Supplemental Material can be found at:

http://jn.nutrition.org/content/suppl/2013/02/10/jn.112.16901
1.DCSupplemental.html
The Journal of Nutrition
Methodology and Mathematical Modeling

Factor Analysis Is More Appropriate to Identify

Overall Dietary Patterns Associated with
Diabetes When Compared with Treelet
Transform Analysis1–3
Danielle A. J. M. Schoenaker,4 Annette J. Dobson,4 Sabita S. Soedamah-Muthu,5 and Gita D. Mishra4*
4
School of Population Health, University of Queensland, Brisbane, Queensland, Australia; and 5Division of Human Nutrition,
Wageningen University, Wageningen, The Netherlands

Abstract
Treelet transform (TT) is a proposed alternative to factor analysis for deriving dietary patterns. Before applying this method

Downloaded from jn.nutrition.org by guest on November 14, 2017

to nutrition data, further analyses are required to assess its validity in nutritional epidemiology. We aimed to compare
dietary patterns from factor analysis and TT and their associations with diabetes incidence. Complete data were available
for 7349 women (50–55 y at baseline) from the Australian Longitudinal Study on WomenÕs Health. Exploratory factor
analysis and TT were performed to obtain patterns by using dietary data collected from an FFQ. Generalized estimating
equations analyses were used to examine associations between dietary patterns and diabetes incidence. Two patterns
were identified by both methods: a prudent and a Western dietary pattern. Factor analysis factors are a linear combination
of all food items, whereas TT factors also include items with zero loading. The Western pattern identified by factor analysis
showed a significant positive association with diabetes [highest quintile: OR = 1.94 (95% CI: 1.25, 3.00); P-trend = 0.001).
Both factor analysis and TT involve different assumptions and subjective decisions. TT produces clearly interpretable
factors accounting for almost as much variance as factors from factor analysis. However, TT patterns include food items
with zero loading and therefore do not represent overall dietary patterns. The different dietary pattern loading structures
identified by both methods result in different conclusions regarding the relationship with diabetes. Results from this study
indicate that factor analysis might be a more appropriate method for identifying overall dietary patterns associated with
diabetes compared with TT. J. Nutr. 143: 392–398, 2013.

Introduction
A diet consists of a variety of foods with complex combinations
of nutrients that are likely to interact. A way to examine the joint
effect of food intakes and capture overall diet is to derive dietary
patterns using appropriate statistical methods. The identification
of dietary patterns offers a comprehensive approach to study
eating habits and makes it possible to examine the relations with
disease risk in order to propose well-grounded dietary guide-
lines.
1 ri methods, where nutritional variables are grouped according
The Australian Longitudinal Study on WomenÕs Health, which was conceived
and developed by groups of interdisciplinary researchers at the Universities of
Newcastle and Queensland, is funded by the Australian Government
Department of Health and Ageing. Gita D. Mishra is supported by the
Australian National Health and Medical Research Council Centre for Research
Excellence in WomenÕs Health.
2
Author disclosures: D. A. J. M. Schoenaker, A. J. Dobson, S. S.
Soedamah-Muthu, and G. D. Mishra, no conflicts of interest.
3
Supplemental Table 1 is available from the ‘‘Online Supporting Material’’ link in
the online posting of the article and from the same link in the online table of
contents at http://jn.nutrition.org.
* To whom correspondence should be addressed. E-mail: g.mishra@sph.uq.edu.
au.
392 Manuscript received September 4, 2012. Initial review completed October 9, 2012. Revision accepted November 30, 2012.
Type 2 diabetes is a growing public health problem; never-
theless, results from the NursesÕ Health Study (1) show that the
majority of cases could be avoided by behavior modification,
including maintaining a diet high in fiber and low in saturated
and trans fat and glycemic load. A great deal of epidemiological
and clinical research on the role of diet and diabetes has resulted
in a considerable body of evidence relating specific dietary pat-
terns to the risk of diabetes (2–4). However, studies have used
different approaches for identifying dietary patterns. Two gen-
eral approaches have been used in observational studies: a prio-
to prior knowledge or theory of a healthy diet (5,6), and a
posteriori methods, where dietary patterns are derived from
statistical modeling of dietary data allowing for hypothesis-
generating analyses.
Factor analysis is a widely used a posteriori method to
identify dietary patterns (7). The use of factor analysis, however,
remains controversial in the field of nutritional epidemiology
because of subjective choices made throughout the analytical
process. Examples are pregrouping of original food items prior
to analysis, choice of the number of factors to extract, the
ã 2013 American Society for Nutrition.
First published online January 23, 2013; doi:10.3945/jn.112.169011.
 TABLE 1 Factor analysis and TT for identifying dietary patterns: features/aims, assumptions and
decisions associated with each method
Factor analysis
Features/aims Uses the correlation matrix of food
items to derive dietary patterns
Factor loadings on each food item
are used to identify important
foods contributing to each dietary
pattern to extract maximum variance
Assumptions Each factor is a linear combination
of all food items to capture overall
diet
Decisions Pregrouping of original food items
prior to analysis
The number of factors to extract
The method of rotation
Labeling of factors
TT, Treelet transform.
method of rotation, and labeling of factors (8,9). Furthermore,
that each factor is a linear combination of all original food items
may make interpretation complicated.
New estimation methods have been proposed to overcome
these limitations and provide better insight into diet and disease
etiology. Recently, the Treelet transform (TT)6, developed by Lee
et al. (10), was suggested to overcome some of the limitations of
factor analysis mainly by improving the interpretation of factors
(11). Gorst-Rasmussen et al. (11) compared TT and Procrustes-
rotated principal component analysis (PCA) as explanatory
methods to study dietary patterns and the risk of myocardial
infarction in middle-aged men in a Danish prospective cohort
study. Risk estimates were not comparable with those obtained
using PCA, even though they found that TT factors were easier
to interpret due to the graphical representation of the clustering
of food items and the limited number of food items with a factor
loading. The authors concluded that TT may be a useful al-
ternative to factor analysis (11). To further assess the validity of this
approach in nutritional epidemiology, however, comparisons are
needed with other methods of determining dietary patterns and
with the resultant associations for a range of health outcomes.
Therefore, the aim of the present study was to compare
dietary patterns derived by factor analysis, a widely used
method, and the proposed alternative, TT. Our second aim was
to compare the associations between these dietary patterns and
incidence of diabetes. Associations between dietary patterns and
diabetes have been extensively studied in literature (2–4), which
is ideal for critical evaluation of a new proposed analysis
method.
Participants and Methods
The Australian Longitudinal Study on WomenÕs Health. The
Australian Longitudinal Study on WomenÕs Health is a prospective
6
Abbreviations used: MET, total metabolic equivalent; PCA, principal compo-
nent analysis; S1, basline survey; S2, S3, etc., second, third, etc. survey; TT,
Treelet transform.
TT
Uses the correlation matrix of food
items to derive dietary patterns
Factor loadings generally on only a
few food items are used to identify
important foods contributing to each
dietary pattern
Provides a hierarchical
cluster tree to visually identify
dietary patterns
Sparsity: factors contain only a few
food items (omitting other food items
by giving them zero loadings)
Pregrouping of original food items
prior to analysis
The number of factors to extract
Determining the optimal cut-level for
the cluster tree
Labeling of factors
Downloaded from jn.nutrition.org by guest on November 14, 2017
study of factors affecting the health and well-being of 3 cohorts of
Australian women born in 1973–1978 (‘‘young’’), 1946–1951 (‘‘mid-
age’’), and 1921–1926 (‘‘older’’). Women were randomly selected from
the national Medicare health insurance database, which includes all
Australian citizens and permanent residents. Women from rural and
remote areas were intentionally oversampled (12). Since 1996 surveys
have been administered to each cohort every 2–4 y on a rolling basis.
Further details of the recruitment methods and response have been
described elsewhere (13). Informed consent was obtained from all
participants at each survey, with ethical clearance obtained from the
Human Research Ethics Committees of the University of Newcastle and
the University of Queensland.
Participants and surveys. The present study focuses on women in the
mid-age cohort. In 1996, 13,715 women aged 45–50 y participated in
the baseline survey (survey 1). This was estimated to be a 53–56%
response rate for this age cohort (12). Diabetes was assessed at every
survey and dietary intake was first assessed at the third survey (S3). From
the initial mid-age cohort, 11,226 women aged 50–55 y in 2001
completed S3. This study further includes women during follow-up who
responded to the fourth, fifth, and sixth surveys in 2004 (S4, n = 10,905),
2007 (S5, n = 10,638), and 2010 (S6, n = 9748), respectively. Attrition
occurred mainly due to participants not returning the survey or inability
to contact the participant (14). Percentages of women deceased between
surveys are 0.4% at S2, 0.5% at S3, 0.8% at S4, 0.7% at S5, and 0.8% at
S6. Women with history of type 1 or 2 diabetes or impaired glucose
tolerance (n = 745) or a history of cardiovascular disease (n = 703) before
or at S3, or with incomplete dietary data at S3 (n = 1627) were excluded;
the data of 8065 participants were used for obtaining dietary patterns.
Those with missing data on covariates (n = 716) were then excluded,
leaving complete data of 7349 women for analysis of the associations
between dietary patterns and incident diabetes.
Dietary intake. At S3, diet was assessed using an FFQ: the Dietary
Questionnaire for Epidemiological Studies version 2. The development
of the questionnaire (15) and its validation were previously reported
(16). A total of 63 women completed 7-d weighted food records next to
the FFQ. Nutrient intakes were compared and deattenuated correlations
corrected for daily variation in nutrient intake ranged between 0.28 for
total vitamin A and 0.78 for carbohydrate after energy adjustment,
indicating that the FFQ was useful for assessing habitual intake (16).
Participants were asked to report their usual frequency of consumption
Comparing methods for dietary pattern analysis 393
of 74 food and 6 alcoholic beverage items during the previous 12 mo
using 10 response options ranging from ‘‘never’’ to ‘‘three or more times
per day.’’ The FFQ additionally included 21 items about the number of
servings and type of milk, bread, fat spread, sugar, eggs, and cheese
consumed. Ten items on nonalcoholic beverages with similar response
options were also included. For clear visualization of features of factor
analysis and TT, a total number of 111 food items was aggregated into
40 groups based on similarity of nutrient profiles or culinary usage
among the foods for dietary pattern analysis (Supplemental Table 1).
Some original food items were preserved either because it was inappro-
priate to aggregate them into a particular food group (peanut butter,
eggs, coffee, and sugar) or because of the possibility of the items to
represent distinct patterns (potatoes with fat, potatoes without fat, white
bread, nuts, tomato sauce, and flavored milk). All responses were
converted to frequency of consumption per day for analysis. Nutrient
intakes, including total energy and energy from alcohol, were computed
from the national government food composition database of Australian
foods, the NUTTAB95 (17).
Diabetes incidence. Participants were asked at each survey whether a
doctor had told them that they had diabetes. At S4 to S6, they were asked
whether they had been diagnosed with diabetes since the previous survey.
Diagnosis of diabetes was not differentiated into type 1 or type 2 at S4 to
S6 and therefore referred to as diabetes.
all original food items were included in the hierarchical cluster tree. This
cluster tree was ‘‘cut’’ at a given level to provide high variance factors
that indicate related groups of food items that describe an underlying
factor (10). More variation can be explained at the cost of sparsity when
the cluster tree is cut near its roots. We determined an optimal cut-level
by using 10-fold cross-validation (10). To assess sensitivity of the choice
of cut-level, TT analyses were repeated at 63 levels of the optimal level
(11).
For both methods, the number of factors that best represented the
data was based on eigenvalues $1.25, identification of a break point in
the screeplot, and interpretability (25). For each subject, factor scores
were calculated for each of the retained factors by summing the
frequency of consumption multiplied by factor loadings across all food
items with >0 loading. Pearson correlation coefficients were obtained
between dietary patterns derived by factor analysis and TT. The stability
of factors was assessed by performing factor analysis and TT in 2
random subsamples. Correlations between factors from the 2 subsam-
ples were used as a stability measure for both methods.
Generalized estimating equations (GEE) analyses with link function
ÔlogitÕ were used to analyze the longitudinal relationships between
quintiles of dietary pattern scores and incidence of diabetes to take into
account the correlation between repeated measures in the same
individual (26). An unstructured correlation matrix was used for
within-subject correlation. Models were used to produce ORs and

Downloaded from jn.nutrition.org by guest on November 14, 2017

95% CIs for the associations between factors (continuous scores) at S3
Covariates. Lifestyle characteristics available at different surveys were and incidence of diabetes at S4 to S6. Adjustments were made for
used for analysis. Education was only asked at S1 and responses were potential confounders: total energy intake (including energy from
categorized as: no formal qualifications, school or intermediate certif-
icate, higher school certificate, trade/certificate, or university degree.
Smoking status was available at S3 to S6 and defined as: never smoked,
ex-smoker, smoker (<10 cigarettes/d), smoker (10–19 cigarettes/d), or TABLE 2 Baseline characteristics (S3) of 7349 middle-aged
smoker ($20 cigarettes/d). Participants were asked about alcohol study participants in The Australian Longitudinal Study on
consumption at S3 to S6, which was classified according to the National WomenÕs Health1
Health and Medical Research Council classifications as: nondrinker,
low-risk drinker (#14 drinks/wk), or risky drinker (>14 drinks/wk) (18). Characteristic Value
Physical activity scores obtained at S3 to S6 were derived from self-
reported frequency and duration of walking (for recreation or transport) Age, y 53 6 1.5
and from moderate- and vigorous-intensity activity in the last week. The Total energy intake, kJ/d 6960 6 2390
total metabolic equivalent (MET) in minutes/week was calculated as BMI
(walking minutes 3 3.5) + (moderate minutes 3 4) + (vigorous minutes 3 Underweight, BMI ,18.5 kg/m2 87 (1.2)
7.5) (19–21). Physical activity was then categorized as: sedentary (0 to Healthy weight, BMI 18.5– ,25 kg/m2 2896 (39.4)
<40 MET min/wk), low (40 to <600 MET min/wk), moderate (600 to Overweight, BMI 25– ,30 kg/m2 2499 (34.0)
<1200 MET minutes/wk), or high ($1200 MET min/wk). Participants
Obese, BMI $30 kg/m2 1867 (25.4)
were asked at S3 to S6 whether they had been diagnosed with or treated
History of hypertension 1837 (25.0)
for hypertension in the last 3 y, with responses categorized as ‘‘yes’’ or
‘‘no.’’ BMI, available at S3 to S6, was computed as self-reported weight Physical activity
(kg)/height (m2) and categorized as: underweight (BMI <18.5 kg/m2), Sedentary, 0 to ,40 MET min/wk 1110 (15.1)
healthy weight (BMI 18.5– <25 kg/m2), overweight (BMI 25– <30 kg/m2), Low, 40 to ,600 MET min/wk 2058 (28.0)
or obese (BMI $30 kg/m2) according to the WHO classification (22). Moderate, 600 to ,1200 MET min/wk 1631 (22.2)
High, $1200 MET min/wk 2550 (34.7)
Statistical analyses. Dietary patterns at S3 were identified by both Smoking status
exploratory factor analysis and TT for 8065 women. These dimension- Never smoked 4049 (55.1)
reduction methods are comparable, because they aggregate food items Ex-smoker 2322 (31.6)
based on correlation; however, they involve different aims, assumptions, Smoke ,10 cigarettes/d 331 (4.5)
and decisions to be made (Table 1). Smoke 10–19 cigarettes/d 272 (3.7)
Factor analysis aims to identify patterns with a linear composition of
Smoke $20 cigarettes/d 375 (5.1)
all food items that account for the largest amount of variation in diet
Education (S1)
between individuals. Factors obtained were rotated using the (orthog-
onal) Varimax procedure to facilitate interpretability and ensure No formal qualifications 1014 (13.8)
orthogonality (23). To enhance comparability with TT factors that are School or Intermediate certificate 2337 (31.8)
correlated, (oblique) Procrustes-rotated factor analysis was also per- Higher school certificate 1242 (16.9)
formed in secondary analysis (24). Trade/apprenticeship/certificate/diploma 1507 (20.5)
TT aims to explain variation with factor simplicity by introducing University/higher degree 1249 (17.0)
sparsity in factor loadings (i.e., the number of food items with zero Alcohol consumption
loading). TT produces sparse factors as well as a cluster tree to visualize Nondrinker 1279 (17.4)
the grouping structure among food items. Treelets were constructed as Low risk drinker (#14 drinks/wk) 5982 (81.4)
follows: first, the 2 original food items with the highest correlation were
High risk drinker (.14 drinks/wk) 88 (1.2)
grouped together and local PCA was performed. These 2 items were
then replaced by a sum variable, which was retained, and a difference 1
Data are mean 6 SD or n (%). MET, total metabolic equivalent; S1, baseline survey;
variable, which was disregarded (11). This algorithm was repeated until S3, third survey (baseline for the present study).

394 Schoenaker et al.

alcohol) at S3 (kJ/d), education at S1 (5 categories), smoking status at S3
to S6 (5 categories), alcohol consumption at S3 to S6 (3 categories),
physical activity at S3 to S6 (4 categories), and (continuous) factor scores
of other dietary patterns at S3. The potential effect modifiers hyperten-
sion at S3 to S6 (yes or no) and BMI at S3 to S6 (4 categories) were
adjusted for in separate models 2 and 3, respectively. RRs are expressed
per 1 SD of the factor scores. Statistical analysis was conducted using
Stata version 11.1 (StataCorp). For TT, an add-on for Stata was used
(27). P values < 0.05 were considered significant.
Results
Dietary patterns. Characteristics of the women are described in
Table 2. Three factors for factor analysis and 2 factors for TT
had a variance $1.25. Also taking the screeplot and interpret-
ability of factors into account, 2 factors were obtained from
both factor analysis and TT. Results from cross-validation for
TT indicated an optimal cut-level for the cluster tree of 24.
Factors from factor analysis were labeled based on high factor
loadings ($0.20) and are presented in Figure 1A (prudent
patterns) and 1B (Western patterns). For TT, nonzero loadings
were used for labeling (Fig. 1A,B, right) together with the cluster
tree (Fig. 2). The percentage of variance explained by the 2
factors was almost similar for factor analysis and TT (19.1 and
18.6%, respectively).
The structure of factor 1 was comparable between factor
analysis (Fig. 1A,B, left) and TT (Fig. 1A,B, right) (Pearson
correlation r = 0.99; P = <0.0001) and primarily characterized
by fruit, vegetables, high-fiber bread, medium-fat dairy, and fish.
Factor 1 was labeled as a prudent pattern (Fig. 1A). Factor 2
showed high loadings on take-away food, meat, and snacks for
both factor analysis and TT and was labeled as a Western pattern
(Fig. 1B) (Pearson correlation r = 0.70; P = <0.0001). Factor 2
from factor analysis also included high loadings on high-fat
dairy, potatoes with fat, and white bread. The correlation
between scores from factor analysis and TT factor 2 was 0.72.
Factors derived using (oblique) Procrustes-rotated factor anal-
ysis were similar to those derived using (orthogonal) Varimax-
rotated factor analysis (data not shown).
FIGURE 1 Factor loadings for factors derived by factor analysis and TT for prudent patterns (A) and Western patterns (B) for participants in The
Australian Longitudinal Study on WomenÕs Health (n = 8065). TT, Treelet transform.
TT analyses at cut-levels 21 and 27 (24 6 3) produced similar
factors to those described but with slightly different loadings.
Compared with the factor obtained using cut-level 24, the
prudent pattern included zero loadings for high-fiber bread and
tea and water when obtained using cut-level 21.
Results from stability analyses indicate stability for the
prudent pattern from both methods (Pearson correlation
between factors from 2 subsamples: r = 0.98, P = <0.0001 for
factor analysis and r = 0.97, P = <0.0001 for TT). The Western
patterns obtained from factor analysis in the 2 subsamples
correlated highly (Pearson correlation r = 0.97; P = <0.0001),
whereas the Western pattern from TT seemed less stable
[Pearson correlation r = 0.84; P = <0.0001)].
Associations between dietary patterns and incidence of
diabetes. During 9 y of follow-up (S3 to S6), 635 incident
diabetes cases occurred among 7346 participants. The Western
pattern obtained from factor analysis was positively associated
with incidence of diabetes after adjustments were made for
energy intake and lifestyle factors [highest quintile: OR = 1.94
(95% CI: 1.25, 3.00); P-trend: 0.001] (Table 3). This association
was hardly affected by additional adjustment for hypertension
Downloaded from jn.nutrition.org by guest on November 14, 2017
or BMI. No significant associations were found for the prudent
patterns and the Western pattern identified by TT in relation to
incidence of diabetes. Including factor scores of the other dietary
pattern in the models did not change our results after adjusting
for energy intake and lifestyle factors.
Discussion
In our study of dietary patterns and incidence of diabetes among
Australian middle-aged women, we found that 2 dietary
patterns could be identified by factor analysis and TT, explaining
an almost similar amount of variation. Similar labels were given
to both patterns; however, they differ regarding factor loadings
and number of items with a loading. Therefore, relating these
patterns to incidence of diabetes resulted in different conclusions
from the different methods. Our results from factor analysis
Comparing methods for dietary pattern analysis 395

FIGURE 2 Cluster tree produced by TT to identify dietary patterns in
participants in The Australian Longitudinal Study on WomenÕs Health.
The dashed line indicates the selected cut-level (level 24) for the
cluster tree (n = 8065). TT, Treelet transform.
show evidence for an inverse association between the prudent
pattern and incidence of diabetes. The Western pattern identified
by factor analysis was positively associated with diabetes.
The majority of studies that have examined dietary patterns
in relation to diabetes have revealed 2 primary patterns, mostly
labeled as a prudent and a Western pattern (2–4). The prudent
pattern, often characterized by high intakes of fruit, vegetables,
fish, whole grains, and low-fat dairy products, has an inverse
association with diabetes (2–4). These food items load highly on
our prudent patterns, which had a significant inverse association
with diabetes when identified by factor analysis. The pattern
referred to as Western in many studies consists of high con-
sumption of red meat, processed meat, refined grains, high-fat
dairy products, high-sugar drinks, and sweets and is positively
associated with diabetes risk (2–4). This is in line with our
results from factor analysis: we showed a positive association for
the Western pattern with incidence of diabetes. Even though
caution should be taken when comparing results because of the
highly heterogeneous nature of dietary pattern analyses and risk
of chronic diseases, our findings on dietary patterns identified by
TT in relation with incidence of diabetes are not in line with
consistent findings from existing literature (2–4).
396 Schoenaker et al.
Strengths of the present study include the large sample size,
the longitudinal design, the detailed information on lifestyle
factors, and the use of a validated FFQ (16). A limitation in our
study is the reliance on self-reported diabetes. Additionally,
measurement error is a concern in self-reported dietary assess-
ment. Although these limitations are generally thought to be
nondifferential with regard to the outcome, they may have
influenced the magnitude and the direction of the associations.
Also, the amount of missing data is a constraint in our study.
However, the distribution of lifestyle characteristics was similar
for participants included for the present study and participants
excluded for the present study due to missing data, except for
hypertension and physical activity (data now shown). Partici-
pants excluded from the present study were more likely to have
hypertension compared with participants included (28.8 vs.
25.0% at S3, respectively) and to be less physically active (24.9
vs. 34.7% reported a high physical activity level at S3,
respectively). Another limitation is the possibility of residual
confounding resulting from measurement errors of confounders
or factors we did not measure.
In our study, we identified 2 dietary patterns from Varimax-
rotated factor analysis and TT with Pearson correlations of 0.99
Downloaded from jn.nutrition.org by guest on November 14, 2017
and 0.72 between the 2 methods for factor 1 and 2, respectively.
Gorst-Rasmussen et al. (11) found in their study 7 patterns
derived using Procrusted-rotated PCA and TT with correlations
ranging from 0.48 to 0.99. In this study the percentage of
variance explained by factors from PCA was 36.9 and 31.0%
for TT factors (11), while factors obtained in our study explai-
ned an almost similar amount of percentage of variance from
both methods (19.1% for factor analysis vs. 18.6% for TT).
Gorst-Rasmussen et al. (11) showed in their study that associ-
ations between patterns derived by the 2 methods and risk of
myocardial infarction were not identical, which is in line with
our findings on the association with incidence of diabetes, which
were not identical from patterns derived using the 2 methods.
The selected number of food items contributing to dietary
patterns from TT did not result in significant associations with
diabetes incidence, whereas the contribution of all food items to
dietary patterns identified using factor analysis resulted in
significant associations with diabetes, which is in line with the
literature (2–4).
We demonstrated that TT may have several advantages over
factor analysis. First, the key properties of TT are the production
of sparse factors and a cluster tree to visually identify related
groups of food items. The complex loading pattern of factor
analysis factors makes identification of a pattern challenging,
whereas sparse loadings together with the cluster tree provide
simpler characterization of a pattern from TT. To achieve a
simple structure of factors from factor analysis and produce
factors that are easy to interpret, rotation of the factor loadings
is a commonly used method (23). However, the choice of
approach (orthogonal or oblique) is arbitrary, with possible
impact on the final conclusion. To explore the impact of the
rotation method used on the final results or conclusions, both
rotation methods could be used in sensitivity analysis. In our
study, using orthogonal or oblique rotation did not affect the
dietary patterns identified and their associations with diabetes.
We used Varimax rotation to obtain orthogonal factors, which
provides nearly uncorrelated, distinct factors that can be related
to risk of disease. TT has the advantage that it is able to
automatically produce the factors. Also, from a dimension
reduction perspective, TT performs almost as well as factor
analysis; even with sparse factors, TT factors captured almost as
much variation as factors identified from factor analysis.
TABLE 3 Incidence of diabetes and dietary patterns identified by factor analysis and TT in participants in The Australian Longitudinal
Study on WomenÕs Health (n = 7349)1

Quintiles of dietary pattern scores per 1 SD

1 2 3 4 5 P-trend

Factor analysis: prudent pattern

Model 12 Reference 1.05 (0.74, 1.51) 0.99 (0.68, 1.51) 0.83 (0.56, 1.22) 0.87 (0.59, 1.29) 0.26
Model 23 Reference 1.09 (0.76, 1.55) 1.00 (0.70, 1.44) 0.92 (0.63, 1.34) 0.93 (0.63, 1.37) 0.49
Model 34 Reference 1.14 (0.80, 1.63) 1.05 (0.73, 1.51) 0.89 (0.60, 1.31) 0.94 (0.64, 1.38) 0.40
TT: prudent pattern
Model 12 Reference 1.07 (0.74, 1.55) 0.98 (0.67, 1.43) 0.86 (0.58, 1.27) 0.99 (0.67, 1.47) 0.62
Model 23 Reference 1.06 (0.74, 1.52) 0.98 (0.68, 1.43) 0.92 (0.62, 1.34) 1.06 (0.72, 1.55) 0.94
Model 34 Reference 1.11 (0.77, 1.59) 1.01 (0.70, 1.47) 0.91 (0.62, 1.35) 1.04 (0.71, 1.53) 0.81
Factor analysis: Western pattern
Model 12 Reference 1.17 (0.73, 1.85) 2.04 (1.35, 3.10) 1.86 (1.21, 2.84) 1.94 (1.25, 3.00) 0.001
Model 23 Reference 1.20 (0.76, 1.89) 1.99 (1.32, 3.00) 1.69 (1.11, 2.56) 1.86 (1.21, 2.86) 0.002
Model 34 Reference 1.12 (0.71, 1.77) 1.91 (1.27, 2.89) 1.73 (1.14, 2.64) 1.73 (1.12, 2.67) 0.003
TT: Western pattern
Model 12 Reference 0.82 (0.55, 1.22) 1.16 (0.80, 1.69) 1.18 (0.81, 1.70) 0.83 (0.55, 1.25) 0.97
Model 23 Reference 0.76 (0.51, 1.13) 1.13 (0.79, 1.62) 1.01 (0.70, 1.46) 0.75 (0.50, 1.12) 0.56
Model 34 Reference 0.75 (0.50, 1.12) 1.09 (0.75, 1.57) 1.07 (0.74, 1.54) 0.73 (0.49, 1.10) 0.59

Downloaded from jn.nutrition.org by guest on November 14, 2017

1
Values are OR and 95% CI. TT, Treelet transform.
2
Model 1 adjusted for total energy intake (kJ/d), education (no formal qualifications, school or intermediate certificate, higher school certificate, trade/certificate, and university
degree), smoking status (never smoked, ex-smoker, smoke ,10 cigarettes/d, smoke 10–19 cigarettes/d, and smoke $20 cigarettes/d), alcohol consumption [low-risk drinker (#14
drinks/wk), nondrinker, and risky drinker (.14 drinks/wk)], and physical activity (sedentary, low, moderate, and high).
3
Model 2 adjusted for covariates in model 1 and hypertension (yes, no).
4
Model 3 adjusted for covariates in model 1 and BMI (underweight, BMI ,18.5; healthy weight, 18.5 # BMI , 25; overweight, 25 # BMI , 30; and obese, BMI $30).

In contrast to factor analysis, TT does not automatically
produce factors with high variance. A subjective decision is
made when selecting the cut-level for the cluster tree using cross-
validation before high variance factors can be extracted. The
cut-level influences both the sparsity as well as the grouping of
the factors and might therefore affect the results when looking at
associations with disease incidence. Lowering the cut-level
results in increased sparsity, whereas increasing the cut-level
decreases sparsity, showing contributions from all food items
to each factor, comparable to factor analysis. Increasing the
sparsity improves interpretability, but at the same time, the
factor variances increase, which might result in unstable results
(28). Performing TT with different cut-levels reveals the insta-
bilities and helps determine the optimal level (11). Pattern
structures remained comparable in our study when obtained
using cut-levels of 63; however, further decreasing or increasing
the cut-level would most likely have a larger influence on the
structure of the patterns. Instead of cutting the tree at a single
height, another approach could be to start near the root of the
tree and descending deeper into the tree, looking for optimal
identification of patterns regarding number of food items with
a non-zero loading, interpretability, and public health rele-
vance (28).
A major concern when applying TT to nutritional data is
whether it is in line with the original aim of dietary pattern
analysis: to derive dietary patterns that represent the frequency
and amount of all foods consumed to capture overall diet (29).
The combined role of all foods is essential in the biologic in-
fluence of diet on disease as well as for dietary interventions and
public health messages (30). Where factors from factor analysis
comprise all food items, the sparsity feature of TT results in
patterns ignoring foods with zero loading, as in the case of white
bread and potatoes with fat. This may be due to the fact that
these food items were not correlated with the newly formed
variables from the local PCA and hence were subsequently
disregarded (11). These foods, however, show a high loading on
the Western pattern from factor analysis and have strong
individual relationships with diabetes incidence [white bread:
OR = 1.21 (95% CI: 1.12, 1.30); potatoes with fat: OR = 1.69
(95% CI: 1.01, 2.75)]. These different pattern structures,
therefore, result in different conclusions regarding their rela-
tionship with diabetes incidence.
In summary, we demonstrated that the proposal of a new
approach to derive dietary patterns and comparison of method-
ologies gives insight into the importance of aims and assump-
tions in such analyses. Both factor analysis and TT involve
subjective decisions to be made that should be explored in
sensitivity analyses and taken into account when interpreting
results and conclusions for public health messages. Sensitivity
analyses on, e.g., pregrouping of food items and number of
factors to extract can indicate and optimize robustness of results.
TT produces clearly interpretable factors that account for al-
most as much variation as factors from factor analysis, but the
sparse factors do not represent an overall dietary pattern.
Besides, results on the relation between dietary patterns from TT
and incidence of diabetes are not in line with consistent findings
from the literature. Results from this study indicate that factor
analysis might be a more appropriate method for identifying
overall dietary patterns associated with diabetes compared with
TT.
Acknowledgments
The authors thank Professor Graham Giles of the Cancer
Epidemiology Centre of The Cancer Council Victoria for per-
mission to use the Dietary Questionnaire for Epidemiological
Studies (version 2), Melbourne: The Cancer Council Victoria,
1996. D.A.J.M.S., G.D.M., and A.J.D. designed research;
D.A.J.M.S. analyzed data and had primary responsibility for
final content; G.D.M. and A.J.D. contributed to statistical
analysis and interpretation of results and critical revision of the
Comparing methods for dietary pattern analysis 397
manuscript; and S.S.S-M. contributed by critical revision of the
manuscript for important intellectual content. All authors read
and approved the final manuscript.
Literature Cited
1. Hu FB, Manson JE, Stampfer MJ, Colditz G, Liu S, Solomon CG,
Willett WC. Diet, lifestyle, and the risk of type 2 diabetes mellitus in
women. N Engl J Med. 2001;345:790–7.
2. Esposito K, Kastorini CM, Panagiotakos DB, Giugliano D. Prevention
of type 2 diabetes by dietary patterns: a systematic review of prospective
studies and meta-analysis. Metab Syndr Relat Disord. 2010;8:471–6.
3. Kastorini CM, Panagiotakos DB. Dietary patterns and prevention of
type 2 diabetes: from research to clinical practice; a systematic review.
Curr Diabetes Rev. 2009;5:221–7.
4. Salas-Salvadó J, Martinez-González M, Bulló M, Ros E. The role of diet
in the prevention of type 2 diabetes. Nutr Metab Cardiovasc Dis.
2011;21:B32–48.
5. Haines PS, Siega-Riz AM, Popkin BM. The Diet Quality Index revised:
a measurement instrument for populations. J Am Diet Assoc. 1999;
99:697–704.
6. Kennedy ET, Ohls J, Carlson S, Fleming K. The Healthy Eating Index:
design and applications. J Am Diet Assoc. 1995;95:1103–8.
7. Newby PK, Tucker KL. Empirically derived eating patterns using factor
or cluster analysis: a review. Nutr Rev. 2004;62:177–203.
8. Michels KB, Schulze MB. Can dietary patterns help us detect diet and
disease associations? Nutr Res Rev. 2005;18:241–8.
9. Martı́nez ME, Marshall JR, Sechrest L. Invited Commentary: factor
analysis and the search for objectivity. Am J Epidemiol. 1998;148:17–9.
10. Lee AB, Nadler B, Wasserman L. Treelets: an adaptive multi-scale basis
for sparse unordered data. Ann Appl Stat. 2008;2:435–71.
11. Gorst-Rasmussen A, Dahm CC, Dethlefsen C, Scheike T, Overvad K.
Exploring dietary patterns by using the Treelet transform. Am J
Epidemiol. 2011;173:1097–104.
12. Brown WJ, Bryson L, Byles JE, Dobson AJ, Lee C, Mishra G, Schofield
M. Women’s Health Australia: recruitment for a national longitudinal
cohort study. Women Health. 1998;28:23–40.
13. Lee C, Dobson AJ, Brown WJ, Bryson L, Byles J, Warner-Smith P,
Young AF. Cohort profile: the Australian Longitudinal Study on
WomenÕs Health. Int J Epidemiol. 2005;34:987–91.
14. Young AF, Powers JR, Bell SL. Attrition in longitudinal studies: who do
you lose? Aust N Z J Public Health. 2006;30:353–61.
398 Schoenaker et al.
15. Ireland P, Jolley D, Giles G, OÕDea K, Powles J, Rutishauser I, Wahlqvist
ML, Williams J. Development of the Melbourne FFQ: a food frequency
questionnaire for use in an Australian prospective study involving an
ethnically diverse cohort. Asia Pac J Clin Nutr. 1994;3:19–31.
16. Hodge A, Patterson AJ, Brown WJ, Ireland P, Giles G. The Anti Cancer
Council of Victoria FFQ: relative validity of nutrient intakes compared
with weighed food records in young to middleaged women in a study of
iron supplementation. Aust N Z J Public Health. 2000;24:576–83.
17. Lewis J, Milligan G, Hunt A. NUTTAB95 Nutrient Data Table for use
in Australia. Canberra: Australian Government Publishing Service;
1995.
18. National Health and Medical Research Council. Australian Alcohol
Guidelines: health risks and benefits. Canberra: ACT, Commonwealth
of Australia; 2001.
19. Brown WJ, Burton NW, Marshall AL, Miller YD. Reliability and
validity of a modified self-administered version of the Active Australia
physical activity survey in a sample of mid-age women. Aust N Z J
Public Health. 2008;32:535–41.
20. Brown WJ, Bauman AE. Comparison of estimates of population levels
of physical activity using two measures. Aust N Z J Public Health.
2000;24:520–5.
21. Armstrong T, Bauman A, Davis J. Physical activity patterns of
Australian adults. Canberra: Australian Institute of Health and Welfare;
2000.
22. WHO. Obesity: preventing and managing the global epidemic. Geneva:
WHO; 2000.
Downloaded from jn.nutrition.org by guest on November 14, 2017
23. Jolliffe IT. Rotation of principal components: choice of normalization
constraints. J Appl Stat. 1995;22:29–35.
24. Hurley JR, Cattell RB. The procrustes program: producing direct
rotation to test a hypothesized factor structure. Behav Sci. 1962;7:258–
62.
25. Kline P. An easy guide to factor analysis. London: Routledge; 1994.
26. Liang K, Zeger SL. Longitudinal data analysis using generalised linear
models. Biometrika. 1986;73:45–51.
27. Gorst-Rasmussen A. tt. Stata add-on for performing treelet transfor-
mation; 2011 [cited 2012 Jul 1]. Available from: http://people.math.
aau.dk/gorst/software.htm.
28. Meinshausen N, Bühlmann P. Discussion of: Treelets: an adaptive multi-
scale basis for sparse unordered data. Ann Appl Stat. 2008;2:478–81.
29. Hu FB. Dietary pattern analysis: a new direction in nutritional
epidemiology. Curr Opin Lipidol. 2002;13:3–9.
30. Slattery ML. Analysis of dietary patterns in epidemiological research.
Appl Physiol Nutr Metab. 2010;35:207–10.