The study of EC has grown significantly since Boyatzis (1982) introduced the concept of competencies, emphasising the role of knowledge, skills, and abilities (KSA). In entrepreneurship, EC is broadly defined as the capacity to think, act, and adapt effectively in various business situations (Guion, 2011).

Key definitions from Bird (2019), Man et al. (2002) and Mitchelmore and Rowley (2010) converge on the idea that EC comprises fundamental characteristics-including knowledge, skills and personality trait that enable entrepreneurs to succeed. Despite variations in specific components, these three elements consistently form the core of the EC construct.

The measurement of EC in various studies to date has varied considerably. Among the most frequently referenced is the model developed by Man et al. (2008). This instrument has played a central role in shaping EC research by offering a structured approach to assessing six core competency domains: opportunity, relationship, conceptual, organisational, strategic, and commitment. This framework was later expanded to include two supporting domains (Man et al., 2008). However, despite this broader structure, most studies tend to focus only on the six domains, often neglecting the supporting domains, which limit a full understanding of the construct. To enhance its comprehensiveness, Man et al. (2008) added additional rational competencies—such as ethical competence, familism, social responsibility, and technical competence—into the model. The foundational structure of this framework has been widely accepted by other researchers to develop instruments tailored to specific research contexts and cultural settings (Draksler & Sirec, 2021; Lans et al., 2011; Rahman et al., 2015; Wulani et al., 2019).

Beyond the Man et al. model, another significant theoretical foundation for EC assessment is the EntreComp framework developed by Bacigalupo et al. (2016). This framework is gaining popularity, but most assessment tools derived from it do not fully implement its original structure. Instead, researchers often select and adapt its dimensions and sub-dimensions to suit the specific context of their studies, making it difficult to assess their conceptual grounding. The original EntreComp document itself has been excluded from some academic reviews because it was not published in a peer-reviewed journal and lacked a formal psychometric analysis of its dimensions and reliability.

Despite this variation, most EC instruments demonstrate strong internal consistency, with reported reliability coefficients (e.g. Cronbach’s alpha or McDonald’s omega) meeting or exceeding accepted standards. In terms of validity, the evidence presented in various studies ranges from moderate to strong. However, theoretical and structural inconsistencies among these instruments suggest that, to date, there is no universally accepted best measure for assessing EC.

The selection of the SJT to measure EC is based on its superiority in assessing an individual’s ability to apply knowledge and skills in scenarios relevant to real-life contexts. Unlike the Likert scale, which measures self-perception or beliefs (self-efficacy), SJT is designed to measure applied judgement. Participants are not asked ‘can you do it’; they are asked ‘what would you do’ in a given situation. This provides a more accurate picture of actual competence. However, this advantage also brings significant psychometric challenges. Scenario-based SJT items are inherently complex and multidimensional.

A single business problem scenario – for example, facing a new competitor – may simultaneously require competencies in strategic analysis, achievement orientation (ACH), and persuasion to be resolved effectively. Conventional analysis methods, such as CFA, which assume that each item measures only one-factor (no cross-loading), are no longer adequate. As Sorrel et al. (2016) have clearly demonstrated, approaches rooted in Classical Test Theory (CTT) often fail to capture the multidimensional nature of SJT scores. This explains why previous validation studies using CFA on IECI faced issues such as poor model fit and low convergent validity.

The main debate in the EC literature concerns its dimensional structure. The predominant perspective regarding EC is that it is multidimensional in nature. The framework developed by Man et al. (2002), for instance, categorises EC into six areas: opportunity, relationship, conceptual, organisational, strategic, and commitment. The framework developed by Spencer and Spencer (2008), which served as the basis for the development of the IECI instrument, identified 13 distinct competency dimensions. Early EFA studies of the IECI also supported this multidimensional view by identifying three distinct factors (Suryani et al., 2021).

However, the findings of Riyanti et al. (2022) suggest the possibility of a more complex structure, given their indication that a single-factor model is more suitable in CFA analysis. In response to this, editors and scholars in the field of modern literature have proposed the consideration of hierarchical models, such as second-order or bifactor models, as superior alternatives. This perspective is corroborated by numerous studies that demonstrate EC is most effectively conceptualised as a hierarchical structure. For instance, Man et al. (2008) posit the existence of intercompetency relationships that can be modelled hierarchically within their framework. Other studies, including those by Cárdenas-Gutiérrez et al. (2021), which sought to validate an entrepreneurship competency scale for secondary school students, have also found support for a more complex structure than that which can be explained by independent factors alone.

The bifactor model is of particular theoretical interest because of its assumption of a unifying EC factor (g-factor) that influences all items, as well as several specific factors that explain the unique variance of certain groups of items (Reise, 2012). This theoretical framework has the capacity to reconcile findings that initially appear contradictory. Specifically, the g-factor elucidates the rationale behind the emergence of a unidimensional model as a suitable solution, while distinct factors illuminate the factors that give rise to the identification of a multidimensional model. Consequently, the bifactor model does not compel researchers to opt for either unidimensional or multidimensional structures. Rather, it integrates both into a more comprehensive and realistic framework.

Based on the above literature synthesis, this study adopts a conceptual framework that EC is a hierarchical construct. The disparity between the EFA and CFA findings in previous IECI studies is not seen as a contradiction, but rather as an artefact of the methodological limitations of CFA in modelling complex SJT data.

Therefore, this study hypothesises that the bifactor ESEM model will provide the most theoretically valid and accurate representation of the factor structure for the short-form IECI. Specifically, this model hypothesises the existence of:

Testing these hypotheses will provide a deeper and more accurate understanding of what the IECI actually measures and how EC is structured in the context of micro, small, and medium enterprises (MSMEs) in Indonesia.

The short version of the IECI covered three dimensions, comprising 22 items. These dimensions included (1) strategic influence (INF), with a total of 7 items, (2) ACH, with a total of 7 items, and (3) opportunity recognition (OP), with a total of 8 items. The questions were presented as cases, which describe situations experienced by entrepreneurs, followed by answer choices in the form of response-based constructs, particularly behavioural tendencies (Whetzel et al., 2020). Each option represented a different level of competence ranging from 1 (least effective) to 4 (most effective). Participants were invited to select the response option that they believed to be the most probable when confronted with a comparable case or situation while managing a business. Examples of items, scores and rationales can be found in Appendix 1: Table 1-A1.

The following is an example of the question the participants were asked to select the response option: James runs a photo studio and does wedding photography. James did not have much money when he started his business. What should James do?

The survey was conducted among 512 entrepreneurs residing in major cities across the Indonesian island of Sumatra. The criteria for participation were as follows: (1) owners of MSMEs and (2) those that had been actively running their businesses for at least 3 months. The participants were selected using a non-probability method, particularly convenience sampling. The selected participants consisted of 72.65% females and 23.34% males. They were within the age range of 18–72 years (M = 36.59, standard deviation [s.d.] = 15.45), with an average length of business of 8.75 years (s.d. = 8.67). The majority of the participants had a high school diploma (51.17%), followed by university graduates (36.91%). Before participation, participants were provided informed consent after receiving details about the study’s objectives.

The data collection was conducted between July 2024 and August 2024 using both offline and online methods using Google Forms. Combining online and offline data collection methods is quite effective for expanding the scope and depth of research (Roberts & Voorpostel, 2023). All participants were first directed to review the case questions associated with each number before being inquired about their potential actions when confronted with a comparable scenario in the present business operations. Participants were also prompted to choose one from the four response options (A, B, C, or D). It was essential to acknowledge that no responses were correct or incorrect. Furthermore, participants were directed to select the option that most accurately reflected their individual circumstances. The mean duration attached to the completion of the questionnaire was 20 min.

The response options are presented as behavioural tendencies using a rating scale. Respondents are requested to evaluate the effectiveness in addressing the scenario using a 1–4 Likert scale. A high score indicates the most effective answer choice, while a lower score suggests a less effective answer choice. The subject matter experts, including entrepreneurship scholars and accomplished entrepreneurs, already predetermined the effectiveness level of each option (Suryani et al., 2021).

Prior to the main analysis, initial data screening was conducted to assess the fulfilment of statistical assumptions. Missing data were handled using the listwise deletion method (Dong & Peng, 2013). Normality was evaluated through skewness values, while multivariate outliers were identified using Mahalanobis Distance. The data screening process was carried out using Jamovi version 2.6.44 and SPSS version 23.

Given the multidimensional nature of SJT items and issues identified in previous CFA analyses, this study adopted the ESEM approach. Exploratory structural equation modelling offers greater flexibility compared to traditional CFA by allowing cross-loadings, enabling a more realistic representation of complex factor structures (Marsh et al., 2014). Exploratory structural equation modelling and bifactor ESEM modelling in this study utilised the Robust Maximum Likelihood (RML) estimator, which is considered superior to the conventional Maximum Likelihood (ML) estimator. While ML is commonly used in structural equation modelling, RML provides specific advantages when dealing with data that violate the assumption of multivariate normality. It automatically applies the Satorra-Bentler correction to standard errors and Chi-square test statistics, thereby yielding more accurate parameter estimates, particularly in the presence of skewed distributions, excess kurtosis, or outliers. A target rotation was applied in the ESEM analysis to specify approximate zero loadings on non-hypothesised factors while retaining exploratory flexibility (de Beer & van Zyl, 2019). This rotation method allows for a blend of confirmatory and exploratory approaches, making it particularly suitable for complex factor structures. All data analyses were conducted using Mplus version 7 (van Zyl & Ten Klooster, 2021).

Ethical clearance for this study was obtained from the Health Research Ethics Committee at Faculty of Nursing Universitas Airlangga with approval (reference no.: 3106-KEPK). All procedures performed in this study were conducted in accordance with ethical standards. Ethical principles, such as informed consent and confidentiality, were upheld. Prior to participation, all potential participants were provided with a detailed written information sheet outlining the study’s objectives, procedures, potential risks and benefits, the expected duration of their involvement, and their right to withdraw at any point without consequence. We provided a detailed explanation of the study and ensured all participants fully understood the information before providing their consent. Signed consent forms were obtained and securely stored for all participants. Participation in the study was voluntary, and participants could withdraw from the study at any time.

The research data consists of three factors (OP, ACH, and INF) with 512 respondents. The OP is measured by eight items, the ACH by seven items, and the INF by seven items. A total of 24 missing data points were identified, distributed across 14 rows of respondent answers, constituting 2.7% of the total number of respondents. The utilisation of complete cases, which involves the exclusion of respondents with missing data, is permissible under certain conditions. Specifically, this approach is justifiable when the proportion of missing data is less than 5%, as outlined in the study by Dong & Peng (2013). The data employed in the ensuing analysis are derived from 498 subjects. Table 1 shows a list of missing data categorised by item.

Normality checks on complete case data were performed by evaluating the skewness items. According to Hair et al. (2014), skewness values between −1 and +1 are considered very good, but values between −2 and +2 are generally still considered acceptable. Based on Table 2, all items have skewness values between –2 and +2, indicating acceptable normality, except for item INF2 with a skewness of –2.07. This may be because of the options offered by item INF2 being too biased towards one choice that is normatively considered the most appropriate, namely option D. This skewness violation is minor and still acceptable in practice, especially with an adequate sample size.

To ensure the validity of the data distribution assumptions in the bifactor ESEM analysis, a check was conducted to detect the potential existence of outliers. Outlier detection was performed using two approaches, namely Mahalanobis Distance and the total Z-score of respondents.

Firstly, Mahalanobis Distance was calculated to detect multivariate outliers, i.e. respondents who collectively showed answer patterns that were very different from the majority. The calculation was performed on 22 scale items, and the cutoff value was determined based on the Chi-square distribution with 22 degrees of freedom at a significance level of 0.001, which was 48.27. The results showed that there were 10 respondents with Mahalanobis Distance values exceeding the cutoff. This indicates that there were a small number of individuals who deviated multivariately from the general response pattern.

Secondly, to avoid losing data that could reduce the representativeness of the sample, a univariate examination was also conducted using Z-scores for the total scores of each respondent. The Z-score criterion used was ±3.29, in accordance with the general limits for detecting extreme outliers in normal distributions (Tabachnick & Fidell, 2013). The results of the examination showed that only one respondent had a Z-score outside this range, namely −3.41, indicating a very low total score. Based on the screening results, the researcher decided to retain all 498.

The ESEM modelling on these data was able to achieve convergence, whereas the bifactor ESEM modelling did not initially achieve convergence on the data of this study. The failure of convergence in the bifactor ESEM model is likely because of the complexity of the model structure, which requires estimation of multiple parameters (general and specific factors simultaneously), as well as potential model identification issues. After diagnostic evaluation, modifying the model by freeing the parameter θ (residual covariance) between items ACH6 and ACH4 successfully resulted in a converged bifactor ESEM model with adequate goodness-of-fit.

Conceptually, the residual covariance exemption between ACH4 and ACH6 has a strong theoretical justification. The two items do have similar cognitive and affective load characteristics, which both describe real crisis situations that are both financially urgent and concerning business continuity. This situation is fundamentally different from other ACH items that tend to represent internal, administrative, or procedural problems. While the other items present more mundane operational challenges, ACH4 and ACH6 display higher stress as they pose a direct existential threat to business continuity. This qualitative difference explains why the two items share a unique variance beyond the general and specific factor structures established in the model.

From a psychological response perspective, ACH4 and ACH6 have the potential to trigger more intense emotional reactions and require a faster decision-making process than other items. It is these special characteristics that likely cause the two items to require special handling in modelling. The modifications made not only resolve technical convergence issues but also improve the substantive validity of the model by accommodating the psychometric reality of the measured constructs. Path diagrams for ESEM and bifactor ESEM can be seen in Figure 1 and Figure 2.

The ESEM modelling results show the following goodness-of-fit:

Based on Table 3, the ESEM and bifactor ESEM models are generally equally good and fit with slight differences. Both models show good fit based on the root mean square error of approximation (RMSEA) values (0.031 and 0.033 < 0.05) and Tucker-Lewis Index (TLI) (0.918 and 0.906 > 0.90). The CFI index above 0.95 is very good, while both models have CFI = 0.940, which is still classified as good. The Akaike Information Criterion (AIC) value in bifactor ESEM is smaller than the AIC of ESEM, while the BIC value in bifactor ESEM model is larger than ESEM. The slight difference in several fit indices can be caused by the more complex model form in bifactor ESEM (more parameters or factors) than the ESEM model. The p-value in the Chi-square test of both models shows a significant value, indicating that the model does not fit the data. However, Hooper et al. (2008) suggest that the Chi-square statistic is a statistical significance test that is sensitive to sample size, meaning that the Chi-square statistic almost always rejects the model when the sample is large (Bentler & Bonett, 1980; Jöreskog & Sörbom, 1993). Hu and Bentler (1999) recommend CFI, TLI and RMSEA to test model fit rather than the Chi-square criterion.

Table 4 displays factor loadings and cross-loadings in ESEM and bifactor ESEM modelling. The reliability values of Omega ESEM and bifactor ESEM were calculated using factor loadings and residual variance based on Rodriguez et al. (2016) and Raykov & Marcoulides (2016), namely Equation 1:

In the ESEM model with three factors (OP, ACH, INF), most items loaded strongly onto the corresponding factors. For example, items OP3 and OP6 showed high loadings on the OP dimension, and items ACH1–ACH7 also had dominant loadings on the ACH dimension (between 0.24 and 0.65). However, there are some indications of issues, such as items OP2 and OP7 loading negatively or very weakly on the OP factor, as well as relatively high cross-loadings (e.g. OP5 also loads strongly on the ACH factor at 0.43). Overall, the internal reliability (ω) for each factor is good, particularly for ACH (ω = 0.808) and INF (ω = 0.759), while OP remains moderately reliable (ω = 0.536). The overall reliability of this model is 0.73, calculated from Equation 2:

The bifactor ESEM model has a general factor alongside three specific factors. In this model, items OP1, OP3, and OP6 still show strong loadings on the specific factor OP, while several other items (such as OP2, OP7, and OP8) show low or even negative loadings. In the ACH factor, it was found that items ACH1, ACH5, and ACH6 had negative loadings on their specific factors but showed stronger loadings on the general factor. The ACH4 had weak loadings on its specific factor and strong loadings on the general factor. ACH2, ACH3, and ACH7 had sufficient loadings on their specific factors, but the general factor loadings were still more dominant. This indicates that the items in ACH may be too general, less discriminative of ACH, and more representative of the general factor than the specific sub-dimensions. The pattern of weak item loadings on specific factors, non-significant loadings, and more dominant loadings on general factors than specific factors is commonly observed in bifactor models, where specific factors may be less distinct from certain general factors (Burns et al., 2020; Eid et al., 2017; Heinrich et al., 2023). The presence of a general factor influences all items and explains the common variance of all items, while specific factors only explain the additional variance unique to a group of items. This model remains a viable choice when it demonstrates better model fit compared to the regular ESEM model. In the INF factor, the items have adequate loadings on the specific factor, although item INF5 has a stronger loading on ACH. Further checking and refinement are needed for item INF5. The overall reliability of this model is 0.83 (very good), calculated from (Equation 3):

The ESEM model produces a cleaner and more consistent factor loadings structure for specific dimensions, supported by good internal reliability values. On the other hand, the bifactor ESEM model allows us to recognise the contribution of general factors to items, as well as revealing that some items (such as OP2 and ACH1) tend to be more general than specific.

Although this study has addressed several significant weaknesses of previous studies, some limitations need to be acknowledged to guide future research. Firstly, although the bifactor model showed model fit, the parameter analysis indicates the presence of several problematic items (e.g. cross-loading on item INF5 and weak ACH factor specificity). This indicates that further item revision and refinement are needed to improve the psychometric quality of the instrument. Secondly, the sample of this study was limited to entrepreneurs in Sumatera, Indonesia. To increase the generalisability of the findings, cross-validation studies are highly recommended on more diverse samples, both in other regions of Indonesia and in different cultural contexts globally. This will test whether this bifactor structure remains stable across different populations. Thirdly, this study focused on ESEM and bifactor ESEM. Future research could explore more sophisticated psychometric models, such as Multidimensional Item Response Theory (MIRT) or Cognitive Diagnosis Models (CDMs), to gain a deeper understanding of how each individual interacts with each answer choice on SJT items.