Eating pathology and interoceptive sensibility using the Brief Multidimensional Assessment of Interoceptive Awareness-2: Can there be too much of a good thing?

Abstract

This study evaluated the factorial structure and invariance of the Multidimensional Assessment of Interoceptive Awareness-v2 (MAIA-2). We also investigated incremental validity of the MAIA-2 factors for predicting eating pathology beyond appetite-based interoception. US-based online respondents (N=1294; M_age=48.7±18.4; 63% cis women; 78% White) were sampled. We conducted hierarchical stepwise regressions, dominance analysis, and multiple-group confirmatory factor analyses across age, gender, and eating disorder symptoms. An 8-factor, 24-item Brief MAIA-2 (BMAIA-2) model showed optimal fit. Using strict criteria (ΔCFI>.002), configural, metric, and scalar invariance were supported. After controlling for appetite-based interoception, higher scores on body listening, noticing, and emotional awareness unexpectedly predicted worse eating pathology, while higher scores on not worrying, not distracting, and trusting predicted less eating pathology, as hypothesized. Dominance analysis showed no subscales contributed >2% unique variance to global disordered eating beyond appetite-based interoception. For loss of control eating, however, not worrying was the dominant BMAIA-2 predictor, explaining 5% unique variance beyond appetite-based interoception. Research supported the relevance of multiple interoceptive sensibility dimensions captured by the BMAIA-2 to understanding eating-based pathology. Future studies should consider assessing its incremental validity using behavioral tasks and autonomic biomarkers of interoception to better understand the complex interplay among interoceptive skills and eating behavior.

Interoception, or the ability to identify and adaptively integrate signals from the body, is a skill recognized as integral to embodied self-awareness and well-being (Bonaz et al., 2021; Herbert & Pollatos, 2012). While deficits in interoception are increasingly understood as relevant to various forms of psychopathology (Brewer et al., 2021; Murphy et al., 2017), interoception itself has proven difficult to define, and as a result, difficult to measure. Several researchers have laid out theoretical frameworks to help describe the complexity of classifying differences in interoceptive processes and improve methods for operationalizing distinct forms of interoception (e.g., Murphy et al., 2019; Suksasilp & Garfinkel, 2022). For example, according to a tripartite model developed by Garfinkel et al. (2015), the subjective self-evaluation of bodily cues is known as interoceptive sensibility and is measured through self-report questionnaires. The second component, interoceptive accuracy, which is measured through objective performance on behavioral tasks (e.g., heartbeat perception), is independent of interoceptive sensibility. The third component of the Garfinkel et al. (2015) model is referred to as interoceptive awareness and represents the agreement or coherence between the subjective—interoceptive sensibility—and the objective—interoceptive accuracy (i.e., does performance on behavioral tasks correspond to one’s self-evaluation of bodily signals?).

Objective measures of interoceptive accuracy, which vary across body systems, are critical to understanding the construct, but some research suggests interoceptive sensibility is most associated with well-being (Ferentzi et al., 2018, 2019). This stresses the importance of psychometrically strong self-report measures—a focus of this investigation. Despite the presence of well-described conceptual frameworks, ambiguities surrounding the essence of measuring interoceptive sensibility remain (e.g., Desmedt et al., 2022; Vig et al., 2022).

To provide an expert-driven consensus definition, a summit focused on clarifying the construct offered the following description: “Interoception refers to the process by which the nervous system senses, interprets, and integrates signals originating from within the body, providing a moment-by-moment mapping of the body’s internal landscape across conscious and unconscious levels” (p. 501, Khalsa et al., 2018). This definition recognizes the multifaceted nature of interoception and its potential deficits, which could represent distorted physiological sensitivity, maladaptive decision-making, emotion dysregulation, alexithymia, and impaired insight (Khalsa et al., 2018). The relevance of these overlapping yet distinguishable functional deficits is especially striking in eating pathology (e.g., Khalsa et al., 2022).

A meta-analysis of the extant literature showed that interoceptive deficits are present across all three major eating disorders (i.e., anorexia nervosa, bulimia nervosa, and binge eating disorder) as well as those recovered, with a large, pooled effect size of 1.62 (Jenkinson et al., 2018). Indeed, a failure of interoceptive cues, specifically surrounding nutritional needs, has long been understood as pathognomonic to restrictive eating disorders (Bruch, 1962), and evidence shows it is a marker of poorer outcomes in patients with binge eating (Lammers et al., 2015). Moreover, a systematic review of 104 studies including nonclinical samples and a wider array of interoceptive channels, modalities, and measurement methods showed a consistent and convincing link between deficits in interoception and eating pathology (Martin et al., 2019). Ultimately, this evidence led to proposing interoception as a potential endophenotype (i.e., heritable trait) of eating disorders (Jenkinson et al., 2018; Martin et al., 2019).

Eating disorders are complex and difficult to treat illnesses with significant mortality rates (van den Berg et al., 2019; van Hoeken & Hoek, 2020). In effort to intervene early and better understand the developmental psychopathology of eating disorders, researchers have studied interoceptive deficits as a possible indicator, mechanism, and/or maintenance factor of subclinical eating pathology, especially in college students (e.g., Datta et al., 2021; Jeune et al., 2024; Lattimore et al., 2017; Poovey et al., 2022). Importantly, methods of measuring interoceptive deficits in this growing body of literature are highly varied (Martin et al., 2019), and this may obscure the specific nature of how eating pathology and interoception are related.

In fact, the Jenkinson et al. (2018) meta-analysis comparing patients with eating disorders to healthy controls only included research using the Eating Disorder Inventory Interoceptive Awareness/Deficits subscale (EDI-IA, Garner et al., 1983; EDI-ID, Garner, 2004). While individuals with eating disorders consistently reported lower scores on the EDI-IA/ID compared to healthy controls (Jenkinson et al., 2018), the brevity and unidimensionality of the 9 or 10-item measure fails to fully capture the wide array of facets relevant to interoceptive sensibility. For example, only one item of the current EDI-ID taps into hunger cues (e.g., ‘I get confused as to whether or not I am hungry’). The other items target confusion or fear of feelings and emotions that are more characteristic of alexithymia (e.g., ‘I don’t know what’s going on inside of me’ and ‘I get frightened when my feelings are too strong’). By focusing on research using the EDI, we leave other components of interoceptive sensibility less well-understood. To further advance the field, we need more studies involving appetite-based interoceptive cues, as well as adaptive attentiveness to and recognition of somatic sensations and bodily signals, both of which have shown relevance to disordered eating (e.g., DeVille et al., 2021; Monteleone et al., 2021; Poovey et al., 2022). Indeed, as measuring interoceptive processes beyond treatment-seeking samples expands, a variety of research-based tools of interoceptive sensibility have been adopted. Unfortunately, multiple analytic approaches of the most used measures suggest uncertain construct overlap (e.g., Desmedt et al., 2022; Vig et al., 2022), calling for more attention to the underlying meaning and psychometrics of these scales.

The Multidimensional Assessment of Interoceptive Awareness (MAIA) (Mehling et al., 2012; 2018) is one such tool and is among the most widely adopted measures of interoceptive sensibility across varied clinical, nonclinical, and international samples. The MAIA’s updated version (MAIA-2, Mehling et al., 2018) includes 37 items and 8 dimensions with demonstrated construct and structural validity in both clinical (Eggart et al., 2021) and community samples (Fiskum et al., 2023; Scheffers et al., 2024). The MAIA-2 contains subscales for noticing bodily sensations, not distracting from or worrying about bodily sensations of pain or discomfort, awareness of the connection between the body and emotional states, and the ability to sustain and control attention to bodily sensations, as well as regulate psychological distress through attending to those sensations. It also taps into listening to the body for insight and experiencing the body as safe and trustworthy. The unique multidimensionality of the MAIA strengthens the utility of the tool and has led to its recent recognition as the most precise of seven instruments tested for best capturing the core elements of interoceptive sensibility (Todd et al., 2022).

In a review of randomized controlled trials of interoception-based interventions across psychiatric disorders, the authors lauded the MAIA for measuring important qualities of interoceptive attention (Khoury et al., 2018). This includes self-efficacy, adaptively regulating distress through attending to body sensations, and connecting emotional states to the body—all of which highlight the tool’s strengths in generality and diagnostic non-specificity. Multiple intervention trials indicated the MAIA subscales were indeed sensitive to change, and the review’s authors noted the key adaptive elements of improved interoceptive regulation measured by the MAIA (Khoury et al., 2018). More recently, a pilot trial of a heart rate variability biofeedback mobile app (an example of an interoceptive-based intervention) showed improvements in nearly all MAIA-2 facets and disordered eating behaviors in healthcare workers with elevated eating distress during the COVID-19 pandemic (Mensinger et al., 2024).

However, the MAIA is not without limitations. Global fit statistics of the tool’s proposed eight-factor model have been suboptimal in some studies with multiple unacceptably low internal consistency reliabilities, especially on the ‘not worrying’ and ‘not distracting’ factors (e.g., Borlimi et al., 2023; Da Costa Silva et al., 2022; Rogers et al., 2021). We have seen widespread use of the MAIA for understanding the relationship between general interoceptive sensibility and varying psychophysiological symptoms such as pain, sexual dysfunction, and non-suicidal self-injury (de Jong et al., 2016; Fazia et al., 2024; Forrest & Smith, 2021; Poovey et al., 2023). Yet, further testing of its factor model and clarifying how the various components relate to eating pathology, particularly in older nonclinical samples, is warranted.

Considering the connections between appetite and eating behavior (e.g., Brown et al., 2010; Bullock et al., 2023; Denny et al., 2013), Poovey and colleagues (2022) recently investigated the centrality of appetite-based interoception in various symptoms of disordered eating. Specifically, they aimed to determine if dysfunctional attentiveness to hunger and satiety cues explained variance in disordered eating behaviors among undergraduates after accounting for the most dominant facets of general interoceptive sensibility measured by the MAIA-2 and the EDI-IA. As hypothesized, their study demonstrated the centrality of appetite-based interoception for three of four symptom clusters tested—binge eating, purging, and cognitive restraint. We sought to further expand Poovey et al.’s (2022) findings in an older and economically diverse sample focusing instead on the unique associations of the MAIA-2 factors with eating pathology after controlling for appetite-based interoception.

The aims of the present study were to: 1) re-evaluate the measurement model of the MAIA-2, and 2) examine the incremental validity of the resulting MAIA-2 factors on eating pathology after accounting for the variance explained by appetite-based interoception (while covarying age and gender) in a community-based sample. More specifically, we hypothesized people (i.e., younger/women) with lower attentiveness to hunger and satiety would have greater eating pathology. After accounting for the contributions of gender, age, and appetite-based interoception, we expected lower levels of interoceptive sensibility on the MAIA-2 factors to be associated with more eating pathology. We did not have hypotheses about specific MAIA-2 factors. Although multiple studies have found associations between lower scores on specific MAIA factors and greater eating disorder symptoms (e.g., trusting, not worrying, not distracting, self-regulation, body listening), findings have varied, and most previous research involved eating disorder patients or undergraduates (e.g., Brown et al., 2017; Monteleone et al., 2021; Philipou et al., 2022; Poovey et al., 2022).

Method

Participants and Procedures

Data were collected from a convenience sample using Prime Panels by CloudResearch^® in November 2021. Participants provided informed consent prior to responding to the online survey, which included a series of demographic questions and validated tools requiring full completion. Procedures were approved by the first author’s Institutional Review Board (IRB) at the time of data collection, after which an exempt IRB determination was provided by the authors’ current institution. Of the 1915 US-residing respondents who entered the survey, 1880 passed Qualtrics fraud filters. We required participants to pass all 3 attention checks—filtering out an additional 375 participants—and spend a minimum duration of 1/3 the expected time (i.e., ≥5 minutes) on the survey—filtering out 211 more participants. This gave a final sample size of 1294 participants, with no missing data. Sample characteristics can be found in Table 1.

Table 1.

Sample Demographics

Characteristic	Mean (SD) (range)
Mean Age, years (SD) (range)	48.7 (18.4) (18–95)
Gender	n (%)
Cis women	813 (62.8)
Cis men	468 (36.2)
Gender non-conforming/other gender	13 (1.0)
Race/Ethnicity
Non-Hispanic White	1014 (78.4)
Black/African American	131 (10.1)
Latinx/Hispanic	68 (5.3)
Asian	35 (2.7)
Native American, Native Hawaiian, Pacific Islander	15 (1.2)
Multi-racial, other racial/ethnic identity not listed	31 (2.4)
Education
Less than high school degree	67 (5.2)
High school or equivalent (GED)	401 (31.0)
Some college or trade school	324 (25.0)
Associates or Technical degree	167 (12.9)
Bachelor’s degree	211 (16.3)
Master’s degree or higher	124 (9.6)
Annual Household Income, USD
Less than $20,000	312 (24.1)
$20,000–$39,999	417 (32.2)
$40,000–$59,999	235 (18.2)
$60,000–$79,999	134 (10.4)
$80,000–$99,999	65 (5.0)
>$100,000	131 (10.1)
Marital status
Married or cohabiting with partner	611 (47.2)
Widowed	76 (5.9)
Divorced or separated	234 (18.1)
Single/Never married	373 (28.8)
Employment status
Working full-time	366 (28.3)
Working part-time	168 (13.0)
Looking for work	125 (9.7)
Retired	338 (26.1)
Not working due to disability	130 (10.0)
Stay at home caretaker	106 (8.2)
Not working due to being student/other reason	61 (4.7)

Note. N=1294

Measures

Appetite-based interoception was measured using the Reliance on Internal hunger and satiety Cues (RIC) subscale of the Intuitive Eating Scale-2 (IES-2, Tylka & Kroon Van Diest, 2013). This subscale contains 6 items rated on a 5-point Likert scale ranging from 1 (strongly disagree) to 5 (strongly agree). Total scores were created by averaging the six items. Higher scores indicate better appetite-based interoception through greater reliance on internal hunger and satiety cues from the body to determine when, what, and how much to eat. Internal consistency reliability in the present sample was good with a McDonald’s omega (ω) of .88.

General interoceptive sensibility was measured with the Multidimensional Assessment of Interoceptive Awareness, version 2 (MAIA-2, Mehling et al., 2018). The MAIA-2 is a 37-item questionnaire including 8 subscales: noticing (4 items), not distracting (6 items), not worrying (5 items), attention regulation (7 items), emotional awareness (5 items), self-regulation (4 items), body listening (3 items), and trusting (3 items). Items are rated on a 6-point Likert scale ranging from 0 (never) to 5 (always). Total scores were created by averaging the items associated with each subscale so higher scores are considered more adaptive. Internal consistency reliability of the subscales from the chosen model ranged from .75 to .90 in the present dataset.

To evaluate global disordered eating attitudes and behaviors, we used the 7-item short version of the Eating Disorder Examination Questionnaire (EDE-Q7, Grilo et al., 2015). This widely adopted and validated scale assesses dietary restraint, overvaluation of shape and weight concerns, as well as body dissatisfaction over the past four weeks. Responses are on a 7-point Likert scale ranging from 0 (no days or not at all) to 6 (every day or markedly) with higher scores representing more eating disorder symptoms. Total scores were derived by averaging the items which yielded good internal consistency reliability (ω=.86) in the present sample.

We used the 7-item brief Loss of Control Eating Scale (LOCES, Latner et al., 2014) to capture binge eating symptoms. The LOCES assesses respondents’ perceptions of being compelled to eat, and/or unable to resist or stop eating. Importantly, loss of control eating behavior can occur in the presence or absence of eating objectively large amounts of food, making it a scale that taps into symptoms of both objective and subjective binge eating. Responses are on a 5-point Likert scale ranging from 1 (never) to 5 (always) with higher scores indicating greater perceived loss of control over eating. Composite scores were created by averaging the items, and internal consistency reliability was excellent in this sample (ω=.94).

Statistical Analyses

Data were cleaned and summarized using SPSSv28 (IBM Corp., Armonk, NY). We plotted histograms (raw and multivariate residuals) and examined box and whisker plots to ascertain that distributional assumptions were met. We also performed bivariate/zero-order Pearson correlations of the main study variables as preliminary analyses. To answer aim 1 (i.e., the re-evaluation of the MAIA-2 measurement model), we conducted confirmatory factor analyses (CFA) using the robust maximum likelihood estimator in Mplus v8.6 and compared the fit of 3 models: 1) the 37-item 8-factor MAIA-2 model (Mehling et al., 2018); 2) the 26-item 6-factor model with a higher-order general interoception factor excluding the not distracting and not worrying subscales (Ferentzi et al., 2021); and 3) the 24-item 8-factor Brief MAIA-2 (Rogowska et al., 2023). Optimal model fit was determined by requiring the Comparative Fit Index (CFI) and Tucker-Lewis Index (TLI) ≥0.95, the Root-Mean-Square Error of Approximation (RMSEA) ≤0.06, and the Standardized Root-Mean-Square Residual (SRMR) ≤0.08 (Hu & Bentler, 1999).

To answer aim 2, we used hierarchical stepwise linear regressions with EDE-Q7 and LOCES as dependent variables in two separate models including gender, age, and appetite-based interoception in step 1. Using the forward stepwise procedure, on the next block we entered the eight MAIA-2 factors as predictor variables, one at a time, until all significant factors were in the model, with .05 as the p-value criterion for model entry. We chose the forward stepwise approach despite concerns over potential lack of replicability (Nathans et al., 2012), because this aim was exploratory and lacked specific hypotheses. Unlike simultaneous linear regression, the dynamic nature of forward stepwise allowed for evaluating how the effects of the strongest predictors—entered first (sequentially, after the inclusion of the predetermined covariates)—were impacted by the addition of each predictor in later steps. We provide simultaneous linear regressions as supplementary analyses to show the similarity in the end results. Lastly, given potential susceptibility to model misspecification, and to serve as a validation check of the forward stepwise approach, we performed dominance analysis with the “dominanceanalysis” package in R (Budescu, 1993; Navarette & Soares, 2020). Dominance analysis is recommended as a supplement to regression to see patterns of importance after fully accounting for multicollinearity of predictor variables (Tonidandel & LeBreton, 2011). We fit two sets of dominance models for each dependent variable—one model included appetite-based interoception and the other excluded it. Gender and age had little impact on patterns of importance and were excluded to facilitate visual interpretation. Plots with dominance weights (DWs) are presented for each predictor. DWs represent the average difference in fit between all subset models that both include and exclude the predictor and may be interpreted as an incremental R² where <.01 is considered negligible (Tonidandel & LeBreton, 2011).

Results

Means, standard deviations, internal consistency reliabilities (McDonald’s omega), and Pearson correlation coefficients between all study variables are presented in Table 2.

Table 2.

Means, standard deviations, internal consistency reliabilities (ω), and Pearson (zero-order) correlations between Brief MAIA-2 subscales, eating pathology (EDE-Q7 and LOCES), and appetite-based interoception (IES-2 RIC)

Variable	M (SD)	ω	1	2	3	4	5	6	7	8	9	10	11
1. Noticing	3.64 (1.04)	.75	--
2. Not Distracting	1.90 (1.15)	.85	−.23***	--
3. Not Worrying	2.23 (1.13)	.77	−.35***	.21***	--
4. Attention Regulation	3.11 (1.00)	.83	.44***	−.26***	−.18***	--
5. Emotional Awareness	3.71 (1.10)	.88	.56***	−.23***	−.29***	.48***	--
6. Self-Regulation	3.08 (1.12)	.86	.35***	−.19***	−.09**	.57***	.51***	--
7. Body Listening	2.84 (1.20)	.90	.44***	−.14***	−.27***	.57***	.51***	.67***	--
8. Trusting	3.47 (1.16)	.90	.24***	−.03	−.03	.45***	.34***	.56***	.48***	--
9. EDE-Q7	2.61 (1.67)	.86	.19***	−.17***	−.19***	.10***	.16***	.03	.13***	−.10a**	--
10. LOCES	2.24 (0.98)	.94	.13***	−.16***	−.26***	.00	.05	−.02	.09**	−.18***	.49***	--
11. IES-2 RIC	3.48 (0.84)	.88	.15***	.06*	−.04	.26***	.21***	.32***	.32***	.40***	−.19***	−.40***	--

Note. N=1294; MAIA-2=Multidimensional Assessment of Interoceptive Awareness-2; EDE-Q7=Eating Disorder Examination-Questionnaire 7-item short version; LOCES=Loss of Control Eating Scale 7-item brief version; IES-2 RIC=Intuitive Eating Scale-2 Reliance on Internal hunger and satiety Cues subscale;

^*p<.05,

^**p<.01,

^***p<.001;

^aexact r=.096, p=.001

Aim 1 - Confirmatory Factor Analyses

See Table 3 for the global fit statistics of the measurement models tested. Considering the superiority of the fit statistics for the 24-item 8-factor model, we retained the Brief MAIA-2 (BMAIA-2) model validated by Rogowska et al. (2023). See supplementary Tables S1 and S2 for the standardized factor loadings of the original MAIA-2 and the BMAIA-2 items, respectively.

Table 3.

Confirmatory Factor Analysis models and global fit statistics of the Multidimensional Assessment of Interoceptive Awareness-2

Model	χ2 (df)	χ2/df	CFI	TLI	RMSEA	90% CI RMSEA	p-valuea	SRMR
Mehling’s original 37-item 8 factorsb	2539.72 (601)	4.226	0.898	0.887	0.050	0.048–0.052	0.519	0.080
Ferentzi et al. 6 factors (higher-order)c	1401.86 (293)	4.784	0.916	0.906	0.054	0.051–0.057	0.009	0.065
Rogowska et al. MAIA-2 Brief 8 factorsd	585.19 (224)	2.612	0.969	0.961	0.035	0.032–0.039	1.000	0.040

Note. N=1294; All models used the Robust Maximum Likelihood Estimator in Mplus v8.6;

^ap-value is associated with probability of RMSEA being <.05, which is indicative of good model fit;

^bMehling et al. (2018) 37-item 8-factor MAIA-2 model;

^cFerentzi et al. (2021) 26-item 6-factor higher-order model containing a general interoception factor without ‘not worrying’ and ‘not distracting’ factors;

^dRogowska et al. (2023) Brief MAIA-2 24-item 8-factor model (chosen for remaining analyses in paper)

Aim 2 - Hierarchical Forward Stepwise Linear Regression Analyses

Results of the forward stepwise regressions for the BMAIA-2 factors on global disordered eating attitudes (EDE-Q7) and loss of control eating behavior (LOCES) are reported in Tables 4 and 5, respectively. After controlling for gender, age, and appetite-based interoception on step 1, as predicted, higher scores on trusting, not worrying, and not distracting predicted lower scores on the EDE-Q7. However, unexpectedly, higher scores on body listening, emotional awareness, and noticing predicted higher scores on the EDE-Q7. Similarly, for loss of control eating behavior, higher scores on trusting, not worrying, and not distracting predicted lower scores on the LOCES. However, higher scores on body listening predicted higher scores on the LOCES. Simultaneous linear regression analyses validated the findings of the forward stepwise models and are presented in supplementary Table S3.

Table 4.

Hierarchical forward stepwise regression analysis for the Brief MAIA-2 factors predicting global disordered eating attitudes (EDE-Q7)

Model step	Predictor variable	Std. β	t-statistic	p-valuec	Adjusted R2
Step 1	IES-2 RICa	−.179	−6.59	<.001	.055
	Gender (cis men=1)b	−.140	−5.10	<.001
	Age	−.015	−0.57	.570
Step 2	IES-2 RICa	−.248	−8.81	<.001	.095
	Gender (cis-men=1)b	−.139	−5.18	<.001
	Age	.011	0.40	.688
	Body Listening	.213	7.57	<.001
Step 3	IES-2-RICa	−.243	−8.71	<.001	.112
	Gender (cis men=1)b	−.126	−4.71	<.001
	Age	.035	1.31	.190
	Body Listening	.176	6.11	<.001
	Not Worrying	−.141	−5.04	<.001
Step 4	IES-2 RICa	−.247	−8.92	<.001	.121
	Gender (cis men=1)b	−.111	−4.14	<.001
	Age	.040	1.50	.134
	Body Listening	.135	4.41	<.001
	Not Worrying	−.114	−4.00	<.001
	Noticing	.115	3.77	<.001
Step 5	IES-2 RICa	−.215	−7.49	<.001	.130
	Gender (cis men=1)b	−.100	−3.72	<.001
	Age	.058	2.13	.034
	Body Listening	.185	5.59	<.001
	Not Worrying	−.105	−3.66	<.001
	Noticing	.127	4.16	<.001
	Trusting	−.125	−3.90	<.001
Step 6	IES-2 RICa	−.207	−7.17	<.001	.136
	Gender (cis men=1)b	−.097	−3.62	<.001
	Age	.057	2.10	.036
	Body Listening	.180	5.45	<.001
	Not Worrying	−.093	−3.25	.001
	Noticing	.112	3.66	<.001
	Trusting	−.125	−3.92	<.001
	Not Distracting	−.085	−3.15	.002
Step 7	IES-2 RICa	−.209	−7.25	<.001	.138
	Gender (cis men=1)b	−.094	−3.50	<.001
	Age	.054	2.01	.045
	Body Listening	.162	4.73	<.001
	Not Worrying	−.088	−3.06	.002
	Noticing	.088	2.65	.008
	Trusting	−.132	−4.13	<.001
	Not Distracting	−.079	−2.89	.004
	Emotional Awareness	.069	2.02	.044

Note. N=1294; MAIA-2=Multidimensional Assessment of Interoceptive Awareness-2; EDE-Q7=Eating Disorder Examination Questionnaire 7-item short;

^aIES-2 RIC=Intuitive Eating Scale-2 Reliance on Internal hunger and satiety Cues subscale;

^bcis women & gender non-conforming people=0;

^cp-value criterion for entry into the model was .05; ‘self-regulation’ and ‘attention regulation’ were excluded due to contributing negligible variance.

Table 5.

Hierarchical forward stepwise regression analysis for the Brief MAIA-2 factors predicting loss of control over eating (LOCES)

Model Step	Predictor variable	Std. β	t-statistic	p-valuec	Adjusted R2
Step 1	IES-2 RICa	−.392	−15.42	<.001	.175
	Gender (cis men=1)b	−.064	−2.52	.012
	Age	−.109	−4.29	<.001
Step 2	IES-2 RICa	−.404	−16.56	<.001	.239
	Gender (cis men=1)b	−.040	−1.60	.109
	Age	−.055	−2.19	.029
	Not Worrying	−.262	−10.48	<.001
Step 3	IES-2 RICa	−.456	−17.97	<.001	.262
	Gender (cis men=1)b	−.043	−1.76	.079
	Age	−.043	−1.72	.085
	Not Worrying	−.221	−8.66	<.001
	Body Listening	.168	6.38	<.001
Step 4	IES-2 RICa	−.428	−16.23	<.001	.269
	Gender (cis men=1)b	−.034	−1.42	.157
	Age	−.028	−1.12	.261
	Not Worrying	−.214	−8.45	<.001
	Body Listening	.214	7.38	<.001
	Trusting	−.108	−3.70	<.001
Step 5	IES-2 RICa	−.422	−15.96	<.001	.272
	Gender (cis men=1)b	−.031	−1.27	.203
	Age	−.028	−1.13	.261
	Not Worrying	−.204	−7.94	<.001
	Body Listening	.207	7.13	<.001
	Trusting	−.109	−3.75	<.001
	Not Distracting	−.061	−2.49	.013

Note. N=1294; MAIA-2=Multidimensional Assessment of Interoceptive Awareness-2; LOCES=Loss of Control Eating Scale 7-item brief version;

^aIES-2 RIC=Intuitive Eating Scale-2 Reliance on Internal hunger and satiety Cues subscale;

^bcis women & gender non-conforming people=0;

^cp-value criterion for entry into the model was .05; ‘self-regulation’, ‘attention regulation’, ‘noticing’ and ‘emotional awareness’ were excluded from the model due to contributing negligible variance.

Dominance Analyses.

Results show global disordered eating attitudes in Panel A (without appetite-based interoception) and B (with appetite-based interoception) and then, loss of control eating in Panels C and D, without and with appetite-based interoception in the models, respectively (see Figure 1). Trusting (DW=.023) was the most dominant predictor of global disordered eating without appetite-based interoception in the model. With appetite-based interoception (DW=.043) in the model, the unique contributions of the BMAIA-2 factors to global disordered eating were as follows: not worrying (DW=.016), noticing (DW=.016), trusting (DW=.015), not distracting (DW=.013), body listening (DW=.013), and emotional awareness (DW=.012). Like the stepwise regression model suggested, self-regulation and attention regulation contributed negligible variance to global disordered eating (DWs<.01).

Figure 1.

Dominance weights for the Brief MulBdimensional Assessment of InterocepBve Awareness-2 factors predicBng eaBng pathology

Note. N=1294; Values above the bars are dominance weights (DWs) and show the factor’s average R² for all possible subsets of models; AppeEte-based interocepEon is the Reliance on Internal hunger and saEety Cues of the IntuiEve EaEng Scale-2 (IES2-RIC); * (red font) indicates weights are contributing to the model in the opposite direction to expectations; weights shown in gray contribute negligible variance; gender and age were excluded because they contributed liMle variance and did not alter the paMern of results

For loss of control eating behavior, the dominant BMAIA-2-based predictor was not worrying (DW=.051). With appetite-based interoception (DW=.158) in the model, not worrying (DW=.050) was still the dominant BMAIA-2 predictor; however, trusting (DW=.025), not distracting (DW=.012), and body listening (DW=.018) also contributed significant unique variance. Replicating the stepwise regression, noticing, emotional awareness, self-regulation, and attention regulation contributed negligible variance to loss of control eating (DWs<.01).

Post Hoc Measurement Invariance

Measurement invariance (MI), or the assumption that a scale has the same interpretive meaning to different groups, is essential to comparing mean scores on that scale (Putnick & Bornstein, 2016). Given the unexpected positive associations between eating pathology and several BMAIA-2 factors, we fit multiple-group CFAs (MGCFA) to test for MI based on age (>55 years vs. ≤55 years), gender (cis men vs. cis women), and eating disorder symptoms (high vs. low). The cutoff of 55 years was used because it separated the older (i.e., baby boomers) from younger generations, and it marks the likely completion of bodily changes associated with a menopausal transition. We defined eating disorder symptoms as high when the EDE-Q7 total score exceeded 3.72, as validated by Machado et al. (2020).

To test the MI assumption, equality constraints across groups are imposed and the more constrained model is compared to the less constrained model in steps sequentially evaluating configural, metric, and scalar invariance (Putnick & Bornstein, 2016). Configural invariance is the equivalence of the model’s structural form across groups. Metric invariance is the equivalence of the item factor loadings, and scalar invariance is the equivalence of item intercepts across groups. Although researchers can use a significant chi-square difference test to determine non-invariance at each step, Cheung and Rensvold (2002) suggested the addition of change in CFI due to the chi-square test’s potential over-sensitivity to rejecting MI with large sample sizes. Considering the sample size and number of factors in our model, our criteria for non-invariance were a significant (p<.05) scaled difference chi-square test (Satorra & Bentler, 2001) and a change in CFI of >.002—which Meade et al. (2008) found to perform well at identifying non-invariance in models that were complex and had large sample sizes (see Table 6 for the MGCFAs). MI was supported at the configural, metric, and scalar level for age, gender, and eating disorder symptom groups. To provide an additional effect size estimate of the differences in BMAIA-2 latent variable scores by eating disorder groups, we calculated Cohen’s d based on the scalar invariance model. The group with high eating pathology had significantly higher noticing (0.366; p<.001), emotional awareness (0.249; p<.001), and body listening (0.185; p=.008) scores and significantly lower trusting (−0.193; p=.009), not worrying (−0.260; p=.001), and not distracting (−0.333; p<.001) scores. Latent means for attention regulation (0.094; p=.198) and self-regulation (.057; p=.389) did not differ between high and low eating pathology groups. All Cohen’s d estimates of latent mean differences showed consistency with our findings from the regression and dominance analyses, further validating the results.

Table 6.

Baseline models and measurement invariance tests of the Brief Multidimensional Assessment of Interoceptive Awareness-2

Sample (N)	χ2 (df)	χ2/df	CFI	TLI	RMSEA	90% CI RMSEA	p-valueb	SRMR
Total Sample (N=1294)a	585.190 (224)	2.612	0.969	0.961	0.035	0.032–0.039	1.000	0.040
Cis Men (N=468)	342.530 (224)	1.529	0.972	0.966	0.034	0.026–0.041	1.000	0.037
Cis Women (N=813)	554.377 (224)	2.474	0.957	0.947	0.043	0.038–0.047	.997	0.048
Over 55 years (N=514)	417.677 (224)	1.865	0.964	0.956	0.041	0.035–0.047	.993	0.043
Aged 18–55 (N=780)	440.755 (224)	1.968	0.966	0.959	0.035	0.030–0.040	1.000	0.043
Hi ED (N=339)	372.723 (224)	1.664	0.959	0.949	0.044	0.036–0.052	.884	0.048
Low ED (N=955)	462.440 (224)	2.064	0.971	0.964	0.033	0.029–0.038	1.000	0.040
Gender MI	χ2 (df)	Δχ2 (Δdf)	CFI	ΔCFI	RMSEA	ΔRMSEA	SRMR	ΔSRMR
Configural Model	899.873 (448)	--	0.962	--	0.040	--	0.045	--
Metric Model	929.440 (464)	29.364 (16)*	0.961	−0.001	0.040	0.000	0.046	0.001
Scalar Model	972.455 (480)	46.482 (16)***	0.959	−0.002	0.040	0.000	0.048	0.002
Age MI	χ2 (df)	Δχ2 (Δdf)	CFI	ΔCFI	RMSEA	ΔRMSEA	SRMR	ΔSRMR
Configural Model	858.460 (448)	--	0.966	--	0.038	--	0.043	--
Metric Model	881.713 (464)	22.472 (16)	0.965	−0.001	0.037	−0.001	0.044	0.001
Scalar Model	917.203 (480)	37.143 (16)**	0.963	−0.002	0.038	0.001	0.045	0.001
Eating Disorder MI	χ2 (df)	Δχ2 (Δdf)	CFI	ΔCFI	RMSEA	ΔRMSEA	SRMR	ΔSRMR
Configural Model	839.616(448)	--	0.967	--	0.037	--	0.042	--
Metric Model	865.898(464)	26.068 (16)	0.966	−0.001	0.037	0.000	0.044	0.002
Scalar Model	893.662(480)	27.081 (16)*	0.965	−0.001	0.036	−0.001	0.046	0.002

Note.

^aAll models used the Rogowska et al., 2023 Brief MAIA-2 and were fit in MPlus v8.6 with the Robust Maximum Likelihood Estimator and the Satorra-Bentler correction on the chi-square difference test;

^bp-value is associated with probability of RMSEA being <.05, which is indicative of good model fit; MI=Measurement Invariance; ED=Eating Disorder;

^*p<.05,

^**p<.01,

^***p<.001

Discussion

In an age and economically diverse US-based sample, this study supported the newly validated 24-item BMAIA-2 and the incremental validity of multiple BMAIA-2 factors for predicting lower eating-based pathology beyond the contribution of appetite-based interoception, which is recognized as the dominant interoceptive contributor to disordered eating behavior (e.g., DeVille et al., 2021; Poovey et al., 2022; van Dyck et al., 2021). This was particularly true for trusting one’s body and not worrying about uncomfortable/painful body sensations, extending the work of Poovey et al. (2022) where the dominant MAIA factors for disordered eating in college students were also trusting and not worrying. Considering the connections between eating and anxiety disorders (Levinson et al., 2018; Swinbourne et al., 2012), the importance of not worrying to eating-based pathology—especially perceived loss of control eating—is unsurprising. The not worrying subscale taps into the anxious temperament (e.g., ‘I start to worry that something is wrong if I feel any discomfort’, reversed scored), implicated by its strong negative association with anxiety symptoms (Robinson et al., 2024) and related factors such as pain catastrophizing (Eggart et al., 2021; Ferentzi et al., 2021).

It is interesting that although trusting was one of the most important BMAIA-2 predictors of eating pathology, the dominance analysis patterns showed the variance explained by trusting decreased considerably after accounting for appetite-based interoception in Panels B and D of Figure 1. This was likely due to shared variance between items from the IES-2 RIC scale about trusting the body’s appetite cues and the more generalized body trust captured by the BMAIA-2. Relatedly, our findings overlap with a network analysis of MAIA factors in a clinical sample of eating disorder patients (Brown et al., 2020). Specifically, feeling safe in one’s body, an item from the trusting factor, showed the strongest link overall to disordered eating symptoms. Additionally, ignoring physical sensations of discomfort from the not distracting factor was most strongly connected to symptoms of restraint and fear of losing control of eating (Brown et al., 2020). As shown in an earlier study examining the MAIA in clinical eating disorders (Brown et al., 2017), the tendency to distract oneself from pain was a predictor of greater eating pathology in all models we tested, providing a possible explanation to evidence that people with eating disorders have altered pain perception (Fazia et al., 2024; Papežová et al., 2005). Supporting this connection, lower not distracting scores were the only MAIA factor to predict greater pain tolerance in a study of self-injurious behavior (Rogers et al., 2021).

The present study also revealed several unexpected associations with eating-based pathology. For instance, the tendency to listen to the body for insight was associated with greater global disordered eating attitudes and loss of control eating behaviors. This contradicted prior research in women with anorexia that showed body listening was negatively associated with drive for thinness and body dissatisfaction, and was one of the three MAIA factors (along with trusting and self-regulation) explaining the connection between anxious attachment and eating disorder symptoms (Monteleone et al., 2021). Moreover, in another large clinical sample of transdiagnostic eating disorder patients, body listening scores were significantly lower (i.e., M=1.71±1.35; N=362; Brown et al., 2017) than mean scores reported in a large community-based sample (M=2.20±1.17; N=1090; Mehling et al., 2018). In a culture that rewards dietary restraint and thinness, perhaps listening to the body for insight when lacking a strong sense of agency, positive body connection, and the ability to resist appearance standards—all notable dimensions of the Experiences of Embodiment Scale related to lower eating pathology (Piran et al., 2020)—shapes disordered attitudes and behavior. Of course, the findings on body listening are preliminary and require replication in futures studies.

Similarly, the positive associations between global disordered eating attitudes and scores on the emotional awareness subscale were somewhat surprising, albeit effects were small. When comparing a clinical sample’s mean of 2.88±1.22 (Brown et al., 2017) to Mehling et al.’s (2018) mean of 3.44±0.96 derived from the general population, there are significant differences. However, other studies using nonclinical and clinical samples found no relationship between emotional awareness and eating disorder symptoms (e.g., Monteleone et al., 2021; Poovey et al., 2022). Also, it is important to note that low scores on emotional awareness from the BMAIA-2 are more related to one’s inability to connect bodily states to emotions (e.g., ‘I notice how my body changes when I feel happy/joyful’) than lack of emotional awareness per se—as we might associate with alexithymia. In fact, evidence suggests emotions may be experienced more intensely in people with disordered eating (Bicaker et al., 2022; Svaldi et al., 2012), potentially explaining the positive association we found between eating pathology and connecting emotions with bodily sensations. For some people, recognizing intense emotions in the body could be intolerable, requiring coping strategies like disordered eating. Such negative appraisals could also relate to why better emotional awareness on the MAIA predicted greater pain-related premenstrual symptoms in women (Borlimi et al., 2023).

Noticing bodily sensations also emerged as a positive predictor of global eating disorder symptoms, and though initially unexpected, this is consistent with evidence of elevated noticing scores in people with current or former diagnoses of anorexia nervosa, as well as siblings of patients, compared to healthy controls (Phillipou et al., 2022). Another consideration might be that the noticing factor (e.g., ‘I notice when I am comfortable/uncomfortable in my body’) can accompany a dysfunctional appraisal that tends to catastrophize discomfort and signals the hypervigilance of somatic feelings that is repeatedly found in patients with anxiety and panic disorders, as well as autism (Domschke et al., 2010; Garfinkel et al., 2016). Indeed, Murphy and colleagues (2017) directly addressed the issue of atypical (high) interoception and proposed the potential for anxiety symptoms to arise from prediction errors about body signals that represent inconsistencies between a person’s expected bodily state and their actual physical state. Corroborating this theory, there is recent evidence showing the MAIA’s noticing factor was positively associated with anxiety, depression, insomnia, and pain-related premenstrual symptoms in community samples (Borlimi et al., 2023; Vabba et al., 2023).

Understanding the potential maladaptive aspects of the MAIA-2’s noticing factor may also relate to a condition known as central sensitization. Central sensitization is characterized by an amplification of neural signals resulting in intensified pain due to heightened sensitivity to external and internal stimuli (Woolf, 2011). Paralleling several unexpected associations between BMAIA-2 factors and eating pathology in the present study, Colgan et al. (2022) also found evidence that noticing one’s body signals, as measured by the MAIA-2, could be problematic. More specifically, their study showed both the noticing and emotional awareness factors from the MAIA-2 predicted worse symptoms of central sensitization while not distracting, not worrying, and trusting were associated with fewer symptoms in a sample of chronic pain patients (Colgan et al., 2022). In fact, given the connection between sensitivity to gastrointestinal (GI) discomfort and eating pathology, especially in pediatric populations (e.g., Beckmann et al., 2023), new frameworks for a potential shared pathogenesis between chronic pain and eating disorders via central sensitization are in development (Sim et al., 2021). Similar processes were captured in research using the Visceral Sensitivity Index (VSI)—a self-report measure of one’s sensitivity to GI discomfort—where higher VSI scores were linked to multiple symptoms of disordered eating (Brown et al., 2021; Poovey & Rancourt, 2024). Moreover, Ahlich et al. (2023) found greater sensitivity to GI cues was associated with greater body preoccupation and restriction in a recent study of eating concerns and gastric interoceptive sensitivity using van Dyck et al.’s (2016) Water Load Test-II (WLT-II).

Taken together, these studies suggest more research is needed on the potential connections and interactions between an awareness of body sensations as measured by the MAIA-2 noticing factor and sensitivity to GI cues on determining eating behavior. Perhaps there are nonlinear associations underlying the connections between eating pathology and noticing bodily signals. While current evidence is limited, the unpredicted positive associations between several components of interoceptive sensibility measured by the BMAIA-2 and disordered eating in the present study emphasize the nuances needed in conceptual models of interoception. For example, without trust and confidence in one’s body signals, mere insight and awareness of the body may not always translate to adaptive integration of felt cues (Monteleone et al., 2021). Future research should also extend the incremental validity of the BMAIA-2 factors beyond self-report appetite-based interoception to other models of gastric sensitivity, including performance on behavioral tasks assessing interoceptive accuracy.

Regarding the psychometric findings, the CFA for the original 37-item MAIA-2 showed a below acceptable fit with problems specifically in the not worrying subscale—items 13 and 14, both of which are worded in the opposing direction from the other 3 items on that factor. Importantly, these items had low and/or negative factor loadings, poor local fit statistics, and the not worrying subscale had unacceptable internal consistency reliability, such that an omega coefficient could not be calculated. Other studies have identified similar problems with the not worrying factor (e.g., Brown et al., 2017; Fissler et al., 2016; Mehling et al., 2018; Valenzuela-Moguillanzky & Reyes-Reyes, 2015). These issues have led some researchers to use a higher-order model proposed by Ferentzi et al. (2021), which excludes not worrying and not distracting and combines the other six latent constructs to form a general interoceptive sensibility factor (e.g., Ahlich et al., 2023; Da Costa Silva et al., 2022; Poovey et al., 2023). In our study, the fit of the higher-order six-factor model was an improvement over Mehling’s eight-factor correlational model, but some indices were still below optimal benchmarks. Additionally, opting for the general interoceptive sensibility factor model would not have allowed for our primary aim to explore the unique contributions of all the MAIA-2 factors on eating pathology.

The BMAIA-2, which consisted of three items on each of the eight factors, reached the optimal benchmarks on all global fit indices and showed at least adequate internal consistency (ω ≥.75) on all factors. This provides encouraging evidence for the utility of this new measurement model—recently validated in Poland (Rogowska et al., 2023)—in a US-based English-speaking sample. We also fully supported strong/scalar invariance across gender, age, and eating disorder symptom levels for the BMAIA-2 using the stringent criteria of change in CFI >.002 (to suggest evidence of non-invariance) recommended by Meade et al. (2008). Of note, non-invariance was more pronounced across all groups tested on the original MAIA-2 model we fit (data not shown). Although Todd et al., (2020) supported strong/scalar MI between men and women on the full version of the scale (also using Meade et al.’s 2008 criteria), further research should test for potential measurement biases present in the unabbreviated MAIA-2.

This study had multiple strengths and limitations. The cross-sectional design of our study limits the ability to show the temporal ordering of interoceptive sensibility and eating-based pathology. Conducting large-scale longitudinal studies that include tools like the BMAIA-2 will help elucidate if lower (or higher) scores on the factors manifest prior to eating disorder symptoms. With growing interest in interoceptive deficits as a transdiagnostic feature of psychopathology (Solano Durán et al., 2024), we recommend continued efforts to refine the methods for operationalizing the facets of interoception. The exclusive use of self-report measures also presents possible concerns about limitations surrounding common method variance, such as inflated associations between constructs of interest. Nevertheless, amid many studies of eating and interoception in college student or clinical samples (e.g., Datta et al., 2021; Jeune et al., 2024; Phillipou et al., 2022), our sample of older US-based participants—26% of whom were retired and over half from households earning <40K USD annually—offered a unique opportunity to gain an understanding of interoceptive sensibility and eating pathology in a different setting. Generalizability, however, may still be limited. Interestingly, despite being an older, nonclinical sample recruited from an online platform, there were surprisingly high levels of eating-based pathology, with 26% classified as engaging in clinically significant eating disorder symptoms using a validated cut-off score for the EDE-Q7 (Machado et al., 2020). This rate provides more evidence of the large increases in eating-based pathology seen across the globe during the COVID-19 pandemic, reminding us that many adults struggle with significant disordered eating (Devoe et al., 2023). Considering the age of our participants, the numbers show eating disturbances extend over the lifespan and healthcare providers should remain alert for these problems in patients of all ages (Mangweth-Matzek & Hoek, 2017).

In summary, our study underscores the importance of strengthening measurement models for the MAIA-2 and continuing investigations of how the dimensions of the scale interact to create the most adaptive combination of interceptive skills. Several unexpected relationships cause us to consider when listening to the body for insight, or, paying too much attention to bodily sensations and their connected emotions, might become maladaptive. Though preliminary, our results highlight potential gaps in theory and call for more research on interoceptive cues and eating pathology. As Murphy (2022) noted, perhaps assessing interoception should also involve measuring the propensity to adaptively use the interoceptive signals noticed. Adding a biometric like heart rate variability, which is correlated with emotion regulation and interoceptive accuracy, could move us in this direction (Pinna & Edwards, 2020; Vabba et al., 2023). These results point to future studies of mind-body interventions aimed at improving interoceptive well-being using tools like the BMAIA-2. Such interventions should be sure to contain elements that reinforce a personal sense of embodied safety, agency, and trust in the body to avoid exacerbating potential vulnerability to hyperfocus on bodily sensations.

Data availability statement:

The dataset and corresponding output files for the presented analyses are available at https://osf.io/AB5FS/