International Journal of Obesity (2009) 33, 611620

& 2009 Macmillan Publishers Limited All rights reserved 0307-0565/09 $32.00
www.nature.com/ijo

ORIGINAL ARTICLE
Psychometric analysis of the Three-Factor Eating
Questionnaire-R21: results from a large diverse sample
of obese and non-obese participants
JC Cappelleri1, AG Bushmakin1, RA Gerber1, NK Leidy2, CC Sexton2, MR Lowe3 and J Karlsson4
1
Pfizer Inc., Global Research and Development, New London, CT, USA; 2Center for Health Outcomes Research, United
BioSource Corporation, Bethesda, MD, USA; 3Department of Psychology, Drexel University, Philadelphia, PA, USA and
4
Institute of Health and Care Sciences, Sahlgrenska Academy at University of Gothenburg, Gothenburg, Sweden

Background: The 21-item Three-Factor Eating Questionnaire (TFEQ-R21) is a scale that measures three domains of eating
behavior: cognitive restraint (CR), uncontrolled eating (UE) and emotional eating (EE).
Objectives: To assess the factor structure and reliability of TFEQ-R21 (and if necessary, refine the structure) in diverse
populations of obese and non-obese individuals.
Design: Data were obtained from obese adults in a United States/Canadian clinical trial (n 1741), and overweight, obese and
normal weight adults in a US web-based survey (n 1275). Confirmatory factor analyses were employed to investigate the
structure of TFEQ-R21 using baseline data from the clinical trial. The model was refined to obtain adequate fit and internal
consistency. The refined model was then tested using the web-based data. Relationships between TFEQ domains and body mass
index (BMI) were examined in both populations.
Results: Clinical data indicated that TFEQ-R21 needed refinement. Three items were removed from the CR domain, producing
the revised version TFEQ-R18V2 (Comparative Fit Index (CFI) 0.91). Testing TFEQ-R18V2 in the web-based sample supported
the revised structure (CFI 0.96; Cronbachs coefficient a of 0.780.94). Associations with BMI were small. In the clinical study,
the CR domain showed a significant and negative association with BMI. On the basis of the web-based survey, it was shown that
the relationship between BMI and CR is population-dependent (obese versus non-obese, healthy versus diabetics).
Conclusions: In two independent datasets, the TFEQ-R18V2 showed robust factor structure and good reliability. It may provide
a useful tool for characterizing UE, CR and EE.
International Journal of Obesity (2009) 33, 611620; doi:10.1038/ijo.2009.74; published online 28 April 2009

Keywords: BMI; patient-reported outcomes; eating behavior; Three-Factor Eating Questionnaire; dietary restraint

Introduction its own etiology (derived from theories, such as externality

Obesity is a risk factor for several diseases, such as diabetes,
heart disease, fatty liver, sleep apnea and some cancers,13
and is associated with substantial direct and indirect costs.4
The worldwide prevalence of obesity has increased dramati-
cally over the last decade, and there is an urgent need to
better understand the eating behavior in humans and how
obesity might be treated.5 Various types of eating behavior
have been identified and these include uncontrolled eating
(UE), emotional eating (EE) and restrictive eating, each with
Correspondence: Dr JC Cappelleri, Pfizer Inc., Global Research and Develop-
ment, 50 Pequot Avenue, MS6025-A4225, New London, CT 06320, USA.
E-mail: joseph.c.cappelleri@pfizer.com
Received 13 October 2008; revised 14 February 2009; accepted 23 March
2009; published online 28 April 2009
theory, psychosomatic theory and restraint theory). How-
ever, we need to further understand the link between eating
behaviors and weight-related outcomes that may exist
between obese and non-obese individuals.
The Three-Factor Eating Questionnaire (TFEQ) is a
self-assessment scale used widely in studies of eating
behavior in overweight and normal weight individuals.69
It was designed to assess three cognitive and behavioral
domains (or factors) of eating: cognitive restraint (CR),
disinhibition and hunger. The original TFEQ10 contained 51
items and these three domains. Two subsequent studies were
unable to replicate the factor structure of the original
tool.11,12 The same can be said for a third study, in which
items fell into different types of patterns than the original
TFEQ and hence, into a different factor structure (CR, food
interest, EE).13
 Analysis of the Three-Factor Eating Questionnaire
JC Cappelleri et al
612
On the other hand, in another study, the original three-
factor structure of the TFEQ was essentially confirmed in an
Australian population of undergraduate university women,
but the researchers also suggested that the three main factors
could be further divided into subcategories.14 Another
investigation, whose participants enrolled in a computer-
aided training program for weight reduction, suggested that
the restraint domain could be usefully divided into two
forms, flexible and rigid, and that the rigid form (character-
ized by a dichotomized, all-or-nothing approach to eating)
seemed to be associated more with problematic eating and
diet breakdown.15
A shortened 18-item version (TFEQ-R18) with a revised
three-factor structure on the basis of CR (six items; the
conscious restriction of food intake to control body weight
or to promote weight loss), UE (nine items; the tendency to
eat more than usual because of a loss of control over intake)
and EE (three items; overeating during dysphoric mood
states) was developed using data from severely obese partici-
pants in Sweden.11 The instrument has been used with a
student sample,13 and samples in Sweden and France,16,17
with evidence to suggest that the tool can be used to
characterize eating behaviors in non-obese populations. The
shortened form is also more suited to epidemiological
studies or clinical trials, in which the study participants
may have multiple questionnaires to complete.
The TFEQ-R18 was later refined by adding three additional
items to the EE domain to minimize floor and ceiling effects.
The resulting 21-item version (TFEQ-R21) exhibited improved
psychometric properties, with a stable factor structure and
evidence of construct validity in Swedish studies.18 No studies,
however, have been published to date confirming the factor
structure of the TFEQ-R21 in non-European populations.
The aim of this study was to evaluate the factor structure
and reliability of the TFEQ-R21 in two distinct populations:
(1) a clinical sample of obese patients in the United States and
Canada at baseline (before treatment assignment) and (2) a
non-clinical sample from a web-based survey of obese and
non-obese participants in the United States. The extent of the
relationships between TFEQ domain scores and body mass
index (BMI) in these patients were also examined. In addition,
earlier research suggests that the relationship between
restraint eating behaviors depends on sample characteristics.
Several studies found a positive cross-sectional association
between restrained eating and BMI,19,20 whereas others found
no significant relationships6,21 or inverse relationships.22,23 In
this investigation, we examine the relationship between
restrained eating and BMI using participants with a range of
BMI, as well as clinical and non-clinical samples.
Methods
Objective
The primary objective of this study was to evaluate the factor
structure and reliability of the TFEQ-R21, and its association
International Journal of Obesity
with BMI, in a large obese clinical sample from the United
States and Canada. Subsequently, the aim was to modify the
structure of the TFEQ-R21, if warranted, using the clinical
sample and then test the refined model in a web-based
sample of obese and non-obese healthy individuals from the
United States.
Study populations
Two distinct study populations were used. The factor
structure of the TFEQ-R21 was initially examined using
baseline data, before treatment intervention, from a phase III
clinical trial of a candidate weight management compound
(which was previously but no longer in development). The
dataset included 1741 obese, non-diabetic individuals (obe-
sity defined as BMI 430 kg/m2 for participants without
comorbidities; BMI 427 kg/m2 for participants with comor-
bidities (treated or untreated hypertension or dyslipidemia))
from the United States and Canada. Questionnaires were
administered in the clinic on day one. Data obtained from
these participants are referred to as clinical study data.
The refined structure of the TFEQ-R21 was evaluated using
the data obtained from the participants recruited from the
US arm of the 2006 National Health and Wellness Survey.
The dataset included 1275 participants in total and recruited
participants from the following groups: non-obese healthy
participants (BMI 18.526.9 kg/m2, with no diabetes or
dyslipidemia), overweight and obese participants (BMI
X27.0 kg/m2) with diabetes, and overweight and obese
participants (BMI X27.0 kg/m2) without diabetes. The
National Health and Wellness Survey is an annual study of
the healthcare attitudes and behaviors of nationally repre-
sentative samples of the adult population. In 2006, the
National Health and Wellness Survey was fielded to 60 000
members of the general panel of Lightspeed Research.
Through panel identification numbers, respondents were
identified and recontacted for this study. Any respondent
with a BMI X18.5 kg/m2 was eligible to be recontacted to
participate in this study and they were invited to do so
randomly. As respondents entered the questionnaire, they
were rescreened for eligibility and categorized into three
quota groups: 500 obese participants with diabetes, 500
obese participants without diabetes and 250 non-obese
participants. Within both the obese with diabetes and
obese without diabetes groups, there were minimum
quotas based on BMI: at least 100 respondents must have
a BMI 2729 kg/m2; at least 100 respondents must have a
BMI 3034 kg/m2; at least 100 respondents must have a BMI
X3539 kg/m2; and at least 100 respondents must have
a BMI X40 kg/m2.
Once the quota groups were filled, only respondents to
National Health and Wellness Survey who met the criteria
for the open quota groups received subsequent invitations to
participate in the study. Members of the Lightspeed Research
panel were recruited through opt-in email, co-registration
with Lightspeed Research partners, e-newsletter campaigns,

banner placements, and both internal and external affiliate
networks. The questionnaire was self-administered through
the internet. The questionnaire takes approximately 20 min
to complete. Respondents needed to have internet access to
participate, but that access may or may not have been in
their homes. Information regarding point of internet access
was not collected. The formatting of the screen presentation
was as close as possible to the questionnaire provided in the
clinical study. All data were self-reported by the survey
respondents and were not verified against any clinical
diagnostics. As the study was conducted through the
internet, it was not possible for a respondent to proceed
through the survey without completing all earlier questions.
Therefore, the questionnaire was completed with no missing
data. Data obtained from these participants are referred to as
web-based survey data.
Description of TFEQ-R21
The TFEQ-R21 (see Appendix) asks participants to respond to
21 questions on a four-point Likert scale for items 120 and
on an eight-point numerical rating scale for item 21.
Responses to each of the items are given a score between
1 and 4. Before calculating domain scores, items 116 were
reverse coded and item 21 was recoded as follows: 12 scores
as 1; 34 as 2; 56 as 3; 78 as 4. Domain scores were then
calculated as a mean of all items within each domain; hence,
domain scores also ranged from 1 to 4 (CR (six items), UE
(nine items) and EE (six items)), with higher scores being
indicative of greater CR, UE and EE.
Analysis of the Three-Factor Eating Questionnaire
JC Cappelleri et al
613
categories, we rounded those values to integers from 1 to 4.
This model included BMI as an outcome and TFEQ score
(a categorical variable) as a predictor (controlling for age and
gender). Least-squared means of the TFEQ domain scores
from the model were fitted against BMI to assess and depict
the nature of their relationship. Pearsons correlation
coefficient between BMI and TFEQ domain scores was also
calculated.
Psychometric analyses: web-based dataset
For the web-based dataset, identical psychometric tests to
those undertaken using the clinical dataset were conducted,
except that the confirmatory factor structure of the refined
model was postulated.
Gender and obesity status
We conducted additional analyses and validity tests to
examine the stability of the measurement model for obese
participants (BMI of 30 kg/m2) and non-obese participants
(BMI of o30 kg/m2), as well as for gender in the web-based
sample. In doing so, we compared and tested two multi-
group models: the no-constraint model, which does not
impose any constraint on its parameters in the two
subgroups of interest (for example, obese versus non-obese),
and the invariant measurement model, which constrains
the corresponding factor loadings in the two subgroups to be
equal. Evidence for the invariant measurement model exists
if two conditions are met: the multi-group models exhibits
adequate fit (say, CFI of 0.90 ) and that the difference
between these two models is negligible (p0.01).25,26

Psychometric analyses: clinical dataset

Confirmatory factor analysis was used to test the fit of the
three-factor model to the data, with the final model required
to have a Bentlers Comparative Fit Index (CFI) of 40.9.24
We applied the confirmatory factor analysis on the baseline
data from the clinical trial using the existing three domains
as an initial conceptual model.24 If warranted, this algorithm
allows modification of the existing TFEQ-R21, structured
as a three-factor measurement model, to the model that
adequately fitted the clinical sample data.
The distribution of responses was evaluated at the item
level to identify any significant floor effects (450% of
responses at the lowest end of the scale) and ceiling effects
(450% of responses at the highest end of the scale).
Cronbachs coefficient a was used to estimate the internal
consistency of each domain in the refined model. As
a common rule of thumb, coefficients 40.7 are recom-
mended.24 A regression model, which did not impose
any functional relationship between TFEQ domain score
and BMI, was applied. For this analysis, the TFEQ domain
score was transformed into a categorical variable. This cate-
gorization was achieved by calculating the domain score as a
mean score of its items, then representing that score by a
continuous variable from 1 to 4. To create the original four
Statement of ethics
We certify that all the applicable institutional and govern-
mental regulations concerning the ethical use of human
volunteers/animals were followed during this research. The
research studies were approved by the appropriate Institu-
tional Review Board (IRB).
Results
Participants
There were 1741 patients recruited into the clinical trial.
Mean (s.d.) BMI was 38.66.7 kg/m2; mean (s.d.) age
was 46.311.0 years; 1427 (82%) of the patients were
female. Of the 1275 participants recruited to complete the
web-based survey (mean (s.d.) BMI 33.17.6 kg/m2; mean
(s.d.) age 52.512.8 years; 39% female), 250 participants
had a BMI of 18.526.9 kg/m2 (56% female), 518 participants
had a BMI of 27.075.9 kg/m2 and no diabetes (46% female),
503 participants had a BMI of 27.167.5 kg/m2 and had a
diagnosis of diabetes (25% female) and in four participants,
the diabetic status was unavailable.

International Journal of Obesity

 Analysis of the Three-Factor Eating Questionnaire
JC Cappelleri et al
614
Evaluation of TFEQ-R21 factor structure in the clinical sample
The Bentlers CFI for the original TFEQ-R21 in the baseline
clinical trial dataset was below 0.9 (Table 1), suggesting
that this model needed refinement. Items 17 (How often
do you avoid stocking up on tempting foods?) and 18
(How likely are you to make an effort to eat less than you
want?) were identified as weak items, with standardized
loadings o0.4 (0.32 and 0.39, respectively). When item 17
was deleted, the CFI remained below 0.9 and the loading
of item 18 became 0.37 (Table 1). To further improve the
model, we deleted item 18. The CFI of the resulting model
remained marginally below 0.9 (0.8958; Table 1), with item
21 (On a scale from 1 to 8, where 1 means no restraint in
eating and 8 means total restraint, what number would you
give yourself?) emerging as a weak item in this 19-item
model (standardized loading 0.39; Table 1). When items
17, 18 and 21 were deleted, the CFI of the resulting 18-item
model (called the TFEQ revised 18-item, version 2FTFEQ-
R18V2) was 0.91, indicating an acceptable model fit, with
no weak items being identified (Figure 1). The factor

Table 1 Bentlers Comparative Fit Indices (CFI) and weak items of candidate Three-Factor Eating Questionnaire (TFEQ) models in the clinical study

Model CFI Identified weak itemsa Standardized regression coefficient loadings of weak itemsb

21-item model 0.8887 17 0.32

F 18 0.39
20-item model (item 17 deleted) 0.8904 18 0.37
19-item model (items 17 and 18 deleted) 0.8958 21 0.39
18-item model (items 17, 18 and 21 deleted) 0.9130 0 NA
a
Item number in the TFEQ questionnaire (see Appendix). bAll other items had loading values of X4.0.

Figure 1 Confirmatory factor analysis: standardized regression coefficient loading estimates and domain correlation estimates for Three-Factor Eating
Questionnaire revised 18-item, version 2 (TFEQ-R18V2) in the clinical and web-based studies. i, item; F, factor (domain); s1, clinical sample; s2, web-based survey.

International Journal of Obesity


structure of the TFEQ-R18V2 is depicted in Figure 1 and
Table 2.
The Cronbachs coefficient a was 0.84 for the UE domain
and 0.92 for the EE domain. The coefficient increased from
0.68 to 0.70 after removal of the three items from the CR
domain of the TFEQ-R21.
Confirmation of the TFEQ-R18V2 in a web-based survey
Using data from the web-based study to test the TFEQ-
R18V2, Bentlers CFI was 0.96 and no weak items were
identified. The internal consistency was also acceptable for
the 18-item, three-factor model with Cronbachs coefficient
a being 0.89 for the UE domain, 0.78 for the three-item CR
domain (0.77 for the original six-item CR domain) and 0.94
for the EE domain.
Characteristics of TFEQ-R18V2
Distribution of responses. No evidence of floor or ceiling
effects was observed with TFEQ-R18V2 for either the clinical
sample or the web-based sample, with all percentages at the
lowest end and the highest end being o50% for all items.
Association between BMI and TFEQ domains. Overall, the
association between BMI and the TFEQ domains was small,
Table 2 Domain structure of the revised Three-Factor Eating Questionnaire
revised 18-item, version 2 (TFEQ-R18V2) (reduced CR domain)
Scale Number of Item numbers in questionnairea
items in domain
Uncontrolled eating (UE) 9 3, 6, 8, 9, 12, 13, 15, 19, 20
Cognitive restraint (CR)b 3 1, 5, 11
Emotional eating (EE) 6 2, 4, 7, 10, 14, 16
a
See Appendix. bItems 17, 18 and 21 were deleted from the original TFEQ-R21
for this model.
Figure 2 Plot of least square mean (and 95% confidence interval) of body mass index (BMI) values versus Three-Factor Eating Questionnaire revised 18-item,
version 2 (TFEQ-R18V2) domain scores for the clinical study. UE, uncontrolled eating; EE, emotional eating; CR, cognitive restraint (revised). (a) BMI plotted against
uncontrolled eating score; (b) BMI plotted against emotional eating score; (c) BMI plotted against revised cognitive restraint score.
Analysis of the Three-Factor Eating Questionnaire
JC Cappelleri et al
615
negligible or absent for both the clinical sample and the
web-based sample. In the clinical sample, the UE and EE
domains were not significantly associated with BMI
(Figure 2). For UE, a one-category difference was associated
with a non-significant change in BMI of only 0.23 kg/m2
(P 0.63). For EE, a change of one-category was associated
with a change in BMI of 0.21 kg/m2, which was also non-
significant (P 0.36). Correlations between UE and BMI, and
EE and BMI in clinical sample were close to zero and not
statistically significant. However, the UE and EE showed a
small, positive and approximately linear relationship with
BMI in the web-based sample (Figure 3). For UE, an increase
of one-category was associated with an increase in BMI of
only 1.51 kg/m2 (P 0.055). An increase of one full
category in EE was associated with a small increase in BMI
of 2.09 kg/m2 (Po0.0001). Correlations between UE and
BMI, and EE and BMI in the web-based sample were 0.17
(Po0.0001) and 0.19 (Po0.0001), respectively.
The relationship between the revised CR domain and BMI
differed across samples. In the clinical study, the revised CR
domain was inversely related to BMIFthat is, as CR in these
obese individuals increased, BMI decreased. On average,
a one-category increase on the CR domain resulted in a
difference of 1.54 kg/m2 (Po0.0001). The estimated corre-
lation between the revised CR domain and BMI in the
clinical sample was 0.15 (Po0.0001). In contrast, a small
positive relationship was found between CR and BMI in the
web-based study. On average, a one-category increase on
the CR domain was associated with an increase in BMI of
0.86 kg m2 (P 0.012). The estimated correlation between
revised CR domain and BMI in the web-based sample was
0.07 (P 0.02).
When the results from the web-based survey were analyzed
by obesity and diabetes status, it was found that the relation-
ship between revised CR domain and BMI was sample-
dependent (Figure 3). In the obese with diabetes subgroup
and the non-obese without diabetes subgroup, there was no
International Journal of Obesity
 Analysis of the Three-Factor Eating Questionnaire
JC Cappelleri et al
616

Figure 3 Plot of least square mean (and 95% confidence interval) of body mass index (BMI) values versus Three-Factor Eating Questionnaire revised 18-item,
version 2 (TFEQ-R18V2) domain scores (web-based survey), with plots for revised CR domain stratified by obesity and diabetes status. UE, uncontrolled eating; EE,
emotional eating; CR, cognitive restraint (revised). (a) BMI plotted against uncontrolled eating score; (b) BMI plotted against emotional eating score; (c) BMI plotted
against revised cognitive restraint score in an obese sample with diabetes; panel d: BMI plotted against revised cognitive restraint score in an obese sample without
diabetes; panel e: BMI plotted against revised cognitive restraint score in an non-obese sample.

significant relationship between the CR domain and BMI.
In the obese without diabetes subgroup (which most
closely matches the clinical sample), there was an inverse
relationship similar to that found in the clinical sample:
a one-category increase on the CR domain resulted in
a decrease of 0.96 kg/m2 in BMI (P 0.024) (versus
1.54 kg/m2 in clinical sample).
Gender and obesity status. Mean scores on the three
domain scores reported by gender (male, female) and BMI
(non-obese: o30 kg/m2, obese: 30 kg/m2) showed that, for
each study, mean domain scores were generally comparable
and not meaningfully different for male and female samples,
and for obese and non-obese samples (Table 3). The
exception was EE, in which females had higher mean scores.
For the obese and non-obese subgroups, the no-constraint
model gave a CFI of 0.9536 and the invariance model gave a
CFI of 0.952. As these CFI values exceeded 0.90 and their
difference was negligible (0.0016), the invariance and
stability of the TFEQ measurement model relative to obesity
status was supported by the data. For gender subgroups, the
no-constraint model gave a CFI of 0.9538 and the invariance
model gave a CFI of 0.953. As these CFI values exceeded 0.90
and their difference was negligible (0.0008), the invariance
and stability of the TFEQ measurement model relative to
gender was supported by the web-based data. Data from the
clinical sample, which had sufficient numbers of males and
females for meaningful analysis, concurred: the no-con-
straint model gave a CFI of 0.9125 and the invariance model
gave a CFI of 0.9117, a difference of 0.0008.
Discussion
Before this study, a shortened and refined version of one of
the most widely used eating behavior measurement models
had not been tested in non-European populations. The aim
of this study was to evaluate the factor structure and

International Journal of Obesity


Table 3 Mean domain scores (standard deviation) of domains in clinical and
web-based studies by gender and obesity status
Study Domain
UE CR
Web-based
Non-obese (n 458) 1.83 (0.56) 2.29 (0.80)
Obese (n 817) 2.00 (0.60) 2.40 (0.70)
Male (n 772) 1.94 (0.59) 2.32 (0.75)
Female (n 503) 1.94 (0.59) 2.43 (0.73)
Clinical
Non-obese (n 68) 2.09 (0.52) 2.29 (0.64)
Obese (n 1660) 2.23 (0.55) 2.13 (0.66)
Male (n 306) 2.17 (0.59) 2.00 (0.62)
Female (n 1422) 2.23 (0.54) 2.17 (0.66)
Abbreviations: CR, cognitive restraint; EE, emotional eating; UE, uncontrolled
eating.
reliability of the TFEQ-R21, together with its relationship
with BMI, in two independent populations: (1) obese
patients recruited into a US/Canadian clinical trial for a
weight management therapy and (2) a web-based survey
including obese and non-obese individuals. Findings from
the factor analysis on the basis of the clinical study data
indicated that the TFEQ-R21 needed refinement. An 18-item
version of the questionnaire (the TFEQ-R18V2), with three
items deleted from the CR domain, showed satisfactory
measurement properties in the clinical sample. The factor
structure of the TFEQ-R18V2 also fit the data from the
web-based study. Furthermore, the three-item CR domain in
TFEQ-R18V2 had improved internal consistency compared
with the six-item CR domain in the 21-item model, as
measured by Cronbachs coefficient a. Moreover, the factor
structure of TFEQ-R18V2 was found to be stable across these
study samples and different modes of administration.
Across the populations tested in our study, the UE and EE
domains were shown to be robust and stable. It was
surprising that the original six-item CR domain did not
show stability in the clinical study presented here. Earlier
studies of the TFEQ-R18/R21 suggested that the six-item CR
domain is robust.11,18 However, one French study showed
that item 17, which asks the question How often do you
avoid stocking up on tempting foods? failed in the
convergent and discriminant validity tests, suggesting some
underlying weakness.17,27 In our study, deletion of item
17 highlighted further weaknesses in other items of the CR
domain, namely items 18 and 21. As these items seem to
be acceptable in most Swedish populations tested,11,16,18 but
not in a French population17 or these North American
populations, it is possible that cultural differences might
contribute to the weakness of these items seen in the
populations assessed in this study. The stability of the factor
structure identified in this study across distinct samples and
different modes of administration supports the potential
Analysis of the Three-Factor Eating Questionnaire
JC Cappelleri et al
617
utility of the TFEQ-R18V2 in clinical and general population
settings.
The relationship of the EE and UE domains of the TFEQ
with BMI was in the anticipated direction for both the clinical
EE sample and the web-based sample, although this relationship
was not very strong (a one-category change in a domain led to
1.72 (0.72) a modest change in BMI), and mostly not significant. Initial
1.99 (0.78) findings suggested differences in the relationship between the
CR domain and BMI across samples (CR was inversely related
1.80 (0.75)
2.03 (0.79) to BMI in the clinical sample and positively related to BMI in
the web-based sample). However, subsequent subgroup
analysis in the web-based sample indicated that the only
2.27 (0.75)
statistically significant relationship was also an inverse one
2.31 (0.79)
and applied only to the sample subgroup most similar to the
1.90 (0.73) clinical sample: obese individuals without diabetes.
2.40 (0.77) Consistent with the findings presented here, earlier
research has shown that normal weight and obese/
overweight participants differ in the associations between
BMI and eating behavior.22,28 Furthermore, a study by de
Lauzon-Guillain et al.27 found that restrained eating was
associated with adiposity (which included BMI measure-
ments) in normal weight participants, but not in overweight
participants. This finding was not refuted with respect to
the direction of such a relationship in our studyFin the
non-obese population, the association between BMI and CR
was also positive (although not significant).
Mean scores tended to be similar between obese versus non-
obese groups, and between males versus females. The sole
exception is the mean scores on EE, which is higher for
females than for males, also found elsewhere.11 The stability of
the TFEQ measurement model relative to obesity status and to
gender was supported by the clinical and web-based data.
A few other studies reported a significantly higher
(Po0.0001) CR (on the original 21-items TFEQ) in women
than in men,28,29,30,31 which is in contrast to the current
studies, in which similar scores on CR are found between
men and women. Whether these different results between
current and past research on CR between men and women
are because of publication bias (studies with gender differ-
ences on CR tend to be published more than studies without
such gender differences) or something more substantive is
worthy of future research.
Earlier research on the psychometric evaluation of the
original Restraint Scale indicates that it is weight-dependent
and does not seem to be a valid measure in overweight and
obese females.32,33 A simple examination of the Restraint
Scale discloses that only one item relates to dieting strategies;
the other nine items related to weight fluctuation, disin-
hibition, overeating and concern for dieting (which surely
may be different from dieting itself). Unlike the Restraint
Scale, the TFEQ-R18V2 clearly has an internal structure
that is robust and sound. The Restraint Scale and the
TFEQ-R18V2 are indeed different measures conceptually
and psychometrically.
The domains of the revised TFEQ may reflect certain food
intake patterns. The underlying physiological mechanism
International Journal of Obesity
 Analysis of the Three-Factor Eating Questionnaire
JC Cappelleri et al
618
that brings about these eating patterns is yet to be fully
elucidated. Overactivation of the endocannabinoid system
has been proposed as one possible channel that may play a
role as a causative factor in obesity,34,35 with postulations
that endogenous cannabinoids cause an increase in the
sensory response to food and enhances the subsequent
dopamine-mediated reward response.36,37 Physiologically,
this can promote feelings of being unable to control ones
response to eating and food.38,39 Subsequently, this may
increase UE, EE and, in some populations, decrease CR.
Clearly, more research is needed in this area. The use of the
revised TFEQ in future clinical studies may help illuminate
the relationship between eating behaviors and important
clinical outcomes, and thereby clarify the mechanism by
which anorectic agents act to bring about weight loss. Its use
also has the potential to support the development of new
strategies to effectively manage weight and treat obesity.
Some caution is required in interpreting our observations
because of the lack of longitudinal data from our ongoing
clinical trial, which could reveal a relationship between
changes in TFEQ and changes in BMI over time. In addition,
no data were analyzed yet on the sensitivity (within-group
change) and responsiveness (between-group change) to a
treatment known to be beneficial. Testretest reliability and
stability of responses in the absence of change (correlation of
TFEQ-R18V2 at different time points) was also not evaluated,
because such data have not been collected at this time. The
two studies also used different methods for completing the
questionnaire: written completion for the clinical study and
computer screen for the web-based study, which may have
affected the results.
Another limitation of this study is that the clinical and
web-based studies enrolled differently defined patient popu-
lations and differences seen in the results may reflect this.
However, these varied patient populations also lend to the
strength of the data offering a more complete understanding
across a diverse group of individuals. Moreover, information
collected in the web-based survey relating to BMI and
comorbid conditions are self-reported and were not verified
by physicians or clinical records.
Finally, implications of the shorter 18-item version need to
be considered relative to the original 51-item version with
respect to identification of subgroups of restraint, disinhibi-
tion and hunger.14,15 Despite it being more practicable in
clinical trial settings, the shorter version may prevent these
distinctions among subgroups (if they in fact exist) from
surfacing and compromise the characterization of indivi-
duals by generalizing across different subgroups.
To summarize, our study supported the three-factor
structure of the TFEQ-R21 in these two populations. How-
ever, we identified a refined and reduced 18-item model for
use in these populations, which engendered a robust factor
structure, good internal reliability, and no significant ceiling
and floor effects. The TFEQ-R18V2 should be further tested
in other geographical populations to confirm the general-
izability of its factor structure in other samples. The TFEQ
International Journal of Obesity
may be of benefit in understanding the mechanism by which
anorectic agents act and measuring their effectiveness, and
may therefore help in developing strategies to effectively
manage weight and treat obesity.
Acknowledgements
This study was sponsored by Pfizer Inc. Editorial Support was
provided by Joyce A Healey at Pfizer Inc., New London, CT,
USA; Zoe Thornton-Jones, PhD, at Envision Pharma; the
study was funded by Pfizer Inc. Joseph C Cappelleri, Andrew
G Bushmakin, Robert A Gerber, Chris Sexton and Nancy
Kline Leidy participated in the conception and design of the
study, and data interpretation and writing of the manuscript.
Michael R Lowe and Jan Karlsson provided significant advice
and consultation, and participated in the data interpretation
and writing of the manuscript.
References
1 Bray GA. Medical consequences of obesity. J Clin Endocrinol Metab
2004; 89: 25832589.
2 Pillar G, Shehadeh N. Abdominal fat and sleep apnea: the chicken
or the egg? Diabetes Care 2008; 31 (Suppl 2): S303S309.
3 Saadeh S. Nonalcoholic fatty liver disease and obesity. Nutr Clin
Pract 2007; 22: 110.
4 Raebel MA, Malone DC, Conner DA, Xu S, Porter JA, Lanty FA.
Health services use and health care costs of obese and nonobese
individuals. Arch Intern Med 2004; 164: 21352140.
5 Ogden CL, Yanovski SZ, Carroll MD, Flegal KM. The epidemiology
of obesity. Gastroenterology 2007; 132: 20872102.
6 Lindroos AK, Lissner L, Mathiassen ME, Karlsson J, Sullivan M,
Bengtsson C et al. Dietary intake in relation to restrained eating,
disinhibition, and hunger in obese and nonobese Swedish
women. Obes Res 1997; 5: 175182.
7 Svendsen M, Rissanen A, Richelsen B, Rossner S, Hansson F,
Tonstad S. Effect of Orlistat on eating behavior among partici-
pants in a 3-year weight maintenance trial. Obesity (Silver Spring)
2008; 16: 327333.
8 Yeomans MR, Leitch M, Mobini S. Impulsivity is associated with
the disinhibition but not restraint factor from the Three Factor
Eating Questionnaire. Appetite 2008; 50: 469476.
9 Annunziato RA, Lee JN, Lowe MR. A comparison of weight-
control behaviors in African American and Caucasian women.
Ethn Dis 2007; 17: 262267.
10 Stunkard AJ, Messick S. The three-factor eating questionnaire to
measure dietary restraint, disinhibition and hunger. J Psychosom
Res 1985; 29: 7183.
11 Karlsson J, Persson LO, Sjostrom L, Sullivan M. Psychometric
properties and factor structure of the Three-Factor Eating
Questionnaire (TFEQ) in obese men and women. Results from
the Swedish Obese Subjects (SOS) study. Int J Obes Relat Metab
Disord 2000; 24: 17151725.
12 Mazzeo SE, Aggen SH, Anderson C, Tozzi F, Bulik CM. Investigat-
ing the structure of the eating inventory (three-factor eating
questionnaire): a confirmatory approach. Int J Eat Disord 2003;
34: 255264.
13 Hyland M, Irvine S, Thacker C, Dann P, Dennis I. Psychometric
analysis of the Stunkard-Messick Eating Questionnaire (SMEQ)
and comparison with the Dutch Eating Behaviour Questionnaire
(DEBQ). Curr Psychol Res Rev 1989; 8: 228233.
14 Bond MJ, McDowell AJ, Wilkinson JY. The measurement of
dietary restraint, disinhibition and hunger: an examination of

the factor structure of the Three Factor Eating Questionnaire
(TFEQ). Int J Obes 2001; 25: 900906.
15 Westenhoefer J. Dietary restraint and disinhibition: is restraint a
homogeneous construct? Appetite 1991; 16: 4555.
16 Elfhag K, Linne Y. Gender differences in associations of eating
pathology between mothers and their adolescent offspring. Obes
Res 2005; 13: 10701076.
17 de Lauzon B, Romon M, Deschamps V, Lafay L, Borys JM,
Karlsson J et al. The Three-Factor Eating Questionnaire-R18 is
able to distinguish among different eating patterns in a general
population. J Nutr 2004; 134: 23722380.
18 Tholin S, Rasmussen F, Tynelius P, Karlsson J. Genetic and
environmental influences on eating behavior: the Swedish Young
Male Twins Study. Am J Clin Nutr 2005; 81: 564569.
19 Beiseigel JM, Nickols-Richardson SM. Cognitive eating restraint
scores are associated with body fatness but not with other
measures of dieting in women. Appetite 2004; 43: 4753.
20 Lluch A, Herbeth B, Mejean L, Siest G. Dietary intakes, eating
style and overweight in the Stanislas Family Study. Int J Obes Relat
Metab Disord 2000; 24: 14931499.
21 Lawson OJ, Williamson DA, Champagne CM, DeLany JP,
Brooks ER, Howat PM et al. The association of body weight,
dietary intake, and energy expenditure with dietary restraint
and disinhibition. Obes Res 1995; 3: 153161.
22 Boschi V, Iorio D, Margiotta N, DOrsi P, Falconi C. The three-
factor eating questionnaire in the evaluation of eating behaviour
in subjects seeking participation in a dietotherapy programme.
Ann Nutr Metab 2001; 45: 7277.
23 Foster GD, Wadden TA, Swain RM, Stunkard AJ, Platte P, Vogt RA.
The Eating Inventory in obese women: clinical correlates and
relationship to weight loss. Int J Obes Relat Metab Disord 1998; 22:
778785.
24 Hatcher L. A Step-by-Step Approach To Using The SAS System for
Factor Analysis and Structural Equation Modeling. The SAS Institute
Inc: Cary, North Carolina, 1994.
25 Cheung GW, Rensvold RB. Evaluating goodness-of-fit indexes for
testing measurement invariance. Struct Equation Model: A Multi-
disciplinary Journal 2002; 9: 233255.
26 Byrne BM. Structural Equation Modelling with EQS: Basic Concepts,
Applications, and Programming, 2nd edn, Lawrence Erlbaum
Associates: Mahwah, NJ, 2006.
27 de Lauzon-Guillain B, Basdevant A, Romon M, Karlsson J,
Borys JM, Charles MA. Is restrained eating a risk factor for
Appendix
The Three-Factor Eating QuestionnaireFRevised 21-Item
(TFEQ-R21)
1. I deliberately take small helpings to control my weight.
(1) Definitely true, (2) Mostly true, (3) Mostly false, (4)
Definitely false
2. I start to eat when I feel anxious.
(1) Definitely true, (2) Mostly true, (3) Mostly false, (4)
Definitely false
3. Sometimes when I start eating, I just cant seem to
stop.
(1) Definitely true, (2) Mostly true, (3) Mostly false, (4)
Definitely false
4. When I feel sad, I often eat too much.
(1) Definitely true, (2) Mostly true, (3) Mostly false, (4)
Definitely false
5. I dont eat some foods because they make me fat.
Analysis of the Three-Factor Eating Questionnaire
JC Cappelleri et al
619
weight gain in a general population? Am J Clin Nutr 2006; 83:
132138.
28 Provencher V, Drapeau V, Tremblay A, Despres JP, Lemieux S.
Eating behaviors and indexes of body composition in men and
women from the Quebec family study. Obes Res 2003; 11: 783792.
29 Boerner LM, Spillane NS, Anderson KG, Smith GT. Similarities
and differences between women and men on eating disorder risk
factors and symptom measures. Eat Behav 2004; 5: 209222.
30 Hainer V, Kunesova M, Bellisle F, Parizkova J, Braunerova R,
Wagenknecht M et al. The Eating Inventory, body adiposity
and prevalence of diseases in a quota sample of Czech adults.
Int J Obes 2006; 30: 830836.
31 Lowe MR, Thomas JG. Measures of restrained eating: conceptual
evolution and psychometric update. In: Allison D, Baskin ML
(eds). Handbook of Assessment Methods for Obesity and Eating
Behaviors. Sage: New York, in press.
32 Ruderman AJ. Dietary restraint: a theoretical and empirical
review. Psychol Bull 1986; 99: 247262.
33 van Strien T, Herman CP, Engels RCME, Larsen JK, van Leeuwe
JFJ. Construct validation of the Restraint Scale in normal-weight
and overweight females. Appetite 2007; 49: 109121.
34 Rawls SM, Ding Z, Cowan A. Role of TRPV1 and cannabinoid CB1
receptors in AM 404-evoked hypothermia in rats. Pharmacol
Biochem Behav 2006; 83: 508516.
35 Engeli S, Bohnke J, Feldpausch M, Gorzelniak K, Janke J, Batkai S
et al. Activation of the peripheral endocannabinoid system in
human obesity. Diabetes 2005; 54: 28382843.
36 Melis T, Succu S, Sanna F, Boi A, Argiolas A, Melis MR. The
cannabinoid antagonist SR 141716A (Rimonabant) reduces the
increase of extra-cellular dopamine release in the rat nucleus
accumbens induced by a novel high palatable food. Neurosci Lett
2007; 419: 231235.
37 Solinas M, Justinova Z, Goldberg SR, Tanda G. Anandamide
administration alone and after inhibition of fatty acid amide
hydrolase (FAAH) increases dopamine levels in the nucleus
accumbens shell in rats. J Neurochem 2006; 98: 408419.
38 Kirkham TC. Endocannabinoids in the regulation of appetite and
body weight. Behav Pharmacol 2005; 16: 297313.
39 Monteleone P, Matias I, Martiadis V, De Petrocellis L, Maj M,
Di Marzo V. Blood levels of the endocannabinoid anandamide
are increased in anorexia nervosa and in binge-eating disorder,
but not in bulimia nervosa. Neuropsychopharmacology 2005; 30:
12161221.
(1) Definitely true, (2) Mostly true, (3) Mostly false, (4)
Definitely false
6. Being with someone who is eating, often makes me
want to also eat.
(1) Definitely true, (2) Mostly true, (3) Mostly false, (4)
Definitely false
7. When I feel tense or wound up, I often feel I need to
eat.
(1) Definitely true, (2) Mostly true, (3) Mostly false, (4)
Definitely false
8. I often get so hungry that my stomach feels like a
bottomless pit.
(1) Definitely true, (2) Mostly true, (3) Mostly false, (4)
Definitely false
9. Im always so hungry that its hard for me to stop eating
before finishing all of the food on my plate.
(1) Definitely true, (2) Mostly true, (3) Mostly false, (4)
Definitely false
International Journal of Obesity
 Analysis of the Three-Factor Eating Questionnaire
JC Cappelleri et al
620
10. When I feel lonely, I console myself by eating.
(1) Definitely true, (2) Mostly true, (3) Mostly false, (4)
Definitely false
11. I consciously hold back on how much I eat at meals to
keep from gaining weight.
(1) Definitely true, (2) Mostly true, (3) Mostly false, (4)
Definitely false
12. When I smell a sizzling steak or see a juicy piece of
meat, I find it very difficult to keep from eatingFeven if
Ive just finished a meal.
(1) Definitely true, (2) Mostly true, (3) Mostly false, (4)
Definitely false
13. Im always hungry enough to eat at any time.
(1) Definitely true, (2) Mostly true, (3) Mostly false, (4)
Definitely false
14. If I feel nervous, I try to calm down by eating.
(1) Definitely true, (2) Mostly true, (3) Mostly false, (4)
Definitely false
15. When I see something that looks very delicious, I often
get so hungry that I have to eat right away.
(1) Definitely true, (2) Mostly true, (3) Mostly false, (4)
Definitely false
16. When I feel depressed, I want to eat.
(1) Definitely true, (2) Mostly true, (3) Mostly false, (4)
Definitely false
17. How often do you avoid stocking up on tempting foods?
(1) Almost never, (2) Seldom, (3) Usually, (4) Almost always
International Journal of Obesity
18. How likely are you to make an effort to eat less than you
want?
(1) Unlikely, (2) A little likely, (3) Somewhat likely, (4) Very
likely.
19. Do you go on eating binges even though youre not
hungry?
(1) Never, (2) Rarely, (3) Sometimes, (4) At least once a week
20. How often do you feel hungry?
(1) Only at mealtimes, (2) Sometimes between meals (3) Often
between meals (4) Almost always
21. On a scale from 1 to 8, where 1 means no restraint in eating
and 8 means total restraint, what number would you give
yourself?
Mark the number that best applies to you: 1 2 3 4 5 6 7 8.
r2000. HRQL Group HBFGoteborg University on-
Campus Company, Goteborg, Sweden.
The uncontrolled eating domain was composed of
items 3, 6, 8, 9, 12, 13, 15, 19, 20. The cognitive restraint
domain was composed of items 1, 5, 11, 17, 18, 21. The
emotional eating domain was composed of items 2, 4, 7, 10,
14, 16.
Before calculating the domain scores, items 116 should be
reverse coded and item 21 should be recoded as follows: 12
scores as 1; 34 as 2; 56 as 3; 78 as 4.
Note: Items 17, 18 and 21 are not part of the Three-Factor
Eating Questionnaire revised 18-item, version 2 (TFEQ-
R18V2).