Comorbidity Between Internalizing Symptoms and Disordered Eating Is Primarily Driven by Genetic Influences on Emotion Regulation in Adult Female Twins

Internalizing

Internalizing symptoms were measured using total scores from the second edition of the Beck Depression Inventory (BDI-II; which captures common depression symptoms such as sadness, difficulty concentrating, low self-esteem, and sleep and appetite problems; Beck et al., 1996), Obsessive Compulsive Inventory–Revised (OCI-R; which assesses obsessive-compulsive symptoms such as intrusive thoughts and checking behavior; Foa et al., 2002), State-Trait Anxiety Inventory trait subscale (STAI-T; which measures trait-level experiences of anxious thoughts and feelings; Spielberger et al., 1983), and Penn State Worry Questionnaire (PSWQ; which assesses an individual’s tendency to experience excessive worry; Meyer et al., 1990). The BDI-II, OCI-R, and STAI-T were administered at the intermediate assessment, and the PSWQ was administered on daily questionnaires and averaged across days for analyses. These four scales represent core domains of internalizing, are commonly used in the literature (Abramowitz & Deacon, 2006; Bieling et al., 1998; Dozois et al., 1998; Startup & Erickson, 2006), and showed excellent internal consistency in our sample (αs = .90–.97). Intercorrelations between these scales were moderate to high (rs = .37–.59), consistent with transdiagnostic models suggesting an underlying predisposition to internalizing that can give rise to diverse symptoms of depression and anxiety (Eaton et al., 2013).

Disordered eating

Because the parent study focused on disordered eating, we had access to several different measures of ED symptoms, including the Minnesota Eating Behavior Survey (MEBS ¹ ; von Ranson et al., 2005) and Eating Disorder Examination Questionnaire (EDE-Q; Fairburn & Beglin, 1994) administered at intake, as well as reports of dieting and objective binge eating episodes (OBEs; eating a large amount of food in a short period of time with subjective loss of control over eating) on daily questionnaires. Although the study also collects daily reports of purging to control weight/shape, we did not include those reports in current analyses because rates of item endorsement were extremely low (i.e., no purging reported on 99.4% of surveys).

The MEBS has four subscales that assess weight preoccupation (e.g., fear of weight gain), body dissatisfaction (e.g., feeling that body parts are too big), binge eating (e.g., experiencing loss of control over eating), and compensatory behavior (e.g., vomiting to control weight/shape). The MEBS subscales have adequate to good internal consistency (αs = .68–.81 for all subscales in the current sample except the compensatory behavior scale, which has lower internal consistency [α = .51] because of relatively infrequent item endorsement in a nonclinical sample). Girls and women with both AN and BN score higher on the MEBS than their same-age peers without an ED (von Ranson et al., 2005). This suggests the MEBS captures disordered-eating behaviors relevant to individuals with EDs characterized by both restriction (e.g., AN) and binge eating (e.g., BN).

The EDE-Q likewise has four subscales measuring shape concerns (e.g., preoccupation and dissatisfaction with body shape), weight concerns (e.g., preoccupation and dissatisfaction with body weight), eating concerns (e.g., preoccupation with food, fear of losing control over eating), and restraint (e.g., trying to limit food intake) over the past 28 days. The EDE-Q has good concurrent validity with interview-based measures of disordered eating in community samples (e.g., the Eating Disorder Examination interview; Mond et al., 2004), and internal consistency for all four EDE-Q subscales was good in the current sample (αs = .82–.93). Scores on the EDE-Q have good discriminant validity and are significantly higher in women with AN, BN, and DSM-IV ED not otherwise specified (including BED; American Psychiatric Association, 1994) relative to women without EDs (Rø et al., 2015). We also included two individual EDE-Q items assessing the number of subjective binge-eating episodes (SBEs; i.e., feeling a loss of control when eating an amount of food that is not objectively large) across the prior 28 days and number of days the participant “exercised hard” to influence weight/shape over the prior 28 days. SBEs and excessive exercise are commonly present in people with EDs and predict the presence of clinically significant levels of disordered eating in population-based samples (Brownstone & Bardone-Cone, 2021; Mond et al., 2004, 2006). These items are also not captured by the EDE-Q subscale scores because they are not included in subscale score calculations and were not assessed on daily questionnaires. As with daily reports of purging, we did not include individual EDE-Q items assessing purging or use of laxatives/diuretics to control weight/shape over the past 28 days because levels of endorsement were very low (i.e., < 2% of participants reporting any engagement in these behaviors, with typically low frequency [i.e., one time]).

Finally, participants reported whether they had dieted (yes/no) and the number of OBEs (0 to ≥ 9) they had experienced each day on daily questionnaires, which were averaged across days for analyses. Participants were given detailed definitions of OBEs before daily questionnaires and quizzed on their understanding at the intake and intermediate assessments, which has been shown to improve the validity of self-reported OBEs (Celio et al., 2004).

ER difficulties

ER difficulties were assessed using the six subscales of the Difficulties in Emotion Regulation Scale (DERS; Gratz & Roemer, 2004) administered during the intermediate assessment. Because the DERS was added after the study was already underway, data were available for only 498 participants (72.4%). However, participants could be included in the trivariate twin model even if they were missing data on the DERS because full information maximum likelihood estimation (FIML) makes use of all available data to determine model parameters, including data from cases with missing data for some predictors (Enders & Bandalos, 2001).

DERS subscales assess nonacceptance of emotions (e.g., feeling ashamed of experiencing negative emotions), difficulties engaging in goal-directed behavior (e.g., having difficulty getting work done when upset), impulse control difficulties (e.g., becoming “out of control” when experiencing negative emotions), limited emotional awareness (e.g., not paying attention to emotions), limited emotional clarity (e.g., not knowing what you feel), and limited access to ER strategies (e.g., feeling unable to regulate negative emotions). The DERS does not include items directly assessing internalizing or disordered-eating symptoms (e.g., no items regarding worry, anxiety, depressed mood, emotional eating, or binge eating) and thus has no explicit content overlap with any of the internalizing or disordered-eating measures included in this study.

As shown in Table 1, all DERS subscales had good internal consistency in this sample (αs = .82–.90), and means were very similar to those previously observed for women in population-based samples (Gratz & Roemer, 2004). Scores on the DERS are significantly correlated with behaviors/traits associated with emotion dysregulation, including nonsuicidal self-injury (Gratz & Roemer, 2004), anxiety and depressive symptoms (Neumann et al., 2010; Ruan et al., 2023), binge eating (Mikhail et al., 2022; Weinbach et al., 2018; Whiteside et al., 2007), and emotional eating (Gianini et al., 2013), and predict disordered eating above and beyond overall negative affect (Mikhail et al., 2022).

General analytic strategy

All analyses were conducted in Mplus (Version 8.6; Muthén & Muthén, 2021) using raw data and FIML estimation to account for missingness. Variables with skew more than 1 were log transformed.

Estimation of latent factors

To prevent estimation difficulties, latent factors for ER difficulties, internalizing symptoms, and disordered eating were estimated separately with cluster robust standard errors, and then factor scores were extracted for subsequent twin analyses. Latent factor variances were fixed to 1 and all observed variable loadings were freely estimated. Model-fit evaluation for the latent factors was based on combinational rules recommended in the literature (Hu & Bentler, 1999) for root mean square error of approximation (RMSEA), Tucker-Lewis index (TLI), and standardized root mean square residual (SRMR; the latter estimated with maximum likelihood excluding missing values). Model fit was deemed inadequate if both TLI < .96 and SRMR > .09 or both RMSEA > .06 and SRMR > .09. Model-fit assessment based on these combination rules has been shown to have a superior balance of Type I and Type II error rates relative to use of a single fit index (Hu & Bentler, 1999). Although chi-square is reported, it was not used as a primary indicator of absolute model fit for the latent factors because absolute chi-square values are often significant even in well-fitting models when the sample size is relatively large (Bentler & Bonett, 1980).

Internalizing symptoms and ER difficulties were initially each modeled as a single latent factor with no residual covariances between indicators. For internalizing, a single factor model provided adequate fit to the data with no added residual covariances, χ²(2) = 5.981, p = .050; RMSEA = .054, 95% confidence interval [CI] = [.000, .106], TLI = .966, SRMR = .023 (see Fig. 1). Although the initial model for ER difficulties met our fit criteria, it showed indications of less than ideal fit (i.e., both high RMSEA and low TLI), χ²(9) = 139.729, p < .001; RMSEA = .171, 95% CI = [.147, .197], TLI = .841, SRMR = .077. Modification indices suggested adding a residual covariance between the clarity and awareness subscales, which are conceptually more closely related than the other subscales. Addition of this residual covariance led to a substantial improvement in model fit, χ²(8) = 27.329, p = .001; RMSEA = .070, 95% CI = [.042, .099], TLI = .973, SRMR = .030, and factor scores from this modified model of ER difficulties (depicted in Fig. 1) were used for subsequent analyses.

OPEN IN VIEWER

Fig. 1.

Factor models and observed variable loadings for (a) emotion-regulation difficulties, (b) internalizing, and (c) eating disorder (ED) symptoms. Factor loadings are standardized. Strategies = Difficulties in Emotion Regulation Scale (DERS) limited access to emotion regulation strategies subscale; goals = DERS difficulties engaging in goal-directed behavior subscale; impulse = DERS impulse control difficulties subscale; nonaccept = DERS nonacceptance of emotions subscale; clarity = DERS limited emotional clarity subscale; aware = DERS limited emotional awareness subscale; BDI-II = second edition of the Beck Depression Inventory; OCI-R = Obsessive Compulsive Inventory–Revised; PSWQ = Penn State Worry Questionnaire; STAI-T = State-Trait Anxiety Inventory trait subscale; EDE-Q = Eating Disorder Examination Questionnaire; WC = weight concerns; SC = shape concerns; BD = body dissatisfaction; WP = weight preoccupation; R = restraint; daily diet = proportion of daily questionnaires on which participant reported dieting; BE = binge eating; SBE = subjective binge eating episodes; OBE = objective binge eating episodes; EC = eating concern; CB = compensatory behaviors.

We fit a bifactor model (F. F. Chen et al., 2006) to account for general and domain-specific variance across the wide range of ED scales assessed in our sample. All items loaded onto a general disordered-eating latent factor. We focused on this general factor in analyses because it best captures transdiagnostic aspects of disordered eating that contribute to all ED symptom presentations. Items were also specified as loading on domain-specific factors, including binge eating (MEBS binge eating, EDE-Q SBEs, daily OBEs, and EDE-Q eating concerns), dieting/restrictive (EDE-Q restraint, EDE-Q exercise, and daily dieting), and weight/shape concerns (EDE-Q weight concerns and shape concerns and MEBS body dissatisfaction and weight preoccupation), based on their conceptual underpinnings. As is standard for bifactor models (F. F. Chen et al., 2006), these domain-specific factors were specified to be uncorrelated with each other and with the general disordered-eating factor and thus can be conceptualized as representing variance that is not shared with other ED items (e.g., variance unique to binge eating). Fit for this model (see Fig. 1) was adequate, χ²(43) = 147.814, p < .001; RMSEA = .060, 95% CI = [.049, .070], TLI = .965, SRMR = .032.

Twin modeling

Twin models estimate additive genetic influences (A; genetic influences that sum across genes), shared environmental influences (C; environmental influences that increase similarity between co-twins, e.g., family socioeconomic background), and nonshared environmental influences (E; environmental influences that are not shared by co-twins, e.g., individual experiences of bullying) on individual traits and the extent to which these influences are shared between traits. Note that because each construct was modeled using a latent factor and is therefore error free, E does not contain measurement error in these models.

We fit univariate models for each construct and bivariate models for each pair of constructs before fitting the full trivariate model. In addition to streamlining model fitting for the trivariate model, which has numerous possible submodels, these initial analyses provided important information regarding the etiologic influences on each pair of constructs. This was particularly important because prior research on the etiologic overlap between internalizing and EDs has relied on single scales or categorical indicators that typically have not captured the full possible range of internalizing or ED symptoms (e.g., a single measure of anxiety symptoms or depressive symptoms or a measure capturing a specific facet of disordered eating such as binge eating; Fairweather-Schmidt & Wade, 2020; Munn-Chernoff et al., 2015; Slane et al., 2011). Thus, it was important to establish that the pattern of etiologic overlap between internalizing and disordered eating was similar to that observed in past research using single indicators when these constructs were instead modeled using latent factors before examining the impact of ER difficulties on their overlap. In addition, these initial models provided context regarding the etiology of ER difficulties and their relation to internalizing and disordered eating individually, which to our knowledge has not yet been analyzed in a twin-modeling framework.

The A paths for the full trivariate model are presented in Figure S1 in the Supplemental Material available online; C and E paths are identical to those depicted for A but are not shown for simplicity. ER difficulties were entered into the model first to allow for estimation of genetic/environmental influences shared between internalizing symptoms and disordered eating independent of ER difficulties. The model first estimates total additive genetic (a₁₁ in Fig. S1 in the Supplemental Material), shared environmental, and nonshared environmental influences on ER difficulties, as well as genetic/environmental influences on ER difficulties that overlap with those on internalizing symptoms (a₁₂ in Fig. S1 in the Supplemental Material) and disordered eating (a₁₃ in Fig. S1 in the Supplemental Material). The model then estimates genetic (a₂₂ in Fig. S1 in the Supplemental Material) and environmental influences on internalizing symptoms independent of ER difficulties and the extent to which these residual genetic/environmental influences contribute to disordered eating (a₂₃ in Fig. S1 in the Supplemental Material). If the overlap between internalizing symptoms and disordered eating is primarily driven by genetic/environmental influences on ER difficulties, we would expect the extent of the residual overlap between internalizing symptoms and disordered eating after accounting for ER difficulties to be small. Finally, the model estimates genetic/environmental influences on disordered eating (a₃₃ in Fig. S1 in the Supplemental Material) independent of both ER difficulties and internalizing symptoms. These estimates indicate the extent to which there are unique genetic/environmental factors that influence disordered eating alone. Although not directly estimated in the model, correlations between genetic and environmental influences on each construct (e.g., the correlation between genetic influences on internalizing symptoms and genetic influences on disordered eating) can be computed from other model parameters following estimation and are reported in Results.

Best-fitting models were identified as those that had a nonsignificant difference in minus twice the log-likelihood (−2lnL) between the full and nested model and minimized Akaike information criterion (AIC), Bayesian information criterion (BIC), and sample-size adjusted BIC (SABIC). If AIC, BIC, and SABIC identified different models as best fitting, the model that optimized two out of three fit indices was selected as best fitting. CIs for the best-fitting model parameters were determined using percentile bootstrapped CIs with 1,000 random samples.