Académique Documents
Professionnel Documents
Culture Documents
Clusters of Healthy and Unhealthy Eating Behaviors Are Associated With Body Mass Index
Among Adults William J. Heerman, MD, MPH1,2,3; Natalie Jackson, MPH2,3; Margaret
Hargreaves, MD4; Shelagh A. Mulvaney, PhD2,5,6; David Schlundt, PhD3; Kenneth A.
Wallston, PhD5; Russell L. Rothman, MD, MPP2,3
ABSTRACT Objective: To identify eating styles from 6 eating behaviors and test their association with body mass
in- dex (BMI) among adults. Design: Cross-sectional analysis of self-report survey data. Setting: Twelve primary
care and specialty clinics in 5 states. Participants: Of 11,776 adult patients who consented to participate, 9,977
completed survey questions. Variables Measured: Frequency of eating healthy food, frequency of eating unhealthy
food, breakfast frequency, frequency of snacking, overall diet quality, and problem eating behaviors. The primary
depen- dent variable was BMI, calculated from self-reported height and weight data. Analysis: k-Means cluster
analysis of eating behaviors was used to determine eating styles. A categorical variable representing each eating
style cluster was entered in a multivariate linear regression predicting BMI, controlling for covariates. Results: Four
eating styles were identified and defined by healthy vs unhealthy diet patterns and engage- ment in problem eating
behaviors. Each group had significantly higher average BMI than the healthy eating style: healthy with problem
eating behaviors (b 1⁄4 1.9; P < .001), unhealthy (b 1⁄4 2.5; P < .001), and unhealthy with problem eating behaviors
(b 1⁄4 5.1; P < .001). Conclusions and Implications: Future attempts to improve eating styles should address not only
the consumption of healthy foods but also snacking behaviors and the emotional component of eating. Key Words:
obesity, body mass index, nutrition, behavioral science, eating patterns (J Nutr Educ Behav. 2017;-:1-7.) Accepted
February 5, 2017.
INTRODUCTION
Fighting obesity requires more than eatinglessandexercisingmore.Certainly caloric intake and physical activity fre- quency and
intensity are the primary determinants of energy balance. Howev- er, both weight loss and maintenance
of healthy weight are best achieved through sustained adherence to a broader range of healthy eating (eg, increased
fruit/vegetable intake) and physical activity (eg, reduced seden- tary time) behaviors. Based on a foun- dational understanding of
the complex and multilevel determinants of healthy
eating and healthy physical activity, much work has been done to develop interventions that facilitate these healthy behaviors.1,2
In the US obesity affects 36.5% of adults and 17.0% of children.3 With the long- term health complications of obesity, including
diabetes, heart disease, and cancer, the continued effort to under- stand which eating behaviors support 1General Pediatrics,
Vanderbilt University Medical Center, Nashville, TN
achieving a healthy weight is of
para- 2Department of Internal Medicine and Public Health, Vanderbilt University, Nashville, TN
mount importance.4,5 3Center
for Health Services Research, Vanderbilt University, Nashville, TN
Changes in eating behaviors
have 4Department of Internal Medicine, Meharry Medical College, Nashville, TN
been independently associated
with 5School of Nursing, Vanderbilt University, Nashville, TN
long-term changes in weight.6 In
6Department of Biomedical Informatics, Vanderbilt University, Nashville, TN
particular, behaviors such as
skipping Conflict of Interest Disclosure: The authors’ conflict of interest disclosures can be found online
meals, snacking, drinking
sweetened with this article on www.jneb.org.
beverages, and eating fast food
were Address for correspondence: William J. Heerman, MD, MPH, General Pediatrics, Vander-
frequently studied as potential
contrib- bilt University Medical Center, 2146 Belcourt Ave, 2nd Fl, Nashville, TN 37212; Phone:
utors to obesity.7-9Strong
evidence links (615) 322-7080; Fax: (615) 343-7650; E-mail: Bill.Heerman@vanderbilt.edu
skipping breakfast with obesity,
particularly ©2017 Society for Nutrition Education and Behavior. Published by Elsevier, Inc. All rights
among children.10,11However, a
recent reserved.
meta-analysis of 153 articles
exam- http://dx.doi.org/10.1016/j.jneb.2017.02.001
ining the association between eating
Journal of Nutrition Education and Behavior Volume -, Number -, 2017 1
behaviors and obesity among adults and children concluded that evidence was insufficient to draw meaningful conclusions owing
to 2 major limita- tions.12 First, most existing studies did not adequately consider potential confounders of the proposed associa-
tions. Second, most existing studies considered only how a single eating behavior was associated with obesity and did not
account for the contribu- tion of other potentially related eating behaviors.
Therefore, the objective of this study was to address the existing gap in the literature by considering how a broad range of
eating behaviors relate to body mass index (BMI), controlling for po- tential confounders including socio- demographics (ie, age,
gender, race/ ethnicity) and physical activity in a large sample of adults. This was accom- plished by first considering how 6
measures of diet quality and eating be- haviors clustered together into patterns of eating styles, and then testing whether those
eating styles were asso- ciated with BMI.
Measures
The primary independent (predictor) variables consisted of 6 self-report eating behaviors: (1) frequency of eating healthy food (2
items); (2) frequency of eating unhealthy food (3 items); (3) breakfast frequency (1 item); (4) fre- quency of snacking (1 item);
(5) overall diet quality (1 item); and (6) problem eating behaviors (4 items). A complete list of survey items included in this anal-
ysis and how they were scored is avail- able in the Supplementary Material. The frequency of healthy and un- healthy food scores
as well as the mea- sure of problem eating behaviors were developed based on a factor analysis of 9 items from the survey, each
of which was answered on a 4-point response scale from 1 1⁄4 never to 4 1⁄4 often; once a day or more. The healthy food variable
was a composite frequency of eating fruits and vegetables (range, 2–8). The unhealthy food variable was a
composite frequency of consuming fast food, sugary drinks, and desserts (range, 3–12). Self-reported frequencies of eating
breakfast and snacking were scored on a 3-point response scale from 0 1⁄4 never to 2 1⁄4 daily. Self-reported diet quality was
scored on a 5-point response scale from 1 1⁄4 poor to 5 1⁄4 excellent. These items were used in previous survey research and
demon- strated validity in those contexts (ie, a prospective cohort of 3,000 hospital- ized patients with acute coronary syn- drome
or heart failure15; and 67,928 participants enrolled in the Southern Community Cohort Study).16 The 4- item measure of
problem eating be- haviors was based on a previously vali- dated scale; in this study it had high internal consistency (Cronbach a
1⁄4 .83).17
The primary dependent variable was BMI, calculated from self-reported height and weight data. Demographic variables were
all self-reported and included age, gender, race/ethnicity, and household income. Self-reported physical activity was assessed as a
potential confounder with a single item that ranged from 11⁄4 I am very inactive to 5 1⁄4I am active most days.
Demographiccharacteristicsincluding age, gender, household income, and race/ethnicity were summarized using mean and SD
for continuous variables and proportions for categorical vari- ables. Annual household income was categorized into 3 mutually
exclusive categories (<$35,000, $35,000 to $75,000, and>$75,000). Race/ethnicity was categorized into 4 mutually exclu- sive
categories (white, non-Hispanic; black, non-Hispanic; Hispanic; and other, non-Hispanic).
Analytic Strategy
k-Means cluster analysis was used for the 6 independent eating behavior variables to determine eating styles. A scree plot of the
weighted sum of squares and the proportional reduc- tion of the weighted sum of squares against a range of potential clusters
(range, 1–20) was used to identify the optimal number of clusters neces- sary to prespecify in the k-means algo- rithm.18 Based
on the scree plot, 4 initial centroids were randomly iden- tified and 10,000 repetitions of the k- means algorithm were conducted
to achieve convergence of the solution
(ie, no observations changed clusters for the final centroids).19 A k-means cluster analysis was first conducted on roughly half of
the sample (re- cruited from 1 of the 3 health systems in the CDRN). The analysis was then repeated on the other half of the sam-
ple. These 2 independent cluster ana-
Journal of Nutrition Education and Behavior Volume -, Number -, 2017 Heerman et al 3
Figure 1. Of the 11,776 participants who consented to participate, 10,807 (91.8%) completed at least 70% of survey items and
9,977 had complete data for analysis. The response rate as defined by the Council of American Survey Research Organiza- tions
was 16.4%.23
lyses generated similar cluster centers with nearly identical relationships with BMI. Consequently, the k-means cluster analysis
was repeated on the total sample (n 1⁄4 9,977), which is re- ported here. Because k-means cluster analysis is sensitive to missing
data, only individuals with complete data
on BMI and the 6 predictor variables were included.
To represent the patterns of eating behaviors in each cluster, each of the 6 predictor variables was standardized and presented
as the mean z-score of each variable as it was grouped per cluster (this was also done within the k-means clustering algorithm).
This allowed for comparison across scales with different numbers of items and score ranges.
A categorical variable with levels rep- resenting each replicable cluster served as the primary independent variable for the
subsequent set of analyses. The primary analysis used multiple linear regression with BMI as the dependent variable, adjusting
for age (continuous), gender (dichotomous), household in- come (ordinal), race/ethnicity (categori- cal), and physical activity
(continuous). A secondary analysis in which 3 sepa- rate logistic regressions compared cate- gorical weight status as the
dependent variable and eating cluster as the inde- pendent variable (controlling for the same covariates) was completed to clarify
the clinical interpretation of the results. Weight status was defined using the following cut points: normal weight (reference)1⁄4
BMI$18.5 and<25 kg/m2; overweight1⁄4BMI$25 and<30 kg/m2; obesity 1⁄4 BMI $30 and <40 kg/m2; morbid obesity 1⁄4 BMI
$40 kg/m2.3 Statistical significance was set at a < .05.
RESULTS
Of the 11,776 respondents who con- sented to participate in the survey, 9,977 (84.7%) had complete data on study variables and
were included in this analysis. The Table lists full demo- graphic characteristics of the overall sample. The majority of
respondents were female and self-identified as non-Hispanic white; 41.0% had an annual household income >$75,000. Although
the average BMI was 29.3 (SD 7.4), over one third of respondents were obese (BMI $30 kg/m2) and almost 20% were morbidly
obese (BMI $40 kg/m2).
DISCUSSION
These data demonstrate 4 coherent pat- ternsofbehaviorsreflecting eating styles associated with BMI independent of other
demographic covariates. The most notable characteristic of these clusters is the discriminant value of combining the frequency of
eating healthy and unhealthy foods with problem eating behaviors (ie, overeating, impulsive eating, and emotional eating).
Specifically, respondents who reported a healthy diet pattern had higher average BMI as well as a higher frequency of obesity
and morbid obesity when they also re- ported engaging in more frequent problem eating behaviors. Respon- dents who reported
an unhealthy diet pattern even in the absence of problem eating behaviors had a higher BMI than did those with a healthy diet
who also avoided problem eating be- haviors. Respondents' BMI and those for the proportion of individuals with obesity or
morbid obesity was highest if respondents had an unhealthy diet pattern and simultaneously engaged in problem eating behaviors.
Thus these results demonstrate that both unhealthy eating patterns and problem eating be- haviors independently increased the
odds of a person being obese or morbidly obese. When unhealthy eating patterns and problem eating behaviors were combined,
the OR of obesity was 5.8 and the OR of morbid obesity was 8.9. Thesefindings suggested that clus- ters of eating behaviors may
be more clinically relevant than any single component dietary assessment.
These data have direct implications for developing interventions and clin- ical recommendations for eating behav- iors to
support healthy weight. One main advantage of the clustering approach used for this analysis is that it moved past the
examination of a sin- gle behavior. For example, skipping
Journal of Nutrition Education and Behavior Volume -, Number -, 2017 Heerman et al 5
Figure 2. Four clusters of eating styles were identified using k-means cluster analysis. Standardized z-scores for each of the
scales included in the k-means analysis are shown within the final cluster assignment. Scales are oriented so that healthier behav-
iors have a positive z-score and unhealthier behaviors have a negative z-score.
breakfast and eating at fast-food restau- rants each were associated with higher BMI, both cross-sectionally and in longi- tudinal
data.12These data were consistent with those findings, in that unhealthy diet patterns including consumption of fast food (eg,
pizza), sugary drinks, and desserts (including candy) were associated with higher BMI. These data also suggested that snacking
behaviors clustered with unhealthy eating pat- terns, which is important for address- ing the potential effect of snacking on an
unhealthy weight trajectory. In addition, these data highlighted the importance of problem eating behav- iors such as overeating,
impulsive eating, and emotional eating. Regard- less of diet quality or the frequency of healthy food consumption, problem eating
behaviors confer an additional risk of higher BMI. Because the prob- lem eating behaviors questionnaire used in this study (The
Personal Dia- betes Questionnaire-4) had only 4 items, it may be possible to incorpo- rate screening for these types of behav- iors
into routine clinical practice.
One main limitations of this study was that it was cross-sectional, which limited causal inference. This was partic- ularly
important in this context because it was unclear whether unhealthy eating patterns precede obesity or vice versa.20
In addition, all of the data were self- reported and therefore could have been affected by a social desirability bias, the direction of
which was uncer- tain because the outcome (BMI) was likely related to the bias. However, chart review on approximately half of
those surveyed showed a high correlation (r 1⁄4 0.94) between self-reported and objectively measured weight, which somewhat
alleviated the concern about social desirability bias. Although the items dealing with diet and exercise were used successfully in
other studies, it is a limitation in that the researchers did not revalidate them with this spe- cific population before including them
in this survey, although the pattern of associations of those items with BMI in this study provides strong evidence of their
construct validity. To minimize participant burden, the measure of physical activity was limited to a single item, which may not
have been suffi- ciently sensitive to detect variation in the behavior. Selection bias may also havebeenaconcern,giventhelargeper-
centage of individuals recruited online (or electronically). Consequently, the sample studied was from a predomi- nantly older,
female population and may not be generalizable to a broader, more racially or ethnically diverse popu- lation. This may suggest
that this type of analysis be confirmed in additional de- mographic groups before broad policy or program implications are imple-
mented. However, the geographic varia- tion of the sample mitigated this concern to a degree, and the percentage of patients with
obesity in this sample (38%) closely mirrored the recently re- ported national prevalence.3 Although the items included in this
survey were validated in other populations, and although the strong associations with BMI in this study provided preliminary
evidence of construct validity, they were not validated in this specific popu- lation, which may have led to a misclas- sification
bias.
ACKNOWLEDGMENTS
The Mid-South CDRN was initiated and funded by the Patient-Centered Outcomes Research Institute through
6 Heerman et al Journal of Nutrition Education and Behavior Volume -, Number -, 2017
Figure 3. Distribution of body mass index (BMI) as measured by kernel density according to cluster of eating behavior. The
default STATA (version 14.2; StatCorp, College Station, TX) settings for kernel density were used, including the Epanechnikov
kernel and the calcu- lated optimal bandwidth of the kernel. The inset figure shows the adjustedband 95% con- fidence interval
(CI) from the linear regression evaluating the association between eating cluster and BMI, controlling for age, gender, household
income, race/ethnicity, and phys- ical activity. The referent group is the healthy eating cluster.
contract CDRN-1306-04869, the Van- derbilt Institute for Clinical and Trans- lational Research with support grant from
ULTR000445, from the National Center for Advancing Translational Sci- ences/National Institutes of Health; and institutional
funding. Dr Heerman's time was supported by a K12 grant fr- om the Agency for Healthcare Research and Quality
(K12HS022990) and a K23 grant from the National Heart, Lung, and Blood Institute (K23 HL127104).
SUPPLEMENTARY DATA
Supplementary data related to this article can be found online at http:// dx.doi.org/10.1016/j.jneb.2017.02.001.
REFERENCES
1. Huang TT, Drewnosksi A, Kumanyika S, Glass TA. A systems-oriented multilevel
framework for addressing obesity in the 21st century. Prev Chronic Dis. 2009;6:A82. 2.Waters E, de Silva-Sanigorski A, Hall BJ,
et al. Interventions for prevent- ing obesity in children. Cochrane Data- base Syst Rev. 2011:CD001871. 3. Ogden CL, Carroll
MD, Fryar CD, Flegal KM. Prevalence of obesity among adults and young: United States, 2011- 2014. Hyattsville, MD: National
Center for Health Statistics; 2015. NCHS Data Brief No. 219. 4. Bhupathiraju SN, Hu FB. Epidemi- ology of obesity and
diabetes and their cardiovascular complications. Circ Res. 2016;118:1723-1735. 5.Calle EE, Rodriguez C, Walker- Thurmond K,
Thun MJ. Overweight, obesity, and mortality from cancer in a prospectively studied cohort of U.S. adults. N Engl J Med.
2003;348: 1625-1638. 6. Mozaffarian D, Hao T, Rimm EB, Willett WC, Hu FB. Changes in diet and lifestyle and long-term
weight gain in women and men. N Engl J Med. 2011;364:2392-2404. 7. Pereira MA, Kartashov AI, Ebbeling CB, et al. Fast-food
habits, weight gain, and insulin resistance (the CARDIA study): 15-year prospective analysis. Lancet. 2005;365:36-42. 8.
Rosenheck R. Fast food consumption and increased caloric intake: a system- atic review of a trajectory towards weight gain and
obesity risk. Obesity Rev. 2008;9:535-547. 9. Barnes TL, French SA, Mitchell NR, Wolfson J. Fast-food consumption, diet
quality and body weight: cross- sectional and prospective associations in a community sample of working adults. Pub Health
Nutr. 2016;19:885-892. 10. Berkey CS, Rockett HR, Gillman MW, Field AE, Colditz GA. Longitudinal study of skipping break-
fast and weight change in adolescents. Int J Obesity Relat Metab Disord. 2003; 27:1258-1266. 11. van der Heijden AA, Hu FB,
Rimm EB, van Dam RM. A prospective study of breakfast consumption and weight gain among U.S. men.Obesity (Silver
Spring). 2007;15:2463-2469. 12. Mesas AE, Munoz-Pareja M, Lopez- Garcia E, Rodriguez-Artalejo F. Selected
eatingbehavioursandexcessbodyweight: a systematic review.Obesity Rev. 2012;13: 106-135. 13. Harris PA, Taylor R, Thielke R,
Payne J, Gonzalez N, Conde JG. Research elec- tronic data capture (REDCap)—a
metadata-driven methodology and workflow process for providing transla- tional research informatics support. J Bio- med
Inform. 2009;42:377-381. 14. Rosenbloom ST, Harris P, Pulley J, et al. The Mid-South clinical Data Research Network. J Am
Med Inform Assoc. 2014;21:627-632. 15. Meyers AG, Salanitro A, Wallston KA, et al. Determinants of health after hos- pital
discharge: rationale and design of the Vanderbilt Inpatient Cohort Study (VICS). BMC Health Services Res. 2014;14:10. 16.
Signorello LB, Munro HM, Buchowski MS, et al. Estimating nutrient intake from a food frequency question- naire: incorporating
the elements of race
Journal of Nutrition Education and Behavior Volume -, Number -, 2017 Heerman et al 7
and geographic region. Am J Epidemiol. 2009;170:104-111. 17. Stetson B, Schlundt D, Rothschild C, Floyd JE, Rogers W,
Mokshagundam SP. Development and validation of The Personal Diabetes Questionnaire (PDQ): a measure of diabetes self-care
behaviors, perceptions and barriers. Diabetes Res Clin Practice. 2011;91:321-332. 18. Makles A. Stata tip 110: how to get the
optimal k-means cluster solution. Stata J. 2012;12:347-351. 19. Tan P-N, Steinbach M, Kumar V. Introduction to Data Mining.
1st ed. Bos- ton, MA: Pearson Addison Wesley; 2006. 20. Ludwig DS, Friedman MI. Increasing adiposity: consequence or cause
of
overeating? JAMA. 2014;311: 2167-2168. 21. Raynor HA, Champagne CM. Posi- tion of the Academy of Nutrition and
Dietetics: interventions for the treat- ment of overweight and obesity in adults. J Acad Nutr Diet. 2016;116: 129-147. 22.
American Diabetes Association. Obe- sity management for the treatment of type 2 diabetes. Diabetes Care. 2017;40: S57-S63.
23. Council of American Survey Research Organizations. On the Definition of Response Rates, 1982. http://c.ymcdn.
com/sites/www.casro.org/resource/resmgr/ docs/casro_on_definitions_of_resp.pdf. Accessed March 6, 2017.
7.e1 Heerman et al Journal of Nutrition Education and Behavior Volume -, Number -, 2017
CONFLICT OF INTEREST
The authors have not stated any con- flicts of interest.