Content uploaded by Eric R Helms
Author content
All content in this area was uploaded by Eric R Helms on Aug 29, 2023
Content may be subject to copyright.
See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/338927605
Interactions between different eating patterns on recurrent binge eating
behavior: A machine learning approach
ArticleinInternational Journal of Eating Disorders · January 2020
DOI: 10.1002/eat.23232
CITATIONS
0
READS
3
6 authors, including:
Some of the authors of this publication are also working on these related projects:
A Meta-Analysis of Dropout Rates in Randomised Controlled Trials of Dialectical Behaviour Therapy View project
Jake Linardon
Deakin University
38 PUBLICATIONS346 CITATIONS
SEE PROFILE
All content following this page was uploaded by Jake Linardon on 31 January 2020.
The user has requested enhancement of the downloaded file.
ORIGINAL ARTICLE
Interactions between different eating patterns on recurrent
binge eating behavior: A machine learning approach
Jake Linardon PhD
1
| Mariel Messer BA (hons)
1
| Eric R. Helms PhD
2
|
Courtney McLean BA (hons)
1
| Lisa Incerti BA (hons)
1
|
Matthew Fuller-Tyszkiewicz PhD
1,3
1
School of Psychology, Deakin University,
Geelong, Victoria, Australia
2
Sports Performance Research Institute
New Zealand (SPRINZ), Faculty of Health and
Environmental Science, Auckland University of
Technology, Auckland, New Zealand
3
Center for Social and Early Emotional
Development, Deakin University, Burwood,
Victoria, Australia
Correspondence
Jake Linardon, School of Psychology, Deakin
University, 221 Burwood, Highway, Burwood,
VIC 3125, Australia.
Email: jake.linardon@deakin.edu.au
Abstract
Objective: Previous research has shown that certain eating patterns (rigid restraint, flexi-
ble restraint, intuitive eating) are differentially related to binge eating. However, despite
the distinctiveness of these eating patterns, evidence suggests that they are not mutually
exclusive. Using a machine learning-based decision tree classification analysis, we exam-
ined the interactions between different eating patterns in distinguishing recurrent
(defined as ≥4 episodes the past month) from nonrecurrent binge eating.
Method: Data were analyzed from 1,341 participants. Participants were classified as
either with (n= 512) or without (n= 829) recurrent binge eating.
Results: Approximately 70% of participants could be accurately classified as with or
without recurrent binge eating. Intuitive eating emerged as the most important classi-
fier of recurrent binge eating, with 75% of those with above-average intuitive eating
scores being classified without recurrent binge eating. Those with concurrently low
intuitive eating and high dichotomous thinking scores were the group most likely to
be classified with recurrent binge eating (84% incidence). Low intuitive eating scores
were associated with low binge eating classification rates only if both dichotomous
thinking and rigid restraint scores were low (33% incidence). Low flexible restraint
scores amplified the relationship between high rigid restraint and recurrent binge eat-
ing (81% incidence), and both a higher and lower BMI further interacted with these
variables to increase recurrent binge eating rates.
Conclusion: Findings suggest that the presence versus absence of recurrent binge
eating may be distinguished by the interaction among multiple eating patterns. Con-
firmatory studies are needed to test the interactive hypotheses generated by these
exploratory analyses.
KEYWORDS
binge eating, decision tree classification, dietary restraint, intuitive eating
1|INTRODUCTION
Binge eating is defined as the excessive consumption of food during a
short period of time while at the same time experiencing a sense of
loss of control (American Psychiatric Association, 2013). Binge eating is
prevalent in around 5–10% of adults (Mitchison, Hay, Slewa-Younan, &
Mond, 2012), and the recurrence of this behavior in community
samples (i.e., usually defined as one episode per week on average
Received: 18 August 2019 Revised: 11 January 2020 Accepted: 11 January 2020
DOI: 10.1002/eat.23232
Int J Eat Disord. 2020;1–8. wileyonlinelibrary.com/journal/eat © 2020 Wiley Periodicals, Inc. 1
during a prespecified time-period), independent of whether one has
a confirmed binge-eating disorder (BED) diagnosis (Mitchison, Touyz,
González-Chica, Stocks, & Hay, 2017), has been linked with psychologi-
cal distress (Becker & Grilo, 2015; Mitchison et al., 2018), functional
impairment (Harrison, Mond, Rieger, & Rodgers, 2015; Mitchison,
Mond, Slewa-Younan, & Hay, 2013), and overweight and obesity
(Da Luz et al., 2017). Thus, efforts to better understand, screen, and
treat regular binge eating are needed.
Increasing attention has been devoted toward understanding which
patterns of eating are and are not associated with binge eating. Dietary
restraint is one eating pattern that has received significant research
attention in this domain. According to some (Westenhoefer, Stunkard, &
Pudel, 1999), dietary restraint is a multifaceted construct comprised of
distinct forms that cannot be categorized as entirely adaptive or mal-
adaptive. Westenhoefer et al. (1999) proposed that dietary restraint
should be classified into a rigid or flexible form. Rigid restraint involves
an all-or-none approach to dieting. Individuals who practice this form of
restraint tend to think dichotomously about food and dieting, set them-
selves multiple demanding diet “rules,”andengageinvariousregimented
dieting behaviors (e.g., calorie counting, fasting, skipping meals;
Westenhoefer et al., 1999). This form of restraint has been consistently
shown in experimental (Knight & Boland, 1989), prospective (Agras &
Telch, 1998), and cross-sectional (Linardon, 2018; Tylka, Calogero, &
Daníelsdóttir, 2015) studies to be strongly associated with more severe
and frequent binge eating. Flexible restraint, however, reflects a more
graded approach to dieting, defined by behaviors such as allowing one-
self to eat a wide variety of food types while still paying attention to
weight/shape, and opting for “healthier”foods if “unhealthier”foods
were consumed earlier. When controlling for rigid restraint, several
cross-sectional studies have reported inverse relationships between flexi-
ble restraint and binge eating (Linardon & Mitchell, 2017; Smith,
Williamson, Bray, & Ryan, 1999; Westenhoefer et al., 1999), and
increases in flexible restraint during BED treatment have been associated
with binge eating abstinence (Blomquist & Grilo, 2011), suggesting that a
flexible form of restraint may be a healthier alternative to a rigid form.
Intuitive eating is another pattern of eating gaining significant
research attention. Intuitive eating is a style of eating characterized by
a strong connection with internal hunger and satiety cues, in which
individuals eat when they feel hungry and stop when they feel full
(Tylka & Kroon Van Diest, 2013). Intuitive eaters recognise that all
foods serve a variety of important functions and are less likely to think
of foods as “good”or “bad.”Cross-sectional studies (for a review, see
Bruce & Ricciardelli, 2016) and randomized controlled trials of inter-
ventions designed to nurture intuitive eating (Bacon & Aphramor,
2011) have reported consistent, strong, and inverse relationships
between intuitive eating, all forms of dietary restraint, and binge eat-
ing behavior, suggesting that promoting eating based on internal cues
may be important for binge eating prevention and early intervention.
Several key trends are evident from this extant literature. First,
behaviors and cognitions that are characteristic of rigid restraint seem
to increase risk for, or correlate highly with, binge eating patterns,
whereas intuitive eating and flexible restraint behaviors seem to
decrease this risk. Second, prior work has reported strong bivariate
correlations between flexible and rigid restraint (e.g., Linardon &
Mitchell, 2017; Tylka et al., 2015), indicating that these purportedly
distinct restraint forms may co-occur to some extent, although the
unique contributions of these variables in prior regression models
demonstrate that this co-occurrence is variable. As it stands, research
on the role of different dietary and intuitive eating patterns has only
focused on examining their unique contributions to binge eating
behavior. This standard regression-based approach provides no infor-
mation about the level of co-occurrence among these eating patterns,
nor on the association between any possible co-occurrence and their
interactions with recurrent binge eating behavior.
Thus, the present study uses machine learning-based, decision tree
classification to explore whether the various behavioral and cognitive
characteristics of distinct eating patterns, and their interactions, can be
used to distinguish those with and without recurrent binge eating. Age
and BMI were also examined as potential classifiers of recurrent binge
eating, given the known association between a higher BMI and binge
eating (Da Luz et al., 2017), and that representative data from large com-
munity samples show that recurrent binge eating is most prevalent in the
late teen and early adult years (Mitchison, Hay, Slewa-Younan, & Mond,
2014). We defined recurrent binge eating as engaging in binge eating at
least four times over the past month (once per week) based on partici-
pant self-report, consistent with earlier work (e.g., Harrison et al., 2015;
Harrison, Mitchison, Rieger, Rodgers, & Mond, 2016; Mitchison et al.,
2018). We acknowledge that an interviewer-based assessment using a
longer time-frame (three months for DSM-V and six months for DSM-IV)
is the preferred method to assess recurrent binge eating or to establish
the presence of a BED diagnosis (e.g., DeBar et al., 2011; Striegel-Moore
et al., 2010). However, prior work has shown that (a) the two methods
of recurrent binge eating classification (i.e., the 28-day self-report versus
six-month interview criteria) are associated with comparable levels of
eating pathology and functional impairment (Harrison et al., 2015), and
(b) those who meet the 28-day self-report criteria report greater eating
and general psychopathology than those who binge eat below this
threshold (Harrison et al., 2015; Linardon, Messer, Lee, & Fuller-
Tyszkiewicz, 2019). Thus, the validity and clinical significance of the self-
report recurrent binge eating criteria have been established.
We also note that a key advantage of using a machine learning-based,
decision tree approach the is that the model optimizes the best combina-
tion of variables to enhance the classification of group membership (i.e.,
presence versus absence of recurrent binge eating), and is thus not reliant
on researchers to specify which variables will co-occur and interact. How-
ever, we caution that, as argued by Stice and Desjardins (2018), this
approach is exploratory hypothesis generating, and can be used in further
research to follow-up risk pathways for key outcomes of concern.
2|METHOD
2.1 |Participants
A total of 1,341 participants (91% female) were recruited for this
study. Of these, 512 participants (38%) reported recurrent binge
2LINARDON ET AL.
eating, which we defined as engaging in binge eating at least four
times over the past month (once per week), consistent with prior work
(e.g., Harrison et al., 2015; Harrison et al., 2016). The remaining
829 participants either did not engage in any binge eating (n= 516) or
engaged in binge eating below the required cut-off fre-
quency (n= 329).
2.2 |Procedure
We used a self-selected convenience sample, where participants were
recruited mostly through social media outlets (Facebook, Twitter,
Instagram), online forums, and through word-of-mouth. Advertise-
ments indicated that the study was investigating how certain dietary
patterns impact attitudes toward food, dieting, and our bodies.
Respondents to the advertisements were provided with a link to the
questionnaire battery. Participants completed the questionnaire bat-
tery online and at a time and place of convenience. The questionnaire
took ~20 min to complete. Participants completed the survey once
(which was checked through any duplicate IP address). Ethics approval
was obtained. Informed consent was provided by all participants.
2.3 |Measures
2.3.1 |Independent variables
Intuitive eating behaviors
The 23-item Intuitive Eating Scale-2 (IES-2; Tylka & Kroon Van Diest,
2013) was used to assess intuitive eating behaviors. Each item is rated
along a five-point scale, ranging from one (strongly disagree) to five
(strongly agree). Sample items include “I trust my body to tell me what
to eat”and “I rely on my fullness signals to tell me when to stop eating.”
Scores on each item are averaged to produce a total intuitive eating
score. Scores range from 1 to 5, with higher scores reflect higher
levels of intuitive eating. The internal consistency (α> .77), test–retest
reliability (intraclass coefficients > .80), construct validity, and incre-
mental validity of the IES-2 have been upheld in student (Tylka &
Kroon Van Diest, 2013) and community samples (Duarte, Gouveia, &
Mendes, 2016). IES-2 total scores have also been shown to discrimi-
nate between those with and without clinically significant binge eating
symptoms (Duarte et al., 2016).
Flexible restraint behaviors
The 12-item flexible control subscale of the cognitive restraint scale
(Westenhoefer et al., 1999) was used as a measure of flexible
restraint. Each item receives one point if a participant provides a
response indicative of flexible restraint. For example, on the sample
flexible restraint item “I pay attention to my figure, but I still enjoy a
variety of foods,”participants are asked to indicate whether this state-
ment is true or false of them. Participants who mark true on this item
receive one point that contributes to their flexible restraint total
score. Scores range from 0 to 12, with higher scores reflecting higher
flexible restraint behaviors. The flexible control subscale has demon-
strated good internal consistency (α> .80) construct validity (e.g., via
its association with lower self-reported energy intake and weight loss),
and incremental validity in community samples (Linardon, 2018), in
individuals who are obese (Westenhoefer et al., 1999), and in individ-
uals with BED (Blomquist & Grilo, 2011).
Rigid restraint behaviors
The 16-item rigid control subscale of the cognitive restraint scale
(Westenhoefer et al., 1999) was used as a measure of rigid restraint
behaviors. Similar to the flexible restraint subscale, each item receives
one point if a participant provides a response indicative of rigid
restraint. Sample items include “Sometimes I skip meals to avoid gaining
weight”and “I alternate between times when I diet strictly and times
when I don't pay much attention to what and how much I eat.”Scores
are summed to produce a total score. Scores range from 0 to 16, with
higher scores reflecting higher rigid restraint behaviors. The rigid con-
trol subscale is internally consistent (α> .82) and has demonstrated
construct validity (via its strong connection to other dietary restraint
measures, binge eating symptomatology, and eating concerns) and
incremental validity in community (Tylka et al., 2015), student
(Timko & Perone, 2005), overweight/obese (Westenhoefer et al.,
1999), and BED samples (Masheb & Grilo, 2002).
Rigid restraint cognitions
Rigid restraint beliefs and cognitions were assessed via the inflexible
eating questionnaire (IEQ; Duarte, Ferreira, Pinto-Gouveia, Trindade, &
Martinho, 2017) and the eating subscale from the dichotomous think-
ing in eating disorder scale (DTES; Byrne, Allen, Dove, Watt, &
Nathan, 2008). The 11-item IEQ assesses an individual's perceived
importance of adhering to a set of arbitrary diet rules, a sense of con-
trol derived from meeting these rules, and the distress experienced
when failing to meet these rules. Each item is rated along a 5-point
scale, ranging from 1 (fully disagree)to5(fully agree), and are summed
to produce a total score (score range = 11 to 55). Sample items
include “not following my eating rules makes me feel inferior”and “even
if I feel satisfied with my weight, I do not allow myself to ease my eating
rules.”The internal consistency (α> .85), 4-week test–retest reliability
(r= .84), unidimensional structure, construct validity, and incremental
validity of the IEQ have been established in community samples of
Australian (Linardon, Incerti, & McLean, 2019) and Portuguese
(Duarte et al., 2017) adults. IEQ total scores have also been shown to
successfully discriminate between those with and without elevated
eating disorder symptomatology (Duarte et al., 2017). The four-item
eating subscale from the DTES assesses the extent to which an indi-
vidual holds a polarized view toward food, eating, and dieting. Each
item is rated along a 4-point scale, ranging from 1 (never)to4(always),
and averaged (score range = 1 to 4). Sample items include “I think of
food as either good or bad”and “I view my attempts to diet as either suc-
cesses or failures.”Internal consistency (α> .77) and construct validity
of the eating subscale of the DTES have been established in a general
community sample of adults (Linardon & Mitchell, 2017) and in indi-
viduals with obesity and an eating disorder (Byrne et al., 2008).
LINARDON ET AL.3
2.3.2 |Dependent variable
Recurrent binge eating
A single item from the eating disorder examination questionnaire
(Fairburn & Beglin, 1994) was used to measure recurrent binge eating.
This item asks participants to indicate the frequency with which they had
engaged in binge eating (i.e., eating a large amount of food given the cir-
cumstances, accompanied by a sense of loss of control) over the past
month. For this study, we dichotomized binge eating in terms of the pres-
ence versus absence of recurrent binge eating. Recurrent binge eating
was defined as binge eating at least 4 times over the past 4 weeks.
2.4 |Data analytic strategy
Decision tree classification was undertaken using rpart (Therneau &
Atkinson, 2019) and rattle (Williams, 2011) packages in R (R Core
Team, 2013). Decision tree classification is a recursive partitioning
approach to classifying individuals into groups on a target outcome
measure (i.e., recurrent binge eating). The researcher selects a collec-
tion of independent variables to aid classification of group member-
ship, and decision tree classification then uses these independent
variables to maximize separation of participants into groups based on
scores on these variables. The decision tree starts with a parent node
that contains all participants, and then proceeds to split into sub-
groups (child nodes) that increase predictive accuracy beyond this
base rate. This splitting procedure continues until further improve-
ments cannot be achieved.
In decision tree parlance, these splits may constitute main effects
or interactions, with interactions defined as the impact of a “predictor”
on an “outcome”being dependent upon another predictor (Strobl,
Malley, & Tutz, 2009). An interaction may arise, for instance, if the
tree were split into low versus high intuitive eating scores, but only
one of those branches is then further split by a second variable
(e.g., rigid restraint). In contrast, if both low and high intuitive eaters
are split into high versus low rigid restraint and the effect of high ver-
sus low rigid restraint on probability of belonging to the recurrent
binge eating group is similar (i.e., high rigid restraint individuals are
more likely to be classified with recurrent binge eating regardless of
whether they are high or low intuitive eaters), then this would instead
constitute a main effect. We refer the interested reader to Strobl
et al. (2009) for further discussion of the distinction between main
effects and interactions within the context of decision tree analysis.
Becausedecisiontreeclassification seeks to find the best predictive
model for the data, there is risk of overfitting and subsequently poor repli-
cability of results. Several commonly recommended steps were taken to
mitigate this risk. First, the present sample was split into a training sample
(~70% overall sample, n= 922) for model building, and a test sample
(~30% overall sample, n= 419) to cross-validate model performance. Sec-
ond, the optimal solution from the training set was pruned, a process
whereby the number of branches within the overall decision tree are lim-
ited to reduce complexity (and, in turn, overfitting). This pruning was based
on the cost-complexity criterion using a tuning parameter that sought to
strike a balance between model complexity (punishing more complex
models) and misclassification (error in prediction of group membership).
The tuning parameter (alpha) was chosen as the value that resulted in the
lowest error in prediction. Finally, overall prediction accuracy of the pruned
tree was compared against the unpruned tree to ensure that removal of
child nodes does not diminish model predictive value. In this sample,
50 random samples of the test set were used to confirm that the pruned
tree was not worse than the unpruned decision tree.
Several key outputs from the decision tree classification are
reported in the present study (all with respect to the test data set): (a) a
visual representation of the decision tree, with all the branches (nodes)
that remain after pruning, (b) accuracy of prediction for the two catego-
ries of our DV (recurrent binge eating), and (c) variable importance, a
statistic which quantifies how important an independent variable was
for correctly classifying individuals into groups. Accuracy of the model
overall was augmented with several additional statistics often reported
for diagnostic tools: (a) sensitivity (or true positive rate)—the proportion
of individuals with recurrent binge eating who were correctly identified,
and (b) specificity (true negative rate)—the proportion of individuals
who do not meet criteria for recurrent binge eating who are correctly
classified.
TABLE 1 Comparison of groups on study variables
Recurrent binge eating classification
Total sample (n= 1,341) No (n= 829) Yes (n= 512)
Variable M SD MSD αMSD αd[95% CIs]
Age 29.23 8.11 29.42 8.25 –28.94 7.87 –0.06 [−0.05, 0.17]
BMI 24.48 4.29 24.11 3.78 –25.24 4.67 –−0.27 [0.16, 0.38]
Sex (female) 91% –90% ––93% –
Intuitive eating 3.09 0.34 3.20 0.31 .72 2.92 0.30 .71 0.93 [0.81, 1.04]
Flexible restraint 6.88 2.88 6.80 2.94 .76 7.02 2.79 .73 0.07 [0.04, 0.18]
Rigid restraint 8.10 3.34 7.24 3.27 .77 9.50 2.98 .74 0.71 [0.60, 0.83]
Inflexible eating beliefs 35.30 10.10 33.07 10.31 .90 38.93 8.62 .85 0.60 [0.49, 0.72]
Dichotomous thinking 2.45 0.93 2.18 0.84 .84 2.90 0.82 .85 0.86 [0.74, 0.97]
4LINARDON ET AL.
3|RESULTS
3.1 |Preliminary analyses
Descriptive statistics are presented in Table 1 for individuals classified
with and without recurrent binge eating. Those classified with recur-
rent binge eating reported a higher BMI, higher levels of rigid
restraint, inflexible eating beliefs, and dichotomous thinking, and
lower levels of intuitive eating than those classified with non-
recurrent binge eating. Effect sizes were moderate to large. Negligible
differences in flexible restraint scores, mean age, and percent female
were observed between the two groups.
Correlations between study variables for those classified with and
without recurrent binge eating are presented in Table 2. As seen, the
TABLE 2 Pearson correlations (95% CI) between study variables
Variable 1 2 3 4 5 6 7
1. Intuitive eating –.02 [−.07, .10] −.38 [−.31, −.45] −.22 [−.14, −.30] −.43 [−.35, −.49] −.07 [−.16, .02] −.15 [−.05, −.24]
2. Flexible control −.22 [−.15, −.28] –.50 [.43, .56] .45 [.38, .52] .12 [.04, .21] −.03 [−.12, .06] −.22 [−.13, −.30]
3. Rigid control −.50 [−.44, −.55] .63 [.58, .67] –.58 [.52, .64] .48 [.41, .54] .01 [−.08, .10] .05 [−.04, .14]
4. Inflexible
eating beliefs
−.40 [−.34, −.45] .54 [.49, .59] .68 [.64, .71] –.46 [.39, .53] −.08 [−.17, .01] −.08 [−.17, .01]
5. Dichotomous
thinking
−.49 [−.44, −.54] .34 [.27, .39] .61 [.56, .65] .60 [.55, .64] –−.01 [−.10, .08] .13 [.03, .21]
6. Age −.07 [.00, −.14] .05 [−.02, .11] .02 [−.05. .09] .02 [−.05. .09] .34 [.27, .40] –.16 [.06, .25]
7. BMI −.19 [−.12, −.25] −.01 [−.08, .06] .14 [.07, .20] .08 [.01, .15] .21 [.14, .27] .13 [.06, .19] –
Note: Correlations are presented above the main diagonal for those classified with recurrent binge eating, and below the diagonal for those classifed with
nonrecurrent binge eating.
FIGURE 1 Decision Tree for Classifying Recurrent Binge Eating. Note: White boxes indicate a subgrouping where nonrecurrent binge eating
is more prevalent, whereas gray boxes indicate where the subgroups contain more people with recurrent binge eating. Yes = The percentage of
participants meeting criteria for recurrent binge eating in that split. Reported N at each step reflects total remaining subsample in a given
partition, starting with a full sample of 419 for the parent node. Hence, the first split separates into subsamples of 281 and 138
LINARDON ET AL.5
bivariate correlations ranged from small to large, and correlations
between the same pairs of constructs tended to be larger for the
those classified with nonrecurrent binge eating.
3.2 |Decision tree classification
Figure 1 shows the classification tree for recurrent binge eating based
on the test subsample of the overall data set (n= 419 of 1,341). Intui-
tive eating, dichotomous thinking, rigid dietary restraint, flexible die-
tary restraint, and BMI were identified as the important classifiers of
whether a participant would be categorized with recurrent binge
eating. As seen in Figure 1, participants were split first by intuitive
eating scores. Participants with higher intuitive eating scores (≥2.9)
constituted 67% of the overall test sample and, of these, only 25%
were classified with recurrent binge eating (bottom left-most box).
The remaining 33% of the sample (who reported less than 2.9 on the
IES-2) were next split based on scores on dichotomous thinking. Indi-
viduals with higher scores on dichotomous thinking (≥3.3) constituted
13% of the overall sample, and comprised 84% of recurrent binge eat-
ing classification (see bottom right-most box). This accuracy could not
be improved for this subgroup, so they were not split further.
Individuals with low intuitive eating scores (<2.9) and lower
dichotomous thinking scores (≤3.3) were split further by rigid dietary
restraint, flexible dietary restraint, and BMI. Eighty-one percent of
individuals with low intuitive eating scores (<2.9) and low dichoto-
mous thinking scores (≤3.3) who had high rigid (≥7.5) and low flexible
(<5.5) restraint scores were classified with recurrent binge eating. If,
instead, an individual had low intuitive eating, low dichotomous think-
ing, high rigid restraint, but also high flexible restraint, BMI scores
were needed to determine whether they were classified with recur-
rent binge eating (11% of the overall sample); those with BMI less
than or equal to 23 were more likely to be classified with recurrent
binge eating (70% of this subgroup), whereas those with BMI greater
than 25 were also more likely to be classified with recurrent binge
eating (76%). Thus, for a small band of BMI ranges within the normal
weight category, the algorithm struggled to differentiate recurrent
from nonrecurrent binge eating.
The overall accuracy of this model in classifying recurrent binge
eating was 70%, with specificity of .71 and sensitivity of .68. Accuracy
was slightly higher for classifying nonrecurrent binge eating (71%,
n= 224) than for classifying recurrent binge eating (68%, n= 69; see
Table 3). Finally, variable importance information ranked the variables
(from most to least important for classifying recurrent binge eating
status) as intuitive eating, dichotomous thinking, rigid restraint, inflexi-
ble eating beliefs, BMI, flexible restraint, and then demographic fac-
tors of age and gender.
4|DISCUSSION
We used a machine learning-based, decision tree analysis to explore the
relationships between various eating patterns with recurrent binge eat-
ing. In terms of recurrent binge eating classification, results suggested a
complex five-way interaction between intuitive eating, dichotomous
thinking, rigid restraint, flexible restraint, and BMI. Intuitive eating
emerged as the most important classifier of recurrent binge eating, with
75% of those who scored above average on the IES-2 (>2.9) not being
classified with recurrent binge eating. This finding is consistent with
numerous studies demonstrating that those whose eating is guided by
internal body cues are less likely to exhibit regular binge eating patterns
(Bruce & Ricciardelli, 2016). Intuitive eating's relationship with recurrent
binge eating also interacted with dichotomous thinking and rigid dietary
restraint. Those with concurrently low intuitive eating and high dichoto-
mous thinking scores were the group most likely to receive a recurrent
binge eating classification (84% incidence), while those with low intuitive
eating scores were less likely to receive a recurrent binge eating classifi-
cation (33% incidence rate) only if they also had both low dichotomous
thinking and rigid restraint scores. Thus, it appears that the interaction
between certain cognitive and behavioral characteristics that underpin a
rigid dietary approach are also important features that distinguish recur-
rent from nonrecurrent binge eating, which is consistent with predictions
from the restraint theory (Herman & Mack, 1975) and the cognitive
model of eating disorders (Fairburn, 2008). Flexible restraint and BMI
also contributed to the classification, with low flexible restraint scores
amplifying rigid restraints relationship with binge eating (81% incidence
rate),andbothalower(<23)andhigher (>25) BMI being associated with
recurrent binge eating (70 and 62% incidence rate, respectively). How-
ever, these latter splits with flexible restraint and BMI included few par-
ticipants (<4% of total sample for each of the nodes), so confirming these
findings with larger samples is necessary.
This study highlights the complexity of eating behavior, in terms of
the degree of co-occurrence among purportedly distinct eating patterns
and how they interact with recurrent binge eating behavior. Present find-
ings suggest that it may be beneficial for practitioners to screen, assess,
and enquire about the degree to which one endorses each of these dif-
ferent behavioral and cognitive eating patterns, as this may provide addi-
tional insight toward the nature, function, and frequency of their clients'
binge eating behavior. Gathering this information may, in the long-term,
assist in formulating a treatment plan tailored toward the individual
needs of the client (Macneil, Hasty, Conus, & Berk, 2012).
This study has limitations that should be considered. First, this
was a cross-sectional design, so we cannot make any conclusions
regarding the directions of the modelled relationships. Well-designed
prospective studies are needed to clarify and confirm these explor-
atory findings. Second, the psychometric properties (e.g., test retest
TABLE 3 Classification accuracy
Actual grouping
Nonrecurrent Recurrent
Predicted grouping Nonrecurrent 224 33
Recurrent 93 69
Note: Actual grouping based on self-reported frequency of binge eating;
nonrecurrent = an individual classified with nonrecurrent binge eating;
recurrent = an individual classified with recurrent binge eating.
6LINARDON ET AL.
reliability, unidimensional structure, etc.) of some of the measures
used in the present study have not been clearly established in individ-
uals exhibiting recurrent binge eating. This must be taken into
account. Third, although our model identified which eating styles are
associated with recurrent binge eating, we recognize that different
statistical approaches with more variables to model may produce dif-
ferent results, and might indeed improve classification accuracy (sensi-
tivity and specificity indices). Fourth, participants self-selected to
complete this study, which may have led to biases in the sample, such
that only those with access to the Internet and who were interested
in understanding more about their eating behaviors participated. Fifth,
our criteria for defining recurrent binge eating was based on partici-
pant self-report over the prior 28-days. A semi-structured interview
that assesses binge eating over the prior 3 months is considered the
gold-standard for establishing the presence of recurrent binge eating
and a BED diagnosis. This is because an interviewer has the opportu-
nity to clarify any misunderstandings around the nature of binge eat-
ing and thus gain a more accurate assessment of its occurrence (Berg,
Peterson, Frazier, & Crow, 2011). Even though the clinical significance
of self-reported recurrent binge eating has been established (i.e., via
its comparably strong link to functional impairment to those with an
established BED diagnosis; Harrison et al., 2015), replicating our find-
ings in those with a confirmed BED diagnosis is necessary.
This was the first study to use a decision tree classification analy-
sis to explore the relationships and interactions between various eat-
ing patterns with recurrent binge eating behavior. Present findings
suggest that those exhibiting recurrent binge eating behavior may be
distinguished by the complex interaction among various eating and
weight-related characteristics. It will be important for future confirma-
tory studies to test the interactive hypotheses generated by these
exploratory analyses, as this could have important implications for the
assessment, formulation, and treatment of recurrent binge eating.
DATA AVAILABILITY STATEMENT
We do not have ethical clearance to publicly share our data.
ORCID
Jake Linardon https://orcid.org/0000-0003-4475-7139
REFERENCES
Agras, W. S., & Telch, C. F. (1998). The effects of caloric deprivation and
negative affect on binge eating in obese binge-eating disordered
women. Behavior Therapy,29, 491–503.
American Psychiatric Association. (2013). The diagnostic and statistical
manual of mental disorders, fifth edition.: DSM (p. 5). Virginia: American
Psychiatric Association.
Bacon, L., & Aphramor, L. (2011). Weight science: Evaluating the evidence
for a paradigm shift. Nutrition Journal,10,2–13.
Becker, D. F., & Grilo, C. M. (2015). Comorbidity of mood and substance
use disorders in patients with binge-eating disorder: Associations with
personality disorder and eating disorder pathology. Journal of Psycho-
somatic Research,79, 159–164. https://doi.org/10.1016/j.jpsychores.
2015.01.016
Berg, K. C., Peterson, C. B., Frazier, P., & Crow, S. J. (2011). Convergence
of scores on the interview and questionnaire versions of the eating
disorder examination: A meta-analytic review. Psychological Assess-
ment,23, 714–724. https://doi.org/10.1037/a0023246
Blomquist, K., & Grilo, C. (2011). Predictive significance of changes in die-
tary restraint in obese patients with binge eating disorder during treat-
ment. International Journal of Eating Disorders,44, 515–523. https://
doi.org/10.1002/eat.20849
Bruce, L. J., & Ricciardelli, L. A. (2016). A systematic review of the psycho-
social correlates of intuitive eating among adult women. Appetite,96,
454–472. https://doi.org/10.1016/j.appet.2015.10.012
Byrne, S., Allen, K., Dove, E., Watt, F., & Nathan, P. (2008). The reliability
and validity of the dichotomous thinking in eating disorders scale. Eat-
ing Behaviors,9, 154–162.
Da Luz, F., Sainsbury, A., Mannan, H., Touyz, S., Mitchison, D., & Hay, P.
(2017). Prevalence of obesity and comorbid eating disorder behaviors
in South Australia from 1995 to 2015. International Journal of Obesity,
41, 1148–1153.
DeBar, L. L., Striegel-Moore, R. H., Wilson, G. T., Perrin, N.,
Yarborough, B. J., Dickerson, J., …Kraemer, H. C. (2011). Guided self-
help treatment for recurrent binge eating: Replication and extension.
Psychiatric Services,62, 367–373.
Duarte, C., Ferreira, C., Pinto-Gouveia, J., Trindade, I. A., & Martinho, A.
(2017). What makes dietary restraint problematic? Development and
validation of the inflexible eating questionnaire. Appetite,114,
146–154. https://doi.org/10.1016/j.appet.2017.03.034
Duarte, C., Gouveia, J. P., & Mendes, A. (2016). Psychometric properties of
the intuitive eating scale-2 and association with binge eating symp-
toms in a Portuguese community sample. International Journal of Psy-
chology and Psychological Therapy,16, 329–341.
Fairburn, C. G. (2008). Cognitive behavior therapy and eating disorders.
New York, NY: Guilford Press.
Fairburn, C. G., & Beglin, S. (1994). Assessment of eating disorders: Inter-
view or self-report questionnaire? International Journal of Eating Disor-
ders,16, 363–370. https://doi.org/10.1002/1098-108X
Harrison, C., Mitchison, D., Rieger, E., Rodgers, B., & Mond, J. (2016). Emo-
tion regulation difficulties in binge eating disorder with and without
the overvaluation of weight and shape. Psychiatry Research,245,
436–442. https://doi.org/10.1016/j.psychres.2016.09.005
Harrison, C., Mond, J., Rieger, E., & Rodgers, B. (2015). Generic and eating
disorder-specific impairment in binge eating disorder with and without
overvaluation of weight or shape. Behaviour Research and Therapy,72,
93–99. https://doi.org/10.1016/j.brat.2015.07.002
Herman, C. P., & Mack, D. (1975). Restrained and unrestrained eating.
Journal of Personality,43, 647–660.
Knight, L. J., & Boland, F. J. (1989). Restrained eating: An experimental dis-
entanglement of the disinhibiting variables of perceived calories and
food type. Journal of Abnormal Psychology,98, 412–420.
Linardon, J. (2018). The relationship between dietary restraint and binge
eating: Examining eating-related self-efficacy as a moderator. Appetite,
127, 126–129. https://doi.org/10.1016/j.appet.2018.04.026
Linardon, J., Incerti, L., & McLean, C. (2019). Factor structure and psycho-
metric properties of the inflexible eating questionnaire in a sample of
adult women. Appetite,142, 104380. https://doi.org/10.1016/j.appet.
2019.104380
Linardon, J., Messer, M., Lee, S., & Fuller-Tyszkiewicz, M. (2019). Testing
the measurement invariance of the body image acceptance and action
questionnaire between women with and without binge-eating disorder
symptomatology: Further evidence for an abbreviated five-item ver-
sion. Psychological Assessment,31, 1368–1376. https://doi.org/10.
1037/pas0000761
Linardon, J., & Mitchell, S. (2017). Rigid dietary control, flexible dietary con-
trol, and intuitive eating: Evidencefor their differential relationship to dis-
ordered eating and body image concerns. Eating Behaviors,26,16–22.
Macneil, C. A., Hasty, M. K., Conus, P., & Berk, M. (2012). Is diagnosis
enough to guide interventions in mental health? Using case formula-
tion in clinical practice. BMC Medicine,10, 111.
LINARDON ET AL.7
Masheb, R. M., & Grilo, C. M. (2002). On the relation of flexible and rigid
control of eating to body mass index and overeating in patients with
binge eating disorder. International Journal of Eating Disorders,31,
82–91.
Mitchison, D., Hay, P., Slewa-Younan, S., & Mond, J. (2014). The changing
demographic profile of eating disorder behaviors in the community.
BMC Public Health,14, 943.
Mitchison, D., Hay, P. J., Slewa-Younan, S., & Mond, J. (2012). Time trends
in population prevalence of eating disorder behaviors and their rela-
tionship to quality of life. PLoS ONE,7,1–7. https://doi.org/10.1371/
journal.pone.0048450
Mitchison, D., Mond, J., Slewa-Younan, S., & Hay, P. (2013). Sex differ-
ences in health-related quality of life impairment associated with eat-
ing disorder features: A general population study. International Journal
of Eating Disorders,46, 375–380.
Mitchison, D., Rieger, E., Harrison, C., Murray, S. B., Griffiths, S., &
Mond, J. (2018). Indicators of clinical significance among women in
the community with binge-eating disorder symptoms: Delineating the
roles of binge frequency, body mass index, and overvaluation. Interna-
tional Journal of Eating Disorders,51, 165–169. https://doi.org/10.
1002/eat.22812
Mitchison, D., Touyz, S., González-Chica, D. A., Stocks, N., & Hay, P.
(2017). How abnormal is binge eating? 18-year time trends in popula-
tion prevalence and burden. Acta Psychiatrica Scandinavica,136,
147–155.
R packageversion, 4,1–15. https://CRAN.R-project.org/package=rpart
Smith, C., Williamson, D., Bray, G., & Ryan, D. (1999). Flexible vs. rigid
dieting strategies: Relationship with adverse behavioral outcomes.
Appetite,32, 295–305. https://doi.org/10.1006/appe.1998.0204
Stice, E., & Desjardins, C. D. (2018). Interactions between risk factors
in the prediction of onset of eating disorders: Exploratory hypothe-
sis generating analyses. Behaviour Research and Therapy,105,
52–62.
Striegel-Moore, R. H., Wilson, G. T., DeBar, L., Perrin, N., Lynch, F.,
Rosselli, F., & Kraemer, H. C. (2010). Cognitive behavioral guided self-
help for the treatment of recurrent binge eating. Journal of Consulting
and Clinical Psychology,78, 312–321.
Strobl, C., Malley, J., & Tutz, G. (2009). An introduction to recursive par-
titioning: Rationale, application, and characteristics of classification
and regression trees, bagging, and random forests. Psychological
Methods,14, 323–348.
Terry, T., & Beth, A. (2019). rpart: Recursive Partitioning and Regression Trees.
Timko, C. A., & Perone, J. (2005). Rigid and flexible control of eating
behavior in a college population. Eating Behaviors,6, 119–125.
Tylka, T. L., Calogero, R. M., & Daníelsdóttir, S. (2015). Is intuitive eating
the same as flexible dietary control? Their links to each other and well-
being could provide an answer. Appetite,95, 166–175. https://doi.
org/10.1016/j.appet.2015.07.004
Tylka, T. L., & Kroon Van Diest, A. M. (2013). The intuitive eating scale–2:
Item refinement and psychometric evaluation with college women and
men. Journal of Counseling Psychology,60, 137–153.
Westenhoefer, J., Stunkard, A. J., & Pudel, V. (1999). Validation of the flex-
ible and rigid control dimensions of dietary restraint. International Jour-
nal of Eating Disorders,26,53–64.
Williams, G. J. (2011). Data Mining with Rattle and R: The Art of Excavating
Data for Knowledge Discovery, Use R!, Springer.
How to cite this article: Linardon J, Messer M, Helms ER,
McLean C, Incerti L, Fuller-Tyszkiewicz M. Interactions
between different eating patterns on recurrent binge eating
behavior: A machine learning approach. Int J Eat Disord. 2020;
1–8. https://doi.org/10.1002/eat.23232
8LINARDON ET AL.
View publication statsView publication stats