Journal of Nutrition LabDiet, Your World of Nutritional Answers

Home Help [Feedback] [For Subscribers] [Archive] [Search] [Contents]
QUICK SEARCH:   [advanced]


     

This Article
Right arrow Abstract Freely available
Right arrow Full Text (PDF)
Right arrow Purchase Article
Right arrow View Shopping Cart
Right arrow Alert me when this article is cited
Right arrow Alert me if a correction is posted
Services
Right arrow Similar articles in this journal
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Weismayer, C.
Right arrow Articles by Wolk, A.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Weismayer, C.
Right arrow Articles by Wolk, A.
© 2006 American Society for Nutrition J. Nutr. 136:1582-1587, June 2006

Nutritional Epidemiology

Changes in the Stability of Dietary Patterns in a Study of Middle-Aged Swedish Women1

Christoph Weismayer*,2, James G. Anderson{dagger} and Alicja Wolk*

* Division of Nutritional Epidemiology, The National Institute for Environmental Medicine, Karolinska Institutet, Stockholm, Sweden and {dagger} Department of Sociology and Anthropology, Purdue University, Lafayette, IN

2 To whom correspondence should be addressed. E-mail: weismayerc{at}gmail.com.


   ABSTRACT
TOP
 ABSTRACT
SUBJECTS AND METHODS
 RESULTS
 DISCUSSION
 LITERATURE CITED
 
Dietary patterns reflecting food habits may be associated with chronic diseases, yet little is known about the stability of these patterns. The objective of this study was to observe over time the stability of dietary patterns measured with exploratory and confirmatory factor analysis. Four random subsamples of 1000 women between 49 and 70 y old were chosen from >60,000 women included in the Swedish Mammography Cohort. Subjects in these subsamples were administered a FFQ 4, 5, 6, or 7 y after the baseline questionnaire; 3607 of the women responded (90% response rate). The stability of dietary patterns was evaluated with Spearman correlation coefficients between pattern scores at baseline and follow-ups and by a test of internal stability, which evaluated the significance of changes within patterns between baseline and follow-up. We found 3 major dietary patterns: a healthy pattern, a Western pattern, and an alcohol pattern. Correlations between explored dietary pattern scores at baseline and at follow-up decreased from 0.59 (P < 0.01) after 4 y to 0.50 (P < 0.01) after 7 y for the healthy pattern, from 0.47 (P < 0.01) to 0.39 (P < 0.01) for the Western pattern and from 0.54 (P < 0.01) to 0.46 (P < 0.01) for the alcohol pattern. After 4 and 5 y, there was no evidence for internal instability in any of the 3 patterns. The Western pattern became internally unstable after 6 and 7 y and the alcohol pattern was unstable after 7 y. Our findings for this specific population suggest that in longitudinal studies, dietary exposures should be updated after at least 7 y.

KEY WORDS: • diet • factor analysis • dietary pattern • food frequency questionnaire • stability

Factor analysis has been used increasingly to help explain the occurrence of chronic diseases associated with dietary factors because factor analysis by-passes the problem that plagues single-food analysis, i.e., multiple dietary confounders (1–5). The stability of dietary patterns over time is important because stable dietary pattern scores, which aim to measure dietary patterns, should reflect eating behavior with accuracy even several years after baseline data collection without introducing an increasing measurement error with time. Consequently, the cost of maintaining large prospective cohorts would decrease because stable dietary patterns would not require frequent updates of dietary exposures. The question is important because recent publications use dietary pattern scores to predict health outcomes. One study updated exposures every 2 y (6), whereas 3 studies did not update exposures but followed up disease outcomes 9.3 (7), 10 (8), and 23 y (9) after baseline. Thus, assumptions about the stability of pattern scores vary widely in current research.

Despite these important implications, very little is known about the stability of diet over time. Shorter-term (up to 1 y of follow-up) methodological studies of FFQ focused on shorter-term reproducibility of pattern scores, and observed high correlations between pattern scores at baseline and after a 1-y follow-up (10,11). Khani et al. (10) identified healthy, Western, and drinker patterns and correlations of 0.63, 0.68, and 0.73 after 1 y, respectively, by analyzing a different subsample of the Swedish Mammography Cohort. Hu et al. (11) identified a prudent pattern and a Western pattern; the correlations between the 2 FFQ were 0.70 for the prudent pattern and 0.67 for the Western pattern, also after a 1-y follow-up. Both of these studies were smaller and analyzed food groups of nutritionally similar items.

None of the previous studies focused on the internal stability of dietary patterns, although it is possible that high correlations between pattern scores are found for a population at baseline and at follow-up, despite large changes in the internal stability of a pattern. We define internal stability as a lack of change in the correlations between the specific food items that define a pattern as well as stability of the means and SD of these items. As these variables change, these changes can offset each other to produce identical dietary pattern scores in a population and give an impression of stable patterns. None of the previous studies used confirmatory factor analysis, a more advanced method of analyzing factor structures, which is appropriate after the existence of a factor has been explored (12). The additional value of the present study is that with confirmatory factor analysis, these instabilities can be quantified and tested for significance.


   SUBJECTS AND METHODS
TOP
 ABSTRACT
 SUBJECTS AND METHODS
RESULTS
 DISCUSSION
 LITERATURE CITED
 
    Subjects. From 1987 to 1990, a population-based mammography-screening program was introduced in 2 counties in central Sweden (Västmanland and Uppsala County). All women born between 1914 and 1948 were invited to participate in the screening and received a FFQ. The women who answered the FFQ (n = 66,651, response rate = 73.4%) were included in the Swedish Mammography Cohort.

The data analyzed in this study comprise a subset of the Swedish Mammography Cohort. To evaluate the stability of dietary patterns (identified by exploratory and validated with confirmatory analysis) over time, we analyzed data from 4 randomly selected subsets with 1000 participants each, who completed 2 identical FFQ (at baseline and after 4, 5, 6, or 7 y). This design was chosen over 1 subgroup with multiple follow-ups to avoid survey learning effects.

    Food groupings. The self-administrated FFQ asked women about their mean intake of 67 food items over the past 6 mo. In general, the use of self-administered FFQ was found to yield valid and reliable results in an elderly Dutch population (13). Factor analysis with this specific survey also found yielded reproducible and valid factor scores (10).

For each food item, there were 8 frequency response categories, defined as follows: never/seldom, 1–3 times/mo, 1 time/wk, 2–3 times/wk, 4–6 times/wk, 1 time/d, 2–3 times/d, and ≥4 times/d. To reduce the complexity of the data, we combined some of the food items into groups. After recoding the responses as frequencies per day and grouping food items, 25 food groups were retained for analysis. The grouping scheme was based on the similarity of nutrient profiles or culinary usage among the foods and was the same as that used in a previous study of the Swedish Mammography Cohort (10) (Appendix 1).


View this table:
[in this window]
[in a new window]
  		APPENDIX 1 Grouping scheme of individual items from the FFQ

 
    Statistical analysis. We conducted exploratory factor analysis to simplify the data and combine correlated food groups into factor scores. Factor analysis was performed for each of the 4 subgroups separately at baseline and at follow-up. Factors were rotated with an orthogonal transformation (Verimax rotation function in LISREL 8.0) to achieve a factor structure that is easier to interpret. To decide how many factors to extract, we considered the scree plot and the factor structure as well as the interpretability of the structure (14). As a starting value, we ran factor analysis by extracting 2–8 factors. We then performed confirmatory factor analysis to validate the explored findings. Confirmatory factor analysis is performed after exploratory analysis and is a step-by-step procedure that aims to exclude items with low loadings or items that are not hypothesized to contribute to a factor; all loadings are recalculated based only on those items that are retained (12). Confirmation was attempted only for those loadings >0.2.

After confirmed factor loadings were calculated, confirmed factor scores were calculated by standardizing each food group frequency and multiplying these standardized scores by the confirmed loadings for each food group. The results of these multiplications were then summed for each pattern and each person to obtain dietary pattern scores for the population.

To measure the stability of the factor scores over time, we calculated a Spearman correlation coefficient between factor scores at baseline and at follow-up for all 4 groups and all identified patterns for both explored and confirmed factor scores.

To measure internal stability of patterns, we tested the significance of changes in the covariance matrix between baseline and follow-up. Covariance matrices take into account correlations between the specific food items that define a pattern as well as the means and SD of these items. In the hypothesized covariance matrix, those items that are not included in a confirmed factor are hypothesized to be 0, i.e., the more the observed covariance deviates from this assumption, the higher the {chi}2 and the lower the overall fit of the model and the higher the statistical significance of the difference between the observed and hypothesized matrices. We evaluated only the size of the differences in the covariance matrices at baseline and at follow-up for all of those food groups with an explored or confirmed loading >0.2 in any of the 4 groups at either baseline or follow-up. We then tested whether the difference between the hypothesized and observed covariance was significant. A significant difference between the 2 matrices indicates that a pattern is not internally stable (15). The test of equal covariance matrices complements the information obtained from calculating correlation coefficients between factor scores, because a factor score summarizes a complex dietary pattern with a single score. Pattern scores can be highly correlated even if the underlying factor loadings, means, and SD of the food groups that define the pattern undergo drastic changes, if these changes offset each other. Therefore, highly correlated factor scores at baseline and at follow-up may give an incomplete picture and may be misunderstood to represent stable and thus unchanged behavior.


   RESULTS
TOP
 ABSTRACT
 SUBJECTS AND METHODS
 RESULTS
DISCUSSION
 LITERATURE CITED
 
At baseline and at all 4 follow-ups, factor analysis suggested that there were 3 major patterns for all groups, i.e., a pattern made up of mainly of healthy items such as fruits, tomatoes, vegetables, cereal, and fish (a pattern we labeled the healthy pattern); a pattern that consisted of meat, processed meat, fried potatoes, soft drinks, and sweets (Western pattern); and a pattern made up of beer, wine, and liquor consumption as well as snack consumption (alcohol pattern) (Table 1). Additional factors could have been extracted but were either difficult to interpret or were dominated by only one high loading. Tea dominated the 4th factor as the only food group with a high loading, and high-fat dairy products dominated the 5th factor, also with no high loadings from any other food group. We saved the rotated solution as 3 major factor scores, which measured underlying latent dietary patterns.


View this table:
[in this window]
[in a new window]
  		TABLE 1 Exploratory factor loadings for the 4 study groups of middle-aged and elderly Swedish women at baseline (BL) and follow-up (FU)1–3

 
After factor loadings were recalculated using confirmatory analysis (Table 2), Spearman correlation coefficients for each of the dietary patterns between the baseline and at follow-ups for each of the 4 groups were calculated. The correlations between the explored factor scores were 0.59, 0.57, 0.59, and 0.50 for the healthy pattern; 0.47, 0.48, 0.51, and 0.39 for the Western pattern; and 0.54, 0.66, 0.58, and 0.46 for the alcohol pattern after 4, 5, 6, and 7 y, respectively. The correlations for the confirmed factor scores were 0.63, 0.63, 0.62, and 0.54 for the healthy pattern; 0.60, 0.54, 0.56, and 0.57 for the Western pattern; and 0.73, 0.76, 0.70, and 0.75 for the alcohol pattern after 4, 5, 6, and 7 y, respectively.


View this table:
[in this window]
[in a new window]
  		TABLE 2 Confirmatory factor loadings for the 4 study groups of middle-aged and elderly Swedish women at baseline (BL) and follow-up (FU)12

 
We calculated the mean food group consumption for each food group that contributed to the core of any dietary pattern at baseline and at follow-ups (Table 3). There was no evidence of a difference in the means for 10, 6, 6, and 2 of 25 food groups after 4, 5, 6, and 7 y respectively. Conversely, there is evidence (P ≤ 0.01) that 3, 7, 8, and 11 of the 18 food groups underwent significant changes after 4, 5, 6, and 7 y.


View this table:
[in this window]
[in a new window]
  		TABLE 3 Two-tailed t test of baseline and follow-up frequencies of 4 subsamples of middle-aged and elderly Swedish women with 3 major dietary patterns1

 
The correlations between the individual food groups also declined in the size of the correlations over time because the range of correlations decreased from 0.47–0.76 after 4 y to 0.38–0.71 after 7 y, with no correlation after 7 y exceeding the size of the correlation after the 4-y follow-up (Table 4).


View this table:
[in this window]
[in a new window]
  		TABLE 4 Spearman correlations between 18 food groups at baseline and follow-up (FU) in the 4 study groups of middle-aged and elderly Swedish women following 3 dietary patterns12

 
We evaluated the internal stability of the dietary patterns, e.g., the stability of the covariance matrix for each confirmed pattern between baseline and follow-up. There were no significant signs of instability within any of the 3 patterns after 4 and 5 y of follow-up. However, significant instabilities arose for the Western pattern after 6 y (P = 0.01) and for the Western (P = 0.02) and alcohol patterns (P = 0.01) after 7 y (Table 5).


View this table:
[in this window]
[in a new window]
  		TABLE 5 Stability of dietary patterns after 4–7 y measured by absolute values of Spearman correlations between explored and confirmed pattern scores at baseline and at follow-up (FU)

 

   DISCUSSION
TOP
 ABSTRACT
 SUBJECTS AND METHODS
 RESULTS
 DISCUSSION
LITERATURE CITED
 
The results for the studied population of middle-aged and elderly Swedish women provided evidence that dietary behavior changes over time. The size of the Spearman correlations between the explored factors at baseline and at follow-up decreased from 0.47–0.59 after 4 y to 0.39–0.50 after 7 y. The range of the observed correlations after 4 y from the baseline indicates that these correlations are lower than those observed in reproducibility studies in which the 1-y follow-up was 0.63–0.73 for Khani et al. (10), using a subsample of the Swedish Mammography Cohort and 0.67–0.70 in the study of Hu et al. (11). The frequencies reported changed gradually over time, and correlations between food group frequencies decreased gradually. We found a consistent decrease in the size of the Spearman correlations between the factors explored at baseline and at follow-up. Interestingly, the confirmed factor scores (which excluded those food groups that did not have at least 1 loading in any group >0.2) were relatively stable, with Spearman correlations ranging from 0.60 to 0.73 after 4 y and from 0.54 to 0.75 after 7 y. We speculated that by focusing on only those food groups with a loading > 0.2 when we performed the confirmatory analysis, we eliminated some potential sources of variation within the patterns, and therefore observed generally higher correlations over time compared with the lower correlations the of explored scores.

However, the confirmed alcohol pattern and Western pattern (which had the highest Spearman correlations between pattern scores after 7 y) showed significant internal instabilities, indicating that the high confirmed correlations do not reveal shifts in the composition of the pattern. It is important to note that a similar conclusion would be difficult to draw with traditional methods because the conventional focus on correlations between dietary pattern scores is incapable of detecting shifts that offset each other within a dietary pattern. Our findings highlight the existence of a difference between the stability of an overall pattern score (explored or confirmed) and the stabilities of those coefficients that are needed to calculate the score. Ideally, longitudinal studies should base their conclusions not only on highly correlated dietary pattern scores but also on scores that were derived from a stable covariance matrix, e.g., the nutritional behavior we are attempting to measure with the score should closely resemble the behavior at baseline.

Overlooking such conclusions can have important consequences, especially for those studies in which these shifts change the association of a pattern with an outcome, e.g., an identical score on the alcohol pattern may produce different risk ratios for certain diseases when beer consumption replaces spirits and correlations between different alcoholic drinks change. In contrast, the internal stability of the healthy pattern remained high even after 7 y, whereas the correlations between explored and confirmed factor scores decreased. This result was unexpected. We expected the results of the healthy dietary pattern to show higher stability on all such measures because of the social desirability to give answers that closely resemble recommended levels. It is possible that lower levels of internal stability could be a reflection of actual changes in diet or changing perception of what a healthy diet entails.

Based on this population, frequent updates of dietary exposure are likely not necessary, but become increasingly necessary after 6–7 y. Studies that are updated only after ≥7 y risk the validity of their dietary exposures because complex changes in the diet increase with time.

The biggest strength of this study is also a weakness because the above results were not obtained from 1 subgroup of a general population that was followed up every year but from 4 independent subgroups that were followed up at different times. Our strategy was chosen to eliminate potential "in-learning" effects if the same group were followed up every year. We assumed that a design with 4 independent groups with 4 different follow-up intervals would eliminate this effect. Further strengths of the study are that it is population based and relatively large in size.

A limitation of the study is that even if it is reasonable to expect similar trends, such as decreased stability of pattern scores and decreased internal stability of patterns over time, in other populations, our results, based on a population of middle-aged and elderly women, are not directly generalizeable to other populations. Changes in diet over time may depend on sex, age of study participants, secular trends in local food markets, and many other factors. Therefore, a decision concerning how often diet should be measured in prospective cohorts requires stability studies within the populations studied; such studies should include measures of the stability of the pattern scores as well as their internal stability.


   FOOTNOTES
 
1 Supported by grants from the Cancer Research Fund International and the Swedish Cancer Society. Back

Manuscript received 29 June 2005. Initial review completed 15 August 2005. Revision accepted 14 February 2006.


   LITERATURE CITED
TOP
 ABSTRACT
 SUBJECTS AND METHODS
 RESULTS
 DISCUSSION
 LITERATURE CITED
 

1. Kant AK, Schatzkin A, Block G, Ziegler RG, Nestle M. Food group intake patterns and associated nutrient profiles of the US population. J Am Diet Assoc. 1991;91:1532–7.[Medline]

2. Ursin G, Ziegler RG, Subar RF, Graubard BI, Haile RW, Hoover R. Dietary patterns associated with a low fat diet in the national health examination follow-up study. Identification of potential confounders for epidemiologic analysis. Am J Epidemiol. 1993;137:916–27.[Abstract/Free Full Text]

3. Hu FB, Rimm EB, Stampfer MJ, Ascheiro A, Spiegelmann D, Willett WC. Prospective study of major dietary patterns and risk of coronary heart disease in men. Am J Clin Nutr. 2000;72:912–21.[Abstract/Free Full Text]

4. Hu FB. Dietary pattern analysis: a new direction in nutritional epidemiology. Curr Opin Lipidol. 2002;13:3–9.[Medline]

5. Newby PK, Tucker KL. Empirically derived eating patterns using cluster or factor analysis: a review. Nutr Rev. 2004;62:177–203.[Medline]

6. Schulz M, Nothlings U, Hoffmann K, Bergmann MM, Boeing H. Identification of a food pattern characterized by high-fiber and low-fat food choices associated with low prospective weight change in the EPIC-Potsdam Cohort. J Nutr. 2005;135:1183–9.[Abstract/Free Full Text]

7. Balder HF, Goldbohm RA, van den Brandt PA. Dietary patterns associated with male lung cancer risk in the Netherlands Cohort Study. Cancer Epidemiol Biomarkers Prev. 2005;14:483–90.[Abstract/Free Full Text]

8. Kim MK, Sasaki S, Otani T, Tsugane S, for the Japan Public Health Center-based Prospective Study Group. Dietary patterns and subsequent colorectal cancer risk by subsite: a prospective cohort study. Int J Cancer. 2005;150:790–8.

9. Montonen J, Knekt P, Harkanen T, Jarvinen R, Heliovaara M, Aromaa A, Reunanen A. Dietary patterns and the incidence of type 2 diabetes. Am J Epidemiol. 2005;161:219–27.[Abstract/Free Full Text]

10. Khani BR, Ye W, Terry P, Wolk A. Reproducibility and validity of major dietary patterns among Swedish women assessed with a food-frequency questionnaire. J Nutr. 2004;134:1541–5.[Abstract/Free Full Text]

11. Hu FB, Rimm E, Smith Warner SA, Feskanich D, Stampfer MJ, Ascherio A, Sampson L, Willett WC. Reproducibility and validity of dietary patterns assessed with a food-frequency questionnaire. Am J Clin Nutr. 1999;69:243–9.[Abstract/Free Full Text]

12. Byrne BM. Structural equation modeling with LISREL, PRELIS, and SIMPLIS: basic concepts, applications and programming. Mahway, NJ: Lawrence Erlbaum Associates; 1998.

13. Klipstein-Grobusch K, den Breeijen JH, Goldbohm RA, Geleijnse JM, Hoffman A, Grobbee DE, Witteman JC. Dietary assessment in the elderly: validation of a semiquantitative food frequency questionnaire. Eur J Clin Nutr. 1998;52:588–96.[Medline]

14. Terry P, Frank B, Hansen H, Wolk A. A prospective study of major dietary patterns and colorectal cancer risk in women. Am J Epidemiol. 2001;154:1143–9.[Abstract/Free Full Text]

15. Johnson DE. Applied multivariate methods for data analysts. Pacific Grove, CA: Brooks/Cole Publishing Company; 1998.




This article has been cited by other articles:


Home page
 J. Nutr.Home page
S. R. Crozier, S. M. Robinson, K. M. Godfrey, C. Cooper, and H. M. Inskip
Women's Dietary Patterns Change Little from Before to During Pregnancy
J. Nutr., October 1, 2009; 139(10): 1956 - 1963.
[Abstract] [Full Text] [PDF]

Home page
 Arch Intern MedHome page
A. Akesson, C. Weismayer, P. K. Newby, and A. Wolk
Combined Effect of Low-Risk Dietary and Lifestyle Behaviors in Primary Prevention of Myocardial Infarction in Women
Arch Intern Med, October 22, 2007; 167(19): 2122 - 2127.
[Abstract] [Full Text] [PDF]

Home page
 J. Nutr.Home page
S. A. McNaughton, G. D. Mishra, A. M. Stephen, and M. E. J. Wadsworth
Dietary Patterns Throughout Adult Life Are Associated with Body Mass Index, Waist Circumference, Blood Pressure, and Red Cell Folate
J. Nutr., January 1, 2007; 137(1): 99 - 105.
[Abstract] [Full Text] [PDF]

This Article
Right arrow Abstract Freely available
Right arrow Full Text (PDF)
Right arrow Purchase Article
Right arrow View Shopping Cart
Right arrow Alert me when this article is cited
Right arrow Alert me if a correction is posted
Services
Right arrow Similar articles in this journal
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Weismayer, C.
Right arrow Articles by Wolk, A.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Weismayer, C.
Right arrow Articles by Wolk, A.

Home Help [Feedback] [For Subscribers] [Archive] [Search] [Contents]
Copyright © 2006 by American Society for Nutrition