Czech Version of the Multidimensional Assessment of Interoceptive Awareness (MAIA): Psychometric Evaluation and Network Model

Participants were recruited from seven clinical sites in the Czech Republic between January 2018 and December 2019. Patients were undergoing multi-component treatment, predominantly involving group therapy. The therapy’s duration ranged from 4 to 12 weeks, with a median duration of 6 weeks. Group therapy sessions were integrative and similar across all clinical sites, primarily focusing on psychodynamic orientation and lasting 90 min. The data were collected in a paper-and-pencil form before treatment and once a week during treatment (usually after each therapeutic session). For more information about the treatment and therapists, please refer to Pourová et al. (2024).

The initial number of patients participating in the study at the baseline was 444. However, 13 patients did not respond to any of the items, so they were excluded from the dataset. This resulted in a baseline sample size of 431 patients (74.5% female; 95% Czech nationality; aged between 18 and 71 years, M = 39.20; SD = 11.04). Patients’ psychiatric diagnoses according to the International Classification of Diseases 10 (World Health Organization, 2004) included primarily F4# (n = 305; where the # symbol represents any number of specific diagnoses type, e.g., F41.1, Generalized anxiety disorder), F3# (n = 81, e.g., affective disorders), F6# (n = 66, e.g., personality disorders), F5# (n = 9, e.g., behavioral syndromes with physiological problems), and F1# (n = 8, e.g., psychoactive substance abuse disorders). Comorbidity was reported in 40 patients, with combinations of F4# and F6# (n = 18), F3# and F4# (n = 12), and F3# and F6# (n = 9). For more detailed demographic characteristics of the sample, please see Table 1. Furthermore, data from patients who completed the questionnaire up to the 8th week during their therapy were included in the longitudinal analyses, as there was a rapid reduction in the sample size from the 9th week onward. These 8 weeks within therapy represent the 8 measurement waves in the present study.

MAIA is a 32-item measure developed to assess interoceptive awareness (Mehling et al., 2012). Each item is rated on a 6-point Likert scale, ranging from 0 (strongly disagree) to 5 (strongly agree). MAIA is organized into eight distinct subscales, each focusing on different aspects of interoceptive awareness: Noticing (Items 1, 2, 3, and 4), Not distracting (Items 5, 6, and 7), Not worrying (Items 8, 9, and 10), Attention regulation (Items 11, 12, 13, 14, 15, 16, and 17), Emotional awareness (Items 18, 19, 20, 21, and 22), Self-regulation (Items 23, 24, 25, and 26), Body listening (Items 27, 28, and 29), and Trusting (Items 30, 31, and 32).

The scale was translated from English into Czech by employing a following process. First, five independent Czech translations were prepared by native Czech speakers, including a psychology student, two psychologists, and two laypeople. Second, a subset group consisting of two psychologists and the psychology student discussed these translations and synthesized them into a single version. Third, this version was back-translated into English by a bilingual native English speaker, and the translated version was compared to the original English one. Finally, the final Czech version was field-tested with five respondents to ensure the comprehensibility of the items. For the Czech and original English wordings of all 32 items, as well as item means and standard deviations, please refer to Supplement 1 in the online supplemental materials.

Outcome Rating Scale (ORS; Miller et al., 2003) is a very brief measure of psychological wellbeing suitable for repeated measurement of psychotherapy outcomes. It allows patients to rate their individual, relational, social, and global wellbeing. The assessment employs four visual analog scales, each ranging from 0 to 100. Seryjová Juhová et al. (2021) reported a unidimensional factor structure for the ORS, indicating that it measures a single underlying construct. In the current study, the McDonald’s ω-coefficient for the ORS at baseline was found to be ω = 0.85, demonstrating good internal consistency.

The nine items of Patient Health Questionnaire-9 (PHQ-9) are scored on a Likert scale ranging from 0 (Not at all) to 3 (Nearly every day). Higher scores represent higher prevalence of depressive symptoms. Cronbach’s α of the scale was 0.89 (Kroenke et al., 2001). Even though the Czech translation of PHQ-9 demonstrated a two-dimensional solution with somatic and affective domains (Daňsová et al., 2016), the scale could be considered unidimensional (Kocalevent et al., 2013) because both factors are strongly correlated and the overall fit of Czech version was satisfactory even for the unidimensional structure (Daňsová et al., 2016).

The seven items of Generalized Anxiety Disorder-7 (GAD-7) are scored on a Likert scale ranging from 0 (Not at all) to 3 (Nearly every day). Higher scores represent higher generalized anxiety disorder severity. The scale is believed to be unidimensional with a Cronbach’s α of 0.92 (Spitzer et al., 2006).

The 20 items of Chronic Pain Acceptance Questionnaire-20 (CPAQ-20) are scored on a Likert scale ranging from 0 (Never true) to 6 (Always true). The scale is believed to be two-dimensional divided into activity engagement and pain willingness. All pain willingness items were negatively worded. Higher scores represent higher acceptance of pain. Cronbach’s α of both scales was shown to be satisfactory: activity engagement of 0.78 and pain willingness of 0.82 (McCracken et al., 2004; Vowles et al., 2008). In this study, a modified translation regarding acceptance to symptom, which is more general than just pain, was used (Klocek et al., 2023).

The eight items of Brief Dissociation Experiences Scale (DES-B) are scored on a 5-point Likert scale ranging from 0 (Not at all) to 4 (More than once a day). Higher scores represent higher severity of dissociative symptoms. This scale DES-B is a modification for DSM-5 from the larger DES-R scale and psychometric characteristics have not been provided yet. However, DES-R Cronbach’s α was 0.96 (Dalenberg & Carlson, 2010).

The 14 items of Balanced Index of Psychological Mindedness (BIPM) are scored on a Likert scale ranging from 0 (Not true) to 4 (Very much true). The scale is believed to be two-dimensional divided into Interest in attending to one’s psychological phenomena and Insight into these phenomena. All the insight items were negatively keyed. Cronbach’s α of Interest subscale was 0.85 and Insight subscale was 0.76. Higher scores represent higher psychological mindedness (Nyklíček & Denollet, 2009).

The 6-item version of Whiteley Index (WI) is scored on a Likert scale ranging from 1 (Not at all) to 5 (To a great extent). Higher scores represent higher level of health anxiety or hypochondriasis. Internal consistency for the 6-item version was not reported; however, in the present study, McDonald’s ω was 0.87. The scale is believed to be unidimensional (Veddegjærde et al., 2014).

The five items of Psychological Treatment Inventory–Alexithymia Scale (PTIAS) are scored on Likert scale ranging from 1 (Not at all) to 5 (To a great extent). Higher scores represent higher alexithymia level. Cronbach’s α was 0.88. The scale is believed to be unidimensional (Gori et al., 2012).

Statistical analyses were conducted using R, version 4.0.3 (R core team, 2020). Prior to analysis, items with negative wording (specifically, MAIA^CZ Items 5, 6, 7, 8, and 9) were reversed. Before proceeding with any further analyses, patterns of missing data were carefully examined using the “visdat” R package (Tierney et al., 2019). Moreover, missing data was handled using full information maximum likelihood (FIML) in the individual analyses. For transparency and accessibility, anonymized open data are available in the Open Science Framework repository (Řiháček, 2019).

To assess factorial validity and measurement invariance, we conducted a confirmatory factor analysis (CFA) using the lavaan R package (Rosseel, 2012) with a robust maximum likelihood estimator (MLR), which is still suitable for ordinal data, such as Likert scale items, and provides more degrees of freedom than specialized estimators like WLSMV or DLS (in MLR, thresholds are not parametrized). Missing data was considered not to be missing completely at random (Little’s test was significant); however, we employed the full information maximum likelihood (FIML) technique to handle also data that may be missing at random.

Covariances of latent variables were freely estimated, and model identification was achieved by fixing the variances of latent variables to 1. We evaluated model fit using several indices, following structural equation modeling fit criteria according to Hooper et al. (2008): (1) standardized root mean residual (SRMR) values optimally below 0.05 but values below 0.08 are still acceptable; (2) root mean square error of approximation (RMSEA) optimally below 0.08; (3) Tucker-Lewis index (TLI), at least above 0.90, optimally above 0.95; (4) Comparative fit index (CFI akin as TLI), and (5) insignificant χ².

Two additional considerations were made before analysis. First, significant χ² is frequent in larger datasets and those without normal distribution (Saris et al., 2009); therefore, we considered any model with satisfactory values in other fit indices, even when χ² is significant. Second, TLI and CFI indices were considered less informative when the baseline model’s RMSEA was very low (below or close to 0.158; Kenny, 2020). We estimated McDonald’s ω total and ω hierarchical coefficients (McDonald, 1999) for each subscale to assess internal consistency.

To test invariance of the final model, multigroup CFA was performed using the “semTools” R package (Jorgensen et al., 2020). Invariance was tested at different levels, including configural (no constraints between groups), metric (loadings), scalar (loadings and intercepts), strict (loadings, intercepts, and residual variance), and factor means (loadings, intercepts, and factor means) levels (Curtis et al., 2020). When increasing equality constraints on the estimated parameters does not excessively change the model fit, the model is considered to be invariant. Invariance was tested on the Models with constrained factor means and residual variances are parallel to each other, wherefore were both compared with the scalar model. The same fit indices (and their change values, Δ) as in the single-group CFA were employed. In samples of at least 300 participants, changes in fit indices compared to less constrained models needed to comply with specific rules to be considered invariant: ΔTLI ≤ 0.010 (all invariance levels), ΔRMSEA ≤ 0.015 (all invariance levels), or ΔSRMR ≤ 0.030 (only metric invariance level) and ≤ 0.010 (scalar and strict invariance level) (Chen, 2007).

Invariance was tested between genders and between two artificially generated age cohorts (younger and older according to median split). Additionally, because the network model assumes stationarity, we performed also a longitudinal measurement invariance across measurement waves, involving series of nested model comparisons with increasingly restrictive parameter equality constraints over measurement waves (Van de Schoot et al., 2012). One additional constraint was present in all longitudinal invariance models—residual auto-covariances were set to be equal across all waves for all items. Nested models were compared using an ANOVA function. Because of increasing missingness in progressive measurement waves, a FIML estimator was used.

In terms of convergent validity, MAIA^CZ subscales were associated with several related constructs using correlation between latent factors free of measurement error. It has been hypothesized that MAIA^CZ subscales are negatively correlated with depression (PHQ-9), anxiety (GAD-7), alexithymia (PTI-AS), hypochondriasis and worries (Whiteley Index), and dissociative experiences (DES-Brief) showing discriminant validity, whereas acceptance (CPAQ-20) and psychological mindedness (BIPM) are positively correlated with MAIA^CZ subscales showing convergent validity. In this study, the second subscale of CPAQ-20 was reversed; therefore, it represents pain willingness. Also, the second subscale of BIPM was reversed; therefore, it represents insight.

Psychological network models are increasingly often utilized in psychometric literature (Epskamp et al., 2020). A network is a graphical representation of a set of variables (labeled as nodes) and interconnections between these variables (labeled as links/edges). Psychometric networks usually interconnect items of a specific scale or symptoms as nodes and estimated partial correlations between them as edges. Unlike factor analysis, network models do not require assumption of any latent variable allowing for the exploration of direct associations between nodes, exploration of various patterns in the data structures and of specific network characteristics such as node centrality (Epskamp et al., 2018a, 2018b). Given the edges are often based on partial correlations, adding or omitting a particular node from a network might yield slightly different results. Network models are relatively sensitive to small effects and rich information might be difficult to interpret. Therefore, some forms of network regularization are often considered to increase the specificity and reduce the potential noise in the data, such as pruning of non-significant edges (Epskamp et al., 2020). To increase the stability of the network structure, non-parametric bootstrapping is often employed.

In the present study, we did not have sufficient power for longitudinal network models computed on the item-level; however, previously extracted factor scores or subscale composite scores could be used as input data to the network model as well (Epskamp, 2020). Moreover, based on the theoretical model from Mehling et al. (2012), MAIA^CZ subscales represent several distinct facets of which IA consists of. Network models are more suitable methods to explore how these subscales influence each other and evolve over time together than structural equation modeling alternatives (Epskamp, 2020; Epskamp et al., 2018a, 2018b). In this study, a dynamic latent variable model for panel data (with lag-1) was used (psychonetrics, Epskamp et al., 2020). Using this model, we considered data across all 8 measurement waves and employed the “nlminb” optimizer. There is an assumption of a multivariate Gaussian density of variables and requirement of stationarity. By adding a wellbeing node into the network model of all eight MAIA^CZ subscales, a predictive validity of the instrument could be tested in an innovative way, exploring which of the subscales are associated with the wellbeing outcome. This method is suitable for demonstrating, which of the MAIA^CZ subscales possess a practical utility by predicting better treatment outcomes over time.

To address the missing data, we used a FIML estimator. The network model was defined as a multi-level graphical vector-autoregression model with random effects on the mean structure (Epskamp, 2020). We computed two network models: (1) one focused on interoceptive awareness alone (all MAIA^CZ subscales) and (2) another where we added wellbeing to the MAIA^CZ subscales. The first network model was reported in Supplement 5a, 5b, and 5c. It was not reported in the main text of the present study because it wielded the same associations as the second network model. We extracted detrended latent factor scores from all MAIA^CZ subscales (Model 2a) and the ORS for each of the 8 measurement waves. These scores served as input to the longitudinal network model, which was initially fitted as a saturated Gaussian graphical model and then pruned of all non-significant associations (α-level = 0.01). To enhance the stability of the estimated parameters, the saturated pruned model was further bootstrapped with 100 resamples with replacement to create 90% confidence intervals around the mean estimates.

The output of the longitudinal network model was divided into three separate networks, each graphically represented in separated plots. First, a temporal network incorporating both vector-autoregressive (van der Krieke et al., 2015) and cross-lagged effects (Selig & Little, 2012) represent within-patient temporal partial directed correlations from a previous measurement wave (Epskamp, 2020). This network shows which IA facets measured in 1 week during therapy temporally precede other IA facets measured the following week during therapy. Second, contemporaneous network incorporates within-patient residual partial correlations in the same measurement wave controlled for all effects from the temporal network (Epskamp et al., 2018a, 2018b). This network shows which IA facets are inter-related at the within-person level regardless of the measurement wave. Third, a between-person network incorporated partial correlations between mean structure of nodes varying across patients. All three networks were assessed for centrality indices (see Supplement 8). The node is understood as central, when (1) having strong relationships with other nodes (edge strength); (2) being often in the shortest path between other nodes (closeness); and (3) being often in between two other nodes (betweenness) (Costantini et al., 2019; Epskamp et al., 2018a, 2018b).

The fit of Models 1 (unidimensional), 3 (single second-order factor), and 4 (two second-order factors) was insufficient, as shown in Table 2. In the unidimensional factor solution, when examining factor loadings, items from Not distracting and Not worrying subscales had notably low and insignificant loadings. In contrast, the eight-factor structure in model 2 provided the best fit of the tested models (see also in Table 2), even though still not satisfactory. All factor loadings were significant, ranging from 0.325 (Item 10) to 0.917 (Item 31). For model parameters, see Table 3, and a corresponding figure of the model is provided in Supplement 2 (online supplemental materials). The TLI and CFI fit indices were slightly lower than 0.90. The RMSEA was in an acceptable range. However, it is worth noting that the baseline model’s RMSEA was unusually low at 0.160, which complicates the interpretation of the TLI and CFI fit indices (for explanation, please see Kenny, 2020).

According to modification indices, assigning Item 18 successively to Body listening or to Attention regulation would improve the model fit. Another suggestion was to allow residual correlation between Items 16 and 17 or between Items 13 and 14, all of which were representing the diverse attention regulation mechanisms. These modifications were considered to make the model align better with the theoretical representation of the IA construct. However, Reis (2019) criticized excessive usage of modification indices to improve the model fit of MAIA. He recommended freeing only one parameter at a time and only with theoretical support to maintain the model sparsity. The residual correlation between Items 13 (“When I am in conversation with someone, I can pay attention to my posture”) and 14 (“I can return awareness to my body if I am distracted”) had been found in a previous validation study (Valenzuela-Moguillansky and Reyes-Reyes, 2015). However, there was no prior validation or additional theoretical support for the residual correlation between Items 16 (“I can maintain awareness of my whole body even when a part of me is in pain or discomfort”) and 17 (“I am able to consciously focus on my body as a whole”). Therefore, the modified Model 2a (i.e., Model 2 with a single residual correlation between Items 13 and 14) was considered the final model for the MAIA^CZ instrument, with an acceptable model fit. The residual correlation between Items 13 and 14 was r = 0.41, and Model 2a significantly outperformed Model 2: χ²(1) = 47.09, p < 0.01.

The internal consistency of the subscales generally was adequate (ω > 0.80, α > 0.80), except for Not distracting (ω = 0.639, α = 0.620), Noticing (ω = 0.713, α = 0.684), and Not worrying (ω = 0.719, α = 0.658). If all 32 items were considered together, internal consistency was also adequate (total ω = 0.866, total α = 0.876). Additionally, regarding Cronbach’s α, Emotional awareness (α = 0.780) and Self-regulation (α = 0.790) did not reach the 0.80 threshold, which can introduce some error proneness. It is important to mention that Cronbach’s α was reported only for comparison purposes with prior literature because Tau-equivalent model was not confirmed, consistent with the studies by Reis (2019) and Todd et al. (2020). Reis (2019) emphasized the use of McDonald’s ω over Cronbach’s α after testing both the model of congeneric tests and Tau-equivalent model.

Model 2a was tested for measurement invariance with respect to age cohorts and gender groups (see Supplement 3 in online supplemental materials). The final model was fully invariant between age groups divided by median split (n = 428), i.e., younger (n = 226, age ≤ 40 years) and older (n = 202, age > 40 years) up to the factor means and strict levels. However, when analyzing invariance among both gender groups, we encountered a negative error variance or a Heywood case. After reviewing the Likert response options, we discovered that Option 5 was rarely used by the study sample, resulting in missing responses. To address this, we have decided to merge both response Options 4 and 5 to reduce the number of estimated parameters. Following this adjustment, Model 2a demonstrated invariance between gender groups (n = 429), male (n = 108), and female (n = 321) up to the factor means and strict level.

Model 2a was further subjected to longitudinal invariance testing (see Supplement 4 in online supplemental materials). Due to the limited data size, the statistical test lacked the power to perform longitudinal measurement invariance analysis within the full model, which included all eight subscales and 8 measurement waves. Therefore, we performed the analysis for each MAIA^CZ subscale separately. According to Chen’s (2007) fit criteria, Attention regulation, Emotional awareness, and Self-regulation subscales were deemed longitudinally invariant up to a strict level. For the remaining MAIA^CZ subscales, invariance was established up to the scalar level, with a caveat regarding Trusting and Not worrying subscales, as these models did not yield positive definite results.

In terms of convergent or discriminant validity (Table 4), MAIA^CZ domains were overall negatively correlated with depression and anxiety. The exception was Noticing subscale which correlated positively with depression and anxiety. Furthermore, Not distracting and Emotional awareness did not correlate with depression and anxiety at all. Alexithymia was also usually negatively correlated with MAIA^CZ subscales, except of Noticing, Emotional awareness, and Not worrying, which were unrelated to alexithymia. Relatively strong negative correlation was found between Not worrying and hypochondriasis. This adds certain credibility to the otherwise less optimally functioning Not worrying subscale. Trusting and Self-regulation were also negatively correlated with hypochondriasis, whereas unexpectedly, the Noticing and Not distracting subscales were related to the hypochondriasis positively. Dissociative experiences correlated negatively only with Trusting and Not distracting (although significant, the correlation of r = 0.15 was expected to be higher) and positively with Noticing. Psychological mindedness subscales, interest and insight, correlated positively with all MAIA^CZ domains, except for Not worrying. Acceptance was associated with all MAIA^CZ domains positively, except of Noticing which was correlated negatively. The pain willingness domain of acceptance showed many non-significant associations. However, this could be related to the overall lower reliability of this subscale.

Latent variables of MAIA^CZ subscales (Model 2a) together with the latent variable of ORS (unidimensional) were included as nodes within the longitudinal network. Longitudinal associations between variables with 1-week lag were estimated using a panel data design. The initial saturated model showed adequate fit, χ²(2520) = 4880.43, p < 0.001, TLI = 0.910, CFI = 0.911, RMSEA = 0.047 95% CI [0.045–0.049], AIC = 35721.77. The model was then pruned of non-significant edges (α = 0.01) by fixing them to zero. The fit of the pruned model, χ²(2605) = 5128.94, p < 0.001, TLI = 0.907, CFI = 0.905, RMSEA = 0.047 95% CI [0.045–0.049], AIC = 35800.27, was significantly worse, compared to the only saturated model, χ²(85) = 248.51, p < 0.001. Nevertheless, we still consider the fit as excellent. Graphs of mean bootstrapped temporal, contemporaneous, and between-person networks are depicted in Fig. 1. Numeric values of edges with bootstrapped confidence intervals are reported in Supplements 6 and 7. Centrality indices of all three networks are reported in Supplement 8.

In the temporal network, only weak relationships were observed. Given the directed nature of this network, we could identify several Granger causal pathways. Not distracting and Not worrying subscales emerged as key independent predictors of other nodes in the network. Patients who were less likely to distract themselves from perceiving bodily sensations (less passive suppression of interoceptive sensations or avoidance) reported increased Self-regulation (r = 0.075), Attention regulation (r = 0.044), and Body listening (r = 0.037) in the following week. Similarly, patients who tended not to worry about bodily sensations (less ruminative cognitive appraisal) reported increases in Self-regulation (r = 0.032), Attention regulation (r = 0.029), Body listening (r = 0.022), and Trusting (r = 0.052) in the subsequent week. Trusting appeared less connected with other IA mechanisms and was considered as potential “causal” endpoint or sink node in the network. The only way to increase the level of seeing bodily sensations as helpful indicators of the condition of the body in our data was through reduction of worrying (ruminative judgmental thinking) of bodily sensations.

For successful connection to one’s body, Body listening, a patient needs the ability of Self-regulation (r = 0.055), which could serve as a potential mediator or transitional node of the influence emerging from Not distracting and Not worrying. Self-regulation and Body listening represent the mind–body integration strategies and were the most central nodes in the temporal network in terms of edge strength and betweenness, respectively. After self-regulation, an interesting feedback loop emerged between Noticing and Body listening. As could be seen from the graph, Noticing is further important for increased Emotional awareness (r = 0.076), another sink node in the temporal network.

Regarding predictive validity, none of the MAIA^CZ subscales as measured in the previous week predicted wellbeing in the following week, indicating limited support for the directed predictive validity of MAIA^CZ in this timeframe. Well-being itself showed stability over time and some independence from the MAIA^CZ construct. There was also a directed association between wellbeing and Not worrying (r = 0.057), suggesting that the relationship could also be vice versa and wellbeing might also predict IA.

In the contemporaneous network, a multitude of associations emerged that were not dependent on the weekly measurement timescale. The most central node was again Self-regulation (in terms of edge strength), integrating the entire network. The highly interconnected core of beneficial IA consisted of a triad of skills identified also as the core of mindfulness: Body listening, Self-regulation, and Attention regulation, which were independently positively associated with Trusting. However, several negative partial correlations were found in the contemporaneous network, particularly for the Noticing, Not distracting, and Not worrying subscales. Clearly, these three nodes behaved differently than the rest of the network. Both subscales representing emotional reaction and attentional response to bodily sensations were situated on the network’s margin rather than in the center (based on centrality indices). When compared to the temporal network, associations between Not distracting and both Self-regulation and Attention regulation changed from positive to negative in the contemporaneous network. Moreover, Noticing was most often in between two other nodes (highest betweenness centrality). Regarding predictive validity in the contemporaneous network, four MAIA^CZ subscales were independently associated with wellbeing: Self-regulation (r = 0.093), Trusting (r = 0.076), Emotional awareness (r = 0.083), and Not worrying (r = 0.094). However, these associations were small.

In the between-person network, many relationships were consistent with the contemporaneous network, but some differences were observed. Noticing and Body listening no longer exhibited an association, and a novel negative association emerged between Emotional awareness and Attention regulation (r = − 0.246). Attention regulation was the most central node in the between-person network and appeared relevant for wellbeing through Trusting. Specifically, patients with higher general levels of Trusting were associated with better wellbeing. In the between-person network, only Trusting was independently associated with wellbeing (r = 0.441).

The first limitation was the study sample. Similar to earlier MAIA psychometric studies (e.g., Machorrinho et al., 2019), the Czech translation was also validated in a female-dominated sample with a model fit that is not convincing. Furthermore, the sample size was relatively small. The analysis was not powered enough to estimate the network models on the item level. Therefore, latent factor means of the subscales of the final model were used instead with the possibility of being underpowered even after using subscales as nodes. Finally, patients were collected in seven clinical sites, and differences between sites were neither investigated, nor controlled for in this study. The potential differences between clinical sites might bias the results.

The second limitation was related to testing of the factorial validity. The final model was modified by allowing a residual covariance between Items 13 and 14. On the one hand, even though the modification was justified, any modification might reduce the overall generalizability of the results and subsequently also the network models utilizing extracted factor scores from this modified factor analysis model. On the other hand, the model fit of the final model was still not satisfactory and further modifications might be required if the future authors intend to use the instrument as a compact multidimensional measure of individual IA facets. Additionally, the estimated internal consistency coefficients of several subscales (i.e., Not distracting, Not worrying, and Noticing) were rather low (< 0.80). By including unreliable nodes in the network, a certain bias of all estimates is unavoidable. However, using factor scores (instead of items) as nodes filtered out part of the measurement error. Moreover, the internal consistency differed considerably across measurement waves. Furthermore, the age groups in measurement invariance testing represented arbitrary categories without any theoretical support. Finally, the recent advancements in testing model fit suggest using dynamic fit indices (see, e.g., McNeish & Wolf, 2023) instead of one-time-fits-all cutoff criteria as we used in the present study. We assume none of the tested models would reach an acceptable fit using these criteria.

The third limitation was related to the stationarity assumption. Trusting and Not worrying subscales did not have positive definite models in the longitudinal measurement invariance testing. Additionally, within-person networks assumed stable mean structure across patients over time and real clinical data seldom conform to this assumption. The stationarity problem affects also the between-person network. On the one hand, because the responses from several time points were averaged, a bias of self-report measurement is reduced. On the other hand, because these data originated from a sample of patients undergoing a therapeutic intervention, average values across all time points might not be interpretable. Also, the lag of 1 week between measurement points might be suboptimal for the IA variables. Perhaps, a longer or shorter timeframe would produce different results. The usage of experience sampling methodology might be relevant in this regard.

The fourth limitation was related to the interpretation of network edges. Edges in temporal and contemporaneous networks may indicate both real causal relationships and unmodeled nonstationarity of the mean. For conclusions about causal relationships, experimental manipulation is still needed (Epskamp et al., 2018a, 2018b). Moreover, the negative edges occurring in the contemporaneous and between-person networks could be caused by a collider included as a node in the network known as a Berkson’s bias (De Ron et al., 2021).