Given the basics of the theory, the core emotional needs and the EMSs are both about our interpersonal life. While most studies have focused on their influence on more complicated constructs such as depression and interpersonal problems, recently some researchers have proposed that schemas may affect our interpersonal functions at a more basic level: our ability to recognize emotions from facial cues (Csukly et al., 2011; Panić et al., 2022). For instance, a person with an Abandonment schema may be more sensitive to facial cues of disgust; another with a Vulnerability schema may be more sensitive to anger; a third person with a Self-Sacrifice schema may be oversensitive to sadness; and one with an Insufficient Self-Control schema who has a high tendency to avoid discomfort (Young et al., 2003, p. 16) may automatically avoid facial cues of sadness or anger. Though these relations make logical sense, they are speculative and hence need empirical validation. If such hypotheses are correct, the maladaptively filtrated nonverbal cues may adversely affect our social life, given the vital role of nonverbal communication (Wegrzyn et al., 2017).

The influence of early experiences on facial emotion recognition fits even biologically-based theories such as Ekman’s theory (Ekman, 1992). While considering emotion recognition as biologically determined, Ekman argues that humans have an open affect program. That is, their program is open to inputs from their experience; once these new inputs are integrated into the program, it runs automatically as it was biologically determined (Ekman and Cordaro, 2011).

Young Schema Questionnaire (YSQ) is the primary tool for assessing the EMSs. Given its theoretical framework and the empirical evidence from factor analytic studies (e.g., Bach et al., 2018), a two-level structure is suggested for the YSQ.

The first-order structure corresponds to the EMSs; each latent factor is assumed to indicate the severity of a schema. As there are many schemas (18 in YSQ-3), it is difficult to study how they relate to other substantive variables. Therefore, many researchers prefer to reduce the schemas to a smaller number of variables, i.e., the second-order factors.

The second-order structure, called schema domains, is theoretically based on categorizing schemas according to unmet needs that give rise to the EMSs. Young et al. (2003, p. 10) tentatively proposed that there are five core emotional needs: secure attachment, autonomy and competence, freedom to express needs and emotion, spontaneity and play, and realistic limits and self-control. Accordingly, the frustration of one need leads to the formation of one or more maladaptive schemas within its corresponding domain. They did not, however, clarify how the frustration of one of these five core needs can lead to not only a specific one but to different schemas within a domain. We can think of at least two hypotheses. The first is that these five needs are further subdivided into facets, and the frustration of each facet of these needs can lead to a specific schema. The second hypothesis is that these needs are unidimensional, but other factors such as temperament determine which EMS to form.

Taking a more in-depth look into the theoretical perspectives, clinical experience, and empirical evidence on core emotional needs, Lockwood and Perris (2012) suggested that for every single schema, there is an unmet need. They further suggested a hierarchical structure for these needs, which they classified into four broad domains. As the schemas stem from the needs, their structure is assumed to parallel the structure of needs. Coauthored by Young, Bach et al. (2018) confirmed this four-domain structure of EMSs on a large sample of clinical and non-clinical participants. Thimm (2022) also provided meta-analytic evidence that this structure is the best fit. Further support came along when a fairly comparable structure was found in early adaptive schemas as well (Louis et al., 2020); that is if the needs are responsible for both maladaptive and adaptive schemas, the same structure is expected to be found in both.

While researchers are attracted to the use of second-order structure (as it reduces the large number of schemas to a more manageable number), there are at least two major issues that need to be addressed. First, the meaning of the higher-order constructs is not well clarified yet. The domain-level scores reflect what is shared between groups of schemas. If we assume their correspondence with the needs hierarchical structure, we then must clarify what the need domains really mean. If we assume the original five core needs, with different schemas rising from a combination of unmet needs and other factors, then the shared component may have a very different meaning. The second problem is, as we will show here, that some domains substantially overlap, to the extent that they may not be easily differentiable; therefore they may not be useful constructs.

Besides the first–and second-order structures, another level of structure also makes sense, as there seems to be a common susceptibility factor across all schemas. Extensive evidence from gene–environment interaction (GxE) indicates that individuals are not similar in how they are affected by the environment. Some are very sensitive to both risky and supporting environments, likely due to higher levels of neuroplasticity; some are less so (Belsky et al., 2009; Boyce, 2016; Assary et al., 2018). On the other hand, neglective environment tends to neglect different aspects of the child’s needs. Hence, a common variability is generated, where sensitive individuals with poor environments develop different maladaptive schemas, while less sensitive individuals do not.

For instance, the development of different schemas is likely affected by similar temperamental and environmental factors such as neuroticism, and poor parenting modes (Lockwood and Perris, 2012). For instance, someone who this is further evident from the positive correlation among all schemas (for a meta-analysis see Thimm, 2022).

Given what we elaborated, and taking a statistical perspective, the reliable variance in schemas can be decomposed into three components: (1) the variance due to the general susceptibility that is shared among all schemas and/or schema domains; (2) the variance due to how the interaction between general susceptibility and environment leads to the frustration of some needs but not others; and (3) the variance in the specific schemas developed in response to the frustration of needs. These different variance components may differentially correlate with other variables of interest; therefore, ignoring the issue can lead to the misinterpretation of empirical findings, hindering the advancement of our understanding of the underlying mechanisms.

In this study, we first explore the structure of schema, with a greater focus on schema domains and the general susceptibility factor. While such a generic factor has been suggested and investigated before (Kriston et al., 2012; Oettingen et al., 2018; Yan et al., 2018), the possibility that it may constitute the major proportion of variance in domain scores has not been explored. This is important for research, given the widespread interest in using domain scores. As a second objective, we used the variance decomposition approach to explore how different variance components of schemas relate to facial emotion recognition, a fairly under investigated topic (Csukly et al., 2011; Panić et al., 2022). We suggest that such a framework may facilitate a more in-depth and less biased understanding of how schemas are related to other constructs of interest, especially with regard to the general susceptibility factor which seemingly has not been explored.

A total of 259 participants were recruited from social media; snowball sampling was implemented when possible. All participants were Iranian adults (> 18) with no severe mental disorders. A data cleaning procedure, using Mahalanobis distance (Desimone et al., 2015), revealed 26 potentially random responders (ps < 0.001). Removing these cases yielded a sample of 233. Of the final sample, the majority were female (88%), unmarried (68%), and with academic degrees of bachelor’s or higher (70%). The age distribution was positively skewed, with a median of 22 and a range of 18–71; 73% were between 18 and 30 and 27% were older.

To ensure sufficient power, we conducted a series of power analyses, using the semPower package (Moshagen and Erdfelder, 2016). These analyses show us the present statistical power to reject the models due to global model misfit, based on a commonly used approximate fit index, root mean square of error approximation (RMSEA). For the original first-order structure, our sample of 233 had excellent power (> 0.99) to detect a misfit of RMSEA = 0.05 (α = 0.05, df = 2,595). For the original second-order structure, our sample of 226 (excluding schema-level outliers) had a power of 0.9 to detect such a misfit (α = 0.05, df = 80). Concerning hypotheses about relations, we had 0.8 power to detect a significant (α = 0.05) correlation of |r| = 0.185 (r²  = 0.034; this power analysis was conducted in G^*Power. Given the complexity of power analysis in the context of SEM, we used the traditional power analysis as an approximation).

The research was conducted in 2022. The data was collected via Google Forms; the link was distributed through social media (e.g., WhatsApp, and Instagram). Participants completed demographics, then YSQ–75, followed by the facial emotion recognition task. They provided consent by clicking a checkbox with a clear description. If they were willing so, they received their results on YSQ via email.

All analyses were conducted in R. A data cleaning procedure was conducted, using Mahalanobis distance, to detect random responders; a probability of p < 0.001 was considered to indicate random responding (DeSimone et al., 2015). We used lavaan (Rosseel, 2012) to perform confirmatory factor analysis (CFA) and structural equation modeling (SEM). Due to moderate violation of parametric assumptions (normality and homoskedasticity), the robust maximum likelihood (MLR) was used for estimation, along with sandwich-type standard error and robust fit indices [χ², comparative fit index (CFI), root mean square of error approximation (RMSEA) with its 90% confidence intervals, and standardized root mean square residual (SRMR)]. For higher-order factor analysis, we included schema scores as indicators to be able to assess the fit exclusively for higher-order structure.

We explore the distributional characteristics of the data to ensure the statistical assumptions. The item-level data on the Yaung’s questionnaire showed skewness indices ranging from −0.1.24 to 3.47; however, only two items showed skewness higher than 3. They also showed kurtosis indices (ranging from −1.37 to 13.52). Expectedly, the item-level data further showed heteroskedastic bivariate distributions. The schema-level data also showed mild to moderate violations of normality (with skewness ranging from 0.01 to 2.37 and kurtosis from −0.56 to 8.9) and homoskedasticity, though these violations were less severe. Due to these data characteristics, to ensure valid results we used the method of Robust Maximum Likelihood with a heteroskedasticity consisted standard error along with robust fit indices.

The overall success rate in the emotion recognition task across all conditions was 88%; it was 94% for fear, 80% for both anger and disgust, 88% for both sadness and neutral, and 99% for happiness. The overall rate in masked and non-masked conditions was 86 and 91%, respectively.

Regarding our first objective, while the 15-factor first-order structure was confirmed, the original and the revised models of the second-order structure showed a problematic fit. The original five-domain model seemed more problematic; it had a poor fit along with very high intercorrelations among domains. Such high intercorrelations raise doubt whether the domains indeed represent distinct factors, and even so, it would be of little use in applied research, as there is very small domain-specific variance to work with.

The revised four-factor model showed a better yet unsatisfactory fit. The unsatisfactory fit of the four-domain model is also shown in a meta-analytic CFA investigation in which the model was fitted to a pooled covariance matrix meta-analytically drawn from 27 samples (Thimm, 2022). Furthermore, in a recent study with a large sample (N ≈ 2,300), the four-, five-, and single-domain models all showed similarly poor fit according to CFI and χ²/df and similarly good fit according to SRMR and RMSEA. However, another study on Iranian psychiatric patients reported an acceptable fit for the original five-domain model and a nearly acceptable fit for the four-domain model (Khosravani et al., 2020).

This study, along with previous evidence, suggests that a single susceptibility factor is consistently extractable from YSQ. At the item level, a single factor accounted for 28% of the total variance; at the schema level, a single factor accounted for 40% of the total variance; and at the domain level, a single factor accounted for 64% of the variance in the four domains (90% of the variance in Disconnection & Rejection). This means that any relation of domains to other constructs can be due to the relation of the general factor to those constructs. The same issue is present in relation between schemas and other outcomes but with a lower magnitude.

The nature of the common susceptibility factor requires further investigation. The shared variance likely reflects the influence of neuroticism. Previous evidence suggests that all EMSs are correlated with neuroticism, except for Entitlement and Self-Sacrifice. Interestingly and in concordance with these findings, in this study, the same two schemas had the lowest loading of the general factor; while the total item variance explained by the general factor was 28% across all items, the total variance explained by the general factor for Entitlement was 11% and for Self-Sacrifice was 9%. The common method variance (Podsakoff et al., 2003) is another possible explanation for the empirical presence of the general factor. Also, a combination of these two is possible.

There is another important issue to consider. If we accept that the bifactor model is the correct one in the population, then YSQ may need further revision to enhance the reliability of schemas when the variance due to the common factor is partialled out. In this and a previous study (Kriston et al., 2012), with the inclusion of a general factor, the reliability of specific factors was substantially reduced, and for some schemas (in this study for most of them), it fell below the conventional threshold. The low reliability makes it difficult to investigate the contribution of a specific schema to other constructs of interest. One potential way for such a revision is to use the original long-form schema questionnaire to perform an item that maximizes the schema-specific variance in the context of the bifactor model. Notably, such a revision may also help with finding the higher-order structure that is of incremental validity and practical utility.

However, ignoring the shared variance among schemas, and exploring the relation between schemas and other constructs can be misleading. A consistently found relation between a schema and an outcome may be indeed due to the influence of general susceptibility (neuroticism perhaps). Also, given that the majority of the variance in domains (especially Disconnection & Rejection and Impaired Autonomy) can be attributed to the general factor, any association between domains and other constructs must be interpreted with great caution.

Before ending our discussion regarding the second-order structure, it is worth noting that we can think of such a classification in a different way. We think of it as a merely practical tool. Regardless of their correspondence with real entities, the domains may provide a structure to the schemas so that the therapists do not get lost between the large number of schemas. Notably, Lockwood and Perris (2012, p. 58) cautioned against firm conclusions about these higher-order factors being real entities.

Regarding our second objective, we had tentative evidence for a positive association between the general factor of schemas and anger recognition. Disconnection & Rejection was also related to anger recognition; however, when the variance of the general factor was partialled out, the correlation reduced to a nonsignificant level. We could not find reliable evidence on the association of other domains or schemas with facial emotion recognition. So far, the evidence from this and the two previous studies do not provide a clear picture.

In the first study on this topic, Csukly et al. (2011) investigated the association between four schema domains (extracted in an exploratory analysis) and seven facial expressions among inpatients with major depression. Three significant negative associations were found: one domain (heavily loaded by Dependence, Failure, and Subjugation) was related to weaker detection of happiness, another (heavily loaded by Entitlement and Insufficient Self-Control) to detecting sadness, and a third (heavily loaded by Self-Sacrifice and Unrelenting Standards) to detecting fear; but the effect sizes were small.

We could not replicate these results. There are several possible explanations. First, their sample were inpatients with severe depression, which can influence the results; according to the same study, depressed individuals scored approximately 1 SD lower than the control group on the emotion recognition task. Second, it seems that a total of 28 tests were done, which inflates the type-I error rate. Third, it is possible that the domain factors were assessing different constructs, as the different loading patterns suggest. Furthermore, we used a confirmatory approach which does not allow for cross-loadings while they used an exploratory one which allows. While the specific schemas were similar, different constructions of latent factors can lead to different relations to emotion recognition.

More recently, Panić et al. (2022) conducted a similar study on a student sample. They reported significant and fairly high associations between all the five domains of schemas (from the original structure) and overall recognition of negative emotions. The magnitude was fairly similar for all the five domains (ranging from β = −0.32 to β = −0.41). Given the high correlation among the five domains and the similarity of their associations with emotion recognition, it is very likely that there is a shared aspect from these domains that is associated with emotion recognition, i.e., the general factor. Notably, in contrast to Csukly et al. (2011), none of the domains were significantly related to the recognition of happiness or sadness. Also, only one domain was correlated with fear; even so, the domains correlated with fear recognition were substantially different across the two studies.

Summing up, current evidence on the association between schemas and facial emotion recognition is mixed. This study, which had a larger sample size, showed mostly null results; the other two, while supporting some significant relations, they support different patterns of associations. The contradiction among pieces of evidence may partly stem from a combination of low power and numerous hypothesis testing. The low power can be partly due to the low reliability in the measurement of emotion recognition tasks. In this study, the reliability of the total scores was acceptable at best (ICC = 0.61), and the reliability for scores on separate emotions was poor (with ICCs ranging from 0.15 to 0.55). The other two studies did not report information on reliability, but they likely had similar or slightly higher reliability coefficients, given the similarity of tasks. Low reliability attenuates the relationship and reduces the power; when combined with multiple hypothesis testing, this can lead to random significant findings that contradict across studies.

Another possible explanation for the null results is the ceiling effect due to the simplicity of the emotion recognition task, which does not match real-life situations. In real-life situations, facial expressions usually last for a few seconds at most (long expressions are likely fake; Ekman and Friesen, 1975, p. 14). Also, in contrast to the tasks conducted in these studies, the observers’ attention is usually divided within real-life situations. This ceiling effect may have affected us more than Csukly et al. (2011), as their overall recognition rate was 48.3%, likely due to their recognition task being more difficult (They used different levels of task intensity). Panić et al. (2022) did not report the recognition rate, so the comparison was not possible.

Finally, and from another perspective, the behavioral response can be different for different individuals with the same schema (Young et al., 2003, p. 32); one with an Abandonment schema may avoid cues of rejection; another may get more vigilance; and a third person may use two different strategies at different time points (p. 33). While these factors are ignored, the relation may not be observable.

Several limitations must be considered. Despite our sample size being fair for our main objectives, complex models such as bifactor models may need larger samples for accurate parameter estimation; this may be why some models resulted in improper solutions. Also, the majority of our sample were women (88%); therefore, we must be cautious in generalizing these results to men. Furthermore, the sampling was done through social media, and the majority were young adults (the median age was 22) with some level of university education. In the emotion recognition task, we used static pictures with fairly intense expressions; this usually is not the case in real-life situations. However, these static pictures are the validated tools that were used in previous similar studies. All these aspects can affect the generalizability of our results to other populations and to real-life situation.

Regarding the internal validity, using extreme static emotional expressions makes the task easy for participants, leading to a high success rate and, hence, the ceiling effect, which can mask the real effects by attenuating the relation. Furthermore, the reliability of scores on emotion recognition was not good. Low reliability can lead to lower statistical power. Also, it can lead to the underestimation of the relationships between constructs. Of course, our analytic methods, i.e., structural equation modeling, can partly overcome these shortcomings by accounting for unreliability.

The literature on the higher-order structure of schemas is in its infancy. Researchers have begun using domain scores that still lack clear meaning, rationale, and validity; especially, as there is evidence that they may represent some sort of general susceptibility. Our discussion on this general factor further points to another word of caution: any bivariate relation of a schema to another construct may be attributed to this general factor. To make sure that this is not the case, this common factor needs to be partialled out, e.g., in a bifactor model. Relevant to this point, the schema questionnaire likely benefits from a revision that enhances the schema-specific variance; such a revision enhances the reliability of schemas when the common factor is partialled out. Furthermore, while we provided notable evidence on the presence of a general factor, its nature is still unclear and requires further exploration. Investigations into the relation of the general factor with other related constructs such as neuroticism may provide a better understanding. Finally, the hypothesis of parallel structure of needs and schemas and their correspondence is interesting, yet further examination is warranted.

Regarding the contribution of schemas to facial emotion recognition, the literature likely benefits from studies with larger samples, better assessment procedures, and more realistic tasks. Using video instead of static pictures may be useful in this regard. Finally, the reliability of emotion recognition assessment warrants further attention.