Examining Changes in Binge-Eating Disorder Network Centrality and Structure in Patients Treated with Cognitive-Behavioral Therapy versus Interpersonal Psychotherapy

Abstract

Objective:

A sizeable minority of patients with binge-eating disorder (BED) do not fully respond to evidence-based treatments. Evidence to guide refinements of treatments is needed. Conceptualizing BED as arising from a network of symptom-to-symptom interactions allows for identification of the most strongly connected symptoms, which could inform intervention targets. This study estimated networks of BED features at pretreatment and posttreatment to assess whether cognitive behavioral therapy (CBT) and interpersonal psychotherapy (IPT) differentially impacted the interrelationships of BED symptoms/features.

Methods:

Participants were 392 adults (83% women, 88% white) with BED who received CBT (n=236) or IPT (n=156) and assessed at pretreatment and posttreatment. Networks were estimated across timepoints and treatments. Expected influence (EI) was calculated; symptoms with the highest EI have the most strong and frequent associations with other symptoms. We also assessed whether the symptom with the highest EI predicted posttreatment BED symptoms.

Results:

In the CBT and IPT networks, shape concern, weight concern, and eating concern had the highest EI at pretreatment and posttreatment. EI significantly increased from pretreatment to posttreatment for some symptoms in CBT but did not change for any symptoms in IPT. Shape concern significantly and positively predicted BED symptoms at posttreatment in CBT and IPT.

Conclusions:

CBT and IPT similarly impacted interrelations among BED features. Pretreatment EI predicted posttreatment BED symptom change, indicating that pretreatment centrality could signal meaningful intervention targets. Clinical implications and avenues for future research are discussed including how personalized network analysis may advance the understanding of the clinical utility of centrality.

Introduction

Binge-eating disorder (BED) is defined by recurrent episodes of binge eating, marked distress, and absence of inappropriate weight-compensatory behaviors (American Psychiatric Association, 2013). BED is associated with elevated rates of psychiatric/medical comorbidity and functional impairments (Udo & Grilo, 2018, 2019).

Several psychological and pharmacological treatments are effective for BED (Hilbert, Petroff, & Herpertz, 2019). Cognitive behavioral therapy (CBT) and interpersonal psychotherapy (IPT) have received strong and reliable support, with studies reporting significant reductions in binge-eating frequency and eating-disorder psychopathology, and improved psychological functioning (Grilo, Masheb, Wilson, Gueorguieva, & White, 2011; Wilson, Wilfley, Agras, & Bryson, 2010). Although leading evidence-based BED interventions (including CBT and IPT) produce meaningful improvements that are well-maintained through two-year follow-up (Wilson et al., 2010), roughly 50% of patients do not attain binge-eating abstinence (Linardon, 2018). While binge-eating abstinence may not be necessary to achieve good quality of life nor necessarily reflect full BED recovery (Slof-Op’t Landt et al., 2019), the fact that 50% of people continue to experience binge eating following treatment highlights the need for treatments to be improved.

It has been long suggested that identifying mediators and predictors/moderators of treatment outcomes could inform intervention refinement (Kraemer, Wilson, Fairburn, & Agras, 2002; Wilson, Grilo, & Vitousek, 2007). To date, however, research with BED has identified no mediators, no reliable predictors (other than rapid response and body-image disturbances) and mainly nonsignificant findings for moderators of treatment outcomes (Linardon, de la Piedad Garcia, & Brennan, 2017). The very few studies that reported significant moderator findings (Linardon et al., 2017) did so for limited outcomes (e.g., Anderson et al., 2020; Grilo, Masheb, & Crosby, 2012) and require broader replication prior to being viewed as reliable indicators of the characteristics for whom a specific treatment works best (Poldrack, Huckins, & Varoquaux, 2020). An alternative approach to understanding BED and its response to treatments might prove informative.

While we know what evidence-based treatments do (i.e., they reduce symptom intensities; Hilbert et al., 2019), we do not yet understand how treatments exert their influences on psychopathology. Viewing psychopathology through the lens of the network theory of psychopathology may advance our understanding of how treatments impact symptoms. The network theory of psychopathology views disorders as a network of interconnected symptoms that influence and maintain one another (Borsboom, 2017; Borsboom & Cramer, 2013). By quantifying symptom interactions, ideally with longitudinal data, it is possible to identify those symptoms that are most central to a disorder. The most central symptoms in a network are thought to be those that either cause most other symptoms or are caused by most other symptoms (Levinson et al., 2022). Theoretically, if an intervention targets a disorder’s most central symptom(s), this may reduce the symptom’s “input/influence” on other symptoms, causing those symptoms to lessen, which would thereby disrupt the symptom network (Borsboom & Cramer, 2013).

Assessing interconnections among psychiatric symptoms before and after treatment can address the question of how treatments impact symptoms and psychopathology as a whole, which may, in turn, contribute to greater understanding of treatment mechanisms and potential targets for refinement (Levinson et al., 2022). For example, if a particular symptom transitioned from being highly central before treatment to not highly central after treatment, this could mean that treatment (1) effectively targeted this central symptom and/or (2) disrupted the symptom-to-symptom interconnections in a potentially meaningful way. It is possible, however, that symptom centrality changes could also occur due to measurement error (Fried et al., 2016); this is discussed further in the Discussion.

To date, network analyses have been applied to BED treatment outcomes in two studies (Hilbert et al., 2020; Forrest & Grilo, 2022). Hilbert and colleagues (2020) explored change in a network of BED and general psychopathology symptoms in 178 patients with subthreshold and/or full BED before and after CBT and at 6-month follow-up. Eating disorder-related impairment and self-esteem were the most central features. Network connectivity overall increased significantly from pretreatment to posttreatment (i.e., the posttreatment network had more and/or stronger edges than the pretreatment network), and binge eating and shape concerns had the greatest increases in strength centrality, meaning these two symptoms had the strongest and most frequent interconnections with other symptoms. Forrest and Grilo (2022) examined network structure and changes in 191 patients with BED before and after behaviorally-based weight-loss treatments (BBWLT) and at 12-month follow-up. At pretreatment, overvaluation of shape/weight and preoccupation with shape/weight and with food/eating had highest strength centrality. At posttreatment and 12-month follow-up, dissatisfaction with shape/weight had highest strength centrality, which significantly increased from pretreatment. Importantly, the network structure following BBWLT resembled networks of people without eating disorders (e.g., Forrest et al., 2019). This shift in centrality—where overvaluation was the most central at pretreatment but dissatisfaction was the most central at posttreatment and follow-up—could indicate that even though CBT proposes that treatment exerts its impact via weakening connections between overvaluation and other symptoms (Fairburn et al., 2003), BBWLT may have amplified connections between dissatisfaction and other eating-disorder symptoms (Forrest & Grilo, 2022). Collectively, these two initial applications of network analyses to CBT (Hilbert et al., 2020) and BBWLT (Forrest & Grilo, 2022) for BED highlight the potential importance of understanding how treatments impact relationships, not just intensities, of symptoms.

One ongoing question within network science is to what extent central symptoms at pretreatment are suggestive of important treatment targets. Four studies have examined this issue, and all suggest that centrality could be an important indicator of treatment outcomes in anorexia nervosa (Elliott et al., 2020a, 2020b), social anxiety disorder (Rodebaugh et al., 2019), PTSD (Papini et al., 2020), and complicated grief (Robinaugh et al., 2016). The association between centrality and treatment outcomes merits investigation in BED both for informing treatment refinements and for guiding nosological issues, given the tension between evidence of the importance of body image as a severity signal in BED vs. the DSM-5 BED diagnostic criteria, which does not include a body-image criterion (Grilo, 2013). For example, if overvaluation is the most central BED symptom (DuBois et al., 2017; Wang et al., 2019), does this suggest that targeting overvaluation (i.e., reducing the symptom’s intensity and/or interconnectivity with other symptoms) is a necessary treatment goal? Moreover, an ongoing question concerning treatment for BED is how interventions that are seemingly so distinct, such as CBT, BBWLT, and IPT, result in relatively similar BED courses and outcomes, given their differential foci and attention devoted to binge eating and body image. Examining symptom centrality change following different BED treatments is needed to understand (1) how treatments impact connectivity of BED symptom networks and (2) how centrality at pretreatment may be associated with BED outcomes.

The current study used data from “Clinical Trials of BED” (CT-BED), which aggregated data from randomized controlled trials testing psychosocial treatments for BED to examine racial/ethnic differences in BED (Franko et al., 2012) and demographic (Shingleton et al., 2015; Thompson-Brenner et al., 2013) and clinical variables (Grilo et al., 2021) as predictors/moderators of outcomes. The current study analyzed data from patients who received CBT or IPT, given the treatments’ conceptual and procedural differences (i.e., cognitive versus interpersonal models of binge-eating; CBT focuses exclusively on eating-related behaviors and cognitions whereas IPT focuses on interpersonal factors while proscribing attention to eating), yet comparable outcomes (Wilfley et al., 2002; Wilson et al., 2010). Aim 1 compared pretreatment and posttreatment centrality of BED features between CBT and IPT to assess whether these two treatments differentially impacted the interrelationships among BED features. Aim 2 assessed whether the symptom with highest centrality at pretreatment predicted indicators of remission status at posttreatment (e.g., Brown et al., 2020). These models also included the symptom with the lowest centrality at pretreatment as a predictor in order to determine the extent to which more central symptoms are more predictive of outcomes compared to least central symptoms. This aim has implications not only for understanding BED treatments but also for network methodology, as there currently exists very limited evidence regarding whether centrality relates meaningfully to treatment outcomes (McNally, 2021; Robinaugh et al., 2020).

Methods

Participants

N=392 patients with BED from the CT-BED database (Franko et al., 2012) received either CBT (n=236; five sites) or IPT (n=156; two sites). Other treatments in CT-BED, including briefer guided-self-help CBT, were excluded. Table 1 summarizes demographic characteristics, which were self-reported. Methods have been previously described (Franko et al., 2012); only a brief overview follows. Institutional Review Boards for each site approved studies.

Table 1.

Sample descriptive statistics summarized by treatment type and timepoint, including results of mixed effects models assessing symptom intensity changes from pretreatment to posttreatment

	CBT N = 236	IPT N = 156
	n (%) or M (SD)	n (%) or M (SD)
Sex
Male	41 (17.4)	25 (16.0)
Female	195 (82.6)	131 (84.0)
Race and ethnicity
White	212 (89.8)	132 (84.6)
African American	14 (5.9)	16 (10.3)
Native American	2 (0.8)	1 (0.6)
Hispanic	6 (2.5)	7 (4.5)
Additional racial/ethnic identity*	2 (0.8)	0 (0)
Education
Some high school or less	5 (2.1)	1 (0.6)
High school graduate or GED	24 (10.2)	21 (13.5)
Some college	74 (31.4)	45 (28.8)
College graduate	86 (36.4)	28 (17.9)
Graduate degree	44 (18.6)	59 (37.8)
Not reported	3 (1.3)	2 (1.3)
Age	44.9 (9.82)	46.72 (10.55)
	CBT				IPT
	Pre	Post			Pre	Post
	M (SD)	M (SD)	t	p	M (SD)	M (SD)	t	p
BMI	38.09 (6.82)	38.07 (6.97)	−0.16	.88	36.73 (5.04)	36.55 (5.26)	−1.04	.30
Restraint	1.79 (1.32)	1.21 (0.94)	−5.80	< .001	1.92 (1.20)	1.51 (1.13)	−3.37	< .001
Shape concerns	3.80 (0.98)	2.68 (1.30)	−14.35	< .001	3.90 (0.95)	2.51 (1.27)	−13.81	< .001
Weight concerns	3.41 (1.09)	2.48 (1.21)	−11.43	< .001	3.31 (1.03)	2.23 (1.20)	−10.07	< .001
Eating concerns	2.30 (1.26)	0.95 (0.94)	−15.60	< .001	2.19 (1.39)	0.65 (0.79)	−13.96	< .001
Binge-eating frequency	21.04 (13.69)	2.98 (6.63)	−20.22	< .001	19.72 (11.54)	3.03 (13.27)	−18.64	< .001
Depression	1.04 (0.75)	0.56 (0.78)	−10.70	< .001	1.25 (0.70)	0.42 (0.73)	−13.53	< .001

Note. CBT = cognitive behavioral therapy, IPT = interpersonal psychotherapy, GED = general educational development, BMI = body mass index. Mixed models were estimated with random intercepts to account for non-independence of observations between timepoints. t and p values indicate the fixed effects of time on the given node.

^*Describing specific racial/ethnic identities is preferable to categorizing any racial/ethnic identity as “other” or “additional.” Regrettably, these data were originally merged such that people who reported a racial/ethnic identity other than White, African American, Native American, or Hispanic were coded only as “other” and we do not have the ability to determine these people’s specific racial/ethnic identities.

Measures

All BED features were assessed at pretreatment and posttreatment. Eating Disorder Examination (EDE; Fairburn & Cooper, 1993) is a semi-structured investigator-based interview that assesses binge-eating frequency and eating-disorder psychopathology, which is reflected in four subscales: restraint, eating concern, weight concern, and shape concern. Hamilton Depression Rating Scale (HAM-D; Hamilton, 1960; Reynolds & Kobak, 1995) assessed depression levels; HAM-D depression severity categories are as follows: 0=no depression, 1=mild depression, 2=mild-to-moderate depression, and 3=severe depression. Body mass index (BMI; kg/m²) was calculated from measured weights and heights.

Our node selection process proceeded as follows. First, only the EDE four subscales and binge-eating frequency were available, so these five variables—which are important CBT constructs—were included as nodes. Second, we included a HAM-D depression variable because depression/negative affect frequently co-occurs with BED and, if present, may signal a more disturbed variant (Grilo, Masheb, & Wilson, 2001). In addition, depression scores correlate moderately with measures of self-esteem (Hrabosky et al., 2007), and self-esteem is relevant in IPT; thus, including a depression node was an imperfect but relevant variable to ensure that networks included measures of constructs relevant in both treatments. Third, consistent with previous BED network studies (e.g., Hilbert et al., 2019; Wang et al., 2020) we included BMI as a node. Although BMI and the issue of weight loss is viewed by some within the eating-disorder field as controversial, the empirical reality is that BED is associated strongly with obesity (Udo & Grilo, 2018), most of the RCT datasets used in these aggregated data either required obesity as an inclusion or comprised participant groups that met criteria for obesity (Franko et al., 2012), and many patients with BED report both wanting to reduce binge eating and lose weight (Brody, Masheb, & Grilo, 2005). With that said about the logic of including BMI data in the network analyses, we highlight that network findings for BED to date have reported BMI to be among the least central BED symptoms (Wang et al., 2019), whereas body image concerns are among the most central symptoms. This supports the position, reviewed by Grilo (2013), that body image concerns in BED are important psychological and subjective constructs that do not simply reflect concerns commensurate with higher BMIs or excess weight.

Indicators of Remission Status

We included three dichotomous indicators of remission status. The first indicator was whether participants had zero binge-eating episodes during the past 28 days. The second indicator was whether participants’ EDE Global scores were within one standard deviation (SD) of community norms (M = 1.41, SD = 1.15; Mond, Hay, Rodgers, & Owen, 2006), which we refer to hereafter as normative EDE Global. The third indicator was whether participants attained both binge-eating abstinence and normative EDE Global.

Analytic Plan

Analyses were performed in R software environment (R Core Team, 2021), using the following packages: mice (van Buuren & Groothuis-Oudshoorn, 2011), lme4 (Bates, Maechler, Bolker, & Walker, 2015), lmerTest (Kuznetsova, Brockhoff, & Christensen, 2017), bootnet (Epskamp, Borsboom, & Fried, 2018), qgraph (Epskamp, Cramer, Waldorp, Schmittmann, & Borsboom, 2012), NetworkComparisonTest (van Borkulo et al., 2017), dplyr (Wickham, François, Henry, & Müller, 2020), and ggplot2 (Wickham, 2016).

Missing Data

The maximum proportion of missing data for EDE variables and BMI was 11%. The depression variable was missing for 41.6% at pretreatment (CBT = 34.7%, IPT = 51.39%) and 52.0% at posttreatment (CBT = 52.1%, IPT = 51.39%), because one of the five CBT sites and one of the two IPT sites did not include the HAM-D. Missing data were imputed (m=50) using predictive mean matching and aggregated for analyses.

Symptom Changes Throughout Treatment

To supplement interpretation of any network changes, we used linear mixed models with random intercepts to evaluate how network symptoms changed throughout treatment.

Network Estimation

Pretreatment and posttreatment BED symptom networks were estimated for CBT and IPT through the graphical least absolute shrinkage and selection operator (LASSO) method. LASSO method includes a tuning parameter to reduce small partial correlations to 0, reducing the likelihood of false positive edges (i.e., partial correlations) from emerging. LASSO’s tuning parameter was set to .50. Given the differing scales of nodes, networks were estimated via Spearman correlations, where edges represent partial Spearman correlation between two nodes (i.e., symptoms).

Centrality

Network centrality was indicated by calculating expected influence, which indicates the sum of the edges present for each node. Expected influence is similar to strength centrality but differs in that strength indicates the sum of the absolute values of the edges present for each node whereas expected influence is calculated by raw values. Given that positive vs. negative edges imply differential impact and that several edges were negative (i.e., red lines in Figure 1), we chose to use expected influence in the current study.

Figure 1.

Binge-eating disorder symptom networks by treatment type and timepoint. Note. CBT = cognitive behavioral therapy, IPT = interpersonal psychotherapy, bmi = body mass index, rest = restraint Eating Disorder Examination (EDE) subscale, shape_con = shape concerns EDE subscale, wt_con = weight concerns EDE subscale, eat_con = eating concern EDE subscale, OBE = binge-eating frequency, dep = depression.

Centrality is sometimes related to node variance (Elliott et al., 2020b; Terluin et al., 2016), which could mean that centrality may be an unnecessary metric. To examine this possibility, we assessed Spearman’s rho and Pearson’s r correlations between expected influence and standard deviation in both treatment groups cross-sectionally and longitudinally, with and without controlling for outliers (Supplemental Table 2). Of the 32 correlations, only two were significant, and one of these was in negative direction, which is the opposite direction of Terluin et al.’s (2016) findings.

Centrality Difference Tests

Centrality difference tests were conducted via nonparametric bootstrapping. These are a useful supplement to centrality analyses because centrality is relative. Centrality difference tests identified whether a given node’s centrality is significantly greater than centrality for other nodes.

Stability

Network stability was indicated by the correlation-stability (CS) coefficient. The CS coefficient is obtained by re-estimating expected influence using progressively smaller subsets of the original sample via bootstrapping. The correlation between the full sample’s expected influence and the bootstrapped subsets’ expected influence is calculated. The CS coefficient indicates the maximum proportion of the sample that can be dropped while retaining an expected influence correlation >0.70 with the original sample. CS coefficients ≥0.50 are considered good (Epskamp et al., 2018).

Network Comparison Tests

Network comparison tests were performed to assess pretreatment to posttreatment comparisons within treatments (dependent samples) and timepoint comparisons between treatments (e.g., pretreatment networks compared for CBT vs. IPT; independent samples). For each comparison, we assessed whether the two networks differed in global strength (i.e., overall network connectivity) and expected influence for each node. Given the number of expected influence comparisons, node centrality comparison p-values were adjusted (Benjamini and Hochberg method).¹

Pretreatment Expected Influence Predicting Indicators of Remission Status

We used logistic regressions to assess whether the symptoms with the highest and lowest expected influence in each treatment predicted indicators of remission status at posttreatment (binge-eating abstinence, normative EDE Global, and binge-eating abstinence plus normative EDE Global).

Results

Symptom Changes

Table 1 shows that for both CBT and IPT, all BED features except BMI were significantly lower at posttreatment than pretreatment.

Symptom Networks

CS coefficients indicated that all networks were estimated reliably (pretreatment CBT=0.75, posttreatment CBT=0.75, pretreatment IPT=0.67, posttreatment IPT=0.75).

Table 1 includes descriptive statistics for all nodes. Figure 1 displays the four networks. Table 2 and Figure 2 show expected influence. Supplemental Figure 1 shows the centrality difference test results. At CBT pretreatment, shape concern had the highest expected influence, which was significantly higher than expected influence for all other nodes, except for weight concern. At CBT posttreatment, weight concern and eating concern had the highest expected influence, which were significantly higher than expected influence for all other nodes at posttreatment.

Table 2.

Expected influence centrality by treatment type and timepoint, with comparisons by pretreatment vs. posttreatment within cognitive behavioral therapy (CBT) and interpersonal psychotherapy (IPT) and comparing pretreatment centrality between CBT and IPT and posttreatment centrality between CBT and IPT

	CBT		IPT		p for centrality differences between networks
	Pre	Post	Pre	Post	CBT pre vs. CBT post	IPT pre vs. IPT post	Pre CBT vs. pre IPT	Post CBT vs. post IPT
BMI	−0.14	0.20	−0.02	0.06	.03	1.00	.61	.25
Restraint	0.00	0.28	0.26	0.13	.11	.71	.47	.25
Shape concerns	1.20	0.97	0.99	1.10	.49	.77	.77	.82
Weight concerns	1.00	1.10	0.80	0.99	.89	.77	.47	.23
Eating concerns	0.67	1.10	0.67	0.95	.01	.77	.59	.04
Binge-eating frequency	0.08	0.49	0.12	0.39	.01	.47	.47	.61
Depression	0.51	0.69	0.79	0.70	.01	.71	.47	.82

Note. CBT = cognitive behavioral therapy, IPT = interpersonal psychotherapy, BMI = body mass index.

Figure 2.

Centrality by treatment type and timepoint Note. CBT = cognitive behavioral therapy, IPT = interpersonal psychotherapy, BMI = body mass index.

At IPT pretreatment, shape concern, weight concern, eating concern, and depression had the highest expected influence. These values were significantly higher than expected influence for the three other nodes but did not differ from one another. At IPT posttreatment, shape concern, weight concern, and eating concern had the highest expected influence. These values were significantly higher than expected influence for the three other nodes but did not differ from one another.

Network Comparison Tests

CBT by Timepoint

Global strength did not significantly differ between pretreatment and posttreatment in CBT (2.01 vs. 2.55, p=.08). After adjusting for multiple comparisons, centrality comparisons revealed that expected influence for BMI, eating concern, binge-eating frequency, and depression was significantly higher at posttreatment vs. pretreatment in CBT.

IPT by Timepoint

Global strength did not significantly differ between pretreatment and posttreatment in IPT (1.49 vs. 1.66, p=.53). After adjusting for multiple comparisons, expected influence did not differ for any nodes between pretreatment and posttreatment in IPT.

Pretreatment by CBT vs. IPT.

Global strength at pretreatment did not differ between CBT and IPT (2.01 vs. 1.49, p=.27). After adjusting for multiple comparisons, expected influence did not differ for any nodes at pretreatment between CBT and IPT.

Posttreatment by CBT vs. IPT

Global strength at posttreatment was significantly higher in CBT vs. IPT (2.55 vs. 1.66, p=.03). After adjusting for multiple comparisons, expected influence for eating concern was significantly higher at posttreatment in CBT vs. IPT.

Centrality Predicting Indicators of Remission Status

Given that shape concern was the node with the highest centrality and BMI was the node with the lowest centrality, shape concern and BMI were entered as predictors of the three indicators of remission status (see Table 3). In CBT, higher shape concern (most central) significantly predicted lower likelihood of attaining binge-eating abstinence, normative EDE Global, and both binge-eating abstinence plus normative EDE Global. BMI (least central) did not significantly predict likelihood of binge-eating abstinence or attaining both binge-eating abstinence plus normative EDE Global, but it did significantly predict normative EDE Global. Even though both the most central and least central symptom predicted normative EDE Global, the effect was stronger for shape concern vs. BMI.

Table 3.

Most and least central symptoms at pretreatment predicting indicators of remission status in both samples

	CBT			IPT
	OR (SE)	z	p	OR (SE)	z	p
Outcome = Binge-eating abstinence
Pre EDE shape concern (most central)	0.68 (0.14)	−2.65	.008	0.92 (0.18)	−0.48	.63
Pre BMI (least central)	0.99 (0.02)	−0.66	.51	1.00 (0.03)	0.01	.99
Outcome = EDE Global within 1 SD of norm
Pre EDE shape concern (most central)	0.45 (0.20)	−4.12	< .001	0.53 (0.25)	−2.53	.01
Pre BMI (least central)	0.95 (0.02)	−2.01	.04	0.95 (0.04)	−1.09	.28
Outcome = Binge-eating abstinence + EDE Global within 1 SD of norm
Pre EDE shape concern (most central)	0.62 (0.14)	−3.29	.001	0.88 (0.17)	−0.72	.47
Pre BMI (least central)	0.98 (0.02)	−1.12	.26	1.01 (0.03)	0.40	.69

Note. CBT = cognitive behavioral therapy, IPT = interpersonal psychotherapy, EDE = Eating Disorder Examination.

In IPT, higher shape concern (most central) significantly predicted lower likelihood of attaining normative EDE Global. Shape concern was not associated with likelihood of attaining binge-eating abstinence or both binge-eating abstinence plus normative EDE Global. BMI (least central) did not significantly predict likelihood of binge-eating abstinence, normative EDE Global, or both binge-eating abstinence plus normative EDE Global.

Discussion

This study estimated networks of BED features at pretreatment and posttreatment to assess whether CBT and IPT differentially impacted interrelationships of BED features. We also assessed whether pretreatment BED centrality corresponded to change in BED symptoms at posttreatment. Findings have relevance to both the eating-disorders field specifically and the network theory of psychopathology more broadly.

Shape concern, weight concern, and eating concern were the most central symptoms at both timepoints in both treatments (CBT and IPT). This is broadly in line with the CBT model of eating disorders (DuBois et al., 2017; Fairburn et al., 2003). These centrality findings also converge with those of other network studies of eating disorders that constructed networks using EDE subscales (Hilbert et al., 2020; Mares et al., 2022).

In terms of prognostic significance, the symptom with the highest expected influence at pretreatment (shape concern) significantly and positively predicted all three indicators of remission status in CBT and predicted one of the three indictors of remission status (normative EDE Global) in IPT. The least central symptom (BMI) was associated with only one indicator of remission status (normative EDE Global) in CBT only, though the effect was stronger for the most central symptom vs. the least central symptom. Overall, this pattern of findings might suggest that the most central symptom in pretreatment BED networks might inform targets for BED treatment, where interventions specifically designed to reduce the most central symptom (e.g., shape overvaluation [Wang et al., 2019]) may result in higher likelihood of experiencing various indicators of remission status. These findings add to those of the few emerging studies suggesting that pretreatment symptom centrality be associated with treatment outcomes (Brown et al., 2020; Elliott et al., 2020a, 2020b; Papini et al., 2021; Rodebaugh et al., 2019). These findings also extend those of Brown et al (2020), by assessing whether the most central symptom is more predictive of outcomes compared to the least central symptom. Two caveats, however, should be noted. One, in some studies, pretreatment centrality did not predict change in symptoms measured by assessments that were not included in the original network (e.g., Rodebaugh et al., 2019). Although our three indicators of remission status were not included in the network, they were all derived from nodes that were included in the network (binge-eating abstinence was derived from binge-eating frequency, EDE Global scores were derived from the four EDE subscales). Ideally, studies determining whether symptom centrality corresponds to treatment outcomes would operationalize treatment outcomes using measures that are completely distinct from measures used to estimate networks. Whether pretreatment centrality predicts change in BED-relevant outcomes not included in the networks here remains an empirical question for future research. Two, symptom centrality may change throughout treatment (Nemesure et al., preprint). Much remains to be learned about network dynamics during treatment and how matching treatments to network dynamics may influence treatment outcomes.

In both treatment samples, pretreatment and posttreatment networks did not differ in global strength. These findings were unexpected and contrast with those reported by Forrest & Grilo (2022) and Hilbert et al. (2020), which found that global strength significantly increased following treatment. Research comparing symptom networks before and after treatment is in its early stages, and we do not yet know whether significant global or node-wise centrality changes have clinical relevance. Indeed, some authors question whether centrality is a meaningful metric overall (Bringmann et al., 2019). Global centrality decreases could mean a significantly disrupted symptom network (Borsboom, 2017) while global centrality increases could indicate increased likelihood for a significant transition toward symptom change (Groen et al. 2019), both of which could have clear clinical relevance. However, global centrality increases could also indicate less clinically relevant information, such as effects of repeated measurements (Fried et al., 2016). Similarly, changes in node centrality following treatment could signal that the symptom network transitioned from an “active disease” state, demarcated by one symptom being highly central, to a “recovered disease” state, demarcated by a different symptom being highly central (c.f. Borsboom, 2017), or could have no association with treatment response or symptom severity. We—and these previous studies—are unable to address these questions, as our data only allow for estimating between-subjects networks. However, these questions could be answered by estimating idiographic networks, constructed from intensive longitudinal data collected throughout treatment (e.g., DuBois et al., in press). This would allow for comparison of treatment outcomes and prognosis between patients whose global centrality and pattern of centrality did vs. did not change throughout treatment. We contend that idiographic networks hold great promise in clarifying the current debate surrounding the clinical utility of centrality (McNally, 2021; Robinaugh et al., 2020).

Comparisons of symptom networks between CBT and IPT revealed no significant differences in global strength or centrality at pretreatment. At posttreatment, the CBT network was more strongly connected than the IPT network but only the eating concern node had significantly different expected influence between samples. One potential interpretation of these quite similar node centrality findings is that CBT and IPT impact relationships amongst BED symptoms in a similar fashion. In other words, the partial correlations for how weight concern relates to depression, shape concern, and restraint are roughly similar in CBT and IPT. This parallels findings that CBT and IPT reduce the intensities of the symptoms quite similarly in time course and outcome (Wilfley et al., 2002; Wilson et al., 2010). We view our findings as very preliminary because our data are not able to address mechanisms of change. We encourage future research to include individual items (as compared to subscales) in networks to allow for more fine-grained understanding of how CBT vs. IPT impact connectivity among specific types of symptoms emphasized in each treatment (e.g., overvaluation vs. dissatisfaction in CBT, negative social evaluation vs. negative self-evaluation in IPT).

Study strengths include having two treatment samples. Our measures of BED symptoms and depression are validated, which increases confidence in the measurement of symptoms. The CBT sample size of 236 participants is the largest sample to compare pretreatment and posttreatment BED symptom networks to date, which further increases confidence in those findings. Although the aggregated treatment data allowed us to perform this first analysis of how distinct treatments impact BED symptom connectivity, the data also have limitations. One, although we included a node assessing depression severity, which has relevance as an IPT construct, an IPT-specific measure was not available across studies. The lack of an IPT-specific measure may limit the interpretation of the IPT sample results. Similarly, the large amount of missingness on depression is a limitation of the study. Two, we did not have access to individual EDE items/constructs, which provide more fine-grained information on how treatment impacts specific BED symptom connectivity (e.g., Forrest & Grilo, 2022). Three, the sample was primarily educated, White women. Findings should not be assumed to be generalizable to people with other intersectional identities. Similarly, we recognize that describing specific racial/ethnic identities is preferable to categorizing any racial/ethnic identity as “other.” Regrettably, these data were originally merged such that people who reported a racial/ethnic identity other than White, Black, Native American, or Hispanic were coded only as “other” and we do not have the ability to determine these people’s specific racial/ethnic identities. Four, one correlation between centrality and variance was significant and positive, though 31 of the other correlations did not support a positive relation between centrality and node variance.

In summary, there were many similarities in global network connectivity and node centrality before and after CBT and IPT for BED. Pretreatment network centrality significantly predicted posttreatment BED symptom severity. Conceptualizing BED as a network of interconnected symptoms sets the stage for new research to understand how distinct treatments impact interrelations among symptoms and what symptoms may be the best targets within and across treatments.

Public Health Significance.

Cognitive behavioral therapy and interpersonal therapy for binge-eating disorder, which are two leading evidence-based treatments for binge-eating disorder that are quite different in their models and approaches, similarly impacted interrelations among binge-eating disorder symptoms. In addition, the most strongly interconnected symptom predicted change in binge-eating disorder symptom severity. Studying the interrelations among symptoms may provide new insight on how treatments impact symptom relationships and inform intervention targets.

Data Availability Statement:

The data that support the findings of this study are available from the corresponding author upon reasonable request.