score was composed of eight components: i.e. the MUFA:SFA ratio, consumption of legumes, cereals, fruits and nuts, vegetables, meat and meat products, milk and dairy products, and alcohol. The components were adjusted for energy intake by standardising intakes for men to 10460 kJ (2500 kcal) and for women to 8368 kJ (2000 kcal). Cut-off values were sex-specific median intakes of the studied population, and scoring was either zero (worst) or one (best) for each component. With equal weights for each component, this led to a total score range of zero (poorest adherence to the Mediterranean diet) to eight (excellent adherence to the Mediterranean diet)( Reference Trichopoulou, Kouris-Blazos and Vassilakou 22 ). The Mediterranean Diet Score has been applied to Mediterranean and non-Mediterranean populations( Reference Sofi, Abbate and Gensini 23 ). Over time, various alternative Mediterranean Diet Scores were developed and tested( Reference Mila-Villarroel, Bach-Faig and Puig 20 , Reference Bach, Serra-Majem and Carrasco 21 ). A recent meta-analysis quantified the protective effect of adherence to the Mediterranean diet for overall mortality and major chronic diseases. A two-point increase in the Mediterranean Diet Score was related to a 8% reduction in mortality( Reference Sofi, Abbate and Gensini 23 ).
Evaluation strategies
A careful development process of a dietary quality score should include an in-depth evaluation. A majority of the dietary quality scores have been evaluated with regard to their nutrition adequacy only( Reference Kant 24 ). Various other aspects are relevant to include in the evaluation. A good example can be found in the paper on the 2005 Healthy Eating Index( Reference Guenther, Reedy and Krebs-Smith 25 ). First, content validity was evaluated by assessing if the index captured all aspects of the dietary guidelines. Then construct validity was judged. This was done by checking if the index gave maximum scores to menus developed by nutrition experts to illustrate high diet quality; and by checking if the index distinguished between groups with known differences in dietary quality. The next step was the evaluation of reliability. In this step, the relationships among the index components was assessed, and it was checked which components had most influence on the overall index score( Reference Guenther, Reedy and Krebs-Smith 25 ). Moreover, it is important to assess the validity of the underlying dietary data of descriptive or epidemiological studies in which a dietary quality score is calculated( Reference Lazarou and Newby 12 ).
For scores that intend to quantify the healthiness of the overall diet, the longitudinal relationship with overall health or total mortality is the ultimate evaluation. The majority of studies that assessed this relationship demonstrated that higher dietary quality was consistently inversely related to all-cause mortality, with a protective effect of moderate magnitude. The associations were stronger for men and for all-cause and CVD mortality( Reference Wirt and Collins 11 ).
Considerations, advantages and limitations
The strength of scores of dietary quality is that they rely on the body of scientific evidence from studies on health and disease prevention. However, this is partly a theoretical strength. In practice, there is insufficient knowledge and consensus on what actually is the healthiest diet. This is clearly shown by the large number of existing scores that attempt to express overall dietary quality( Reference Waijers, Feskens and Ocke 7 , Reference Fransen and Ocke 26 ). Also, interpretations of the dietary guidelines are often needed to construct a dietary quality score, and subjectivity is introduced( Reference Moeller, Reedy and Millen 6 ). This can, for example, be about the definition of the score components (e.g. what is Eat a varied diet?); but particularly the scoring and weighting of the various components is often not defined in the dietary guidelines( Reference Kourlaba and Panagiotakos 27 ). A second advantage of dietary quality scores is that they are usually easy to compute; and thereby easily reproducible and comparable( Reference Moeller, Reedy and Millen 6 ).
A limitation of a dietary quality summary score is that it does not describe the overall diet pattern. This is partly due to the fact that many dietary quality scores focus on selected aspects of the diet and the correlated structure of the components is not considered( Reference Hoffmann, Schulze and Schienkiewitz 8 , Reference Arvaniti and Panagiotakos 10 ). However, especially, persons who have a midrange score can have very different contributing components, and thus different dietary patterns( Reference Moeller, Reedy and Millen 6 ).
An important consideration during the application of dietary quality scores is that they should be tailored to their aim. Obviously, if the diet quality for persons with a high risk of a given disease is aimed at, dietary guidelines for the prevention of this specific disease should be the reference. If the dietary quality of children is to be assessed, the guidelines should be applicable to children( Reference Lazarou and Newby 12 ). It should also be considered that although dietary quality scores are often named hypothesis-driven methods, their application usually relies also on the underlying dietary data. In the case of a score with median values as cut-off values this is obvious. A second example is that many dietary guidelines include a recommendation to limit the intake of salt, while total salt intake cannot be captured well with self-reporting dietary assessment methods and is therefore often missing in food consumption databases. More generally, the type of dietary data used for the scoring of dietary quality should be appropriate for its purpose. In many studies, data from FFQ were used, since this dietary assessment method provides information about usual intake and dietary recommendations are intended to be met over time. Data derived by one or a few 24-h dietary recalls should not be used as such to calculate usual dietary quality, since they include too much day-to-day variation. Recently, a statistical model became available to overcome this challenge. The model was applied to the National Health And Nutritional Examination Survey 24-h dietary recall data and provided estimates of the population distribution of the Healthy Eating Index for the US( Reference Zhang, Midthune and Guenther 28 ). Overall, the general problems of dietary assessment through self-reporting of food consumption( Reference Kipnis, Subar and Midthune 29 ) are also reflected in the calculated dietary quality scores.
Applications
Dietary quality scores can be useful tools to monitor the overall adherence to dietary guidelines, and the dietary quality of a population. Comparisons within and between populations can be made to formulate or evaluate the need for dietary interventions. In addition, dietary quality scores are useful tools to test if current dietary recommendations have a measurable protective effect against diseases, and to get insight into the magnitude of the overall effect( Reference Wirt and Collins 11 ).
An efficient application of dietary quality scores is the combination of a dietary quality score with a short dietary assessment or screening method that only enquires about the relevant dietary components. See, for example, a web-based questionnaire to score the Dietary Approaches to Stop Hypertension Diet( Reference Apovian, Murphy and Cullum-Dugan 30 ). This allows a quick and efficient assessment of overall dietary quality as an alternative to an extensive dietary assessment covering all components of the total diet. In the setting of developing countries, a quick screening method about the number of food groups consumed in a given period is often converted into a dietary variety score. The dietary variety score can subsequently be used as proxy for monitoring the overall dietary quality and household economic access to food( Reference Fransen and Ocke 26 , Reference Kennedy, Berardo and Papavero 31 ).
Empirically derived dietary patterns
Principles
In contrast to dietary quality scores, empirically derived dietary patterns are driven by the dietary data from which they are derived. Two main approaches can be distinguished( Reference Newby and Tucker 1 ). In the first approach, the dietary variables are combined into fewer variables based on their interrelationships. Common methods in this approach are principal component analysis and exploratory factor analysis. In principal component analysis, patterns or components are direct linear relationships of the underlying dietary variables. The created dietary pattern variables explain as much as possible the total variation of the original dietary variables. In exploratory factor analysis dietary patterns are modelled as underlying factors; only the variance that is shared with other variables is accounted for, excluding variance unique to each variable and random error variance( Reference Schulze and Hoffmann 3 ). In dietary pattern analysis, principal component analysis is more commonly used than exploratory factor analysis. The obtained component scores are continuous variables( Reference Moeller, Reedy and Millen 6 ).
In the second approach, i.e. cluster analysis, mutually exclusive non-overlapping clusters of individuals are created( Reference Devlin, McNulty and Nugent 32 ). Individuals within clusters share a similar dietary pattern, whereas individuals in other clusters have food patterns that are far apart. The K-means method is the most often applied method to obtain clusters of people with similar dietary patterns. It is an optimisation method to derive a specified number of clusters, by minimising an error criterion. Alternatively, Ward's minimum variance method is an agglomerative hierarchical clustering method. Although requiring a large computation time, it is also found in the dietary pattern literature( Reference Moeller, Reedy and Millen 6 ).
Other empirical approaches have been applied to obtain dietary patterns. A recent example is the use of the treelet transform( Reference Gorst-Rasmussen, Dahm and Dethlefsen 33 ). This approach combines the quantitative pattern extraction capabilities of principal component analysis with the interpretational advantages of cluster analysis. The end result is a small number of naturally and hierarchical grouped variables. A disadvantage of the treelet transform method for dietary pattern analysis( Reference Gorst-Rasmussen, Dahm and Dethlefsen 33 ) is its assumption that only selected dietary components can contribute to the patterns, allowing no contributions of other dietary factors to the patterns( Reference Imamura and Jacques 34 ).
To obtain empirically derived dietary patterns, the researcher has to make many decisions. The empirically derived dietary patterns are therefore not entirely data driven. Table 2 shows important aspects that should be considered during the preparation phase, statistical analysis and reporting phase, with often chosen implementations. As dietary input variables, most often food groups are used. An advantage of this is that together they can represent the total dietary intake, accounting for interactions between nutrients and other components within the groups( Reference Devlin, McNulty and Nugent 32 ). Many of the decisions are important for the interpretation of the dietary patterns. For example, expressing the input variables as contributions to energy intake has the disadvantage that the analysis is less sensible to detect variations in food group consumption that might be important for health but contribute little to energy intake. This is particularly the case for fruit and vegetable consumption( Reference Bailey, Gutschall and Mitchell 35 ). On the other hand, several studies found little differences in the derived dietary patterns with input variables that were or were not adjusted for energy intake before the dietary pattern analysis( Reference Balder, Virtanen and Brants 36 , Reference Northstone, Ness and Emmett 37 ).
Key aspects to be decided during the process of factor/cluster analysis and commonly used implementations
Often observed patterns
Although dietary patterns will never be exactly the same across studies, it is apparent from the published studies that certain dietary patterns are frequently found. A large number of studies using principal component, exploratory factor or cluster analyses have identified variations of a healthy and a traditional or less-healthful dietary pattern. Also a pattern high in desserts or sweets and patterns high in alcohol appeared repeatedly( Reference Newby and Tucker 1 ). In principal component and exploratory factor analyses, a healthy dietary pattern is often labelled ‘prudent’ and a less-healthful pattern ‘Western’( Reference Newby and Tucker 1 , Reference Devlin, McNulty and Nugent 38 ). Obviously, patterns with the same label can be defined by different food components or by different weights of the components( Reference Reedy, Wirfalt and Flood 39 ). The Western pattern is usually characterised by high loadings of red meat, processed meat, butter, potatoes, refined grains and high-fat dairy. The prudent pattern, in contrast, has high loadings of vegetables, fruit, legumes, fish and seafood, and whole grains( Reference Hu 40 ). In general, the ‘healthy’, compared with the ‘Western’ pattern has been associated with more favourable biological profiles, slower progression of atherosclerosis and reduced incidence of CVD( Reference Newby and Tucker 1 , Reference Hu 40 ).
Evaluation strategies
Evaluation strategies for empirically derived dietary patterns can focus on different aspects. These include the goodness of the solutions (using criteria such as explained variance in principal component analysis and exploratory factor analysis, or internal cluster validity indices), comparison of using dietary data obtained with different dietary assessment methods( Reference Hu, Rimm and Smith-Warner 41 ), comparison of using different types of input variables( Reference Bailey, Gutschall and Mitchell 35 ) or different strategies to derive the dietary patterns( Reference Lo Siou, Yasui and Csizmadi 42 ), and the reproducibility of derived dietary patterns. Reproducibility can be assessed internally using split sample techniques( Reference Lo Siou, Yasui and Csizmadi 42 ), or externally over time for the same population( Reference Hu, Rimm and Smith-Warner 41 ), and in different but similar study populations( Reference Balder, Virtanen and Brants 36 ).
With split samples, for example, splitting the dataset into two equal parts, dietary patterns obtained in one-half of the sample (the derivation sample) can be confirmed in the second half (the validation sample). This can be done either by repeating the exploratory analyses in both samples or by using a confirmatory approach in the validation sample. Dietary patterns derived with principal component analysis or exploratory factor analysis can thus be validated using confirmatory factor analysis( Reference Lau, Glumer and Toft 43 ) and those derived by cluster analysis using discriminant analysis( Reference Quatromoni, Copenhafer and Demissie 44 ).
Some researchers indicate that empirically derived dietary patterns should be validated by assessing whether the dietary patterns can reliably predict diseases or mortality( Reference Moeller, Reedy and Millen 6 ). However, an empirically derived dietary pattern might be perfectly valid, i.e. existing in a study population, but without a relationship with health and disease.
Considerations, advantages and limitations
Empirically derived dietary patterns have the advantage that they are independent of definitions of what is a healthy pattern, and they are multidimensional in nature. However, principal component, exploratory factor and cluster analyses are no prediction techniques and are study-population- and data-specific. The derived patterns ‘simply’ explain the variation in intake. There is no guarantee that the identified patterns will be related to specific health outcomes( Reference Schulze and Hoffmann 3 ). Moreover, the application of principal component, exploratory factor and cluster analyses relies on various subjective decisions to be taken by the researcher( Reference Newby and Tucker 1 , Reference Moeller, Reedy and Millen 6 ). See Table 2 for an overview.
Specific advantages of principal component and exploratory factor analysis are that they have good statistical power and the resulting dietary patterns show the interrelationships between the dietary components. In contrast, translation of the obtained dietary patterns to the individual is difficult since each individual scores on all the obtained dietary patterns( Reference Hearty and Gibney 45 ). In practice, the obtained dietary patterns usually explain a limited part of the variation in food intake( Reference Kant 2 ).
For cluster analysis, the translation to individuals is very easy to make since the dietary patterns are mutually exclusive( Reference Hearty and Gibney 45 ). In most cluster analysis procedures, food components with high variance and outliers have large impacts on the results. For this reason, standardised input variables or the percentage of energy contributed by the food groups are often used as input variables( Reference Newby and Tucker 1 ). However, using standardised input variables might give minor food groups undue influence and the differences in the dietary patterns might be diluted( Reference Moeller, Reedy and Millen 6 ), whereas expressing foods as their contribution to energy intake might give too little weight to health-related food groups such as fruit and vegetables( Reference Bailey, Gutschall and Mitchell 35 ). Clusters obtained with the K-means method produced cluster solutions that were more reproducible than those obtained with Ward's method( Reference Lo Siou, Yasui and Csizmadi 42 ).
It has been suggested that the combination of factor and cluster analyses is the ultimate way of empirical dietary pattern analysis, since they are complementary and give a better perspective and understanding of dietary habits.( Reference Engeset, Alsaker and Ciampi 46 )
The type of dietary assessment used to collect the dietary data is important to be considered. Interest will mostly be on usual dietary patterns, and in this case day-to-day variation such as present in dietary data collected with 24-h dietary recalls or diet records will behave like random measurement error( Reference Moeller, Reedy and Millen 6 ). The general problems of measurement error associated with self-reported dietary data transfer to the obtained dietary patterns; and might even be more severe because correlations in measurement error might distort the definition of the patterns( Reference Schulze and Hoffmann 3 ). This would, for example, occur if unhealthy foods are underestimated systematically by study participants. The effects of misreporting of energy intake on the results of dietary pattern analysis need further study( Reference Devlin, McNulty and Nugent 38 ).
Applications
Principal component, exploratory factor and cluster analyses are very useful in obtaining insight into existing dietary patterns within a specific population. Such insight is essential for nutrition education and for developing public health interventions( Reference van Dam, Grievink and Ocke 47 ). Principal component and exploratory factor analysis is of particular importance for insight into combinations of foods, and how people score on this( Reference Reedy, Wirfalt and Flood 39 ). Cluster analysis is more useful in getting insight into different subgroups in the population with different diets, i.e. for identifying groups of people who may be at nutritional risk( Reference Devlin, McNulty and Nugent 38 , Reference Reedy, Wirfalt and Flood 39 ). The thus obtained dietary patterns can be used to explore the combined health effects of commonly existing dietary habits. However, these approaches are not powerful in generating new hypotheses( Reference Kant 2 ). For hypotheses testing, follow-up is needed through confirmatory type analysis or intervention studies.
Hybrid methods
Principles
The principle of hybrid approaches to study the overall diet is, not surprisingly, the combination of the two previous approaches. Hybrid approaches are partly theoretically driven, by using predictor variables that are relevant for the purpose of the researcher. In addition, the hybrid approaches identify multivariate dietary patterns based on the study data, specifically relevant for the study population( Reference Hoffmann, Schulze and Schienkiewitz 8 ). The predictor variables can be biomarkers that are intermediate risk factors for a dietary-related disease( Reference Hoffmann, Zyriax and Boeing 48 ), but also other risk factors can be used such as nutrients that are related to the outcome of interest( Reference Hoffmann, Schulze and Schienkiewitz 8 ), a disease itself( Reference Camp and Slattery 49 ) or an overall dietary quality score based on recommendations for a healthy diet( Reference Hearty and Gibney 9 ).
The most commonly used hybrid approach in the field of dietary pattern analysis is reduced rank regression (e.g.( Reference Hoffmann, Schulze and Schienkiewitz 8 , Reference Hoffmann, Zyriax and Boeing 48 , Reference Vujkovic, Steegers and Looman 50 )). With this approach, linear combinations of food intakes are defined that maximally explain a set of response variables( Reference Schulze and Hoffmann 3 , Reference Hoffmann, Schulze and Schienkiewitz 8 ). The response variables need to be continuous variables( Reference Kroke 51 ), such as levels of a biomarkers or nutrient intakes. The resulting dietary patterns are new dietary variable scores similar to factor scores. Partial least-squares regression is a compromise between principal component analysis and reduced rank regression. With this approach, patterns are obtained that explain both variation in response variables and in the dietary components( Reference Hoffmann, Schulze and Schienkiewitz 8 ).
Also, cluster analysis has a parallel methodology defining distinct subgroups in a population while making use of an outcome variable. Classification and regression tree analysis is a non-parametric decision tree procedure that identifies mutually exclusive and exhaustive subgroups of a population whose members share common characteristics that are associated with the dependent variable of interest( Reference Lemon, Roy and Clark 52 ). In contrast to reduced rank regression, decision tree analysis uses one response variable only, e.g. a disease risk factor or disease outcome( Reference Camp and Slattery 49 ). The dependent, or response variable, can be either categorical (i.e. classification tree analysis) or continuous (i.e. regression tree analysis). In classification and regression tree analysis, independent variables can be any combination of categorical and continuous variables; no data assumptions are required( Reference Lemon, Roy and Clark 52 ). Decision tree analysis produces a visual output that is a multilevel structure that resembles branches of a tree. The results can thus be interpreted as hierarchical dietary patterns. The structure of the classification tree model is a set of nodes from the top to the bottom, in which the terminal nodes show the specific pattern features of the subpopulations in percentage, including the number of people and the probability or mean values of the outcome variable( Reference Teng, Lin and Ho 53 ). Until now, decision tree analysis was seldom applied for dietary pattern analysis( Reference Hearty and Gibney 9 ) or in a broader risk factor pattern analysis including dietary and other variables( Reference Camp and Slattery 49 ).
Other data mining techniques, such as neural network approaches might also be promising to obtain insight into the multiple dietary factors or a combination of diet and other risk factors that predict a disease outcome. Only a few applications including dietary variables have been published( Reference Hearty and Gibney 9 , Reference Cooper and Purcell 54 – Reference Park and Edington 57 ).
Evaluation strategies
The evaluation strategies of the hybrid approaches to study the overall diet are similar to those described for the empirically based type of analysis. In addition, observed relationships between the obtained dietary patterns and outcome variables should be confirmed. It is important to perform this confirmation in independent populations( Reference Hoffmann, Schulze and Schienkiewitz 8 ). In general, more experience is needed with evaluation of the hybrid approaches to study the overall diet( Reference Tucker 58 ).
DiBello et al.( Reference DiBello, Kraft and McGarvey 59 ) compared dietary patterns derived with principal component analysis, reduced rank regression and partial least-squares regression. Response variables for reduced rank regression and partial least-squares regression were adipose tissue levels of α-linolenic and trans-fatty acids and dietary intakes of saturated fat, fibre and folate. All three methods derived a similar vegetable pattern that was associated with myocardial infarction status. However, principal component and partial least-squares regression analysis derived additional dietary patterns that were associated with the health outcome. They conclude that reduced rank regression would have been the most appropriate method if the goal was to test hypotheses limited to the present group of response nutrients. However, to test any dietary pattern relationships with myocardial infarction, partial least-squares regression offered more flexibility( Reference DiBello, Kraft and McGarvey 59 ).
Other studies compared dietary patterns derived by reduced rank regression and principal component or exploratory factor analyses. In three studies, the first dietary pattern derived by reduced rank regression was related to the health outcome, whereas the first dietary pattern obtained by principal component analysis was not( Reference Manios, Kourlaba and Grammatikaki 60 – Reference Hoffmann, Boeing and Boffetta 62 ). In a fourth study, the Mediterranean type dietary patterns derived using both approaches were similar and were both related to the health outcome( Reference Vujkovic, Steegers and Looman 50 ).
Considerations, advantages and limitations
Hybrid approaches have the advantage of building on a priori knowledge of biological relations. In this way the derived dietary patterns should be better able to examine the importance of overall dietary patterns in the aetiology of diseases( Reference Schulze and Hoffmann 3 , Reference Tucker 58 ). The associated disadvantage is that hybrid approaches require a clear picture of the biological mechanism underlying the development of a given disease. They can only provide answers in the current theoretical framework( Reference Kroke 51 ). There is especially incomplete knowledge as to whether a dietary nutrient or biomarker is causal or merely a marker( Reference Kant 63 ).
One of the criticisms of reduced rank regression is that the observed relations between the dietary pattern and outcome of interest may arise due to the dietary pattern acting as a proxy for the biomarker( Reference Tucker 58 ). This requires confirmation of the results in randomly split samples and in independent studies. Confirmation in other studies can be done using the same weight and dietary components as in the original study, hence without actually having the biomarker information available( Reference van Dam 64 ).
The disadvantages of decision tree methods are that one key factor can dominate the model, misclassification can be rather large, and the methods might overfit( Reference Hearty and Gibney 9 ). Further considerations, advantages and limitations of decision trees and other data mining techniques need to be learned through more experience in the field of nutrition science.
Applications
Hybrid approaches to study the overall diet may be particularly useful in identifying combinations of dietary components that are relevant for given health outcomes. The application of reduced rank regression is limited to those health outcomes for which sufficient knowledge about intermediate risk factors is available( Reference Schulze and Hoffmann 3 , Reference Hoffmann, Schulze and Schienkiewitz 8 , Reference Nettleton, Steffen and Schulze 61 ). In the case of partial knowledge about the biochemical pathways, partial least-squares regression might be more appropriate. This technique offers the possibility to obtain dietary patterns that are constrained by the response variables, as well as dietary patterns that are unconstrained by the response variable( Reference DiBello, Kraft and McGarvey 59 ).
In the context of hybrid methods to identify dietary patterns, decision tree type methods seem particularly useful in identifying at-risk subgroups for a health outcome based on combinations of several known dietary and other risk factors (prediction application). In these approaches it is logical to include dietary as well as non-dietary information, because the methodology offers no other option to adjust for non-dietary confounders. It should be noted that decision tree analysis is also a useful technique to generate new hypotheses in the case of no prior hypotheses and many potential risk factors( Reference Dasgupta, Sun and Konig 65 ). This selection application would, however, not be called a hybrid approach for deriving dietary patterns.
Discussion and conclusion
In the past three decades, studies of the overall diet have emerged as an important area of research complementary to single component studies( Reference Jacobs and Steffen 4 ). This paper presents an overview of different approaches used in studies of the overall diet. The described approaches included hypothesis-driven scores of overall dietary quality, data-driven approaches such as principal component, exploratory factor and cluster analysis, and hybrid approaches such as reduced rank regression, partial least-squares regression and decision tree analysis. Several reviews have been published that present comprehensive overviews of existing dietary quality scores, empirically derived dietary patterns, and their relationships with demographic characteristics, risk factors, biomarkers, health and disease( Reference Newby and Tucker 1 – Reference Schulze and Hoffmann 3 , Reference Moeller, Reedy and Millen 6 , Reference Waijers, Feskens and Ocke 7 , Reference Arvaniti and Panagiotakos 10 – Reference Lazarou and Newby 12 , Reference Mila-Villarroel, Bach-Faig and Puig 20 , Reference Bach, Serra-Majem and Carrasco 21 , Reference Sofi, Abbate and Gensini 23 , Reference Kant 24 , Reference Fransen and Ocke 26 , Reference Kourlaba and Panagiotakos 27 , Reference Devlin, McNulty and Nugent 38 , Reference Tucker 58 , Reference Kant 63 ). The present paper did not attempt to update these reviews, but focused particularly on methodological aspects.
The results of studies of the overall diet have great potential for use in nutrition policy, particularly as it demonstrates the importance of total diet in health promotion. Dietary quality scores are primarily important for monitoring the quality of the overall diet, to evaluate the overall effects of dietary interventions( Reference Fransen and Ocke 26 ) and to test the validity of dietary recommendations( Reference van Dam 64 ). Data-driven approaches are particularly important for nutrition education and setting priorities in the planning of nutritional interventions( Reference van Dam 64 ). They show the interrelationships between dietary components and differences in dietary patterns within a population( Reference Nettleton, Steffen and Schulze 61 ). However, the extent to which dietary quality scores and data-driven approaches help to generate new insights into the relationships between dietary intake and diet-related diseases remains debatable( Reference Jacques and Tucker 66 ).
Reduced rank regression seems to have greater potential for testing new hypotheses on diet–disease relationships through specific biological pathways( Reference Schulze and Hoffmann 3 ). The hybrid approach is potentially strong because the derived dietary patterns are relevant for the population and related to health outcomes; whereas the a priori diet pattern scores might have little contrast within a population and the a posteriori derived diet patterns might not be relevant for health. To our knowledge this is the first overview paper that presented and reflected upon alternative hybrid approaches to reduced rank regression. For reduced rank regression Kant concluded that these methods require further development and innovation( Reference Kant 2 ). This is even more the case of the alternative hybrid approaches, which require more applications in the field of dietary patterns before conclusions on their use can be drawn.
The potential for several of the hybrid approaches to study the overall diet strongly depends on the availability of early risk factors for diseases( Reference Kant 63 ). Many chronic diseases develop over a period of many years, and are multi-causal in nature. This makes the studying of diet in relation to disease extremely complicated. Valid (bio)markers that are causal predictors for the development of disease might be an important help in this complex task. They can serve as response variables in the hybrid approaches to study the relationship with the overall diet. In a second and preferably independent step, the thus derived dietary patterns might subsequently be related to the incidence of diseases in long-term prospective studies( Reference Schulze and Hoffmann 3 ).
Few intermediate risk factors such as LDL- and HDL-cholesterol for CVD have long been used as clinical biomarkers. Since the early 1990s, research on the discovery and validation of biomarkers with prognostic values for CVD, cancer, obesity, diabetes and neurodegeneration has expanded considerably( Reference van Ommen, Keijer and Heil 67 ). Although for many diseases valid predictor biomarkers are currently still lacking, several developments work to the advantage of discovering new biomarkers for disease risk. The wish for substantiating health claims is one of the important driving forces for more research on the identification of further relevant markers to measure food functionality in the human body( Reference Gallagher, Meijer and Richardson 68 ). Moreover, recent advances in genomics and systems biology enable researchers to measure and model biomarker profiles and to translate these into dynamic processes( Reference van Ommen, Keijer and Heil 67 ). Especially markers for suboptimal health before clinical signs of disease are of increasing interest. Work on the development of markers for overarching processes such as oxidative, inflammatory, metabolic and psychological stress( Reference van Ommen, Keijer and Heil 67 ) is of great potential value for hybrid approaches for studying the overall diet.
From this overview, it is concluded that the various approaches for studying the overall diet are complementary, and no method can be considered superior to the other methods. Further insight into the utility of conducting studies on the overall diet can be gained if more attention is given to methodological issues. These include clarification of the aims and assumptions of the analyses and a precise description of the make-up of dietary quality scores or derived dietary patterns. Moreover, in-depth evaluations of the derived measures of the overall diet in terms of reproducibility, validity and comparisons of different methodologies are essential. This is particularly the case for the still less often applied hybrid approaches such as reduced rank regression, partial least-squares regression and decision tree analysis.
Acknowledgements
The author declares no conflicts of interest. This research received no specific grant from any funding agency in the public, commercial or not-for-profit sectors.
