Group and Longitudinal Intra-Individual Networks of Eating Disorder Symptoms in Adolescents and Young Adults at-risk for an Eating Disorder

Abstract

Several studies have identified risk factors that predict future onset of eating disorders (ED) in adolescence, however, it is currently unknown how specific ED symptom dynamics operate both across time and within individuals. Advances in network methodologies allow for the study of how dynamic symptoms interrelate and predict each other within-persons and across time. In the current study, we used longitudinal group-level (N=1272) (addressing symptom interrelations across people and across time; between-subjects, contemporaneous, and temporal networks) and intra-individual (symptom interrelations within each person and across time; contemporaneous and temporal networks) network analyses (subset n=50) in prospective 48-month interview data in at-risk adolescents and young adults. We computed between-subject networks (how symptoms are associated on average, for group sample only), contemporaneous networks (how symptoms are associated at one time point, accounting for previous time points), and temporal networks (i.e., examining how symptoms predict each other across time). We replicated prior network results which suggest that overvaluation of weight and shape are central in at-risk adolescents. We found that individual networks (n=1 networks) were highly variable across individuals. Overall, our results show how both group-level and longitudinal intra-individual network analysis can inform our understanding of how ED develop in adolescence and point to the importance of conceptualizing development on an individual level of analysis.

Eating disorders are serious mental illnesses that carry extremely high societal and personal costs (Arcelus, Mitchell, Wales, & Nielsen, 2011; Keel et al., 2013; Swanson, Crow, LeGrange, Swendsen, & Merikangas, 2011). Onset of eating disorders typically occurs in adolescence, between the ages of 13-15 and 18-20, with estimates ranging from 6.1% to 13% of children and adolescents receiving a diagnosis of anorexia nervosa (AN), bulimia nervosa (BN) or binge eating disorder (BED) or a subthreshold diagnosis conferring near equal levels of impairment as a full diagnosis (Allen et al., 2013; Stice et al., 2009, 2013; Swanson, Crow, Le Grange, Swendsen, & Merikangas, 2011). As such, adolescence is considered a period of high risk for the development of an eating disorder, especially for female adolescents and young adults in the 13-21 age range. Characterization of how eating disorder symptoms develop is of crucial relevance, especially in at-risk samples of adolescents and young adults (females between 13-21) (e.g., Stice et al., 2009; 2013). Such characterization could aid the development of targeted prevention efforts focused on specific and important eating disorder symptoms that might alleviate onset of a clinical eating disorder.

Research on the development of eating disorders is complicated by the fact that eating disorder symptom presentation is very heterogeneous (e.g., Eddy et al., 2008; Fairburn & Cooper, 2011; Forbush et al., 2017; Krug et al., 2013; Luo, Donnellan, & Klump, 2016; Stice, 2016; Stice, Marti, Shaw, & Jaconis, 2009; Sysko, Hildebrandt, Wilson, Wilfley, & Agras, 2010). For example, even within similar diagnoses individual symptom profiles drastically differ (Bos et al., 2018; Fisher, Medaglia, & Jeronimus, 2018; Levinson, Vanzhula, & Brosof, 2018). Recently, analytical methods have been developed that permit researchers to identify how specific heterogeneous symptoms operate (e.g., Borsboom & Cramer, 2013) and provide insight into how specific symptom dynamics operate within single individuals (Bringmann et al., 2013). One of these novel methodological techniques is network analysis.

Network analysis is a methodology that assumes that psychopathology (including eating disorders) develop and maintains itself through symptom dynamics (Borsboom & Cramer, 2013; McNally, 2016). For example, it might be assumed that the symptom fasting leads to increased hunger, which then leads to the symptom binge eating, which then directly leads to purging (see Levinson et al., 2018 for a review of network methods in eating disorders). However, the pathways between these symptoms might vary between individuals, such that one individual might engage in purging, whereas another might engage in other compensatory behaviors, such as excessive exercise. These variations are important, because identifying individual symptom interrelations can inform personalized treatment. Network analysis allows researchers to utilize statistics to identify how symptoms dynamically interrelate with each other both across time and within-persons, which may lead to clinical interventions personalized based on the primary (or in network terminology: central or most important) symptoms, dependent on the individual.

Studying longitudinal relationships between symptoms is particularly informative because it can elucidate how observations at one time point predict observations at the next time point. Such analyses illustrate how psychopathology progresses over time and help researchers identify risk factors and vulnerabilities for mental disorders. Network analyses extend such methodologies and allow researchers to conduct a more fine-grained analyses using relationships among symptoms instead of total or average scores on questionnaires (Bringmann et al., 2013). For example, temporal networks illustrate dynamic relationships among symptoms and directly test theorized symptom inter-relations, such as in cognitive behavioral therapy (i.e., changes in cognitions lead to changes in emotions; Beck, 1979). This work has been done in other fields, using network analysis to understand the development and maintenance of depression and posttraumatic stress disorder (Aalbers et al., 2018; Greene et al., 2020; Hoorelbeke et al., 2019; Price et al., 2020). For example experience-sampling methodology has tested how resilience predicted residual symptoms of depression and positive affect over time and how fear conditioning impacts trauma symptoms in between-group temporal and contemporaneous networks (Hoorelbeke et al., 2019; Price et al., 2020), though we should note this work did not present individual network models.

The eating disorder field has recently seen a proliferation in the use of network analytics for the conceptualization of eating disorders, though this work has mostly focused on cross-sectional rather than longitudinal network analyses. This work has primarily found that fear of weight gain and overvaluation of weight and shape are core symptoms (i.e., symptoms that correlate most strongly with all other symptoms and are hypothesized to maintain the eating disorder) of eating disorders (Levinson et al., 2017; Forbush et al., 2016; Dubois et al., 2017; Wang et al., 2018), implying that these symptoms may dynamically contribute the most to other symptoms in the network of eating disorder symptoms. These findings fit with theoretical models of eating disorders that place overvaluation of weight and shape at the center of eating disorder pathology (Fairburn, Cooper, & Shafran, 2003). However, notably, these prior network analyses have relied on cross-sectional methods of analysis that do not account for within-person variations nor do they examine prospective relations between symptoms.

While cross-sectional network analyses can provide important insight in the conceptualization of eating disorders, they are unable to (a) characterize temporal relationships between symptoms and (b) examine the significant heterogeneity within symptom profiles of eating disorder behaviors. Most recently, one paper demonstrated the use of longitudinal and personalized networks of eating disorder symptoms (Levinson, Vanhzula, & Brosof, 2018). These authors showed the utility of longitudinal group-level networks (i.e., showing how symptom dynamics operate over time across individuals), as well as intra-individual networks or networks within one individual, showing complex inter-relationships on the individual level. The overarching finding of this research was that both between and across individuals, eating disorder cognitions were more likely (than behaviors) to drive later eating disorder symptoms and behaviors, whereas eating disorder behaviors were likely to receive the most input (i.e., the eating disorder behaviors that receive feedback from or predicted by the most other cognitive symptoms) from other cognitive symptoms. Furthermore, this research showed that even within individuals with the same diagnosis (AN), there were drastic differences in how common eating disorder symptoms related with each other. More work of this nature is needed, especially with at-risk adolescent samples, which represent the ideal time period for understanding how specific symptoms prospectively predict growth in other eating disorder symptoms, which may then develop into clinical diagnoses.

Therefore, the current study examined both group-level (across person) and longitudinal personalized (or intra-individual: within one individual) networks of common eating disorder symptoms in at-risk adolescents who were participating in either an eating disorder prevention or control condition using interview data (all prior network analyses have used self-report with the exception of Wang et al. (2018) who used the EDE interview version in adult participants with binge eating disorder). Specifically, for group-level (across persons) networks, we created three different types of networks: Between-subject networks, contemporaneous, and temporal. For individual networks, we created contemporaneous and temporal networks (for more details please see types of models in methods).

We hypothesized, in line with prior cross-sectional network analyses and work that has found that cognitions predict behaviors both in the short and long term (Levinson et al., 2018a,b; Mason et al., 2019; Vanzhula et al., 2020; Sala et al., 2019; Christian et al., 2020; DuBois et al.,2017, that (a) on average fear of weight gain and overvaluation of weight and shape would be most important (i.e., central or contribute to the most other symptoms in the network) symptoms in between-persons networks, (b) networks would be highly variable within individuals, showcasing the importance of creating personalized networks for developing precision interventions and (c replicating prior longitudinal (contemporaneous and temporal) networks, cognitions would be more likely to be central in contemporaneous networks and to prospectively predict (give output to in network terminology) other symptoms in temporal networks, whereas behaviors would be less central (contemporaneous) and more likely to be prospectively predicted by (receive input from in network terminology) other symptoms (temporal).

Methods

Participants

Participants were 1272 female adolescents and young adults at-risk for eating disorders. 90% (n = 1138) percent of the sample was female adolescents and the remainder were female college students. Participants were included if they endorsed having body image concerns during a phone screening. Nine percent of participants met criteria for an eating disorder at baseline. They were assessed each month for the 12 months prior to participating in the study and for 36 months after starting the study, in which they were randomly assigned to one of two prevention programs (Body Project or Healthy Weight) or two control conditions (assessment only, expressive writing). We used data across all 48-months and tested for stationarity by plotting each individuals change trajectory to ensure that there were no changes due to intervention effects. In the full sample (N = 1272), mean age was 18.54 (SD = 4.22; Range = 13-55). Most participants were European American (n = 848; 67%). Other ethnicities reported were: Hispanic (n = 141; 11%), Asian American (n = 130; 10%), African American (n = 58; 5%), Alaskan Native (n = 28; 2%), Pacific Islander (n = 9; 1%), multiracial (n = 15, 1%), and other (n = 31; 3%). In the subsample of 50 participants, mean age was 18.30 (SD = 5.08; Range = 14-49). Twenty-six participants were European American (52%), ten were Hispanic (20%), six were Asian American (12%), three were African American (6%), one participant was Alaskan Native (2%), one participant identified as multiracial (2%), and three participants identified as other (6%). Parts of these data have been utilized before in several prior studies (Stice et al., 2006; 2008; 2011; 2013; 2015), and two studies have utilized the entirety of the dataset been used (Stice et al., 2017 and Stice et al., 2018): none of this prior research was focused on individual network analysis.

Procedure

Assessments were conducted by female assessors, who had a B.A., M.A., or PhD. in psychology and had attended 24 hours of training where they were instructed in interview skills, diagnostic criteria for eating disorders, observed simulated interviews, and role played interviews. Female students were recruited by mailings and fliers. There was only one inclusion criterion: that participants answer affirmatively when asked if they had body image concerns during a phone screen (“Do you have body image concerns?”). Participants were excluded if they answered “no.” More information on this process can be found in Stice et al., 2008. All participants participated in one of four prevention or control conditions, including the Body Project and Healthy Weight Intervention. In the Healthy Weight intervention, participants were encouraged to make gradual healthy and lasting changes to their diet and physical activity to balance their energy needs with their energy intake, and thereby achieve a healthier weight and body satisfaction. With support from the facilitator and group members, they initiated an individual lifestyle change plan to reduce intake of fat and sugar and to increase exercise using behavioral modification principles. Food and exercise diaries were used to identify behaviors to target in this lifestyle modification and to monitor change. Motivational enhancement activities were used to promote motivation for behavior change. More details on these interventions can be found in Stice et al., 2008.

Measures

Eating pathology.

The semi-structured Eating Disorder Diagnostic Interview (EDDI; Stice et al., 2013a) assessed eating disorder symptoms over the past 12 months at baseline and since previous interview at follow-ups on a month-by-month basis. Essentially, this allowed for prospective examination of eating disorder symptoms across 48 months. The measure assesses frequency of eating disorder behaviors (i.e., “Over the past 4 weeks, have you made yourself sick as a means of controlling your shape and weight?) and severity of eating disorder cognitions (i.e., Over the last 4 weeks has your weight and/or shape been important in influencing how you feel about (judge, think, evaluate) yourself as a person?”). Due to limitations of time-varying network analysis, items with a dichotomous yes/no response were not included, which included six follow up yes-no items regarding the binge eating item. A total of eight items were used: Over the past 4 weeks, how many days did you have binge episodes? Over the past 4 weeks have you made yourself sick as a means of controlling your shape and weight? Over the past 4 weeks, have you taken laxatives or diuretics as a means of controlling your shape and weight? Over the past 4 weeks have you fasted as a means of controlling your shape and weight? Over the past 4 weeks have you engaged in exercise that was intended to burn calories to compensate for “overconsumption ” or eating or drinking? Over the past 4 weeks has your weight and/or shape been important in influencing how you feel about yourself as a person? Over the past 4 weeks have you been afraid that you might gain weight or become fat? Over the past 4 weeks have you felt fat? EDDI eating disorder diagnoses have shown 1-week test-retest reliability (α = .79) and inter-rater agreement (α = .75) based on the full measure for DSM-5 eating disorders, sensitivity to intervention effects, and participants with versus without EDDI-diagnosed eating disorders show greater functional impairment, emotional distress, and treatment care (Stice et al., 2008b, 2009, 2013a, 2013b). The Cronbach’s alpha for our eight items was acceptable (α = .63).

Data Analyses

Data Selection and Imputation.

The entire dataset comprised of 1272 participants was used to created group-level networks. Five percent of the data was missing and was imputed using multiple imputation by chained equations using the mice package in R (version 3.4.3) and the predictive mean modeling method. For the individual networks, 200 individuals from the original non-imputed dataset were selected for analysis using the following procedure. The participants reported symptoms on a month-by-month basis for a 48-month period on the eight symptoms, creating a prospective 48-month dataset. As such, each participant in the dataset had 384 potential values across eight variables (8 x 12 x 4 = 394). Seventy-six percent (972/1272) of the participants had no missing data. Of the participants who did have missing data, they were missing an average of approximately 25% percent of their data (range is 0.02 – 76% of data missing, by individual). The average number of zero values per person is 276. This means that, on average, 276 out of 384 (72%) potential values per-person has a value of zero, which is to be expected given that most participants did not meet criteria for an eating disorder at baseline. This high percentage of zero-values-per-person makes both imputation and network analysis difficult (sometimes impossible). We calculated the standard deviation per-variable, per-person and averaged that by individual to estimate the total per-person variance in the original data. Then, we found the median of that overall, per-person variance and used that cut point to subset the data into high-variance individuals and low-variance individuals, which is a necessity of these types of analyses (in other words, without variance the models will not compute).

To compute individual networks, from the high-variance subset of the data, we used a random selection algorithm to subset 200 individuals to comprise the final dataset for individual networks. From the subset of 200, we proceeded to create individual networks until we reached a predetermined number of 50 converged models. During this process, 29 models did not converge. This process is a necessary because only high variance data will compute, keeping in mind that one individual essentially comprises her own dataset, in which the number of responses (rather than persons) is what is traditionally considered sample size (in this instance the 48 values would be sample size for one individual). To date the largest ‘sample size’ of this procedure has been computed on 40 individuals with anxiety and mood disorders (Fisher et al., 2017). In the current dataset of N = 50, we performed multiple imputation by chained equations using the mice package in R (version 3.4.3), and the predictive mean modeling method.

We compared if the sub-sample of 50 individuals (sub-sample 1) and the remaining sample (sub-sample 2: n = 1222) differed in demographic and eating disorder symptoms. Subsample 1 and sub-sample 2 did not significantly differ in age (d = .05; p = .73). Both samples had similar ethnic characteristics: Sub-sample 1 (52% European American; 11% Hispanic, 10% Asian, 5% African American, 2% Alaskan Native); Sub-sample 2 (68% European American; 20% Hispanic, 12% Asian, 6% African American, 2% Alaskan Native). Participants in the two samples did not significantly differ in the number of binge eating episodes (d = .18; p = .19), laxative use (d = .14; p = .31), fasting (d = .43; p = .01), excessive exercise (d = .24; p = .08), judging oneself based on shape and weight (d = .21; p = .14) at baseline after adjusting for multiple comparisons (adjusted p-value = .008). Individuals in sub-sample 1 had higher fear of weight gain than individuals in the sub-sample 2 (d = .67; p < .001). Individuals in sub-sample 1 were also higher in feeling fat than individuals in the sub-sample 2 (d = .62; p < .001). Individuals in sub-sample 1 had lower vomiting than individuals in the sub-sample 2 (d = .26; p < .001).

Item-selection.

We selected 8-items from the EDDI to ensure adequate coverage of eating disorder behaviors and cognitions based on CBT theory, as well as empirical data, including prior network analyses, implicating these behaviors and cognitions as central to eating disorder pathology.

Types of Models.

For group-level (across persons) networks, we created three different types of networks. First, we created between-subjects networks, which represent on average how symptoms relate with other symptoms across persons (similar to prior cross-sectional network analyses work). This network answers the question: how do eating disorder symptoms correlate with each other across persons? Second, we created contemporaneous networks. Contemporaneous networks show on average how at one time point, while accounting for previous time points, symptoms correlate with each other. This network answers the question: how do eating disorder symptoms correlate with each other at one time point, while accounting for prospective relationships. In other words how do eating disorder symptoms dynamically operate with each other (across individuals)? Third, we created temporal networks. Temporal networks show how symptoms operate (prospectively predict) from one time point to the next (across four week time periods in this case). This network answers the question: how do eating disorder symptoms across time predict later eating disorder symptoms, while controlling for earlier eating disorder symptoms, across individuals? In addition to computing group networks in the full sample, we computed between-subjects, contemporaneous, and temporal networks for each subset of individuals in each of the conditions.

Finally, we randomly selected a subset of participants to create intra-individual or personalized networks focused within one person. For each individual, we created contemporaneous and temporal networks. These networks address two questions, both within each of the individual networks we created (n = 50). To answer these questions we essentially create an individual dataset for each individual. First we answer: how do eating disorder symptoms in one person (50 times) correlate with all other eating disorder symptoms accounting for prospective relationships (contemporaneous network)? Second, within one person (50 times) how do eating disorder symptoms predict later eating disorder symptoms (while controlling for all other symptoms: temporal networks)? For additional details on each of these types of networks and their utility in eating disorders, please see Levinson, Vanzhula, and Brosof, 2018.

Group-Level Networks.

Vector autoregressive (VAR) modeling was used for the analysis of all time-series data (Bringmann et al., 2013). These analyses can identify the relations among symptoms both between and within individuals and across time. We used the multi-level vector autoregressive (mlVAR) package, version 0.4 in R to estimate three group-level networks: between-subject (undirected partial correlation network between the means of subject’s scores), contemporaneous (an undirected partial correlation network within the same time measurement point, while controlling for prior temporal relations), and temporal four week-lag (a directed network displaying within-person symptom relations across a four week time period). Only edges significantly different from zero at the p < .05 level were displayed in the final models, as recommended by (Epskamp et al., 2016). For a more detailed description of each model, see Epskamp et al. (2016). To assess the most central symptoms in each network, we computed strength centrality, a commonly used index of centrality that tends to be the most stable (see for example: Epskamp, Borsboom, & Fried, 2017). For the temporal group-level network (four week lag), inStrength (how much input a symptom received from other symptoms) and outStrength (how much output a symptom provides to other symptoms) values were calculated. The indices of centrality were calculated using the centralityPlot and centralityTable functions in qgraph (Epskamp et al., 2012). We also computed these networks across prevention conditions (see Supplemental Material).

Individual Networks.

We created 50 individual networks using the graphicalVAR package in R (Wild et al., 2010). We only included 50 individuals because it would be infeasible to attempt to run all networks in the entire sample (N = 1272) (see data analytic selection procedure above). We created both temporal (a directed network displaying symptoms predicting each other across a four week lag across 48 months total, while controlling for all other symptoms in the model at the prior measurement) and contemporaneous networks (an undirected network showing how symptoms relate to each other in the same window of measurement, controlling for prior temporal relationships) for each of these individuals. See Epskamp et al. (2017) for more details about these types of networks. Strength centrality was calculated for contemporaneous networks, and InStrength and OutStrength centralities were calculated for directed networks using the centralityPlot and centralityTable functions in qgraph (Epskamp et al., 2012).

Differences across Treatment Groups.

We also tested if there were differences in central symptoms across the four intervention/control groups.

BMI.

We did not include BMI in our models as growing literature suggests that BMI is not an appropriate index of body mass in adolescent populations (e.g., Karchynskaya et al., 2020), as well as the fact that BMI was only assessed at baseline and therefore methodologically would not fit into networks which need multiple assessments of each included node. We report descriptives on BMI below.

Results

Descriptives

Mean BMI and ED symptoms assessed via the EDDI for the full sample and subsample on which we conducted individual networks are included in Table 1.

Table 1.

Mean BMI and eating disorder symptoms in both samples

	Full sample (n = 1272)		Subsample (n = 50)
	Mean (SD)	Range	Mean (SD)	Range
BMI	24.7 (5.11)	14.4-57.8	25.0 (5.08)	19.3-39.6
Binge Eating (Days per month)	.61(2.66)	0-30	1.15 (2.79	0-15
Vomiting (Episodes per month)	.36 (3.33)	0-84	.30 (1.18)	0-6
Laxative Use (Episodes per month)	.13 (1.49)	0-30	1.00 (3.80)	0-24
Fasting (Episodes per month)	.87 (3.66)	0-30	2.52 (5.47)	0-25
Exercise (Days per month)	.75 (2.85)	0-30	1.28 (2.52)	0-10
Judging Self Based on	3.15 (1.52)	0-6	3.76 (1.34)	0-6
Shape/Weight
Fear of Weight Gain	1.66(2.13)	0-6	3.64 (2.32)	0-6
Feeling Fat	2.17 (2.09)	0-6	3.78 (2.31)	0-6

Note. BMI= body mass index

Group-Level Networks (Hypothesis A)

Network graphs.

Between-subject, contemporaneous, and temporal group-level networks are shown in Figure 1. Strength centrality graphs are shown in Figure 2.

Figure 1.

Group-level networks: Temporal, contemporaneous, and between-subjects.

Notes. Green nodes indicate cognitions and blue nodes indicate behaviors. Green lines indicate positive relationship and red line indicate negative relationship between symptoms. Partial correlations are shown. Thicker lines between nodes represent stronger relationships. Edges that were not significantly different from zero were removed from the networks. Significance was set at alpha = .05. Symptom label descriptions: Binge = binge eating; Vomit = vomiting; Laxative = laxative use; Fast = fasting; Exercise = exercising to compensate for overconsumption of eating or drinking; Judge = weight or shape influence how you evaluate yourself as a person; Feargain = afraid you may gain weight or become fat; Feelfat = feel fat.

Figure 2.

Centrality tables for group-level networks

Notes: Symptom label descriptions: Binge = binge eating; Vomit = vomiting; Laxative = laxative use; Fast = fasting; Exercise = exercising to compensate for overconsumption of eating or drinking; Judge = weight or shape influence how you evaluate yourself as a person; Feargain = afraid you may gain weight or become fat; Feelfat = feel fat.

Between-subjects network (n > 1: 1272).

In the between-subjects network (across participants on average across time), feeling fat had the highest strength centrality (S = 1.72) and was strongly correlated with fear of weight gain (partial r = .53). Feeling fat had a medium correlation with overvaluation of shape and weight (partial r = .31). Binge eating had the second highest centrality (S = .71) and was correlated most strongly with vomiting (partial r = .56).

Contemporaneous network (n >1: 1272).

Feeling fat (S = 1.41) and fear of weight gain (S = 1.38) had the highest centrality in the contemporaneous network. These symptoms were moderately correlated with each other (partial r = .32), and feeling fat was correlated with overvaluation of shape and weight (partial r = .23).

Directed temporal network (n > 1: 1272).

In the temporal network (predicting later symptoms while controlling for prior symptoms), binge eating had the highest OutStrength centrality (S = 1.62), negatively predicting fear of weight gain (β = −.04) and feeling fat (β = −.03) and positively predicting vomiting (β = .02). Symptoms with the next highest OutStrength are feeling fat (S = .94) and fear of weight gain (S = .78) predicting multiple other symptoms in the network. Feeling fat positively predicted exercise (β = .02), binge eating (β = .02), fasting (β = .01), and overvaluation of shape and weight (β = .02). Fear of weight gain positively predicted fasting (β = .02), laxative use (β = .01), binge eating (β = .01), and overvaluation of shape and weight (β = .02). Fasting had the highest InStrength centrality (S = 1.30), being positively predicted by exercise (β = .03), feeling fat (β = .01), and fear of weight gain (β = .02). Fear of weight gain (S = 1.18) and feeling fat (S = .94) had the next highest InStrength centrality. Overvaluation of shape and weight positively predicted both fear of weight gain (β = .01) and feeling fat (β = .01) and laxative use negatively predicted feeling fat (β = −.01).

Group Networks by Treatment Condition/Prevention Intervention

Please see Table 2 and Figure 2, which show group networks in each of the four conditions, as well as central symptoms by condition.

Table 2.

Symptoms with the highest strength centrality across four prevention/control conditions.

	n	Temporal In-Strength	Temporal Out-Strength	Contemporaneous	Between-Subject
Control condition (assessment only or educational brochure)	542	Laxative Feargain	Binge Vomit	Feelfat Feargain	Feelfat Feargain
Body dissonance	495	Exercise Fast	Laxative Feelfat	Feelfat Feargain	Feelfat Feargain
Expressive writing	119	Exercise Feargain	Laxative Binge	Feelfat Feargain	Feargain Fast
Healthy weight	116	Binge Laxative	Vomit Judge	Feargain Feelfat	Vomit Exercise
Total sample	1272	Fast Feargain	Binge Feelfat	Feelfat Feargain	Feelfat Binge

Individual Networks (n = 1) (Hypothesis B and C)

These networks compute longitudinal relationships within one person as one person essentially becomes its own dataset, providing complex insight on the personalized individual level. Temporal and contemporaneous individual networks are displayed in Figures 3 for our example participants and all models are in the supplemental online materials. The frequency of the most common symptoms with highest strength centrality in contemporaneous networks and most common symptoms with highest OutStrength and InStrength centralities in temporal networks are shown in Table 3. Overall, fear of weight gain, feeling fat, and overvaluation of weight and shape were most common central symptoms in contemporaneous networks across 50 individuals (i.e., identifying the most common individual symptom and summing across 50 individuals). In temporal networks, exercise, fear of weight gain, overvaluation of weight and shape, and feeling fat were the most common symptoms with the highest OutStrength centrality across 50 individuals. Overvaluation of weight and shape, feeling fat, and fear of weight gain were most common symptoms with the highest InStrength centrality across 50 individuals. It can also be useful to interpret relationships between symptoms within individual networks. We will do so using illustrative examples from three individuals below. All individual networks can be found in the online supplemental materials.

Figure 3.

Group-level contemporaneous and between-subject networks for each of the intervention conditions.

Notes. Green nodes indicate cognitions and blue nodes indicate behaviors. Green lines indicate positive relationship and red line indicate negative relationship between symptoms. Partial correlations are shown. Thicker lines between nodes represent stronger relationships. Edges that were not significantly different from zero were removed from the networks. Significance was set at alpha = .05. Symptom label descriptions: Binge = binge eating; Vomit = vomiting; Laxative = laxative use; Fast = fasting; Exercise = exercising to compensate for overconsumption of eating or drinking; Judge = weight or shape influence how you evaluate yourself as a person; Feargain = afraid you may gain weight or become fat; Feelfat = feel fat.

Table 3.

Frequency of most central symptoms across 50 individual networks

Strength centrality in contemporaneous networks		OutStrength in temporal networks		InStrength in temporal networks
Fear of weight gain	21	Exercise	13	Overvaluation of Wt/Shape	16
Feeling fat	18	Fear of weight gain	10	Feeling fat	15
Overvaluation of Wt/Shape	13	Overvaluation of Wt/Shape	8	Fear of weight gain	12
Fasting	6	Feeling fat	8	Fasting	5
Binge eating	4	Binge eating	6	Binge eating	2
Exercise	3	Fasting	6	Vomiting	1
Laxative use	2	Laxative use	3	Exercise	1
		Vomiting	2

Individual Network Interpretation: Examples

Participant 31.

In this individual temporal (prospective) network of participant 31, there was a direct path from feeling fat predicting fear of weight gain, which then predicted more exercise a month later. Feeling fat regressive effect predicted itself in a self-loop over time. In the contemporaneous network, feeling fat was associated with overvaluation of shape and weight, and fear of weight gain was associated with exercising.

Participant 33.

In this individual temporal network, exercise predicted a reduction in feeling fat and fear of weight gain over one month. Similarly, fasting led to reduction in future feeling fat. Binge eating predicted higher feeling fat and fear of weight gain predicted higher overvaluation based on shape and weigh. Binge eating, fear of weight gain, and feeling fat predicted themselves (i.e., showed auto-regressive effects) over time. In the contemporaneous network, there was an association between fasting and laxative use, between overvaluation based on shape and weight and fear of weight gain, and feeling fat and exercising.

Participant 41.

For this individual, feeling fat led to more exercise and overvaluation of shape and weight a month later, which is the opposite of Participant 33. Fear of weight gain led to increased feeling fat and more binge eating episodes a month later. Exercise, however, reduced overvaluation of weight and shape at the next time point. Similarly, binge eating at the next time point eating predicted less fear of weight gain, but higher overvaluation of shape and weight a month later. Fear of weight gain, overvaluation of weight and shape and feeling fat predicted each other at the next time point. In the contemporaneous network, we saw an association between feeling fat and overvaluation of shape and weight and fear of weight gain. Please see Figure 3 for the illustration of these networks.

Discussion

Overall we found support for the idea that there are many complex relationships among specific eating disorder symptoms over time in adolescents and young adults at-risk for eating disorders. These networks begin to shed light on how specific eating disorder symptoms relate to each other across time both within and between persons in an at-risk adolescent and young adult sample. We also found that while these symptoms vary greatly in individual networks, overall, as hypothesized in hypothesis C, the most common symptoms that provide input to/prospectively predict other symptoms were cognitions focused on fatness and fears of weight gain and binge eating behaviors, whereas the most common symptoms that are influenced by other symptoms (receive input/prospectively predicted by) are fasting behaviors, as well as cognitions, specifically again cognitions related to feelings of fatness and fear of weight gain. This study builds on prior group-level temporal networks (e.g., Hoorelbeke et al., 2019; Price et al., 2020) utilizing time-series data and is the first large scale demonstration of longitudinal individual networks in the eating disorders.

Group-Level Networks (n > 1)

Between-subjects networks.

As hypothesized in our first hypothesis, we found that attitudinal eating disorder symptoms related to weight were the most central symptom across individuals, specifically we found support for ‘feeling fat’ as the most central symptom, which was moderately related to fears of weight gain. Feelings of fatness are a primary symptom of eating disorders, and has been shown to be a distinct component of eating disorder body dissatisfaction (Linardon et al., 2018). This finding replicates prior cross-sectional network analyses finding that fear of weight gain and overvaluation of weight and shape are highly central symptoms (Levinson et al., 2017; Forbush et al., 2016; Dubois et al., 2017; Wang et al., 2018). Overall this finding is congruent with prior network analyses and continues to support the cognitive-behavioral theory of eating disorder suggesting that overvaluation of weight and shape are central to the maintenance of eating disorders (Fairburn, Cooper, & Shafran, 2003). This finding also suggests that a fruitful area for future research is on ‘feelings of fatness’ specifically. One recent study found that ‘feelings of fatness’ has a unique relationship with eating disorder symptoms, which is not accounted by overall overvaluation of weight and shape nor depression (Linardon et al., 2018). ‘Feelings of fatness’ have received less attention in the literature and our findings suggest that such emotions may be important in the development of eating disorders, at least in at-risk adolescents. These findings are noteworthy because only one other prevention program has been found to reduce future eating disorder onset in multiple trials (Stice et al, 2008; 2013) Interestingly, binge eating was the second most central symptom, which is a new finding, only found in one prior network study (Christian et al., 2019) and warrants additional research. Specifically, most prior network analyses studies find that binge eating is a peripheral and not central symptoms (e.g., Levinson et al., 2017). We hypothesize that binge eating may be more central in at-risk adolescents because it has not become a habitual behavior, therefore requiring more attention and effort, whereas in older individuals with clinical eating disorders, binge eating may have become established as a habit, though this hypothesis is extremely tentative.

Contemporaneous networks.

In the contemporaneous networks we found again, as hypothesized, that feelings of fatness and fears of weight gain were the most central symptoms. Both of these central symptoms were associated with each other and feelings of fatness had a moderate correlation with overvaluation of shape and weight. Keeping in mind that these networks represent associations at the current time point, while controlling for temporal relationships. In other words, as hypothesized from network theory (Epskamp et al., 2016), these networks represent symptoms that could be tested in future experimental research as symptoms that are possible causal mechanisms (pending experimental research support). These dynamics in the current paper occur within the context of seconds (as in contemporaneous networks), reflecting that there may be a developmental process, such that feeling fat influences judging oneself and vice versa. This finding has implications for future preventative efforts, such that disrupting the cycle between feelings of fatness might also weaken overvaluation of weight and shape. Additionally, it is interesting that three cognitions, ‘feelings of fatness,’ ‘fears of weight gain,’ and ‘judgment of oneself/overvaluation of weight and shape’ are most central and occur together. It is possible that these symptoms represent a ‘rumination cycle’ explicitly focused on weight and shape related rumination. This suggestion is congruent with research showing that rumination is elevated in individuals with eating disorders (Sala, Brosof, & Levinson, 2019; Smith, Mason, & Lavender, 2018). Several connections between thoughts and behaviors were identified, supporting the cognitive-behavioral framework of these symptoms reinforcing each other (Fairburn, 2008). Exercise was correlated with overvaluation of shape and weight and fear of weight gain was correlated with fasting and binge eating. It is likely that individuals engage in exercise due to overvaluation of shape and weight and, in turn, excessive exercise may lead to hyper-focus on shape and weight, making them even more important. Binge eating may trigger fear of weight gain because of the large amount of food consumed (or perceived as a large amount) and, in turn, fear of weight gain may lead to more binge eating as a coping mechanism for dealing with the fear (Duarte, Pinto-Gouveia, & Ferreira, 2017). Although traditionally fear of weight gain has been conceptualized as part of restricting pathology, more research on its relationship with binge eating may be warranted.

Temporal networks.

In the group level temporal network (showing how across persons symptoms operate over time), we found that binge eating negatively predicted feelings of fatness and fears of weight gain and positively predicted vomiting. We are unsure of why a relationship was found between binge-purge behaviors and later cognitions and hope that future research will test if this finding replicates and if replicated why this relationship might exist. It is possible that it may have to do with the time lag under study (four-weeks). It would be interesting to test if there are differences in these temporal relations across different time periods. Future research should test how different time periods (i.e., time-lags) impact cyclical relationships, as these findings might change with both a shorter (e.g., using ecological momentary assessment) or longer time frame.

Additionally, and importantly in line with hypothesis, we found that fear of weight gain and feelings of fatness predicted multiple other later behaviors, including binge eating, exercise, fasting, laxative usage, and overvaluation of shape and weight. These findings support prior research suggesting that fear of weight gain might maintain eating disorder symptoms (Calugi, El Ghoch, Conti, & Dalle Grave, 2018; Levinson et al., 2017), and shows that these relationships occur not just at one time point (as in prior network research), but also over time. Furthermore, this finding supports research showing trends in cognitive eating disorder symptoms predicting later eating disorder behaviors (Levinson et al., 2018) and is consistent with theories of how eating disorders develop in adolescence, such that cognitions often precede behaviors (Rohde, Stice, & Marti, 2015; Stice et al., 2017; Stice, Marti, & Durant, 2011). Regarding, in-strength, we found that fasting had the highest centrality and was predicted by earlier levels of excessive exercise, feelings of fatness, and fears of weight gain. This suggests that fasting is predicted by several other eating disorder symptoms and implies that the behavior of fasting is likely influenced by several other processes. For example, fasting was significantly correlated with earlier excessive exercise, which may be reflective of exercise increasing motivation to continue fasting. A close connection between fasting and exercise has been found in rodent models and is often cited as a deadly combination of behaviors (e.g., Lewis & Brett, 2010). Furthermore, some research suggests that individuals who fast and exercise endorse worse eating disorder symptoms than individuals who only engage in one of those behaviors (LePage, Crowther, Harrington, & Engler, 2008) and that exercise reduces hunger hormones (Bowyer, Carson, Davis, & Wang, 2019). Our results suggest more research is needed on the specific relationship between fasting and exercise and suggests that excessive exercise in at-risk adolescents is predicted by fasting behaviors. Other symptoms with high in-strength were feelings of fatness, fears of weight gain, and overvaluation of shape and weight, again showing that these cognitions, from a network perspective, both give and receive input from other symptoms.

We also computed each of these networks in each of the four conditions (prevention and control conditions). In general, regardless of condition the most central symptoms were feeling, fat and fearing weight gain. However, there were some minor differences in temporal networks. Overall, these results suggests that despite prevention interventions, central symptoms in between subjects and contemporaneous networks remain stable. This finding makes sense because we would not expect the structure of ED pathology to change (Borsboom & Cramer, 2013), rather the strength of the symptoms among connections might change over the course of intervention. However, as participants receive interventions there may be changes in how symptoms exert influence on each other across time, which may explain why there were minor deviations in our temporal networks. Future research is needed to explore this issue further. Importantly, it is necessary for research to begin to parse out the impact of treatment and prevention interventions on central symptoms.

Individual Networks (n = 1)

Finally, moving to our individual networks, our second and third hypotheses were supported. First, we found that these networks were highly heterogeneous, in line with our second hypothesis, showing that eating disorder symptoms vary within an adolescent at-risk sample. This finding is consistent with the one prior study that has implemented individual networks of eating disorders (Levinson, Vanzhula, & Brosof, 2018). Our third hypothesis, that cognitive symptoms would be more likely than behavioral symptoms to provide input was also supported. We found that fear of weight gain, overvaluation of weight and shape, and feeling fat had the highest out-strength centrality, which is consistent with prior findings (Levinson, Vanzhula, & Brosof, 2018; Levinson et al., 2018). The three core cognitions identified in group-level networks are also present in the majority of individual networks: ‘feelings of fatness,’ ‘fears of weight gain,’ and ‘judgment of oneself. This finding points to some level of similarity between group-level and individual results.

Interestingly, we also found evidence for excessive exercise as a highly central out-strength symptom, perhaps suggesting that excessive exercise might have a role in the development of eating disorders and converging with evidence that excessive exercise predicted future onset of purging disorder, but not AN, BN, or BED (Krenz & Warschburger, 2014; Stice et al., 2017; Thompson, Petrie, & Anderson, 2017). Partially against hypothesis, we also found that overvaluation of weight and shape, feeling fat, and fear of weight gain were the most likely symptoms to receive (as well as provide) input from other symptoms. Overall, these findings from both the group level and individual networks seem to suggest that feelings of fatness and fear of weight gain play a large role both in terms of input and output of eating disorder symptoms. We hypothesize that these symptoms may be a pre-cursor (or prodromal stage) to the development of eating disorders in at-risk adolescents. If this hypothesis is supported, future research might assess for these specific symptoms and aim prevention efforts at individuals with high endorsement of these fears and feelings.

Given that this sample was an at-risk adolescent sample, we did not include body mass index. Furthermore, we did not include all relevant symptoms of AN because we could only use symptoms with continuous scales. Therefore, future research should consider including additional symptoms inclusive of all eating disorders. Therefore, many of our conclusions may only apply to eating disorders that include such symptoms. In other words, future research is needed to test how individual networks might apply in at-risk adolescence for the development of AN restricting type. We hope future research will continue to explore this question. However, overall, and perhaps most importantly, these results show that eating disorder symptom dynamics are very heterogeneous. As can be seen from our individual networks, some individuals’ networks did not even include all assessed symptoms, because of their lack of endorsement, and central symptoms varied greatly. This finding has relevance for personalized assessment and intervention. For example, not all symptoms may need to be assessed in all individuals. Future research should determine best practices for assessment of which symptoms in which individuals. Our results illustrate the symptom presentation and the relation between symptoms are very idiosyncratic, implying that it might be critical to take a precision medicine approach to eating disorder prevention that is based on different patterns of symptom endorsement across adolescent girls and young women. Future research should test if tailored prevention programs are more effective than untailored prevention programs (such as the Body Project). Finally, because our individual networks were conducted in the context of a prevention study, it is impossible to know if they would have differed if in a non-intervention context. Future research is needed to model individual networks outside of a treatment or prevention context.

Limitations.

There are several limitations to this work. First, we selected a random sample of individuals from a much larger dataset and only created 50 individual networks. Although this sample size may seem small, it is important to remember that these 50 individuals essentially represent their own datasets. There has been a call for more research focused on the individual or precision level (Insel, 2014) and this research represents an example of a shift in thinking, from large datasets to complex, dynamical systems within individuals. However, future research would benefit from using a wider response scale to increase variability and thereby increase the ability of individual network application to more data (see Piccirillo & Rodebaugh, 2019 for recommendations). Second, the data was collected retrospectively and may contain recall bias error. Third, the data was collected over the period of three years while participants took part in eating disorder prevention program, and therefore assumption of stationarity may be violated, meaning that the changes on variables over one month period might be explained by the intervention, though in tests of stationarity we did not find evidence for a trend explainable from treatment. One additional limitation of this method is that it is impossible to determine if individual networks are heterogeneous due to individual differences or a combination of individual differences and prevention condition. Future research is needed to replicate these findings in a non-treatment seeking sample. Additionally, we cannot make causal claims about the data presented here as this study is not experimental. Furthermore, some of our models only included three nodes, pointing to the fact that endorsement of symptoms was highly heterogeneous. Future research is needed to carefully characterize which are the most commonly endorsed symptoms in at-risk youth and if these variations predict outcomes (perhaps using latent class analysis). This issue may also reflect that network analysis might be optimal in samples that have higher endorsement of symptomatology, such as a treatment sample. Future research is needed to test individual networks in an eating disorder treatment sample. Future research may also consider if certain individuals cluster into ‘groups’ of individual networks. Additionally, our time lag was across four weeks, which is a longer time lag than has traditionally been used in temporal networks. However, there is literature that suggests that ED pathology develops over longer time periods, such as multiple weeks to months to years (LeGrange et al., 2016; Stein et al., 2012;; Smink et al., 2018; Peebles et al., 2017; Ranzenhofer et. al., 2014; Williams-Kerver et al., 2020). Future research is needed to test the impact of time on the development of ED symptoms, as time has largely been ignored in our models of ED symptoms. Last, we did not test if these symptoms predicted onset of clinical eating disorders. However, future research should test this question.

Conclusions.

Overall, our findings replicated prior network analyses, such that we showed that fear of weight gain and feelings of fatness were the most important symptoms across individuals, and excessive exercise, fear of weight gain, and feelings of fatness were the most common symptoms within individuals. We also showed that there is high heterogenity in eating disorder symptoms across time and within persons in an adolescent at-risk sample. Furthermore, we found that prospectively, binge eating negatively predicted feelings of fatness and fears of weight gain and positively predicted vomiting and that fear of weight gain and feelings of fatness predicted binge eating, excessive exercise, fasting, laxative usage, and overvaluation of shape and weight. We hope this research is a starting point for additional research pinpointing exactly how eating disorder symptoms develop and persist, and in turn might progress into full clinical syndromes, which can inform future treatment development personalized to the individual.

Figure 4.

Example individual networks (all 50 networks can be found in supplemental materials)

Highlights.

• It is currently unknown how specific eating disorder symptom dynamics operate both across time and within individuals

• The current study constructed group (N = 1272) and individual (N = 50) longitudinal networks of eating disorder symptoms across 48-months

• We found that that overvaluation of weight and shape are central symptoms in at-risk adolescents

• We also found that networks were highly variable across individuals