Guest guest Posted June 29, 2005 Report Share Posted June 29, 2005 Hi All, The usual method to examine the risk of foods that is used in 4,046 studies in Medline uses the principal component analysis method. Using a new method termed reduced rank regression used in only seven studies in Medline, bad foods are worse for all-cause mortality prediction. The below are pdf-available. van Dam RM. New approaches to the study of dietary patterns. Br J Nutr. 2005 May;93(5):573-4. No abstract available. PMID: 15975154 The relationship between diet and health can be examined at the level of food components, foods and dietary patterns. Until recently, the study of food components, particularly nutrients, has been the dominant approach in nutritional epidemiology. This approach has clear advantages. If the development of a disease is causally related to the intake of a food component, the examination of that food component will be the approach with the greatest power to identify its effect. In addition, results for food components can be compared with associations observed in other populations, data from mechanistic studies and health effects found in intervention studies (Willett & Buzzard, 1998). Knowledge on the level of food components can then be used to produce foods with higher or lower levels of the component. For example, reduction of the amount of trans-fatty acids in margarines has probably resulted in substantial health benefits for populations (Oomen et al. 2001). However, the effect of a food component can differ depending on the food that it is derived from due to interactions between food components or physical characteristics of foods. For example, folate from beer may provide less health benefits than the same amount of folate from bread, because alcohol reduces intestinal folate uptake, interferes with folate metabolism and increases urinary loss of folate (Jiang et al. 2003). Effects of food consumption on disease risk can be different from predictions based on known effects of food components, because our knowledge on the myriad possible beneficial or detrimental aspects of foods (s & Murtaugh, 2000) is still limited. The study of foods and food groups accounts for interactions between different components of a food, and for effects of physical characteristics and unknown components. In addition to different components of a food, synergy or antagonism may also exist for components of different foods and drinks that are included in the dietary pattern of an individual. As a result, health effects of dietary patterns may be greater than for individual foods or nutrients. Results from the Dietary Approaches to Stop Hypertension (DASH) intervention study illustrate the importance of considering dietary patterns (Appel et al. 1997). The DASH diet, a diet characterized by high consumption of vegetables, fruits, low-fat dairy products and whole grains, resulted in a greater reduction in blood pressure than had been found for individual minerals in these foods. Another advantage of the study of dietary patterns in epidemiological studies is that potential dietary confounders are largely incorporated in the dietary pattern variable. Different methods have been used to study dietary patterns in epidemiological studies. Two main approaches can be distinguished. First, an exploratory approach can be used that identifies combinations of foods and drinks as they are consumed in reality in a particular population. Principal components analysis is a frequently used exploratory approach to identify dietary patterns (Hu et al. 2000). This data reduction technique constructs new variables that are linear combinations of the original variables and explain as much of the variation in the original variables as possible. Applied to dietary data, new dietary pattern variables are derived on the basis of the correlation matrix of the original food variables. The use of principal components analysis to derive dietary pattern variables requires several arbitrary choices such as the original variables to include and the number of dietary patterns to identify (ez et al. 1998). This underlines the importance of conducting sensitivity analyses to examine the robustness of the findings (Hu et al. 2000). Dietary patterns identified by the explanatory approach reflect dietary behaviour and are not based on known health effects of diet. As a result, the identified dietary patterns are not necessarily relevant for disease risk. However, results can increase insight into possibilities for dietary changes and can provide information for setting priorities for changing dietary patterns in a population by public health initiatives. Second, a hypothesis-oriented approach that uses predefined criteria to construct dietary pattern scores can be used. These scores reflect the degree to which a person's diet conforms to a dietary pattern that was defined a priori based on presumed health effects. Scores based on dietary recommendations (McCullough et al. 2002) and characteristics of the traditional Mediterranean diet (Trichopoulou et al. 2003) have been used. Deciding what individual components to include, the cut-off points and the weights of different components of an a priori score still requires subjective choices. This approach does not have the advantages related to studying existing dietary behaviour or to the identification of new dietary patterns that may affect disease risk. However, the approach can capture the greater effects of the overall diet as compared with individual components, and can be used to test the validity of dietary recommendations. Hoffmann and colleagues have introduced a new method in nutritional epidemiology that combines characteristics of the explanatory and hypothesis-oriented approaches to dietary patterns (Hoffmann et al. 2004a). This reduced rank regression (RRR) method has similarities with principal components analysis, but it uses a set of response variables in addition to a set of food intake variables. The identified dietary patterns are linear combinations of the original food intake variables that maximally explain variation in the response variables (Hoffmann et al. 2004a). A priori knowledge is introduced by using a set of response variables that is known to predict the disease of interest. In a paper in this issue of BJN, Hoffmann et al. (2005) have used macronutrient intakes as response variables, resulting in dietary pattern variables that are linear combinations of the original food variables and maximally explain variation in macronutrient intakes. The identified dietary pattern was characterized by higher intakes of meats, poultry, butter, sauces and eggs, and lower intakes of bread and fruit, and was associated with a higher risk of premature mortality. Although the approach used invites interpretation of this association in terms of macronutrients only, other characteristics of the identified dietary pattern may still have contributed to the observed association. For the RRR method, the number of dietary patterns that can be identified is limited by the number of response variables. However, arbitrary decisions such as the choice of original food intake variables remain necessary, and sensitivity analyses should be conducted. An alternative approach that is of potential interest is the use of biomarkers of dietary intake that are relevant for the disease of interest as response variables. This application would identify combinations of foods that maximally explain variation in biomarkers of dietary intake, avoiding measurement error related to food composition data or to lack of information on bioavailability for different combinations of foods. In a case–control study of coronary heart disease, Hoffmann and colleagues used biomarkers of disease as response variables in the RRR analysis (Hoffmann et al. 2004b). A dietary pattern characterized by higher intakes of meat, margarine, poultry and sauce, and lower intakes of vegetarian dishes, wine, vegetables and whole grains, maximally explained variation in the five selected biomarkers of CVD (HDL-cholesterol, LDL-cholesterol, lipoprotein(a), C-peptide and C-reactive protein concentrations). The identified dietary pattern was strongly associated with CHD. The use of biomarkers of disease as response variables is of interest because it may aid the pathophysiological interpretation of observed associations between dietary patterns and disease. However, the explorative nature of the approach should be kept in mind: in the theoretical situation where the biomarkers used perfectly predict the disease, the method simply identifies the combination of food variables that is most highly correlated with the disease in the study population. Thus, an important consideration is whether findings can be confirmed in diverse populations. After randomly splitting the sample in half, dietary pattern variables identified in the two sub-populations remained associated with CHD in the other sub-population, but the associations were much weaker (Hoffmann et al. 2004b). The RRR method may prove to be a useful tool in nutritional epidemiology, because it can provide insight into the pathophysiological pathway that links dietary patterns to disease. Use of the same set of response variables in different study populations may facilitate the comparison of results from diverse populations. In addition, confirmation of results for RRR or principal components analyses in other populations can be performed by calculating a dietary pattern score using the same weights and food variables or by using a simplified dietary pattern score (Schulze et al. 2003a). An application of dietary pattern analyses is to obtain insight into intercorrelations between dietary variables and the potential for confounding. In addition, adjustment for dietary pattern scores may be an efficient method to address confounding by other dietary variables in studies of individual foods and food components (Schulze et al. 2003b). The combination of analyses of food components, foods and dietary patterns is likely to provide most insight into the relationship between diet and disease risk. Hoffmann K, Boeing H, Boffetta P, Nagel G, Orfanos P, Ferrari P, Bamia C. Comparison of two statistical approaches to predict all-cause mortality by dietary patterns in German elderly subjects. Br J Nutr. 2005 May;93(5):709-16. PMID: 15975171 http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=Retrieve & db=pubmed & dopt=Abstra\ ct & list_uids=15975171 & query_hl=28 Abbreviations: AIC, Akaike's Information Criterion, EPIC, European Prospective Investigation into Cancer and Nutrition, PCA, principal component analysis, RRR, reduced rank regression .... To study the effect on mortality of a specific food or food group consumed regularly is neither promising nor scientifically justified because foods are consumed in combination and intake data are highly correlated. Adjustment for all other foods would result in a vast number of regression parameters to estimate. As a consequence, multicollinearity of the food intakes and the restricted power of a study may lead to unstable parameter estimates with wide confidence intervals. A possible approach is to consider diet-quality scores based on recommended diets or dietary guidelines ( et al. 1994; Kennedy et al. 1995; Trichopoulou et al. 1995, 2003; Huijbregts et al. 1997; Osler & Schroll, 1997; Haines et al. 1999; Kant et al. 2000; McCullough et al. 2002; Seymour et al. 2003). However, diet-quality scores are based on selected dietary components and do not allow for the intake of some food groups. An alternative approach is to construct dietary patterns and to study the combined effect of foods differently weighted in pattern scores (Kumagai et al. 1999; Osler et al. 2001). Principal component analysis (PCA) is the most widely applied statistical method to derive dietary patterns in nutritional science (Randall et al. 1992; Slattery et al. 1998; Hu et al. 2000; Osler et al. 2001; Schulze et al. 2001; Hu, 2002; Schulze & Hu, 2002; Costacou et al. 2003; Van Dam et al. 2003). PCA, or the very similar factor analysis, aims to construct linear combinations of food intakes, which explain a high proportion of the variation in food intakes. If the objective of the study is to describe typical eating patterns in the study population without reference to health outcomes, PCA is a useful tool. However, explaining much variation in food intakes by PCA does not necessarily mean that also much variation in macronutrients such as fatty acids and carbohydrates will be explained. From a theoretical point of view, dietary patterns that focus on the variation in selected nutrients and especially on changes of the mix of energy sources could be more useful in examining the effects of diet on disease incidence or mortality than the classic PCA patterns. The focus on variation in biologically relevant nutrients was the reason to introduce reduced rank regression (RRR) in nutritional epidemiology (Hoffmann et al. 2004a). RRR is a statistical method that is more flexible than PCA because it works with two sets of variables. It aims to construct linear combinations of variables belonging to one set by maximizing the explained variation in variables of the other set. Defining the second set by a few variables that are expected to be related to the health outcome of the study should identify dietary patterns that are possibly associated with the outcome. In nutritional epidemiology, many previous studies have focused on major energy sources because they are quantitatively important in our diets and differences in the mix of energy sources among human populations are correlated with striking variation in rates of many diseases (Willett, 1998). Although large prospective studies on an individual level have not supported an important role of energy sources on cancer (Willett, 2000b) and stroke (He et al. 2003), energy intake from specific types of fat appear to be related to the risk of CVD (Jakobsen et al. 2004; Tanasescu et al. 2004) and type 2 diabetes (Salmeron et al. 2001) that may have small effects on total mortality (Hooper et al. 2001). Therefore, choosing percentages of total energy intake from different macronutrients as a second set of variables in the RRR analysis could be promising in identifying dietary patterns that are meaningful for the development of chronic diseases and for mortality. In the present study, we applied the RRR method to food consumption data collected by means of a food frequency questionnaire and focused on the variation in percentages of energy from saturated fat, monounsaturated fat, polyunsaturated fat, protein and carbohydrates. We also applied the classic PCA method to the same data and compared the performance of the two approaches to predict mortality. We used data from the European Prospective Investigation into Cancer and Nutrition (EPIC) study. This work was involved in the EPIC-Elderly project and considered only dietary patterns in the German EPIC cohorts. Subjects and methods Study population Study participants were volunteers from the two German cohorts in Potsdam and Heidelberg of the EPIC study who were aged 60 years or older at baseline. Participants were recruited between 1994 and 1998. The response rate was 22·7% in Potsdam and 38·3% in Heidelberg (Boeing et al. 1999a). ... Four hundred and four study participants (221 from Potsdam, 183 from Heidelberg) died during the follow-up time, which varied between 4 and 8 years. Because of the high number of 287 diseased who had self-reported prevalent chronic diseases at enrolment we did not exclude these persons from the analysis, but we adjusted for prevalence of chronic diseases in the respective analysis. Data collection Dietary intake information was collected by a self-administered scanner-readable food frequency questionnaire, which included questions on the frequency and portion size of 148 food items eaten during the year preceding enrolment (Schulze et al. 1999). Photographs supported the estimation of portion sizes. The frequency of intake was measured using ten categories, ranging from ‘never’ to ‘five times per d or more’ (Boeing et al. 1999b). Foods were classified into twenty-three food groups based on nutrient profiles or culinary usage according to the common classification of the EPIC project (Slimani et al. 2002). Nutrient intake was calculated using data from the German Food Code BLSII.3 that is a slight modification of BLSII.2 (Dehne et al. 1999). Macronutrient intake estimated from the food frequency questionnaire was validated by the mean macronutrient intake obtained from twelve 24 h dietary recalls. The correlation coefficient between both intakes adjusted for energy varied between 0·58 for carbohydrates and 0·84 for protein (Kroke et al. 1999b). ... .... Results The factor loadings of the first two patterns obtained by PCA and RRR are presented in Table 1. A high positive loading indicates a strong direct association between the food group and the pattern, whereas a high negative loading reflects a strong inverse association. To make visible what the important food groups for each pattern were we omitted all factor loadings between -0·2 and +0·2. The major contributors to the first PCA pattern were potatoes, vegetables, legumes, bread, all types of meat, eggs, sauces and soups which were all positively correlated with the pattern score. These foods are often eaten together at lunch or at supper in German households. The second PCA pattern was characterized by high positive loadings of vegetables, fruits, dairy products, other cereals, vegetable oils and non-alcoholic beverages, as well as by a negative loading of alcoholic beverages other than wine. This pattern reflects prudent nutrition consistent with dietary recommendations. The first RRR pattern had high positive loadings in all types of meat, butter, sauces and eggs, and was inversely associated with bread and fruits. It represents dietary habits expected to be detrimental for human health. Finally, legumes, poultry, fish and margarine were directly associated and butter, sugar and cakes were inversely associated with the second RRR pattern score. Table 1. Factor loadings* of the dietary patterns obtained by different statistical methods in the German cohorts of the European Prospective Investigation into Cancer and Nutrition Elderly study (n 9356) ................................. --------Principal component analysis Reduced rank regression ................................. Food groups----Pattern 1 Pattern 2 Pattern 1 Pattern 2 ................................. Potatoes and other tubers 0·25 - - - Vegetables 0·25 0·44 - - Legumes - 0·23 - c0·20 Fruits - 0·36 -0·20 - - Dairy products - 0·26 - - Pasta, rice and other grain Bread 0·22 - -0·28 - Other cereals - 0·27 - - Red meat 0·42 - 0·35 - Poultry 0·25 - 0·21 0·22 Processed meat 0·35 - 0·51 - Fish and shellfish - - - 0·20 Eggs and egg products 0·21 - 0·20 - Vegetable oils - 0·43 - - Butter - - 0·31 -0·64 Margarine - - - 0·37 Sugar and confectioneary - - - -0·36 Cakes - - - -0·31 Non-alcoholic beverages - 0·31 - - Wine Other alcoholic beverages - -0·31 - - Sauces and condiments 0·38 - 0·26 - Soups and bouillons 0·21 - - - ......................... *Factor loadings that are between -0·2 and +0·2 are not shown. The variation in food groups and energy sources explained by each of the four dietary patterns is summarized in Table 2. The two PCA patterns explained 11·3 and 8·4% of food intake variation, respectively. In contrast, the first and second RRR patterns explained only 4·4 and 6·1% of the total variation in all twenty-three food groups. However, the variation in energy sources accounting for the two RRR patterns were 30·8 and 15·9% and as expected much higher than the proportion of energy sources variation explained by the two PCA patterns (together 7%). Table 2 also gives a more detailed picture of how much variation in every energy source can be explained by the patterns. As indicated, RRR patterns performed better than PCA patterns in explaining the variation in all energy sources. For example, the first RRR pattern alone accounted for 39·1% of variation in energy from saturated fat and simultaneously for much of the variation in monounsaturated fat and carbohydrates. Table 2. Explained variation (%) of food groups and energy sources by dietary patterns obtained by different statistical methods in the German cohorts of the European Prospective Investigation into Cancer and Nutrition Elderly study (n 9356) ........................ ----Principal component analysis Reduced rank regression ----Pattern 1 Pattern 2 Total* Pattern 1 Pattern 2 Total* ........................ Explained variation (%) of: All twenty-three food groups 11·3 8·4 19·7 4·4 6·1 10·5 All five energy sources 4·4 2·6 7·0 30·8 15·9 46·7 Energy from saturated fat (%) 1·5 1·2 2·7 39·1 20·4 59·5 Energy from monounsaturated fat (%) 8·0 0·2 8·2 57·7 0·9 58·6 Energy from polyunsaturated fat (%) 1·0 2·9 3·9 12·2 29·2 41·4 Energy from protein (%) 1·5 0·3 1·8 12·0 28·7 40·7 Energy from carbohydrates (%) 10·0 8·4 18·4 33·2 0·1 33·3 ........................... *Proportion of variation explained by pattern 1 and pattern 2 together. As can be seen in Table 3, mean percentages of energy from different sources varied across quintiles of the two PCA and two RRR dietary pattern scores. Because of the high number of subjects all trends were statistically significant (P<0·0001). However, the most striking variation was that of energy from saturated fat, monounsaturated fat and carbohydrates across quintiles of the first RRR pattern. A high score of the first RRR pattern reflected a diet rich in saturated and monounsaturated fat and poor in carbohydrates. Energy from total fat increased from 25·9% in the lowest quintile to 37·2% in the highest quintile of the pattern score, whereas the proportion of energy from carbohydrates simultaneously decreased from 48·0 to 37·6%. The energy composition in the highest quintile did not meet current dietary recommendations, which advise an intake of less than 35% of energy from fat and more than 45% of energy from carbohydrates (Institute of Medicine of the National Academies, 2002). Table 3. Mean percentages of energy from different sources according to quintiles of dietary pattern scores in the German cohorts of the European Prospective Investigation into Cancer and Nutrition Elderly study (n 9356) ............................... ----Percentage of energy from: Quintiles of dietary pattern scores ----1 2 3 4 5 Trend* ................................ PCA† Pattern 1 Saturated fat 13·4 13·6 13·7 14·0 14·4 > Monounsaturated fat 10·9 11·3 11·6 11·9 12·5 > Polyunsaturated fat 5·6 5·9 5·9 6·0 6·1 > Total fat 29·9 30·8 31·2 31·9 33·0 ? Protein 13·4 13·6 13·7 13·8 14·1 > Carbohydrates 45·8 44·6 43·6 42·6 40·2 < PCA Pattern 2 Saturated fat 14·1 14·3 13·9 13·6 13·2 < Monounsaturated fat 11·8 11·9 11·6 11·6 11·4 < Polyunsaturated fat 5·6 5·8 5·9 6·0 6·2 > Total fat 31·5 32·0 31·4 31·2 30·8 < Protein 13·4 13·7 13·8 13·9 13·8 > Carbohydrates 39·7 42·7 43·9 44·7 45·5 > RRR‡ Pattern 1 Saturated fat 11·3 12·7 13·7 14·8 16·5 > Monounsaturated fat 9·5 10·6 11·6 12·5 14·0 > Polyunsaturated fat 5·1 5·6 5·9 6·3 6·7 > Total fat 25·9 28·9 31·2 33·6 37·2 > Protein 12·6 13·3 13·7 14·1 14·8 > Carbohydrates 48·0 45·8 43·6 41·5 37·6 < RRR Pattern 2 Saturated fat 16·1 14·4 13·3 12·7 12·6 < Monounsaturated fat 12·2 11·7 11·3 11·3 11·7 < Polyunsaturated fat 4·7 5·3 5·8 6·4 7·4 > Total fat 33·0 31·4 30·4 30·4 31·7 < Protein 12·0 13·1 13·7 14·4 15·4 > Carbohydrates 42·1 43·5 44·3 44·2 42·5 > ............................... PCA, principal component analysis; RRR, reduced rank regression. *< = decreasing quintiles. > = increasing quintiles. All trends across quintiles of dietary patterns were significant (P<0·0001). †PCA and RRR were applied to twenty-three food groups defined in Table 1. ‡In the RRR analysis the percentages of energy from saturated fat, monounsaturated fat, polyunsaturated fat, protein and carbohydrates were used as response variables. The relative risks of all-cause mortality associated with an increase in each pattern score by one standard deviation are presented in Table 4. The relative risks adjusted for centre and sex were significantly different from 1·0 for the first PCA pattern (relative risk 1·12; 95% CI 1·01, 1·24) and for the first RRR pattern (relative risk 1·25; 95% CI 1·14, 1·40). After additional adjustment for prevalent cancer, prevalent CHD, prevalent diabetes, prevalent hypertension, BMI and waist:hip ratio, the first PCA pattern was no more significant. In contrast, the first RRR pattern remained significant even after further adjustment for smoking status, educational level, physical activity at work, physical activity at leisure time and total energy intake (relative risk 1·20; 95% CI 1·09, 1·31). Comparing the goodness of fit of the fully adjusted models by Akaike's Information Criterion (AIC) suggests the use of RRR patterns (AIC=5580) instead of PCA patterns (AIC=5592). Also a fully adjusted model with percentages of energy from macronutrients chosen as predictors had a lower goodness of fit (AIC=5594). Actually, no single energy source had a significant effect on mortality. Moreover, substituting the RRR patterns by the Mediterranean diet-quality score (Trichopoulou et al. 1995, 2003) was associated with a decrease of model fit and a loss of significance (data not shown). Table 4. Relative risk and 95% CI of mortality according to standardized* increase of pattern scores obtained by different statistical methods in the German cohorts of the European Prospective Investigation into Cancer and Nutrition Elderly study (n 9356) ................................ ----Principal component analysis (PCA)† Reduced rank regression (RRR)‡ ----Pattern 1 Pattern 2 Pattern 1 Pattern 2 Relative risk 95% CI Relative risk 95% CI Relative risk 95% CI Relative risk 95% CI ................................ Model 1‡‡ 1·12 1·01, 1·24 0·91 0·82, 1·01 1·25 1·14, 1·40 0·98 0·89, 1·07 Model 2 1·11 0·98, 1·23 0·92 0·82, 1·02 1·23 1·12, 1·34 0·94 0·86, 1·04 Model 3 1·10 0·96, 1·28 0·99 0·89, 1·10 1·20 1·09, 1·31 0·96 0·87, 1·06 ............................... *A unit increase of the pattern score is an increment of SD. †PCA and RRR were applied to twenty-three food groups defined in Table 1. ‡In the RRR analysis the percentages of energy from saturated fat, monounsaturated fat, polyunsaturated fat, protein, and carbohydrates were used as response variables. ‡‡Model 1: adjusted for centre and sex. Model 2: adjusted for centre, sex, prevalent cancer, CHD, diabetes and hypertension, BMI and waist:hip ratio. Model 3: adjusted for all variables included in model 2 and additionally for smoking status, education level, physical activity at work, physical activity at leisure time and total energy intake. Table 5 shows the relative risks of death according to quintiles of dietary pattern scores by taking into account all confounders as before. Again, the first RRR pattern was the only significant predictor of total mortality. The relative risks across increasing quintiles of this pattern score were 1·0, 1·01, 0·96, 1·32 and 1·61 (95% CI 1·17, 2·21; P for trend: 0·0004). Interpreting this relationship as effect of energy composition means that only a simultaneous high proportion of energy from total fat and protein and low proportion of energy from carbohydrates was associated with increased mortality risk. Table 5. Relative risk* and 95% CI of mortality according to quintiles of pattern scores in the German cohorts of the European Prospective Investigation into Cancer and Nutrition Elderly study (n 9356) ......................... --------Quintiles of dietary patterns --------1 2 2 4 5 Dietary patterns---Relative risk Relative risk 95% CI Relative risk 95% CI Relative risk 95% CI Relative risk 95% CI P for trend ........................... PCA pattern 1 1·0 0·83 0·57, 1·22 1·00 0·70, 1·45 1·03 0·70, 1·51 1·06 0·68, 1·65 0·50 PCA pattern 2 1·0 0·91 0·68, 1·22 0·90 0·66, 1·23 1·10 0·81, 1·51 0·80 0·55, 1·15 0·61 RRR pattern 1 1·0 1·01 0·70, 1·46 0·96 0·66, 1·38 1·32 0·95, 1·85 1·61 1·17, 2·21 0·0004 RRR pattern 2 1·0 0·87 0·63, 1·21 0·81 0·57, 1·13 1·07 0·78, 1·48 0·96 0·70, 1·33 0·74 Shortened† RRR 1 1·0 1·07 0·73, 1·58 1·25 0·87, 1·80 1·35 0·94, 1·94 1·34 0·93, 1·94 0·062 Simplified‡ RRR 1 1·0 0·99 0·67, 1·45 1·24 0·86, 1·77 1·26 0·88, 1·79 1·31 0·91, 1·87 0·061 .......................... PCA, principal component analysis; RRR, reduced rank regression. *Adjusted for centre, sex, prevalent cancer, CHD, diabetes and hypertension, BMI, waist:hip ratio, smoking status, education level, physical activity at work, physical activity at leisure time and total energy intake. †Derived from the original pattern by neglecting all food groups with factor loadings between -0·2 and +0·2. ‡Derived from the shortened pattern by setting all positive factor loadings equal to +1 and all negative loadings equal to -1. In Table 6, means and percentages of different baseline characteristics according to quintiles of the first RRR pattern are summarized. Individuals with high scores had a higher BMI, were more likely to come from the Potsdam centre, tended to smoke, had a lower education level and were more likely to be unemployed. Confounding effects on relative risk estimates due to these associations were taken into account as before. Table 6. Baseline characteristics according to quintiles of the first reduced rank regression (RRR) pattern score in the German cohorts of the European Prospective Investigation into Cancer and Nutrition Elderly study (n 9356) ......................... ----Quintiles of the first RRR pattern Characteristics 1 2 3 4 5 P for trend ......................... Mean age (years) 62·5 62·5 62·7 62·7 62·6 0·02 Mean BMI (kg/m2) 27·1 27·3 27·7 27·7 28·2 <0·0001 Mean waist:hip ratio 0·90 0·89 0·89 0·90 0·91 0·0007 Potsdam (%) 0·46 0·53 0·57 0·59 0·57 <0·0001 Male (%) 0·43 0·55 0·56 0·51 0·46 0·59 Current smokers (%) 0·11 0·11 0·12 0·16 0·20 <0·0001 Former smokers (%) 0·42 0·36 0·36 0·37 0·38 0·08 Education level (%) None/primary school 0·36 0·40 0·41 0·42 0·43 <0·0001 Technical school 0·30 0·31 0·31 0·30 0·30 0·53 University degree 0·34 0·29 0·28 0·28 0·27 <0·0001 Physical activity at work (%) Unemployed 0·74 0·79 0·81 0·80 0·79 <0·0001 Sedentary occupation 0·16 0·13 0·12 0·11 0·12 <0·0001 Standing occupation 0·07 0·07 0·06 0·07 0·07 0·56 Manual work 0·02 0·01 0·01 0·02 0·02 0·79 Physical activity at leisure time (%) Minimal 0·23 0·22 0·19 0·21 0·22 0·32 Moderate 0·32 0·32 0·31 0·33 0·31 0·56 Heavy 0·45 0·46 0·50 0·46 0·47 0·17 To examine the robustness of our findings and to reduce the dependency of the dietary pattern from the data we simplified the first RRR pattern score. At first, we shortened the score by neglecting all food groups with loadings between -0·2 and +0·2. Then we formed a simplified pattern (Schulze et al. 2003) by defining the coefficients of the remaining food groups as +1 or as -1 depending on the sign of the factor loading. The resulting simplified pattern score is simply the sum of the standardized food groups of red meat, poultry, processed meat, butter, sauces and eggs, minus the sum of the standardized food groups of bread and fruits. The shortened and the simplified versions of the first RRR pattern were separately used as a predictor of mortality (see Table 5). The relative risks for the shortened and simplified patterns were attenuated as compared with the full pattern and the trend across increasing quintiles of the score was only borderline significant in both cases (P=0·062 and P=0·061). As further sensitivity analysis we excluded all participants with prevalent chronic diseases (cancer, CHD, diabetes) and prevalent hypertension. The first RRR pattern only slightly changed after exclusion. The factor loadings were almost the same as before with differences of corresponding loadings varying between -0·01 and +0·01. The proportion of variation explained by the first RRR pattern was 30·7%, which represented a decrease of only 0·1%. The adjusted relative risk associated with an increase of the first RRR score by SD was 1·24 (95% CI 1·05, 1·48) and therefore similar to the result obtained with the total sample, although it was less precise because of the exclusion of a relatively large proportion of deaths. Discussion In the German cohorts of the EPIC-Elderly study we found a significant association between a specific dietary pattern and total mortality. This pattern was derived by the statistical RRR method recently introduced in nutritional epidemiology (Hoffmann et al. 2004a). The RRR pattern score adequately reflects variation in energy sources and corroborates the hypothesis that certain combinations of energy sources affect human health. A high RRR pattern score, which was associated with high intake of fat and protein and low intake of carbohydrates, increased the risk of death. Subjects with a pattern score belonging to the highest quintile obtained on average 37·2% of their energy from fat and 37·6% from carbohydrates and thus did not meet current dietary recommendations (Institute of Medicine of the National Academies, 2002). Food groups that contributed to this unfavourable pattern of energy sources were red meat, poultry, processed meat, butter, sauces and eggs, whereas a high intake of bread and fruits decreased the pattern score. These food groups did not have simultaneously high loadings in one of the first two PCA patterns that reflect a typical German diet and a diet supposed to be prudent. The most appealing advantage of the new RRR method compared with PCA or traditional factor analysis is the capability to incorporate prior information. The type of prior information needed is the knowledge of diet-related variables that are expected to be predictive for the health outcome under study. Previous correlation and epidemiological studies indicated that the mix of energy sources in diet could be important for health (Hu et al. 1997; Willett, 1998; Liu & Manson, 2001; Salmeron et al. 2001; Mann, 2002). Evidence clearly suggests that dietary fat affects the lipid and lipoprotein risk profile (Kris-Etherton et al. 2002; Lichtenstein, 2003; Wolfram, 2003; Pelkman et al. 2004). Because the number of energy sources is markedly smaller than the number of foods, focusing on the set of energy sources as done in the present RRR analysis results in a reduction of dimension. This reduction can be considered gain by prior knowledge since explaining a high proportion of variation in a smaller set of variables is much easier than in a larger set of variables. The result of the present study, that in contrast to PCA patterns the first RRR pattern was significantly associated with all-cause mortality, confirms prior knowledge and corroborates the hypothesis that variation in the mix of energy sources is meaningful for mortality. However, whether the macronutrient composition or other dietary components highly correlated with some macronutrients are linked to disease prevention and higher life expectancy remains unclear (Sacks & Katan, 2002). Randomized clinical trials and intervention studies could help to find out the crucial components and aspects of diet that could be the response variables in future RRR analyses. In the current study several limitations should be considered. First, dietary data were collected by food frequency questionnaire and, therefore, were subject to considerable measurement error. Because dietary patterns use intakes of all foods the effect of measurement error on the pattern score will be high but in general cannot be quantified. Second, dietary intake assessed by a single food frequency questionnaire referred to a period of 1 year and cannot be considered lifetime exposure. Diet may change over lifetime and these changes in dietary habits may have an impact on mortality. Repeated measurements of dietary intake over many years would be necessary to model diet history and to explore the short- and long-term effects on life expectancy. Third, the definition and number of food groups may have an effect on the results of a pattern analysis. Because of adherence to the food group classification of the EPIC project and the necessary consistency and comparability of dietary data from both German EPIC cohorts, we chose fewer food groups in the present study than in previous studies (Hoffmann et al. 2004a,. A drawback of a small number of food groups is that foods with potential different health effects are possibly combined in the same group and important variation in dietary intake can possibly not be reflected by patterns derived subsequently. Fourth, residual confounding due to insufficient consideration of socio-demographic, occupational and other environmental factors may have caused bias in relative risk estimation. For example, bias may probably be attributed to modelling the important exposures, smoking and physical activity, only by categorical variables with few categories. Interactions between nutrition, physical activity and obesity were not considered in the present study although these risk factors could be components of the same sufficient condition for mortality following the concept of Rothman (1976). Fifth, the low response rate of the cohort study calls into question whether the cohort is representative for the German population. Distinctions in dietary habits may lead to patterns derived in the study population that are quite different from those of the target population. Sixth, the findings of an RRR analysis are not completely reproducible by other studies because they depend on the data at hand. The construction of simplified patterns may be helpful to find more robust results and possibly form a basis for dietary recommendations. Seventh, the choice of response variables is not unique and seems to be somewhat arbitrary. Two RRR analyses with different sets of responses will, in general lead to different patterns. However, the non-uniqueness of responses can also be considered a strength of the RRR method because RRR can allow for new findings in research and additional information sources by modifying the set of response variables in future studies. For example, in the case of available blood samples, biomarkers may be more informative than nutrients because they can provide the link between diet and the health outcome. Thus, choosing appropriate biomarkers as responses in the RRR analysis is a promising way to quantify indirect effects of diet on chronic diseases and mortality by specifying possible pathways (Hoffmann et al. 2004b). In conclusion, the RRR method is a powerful tool to derive dietary patterns in nutritional epidemiology. It is a flexible method that combines the strength of PCA to consider inter-correlations of dietary intakes and the advantage of diet-quality scores to allow for use of biologically plausible prior information. The application of RRR in the present study to evaluate the impact of diet on mortality suggests that a diet rich in fat and poor in carbohydrates decreases life expectancy. Al Pater, PhD; email: old542000@... __________________________________________________ Quote Link to comment Share on other sites More sharing options...
Recommended Posts
Join the conversation
You are posting as a guest. If you have an account, sign in now to post with your account.
Note: Your post will require moderator approval before it will be visible.