Although the existence and importance of intelligences beyond memory and problem solving (mostly defined as intelligence and tested by an IQ test) has long been recognised, it was not until relatively recently that serious efforts were made to define emotional intelligence (EI) (Goleman, 1995; Salovey & Mayer, 1990). However, in the last two decades various research processes and efforts have been made to define and measure the impact of emotional intelligence on organisational outcomes (Cherniss, 2000; Cooper, & Sawaf, 1997; Goleman, 1995, 1998; Mayer & Salovey, 1997; Mayer, Salovey & Caruso, 2000; Ryback, 1998; Weisinger, 1998).

In the process and scientific debate to propose an operational model of emotional intelligence, two markedly different, yet related, models of EI have been suggested (Ciarrochi, Chan & Caputi, 2001). The first model is an ‘ability model’, which combines emotion with intelligence, and the second is what is termed a ‘mixed model’, which combines traits with social behaviours and competencies. The ability model has largely evolved from Salovey and Mayer's (1990) original definition of EI as a traditional intelligence and has attracted considerable research attention (Mayer & Salovey, 1997; Salovey & Mayer, 1990). Salovey and Mayer originally defined EI as the ability to deal with one's own emotions and those of others to benefit from problem solving and decision-making. The Mayer Salovey Caruso Emotional Intelligence Test (MSCEIT; Mayer, Salovey & Caruso, 2002) is the most widely used ability measure of EI.

The mixed model (sometimes called the trait model), an approach generally embraced and advocated by Goleman (1995; 1998), arose largely from the work of Bar-On (1997). Bar-On (1997, p. 385) argues that emotional and social intelligence is a ‘multifactorial array of interrelated emotional, personal, and social abilities that influence our overall ability to actively and effectively cope with daily demands and pressures’. The mixed model framework is generally comprised of self-report instruments that measure a combination of cognitive, personality and affective attributes (Papadogiannis, Logan & Sitarenios, 2009). Examples of such instruments include the emotional quotient inventory (EQI; Bar-On, 1997), the Schutte self-report emotional intelligence test (Schutte et al., 1998) and the emotional competence inventory (Sala, 2002).

The last decade has seen a robust debate develop regarding the appropriateness and efficacy of the two models, a debate to which academic literature has greatly contributed (e.g. Emmerling & Goleman, 2003). Advocates of the ability model argue that their models and measurement instruments are scientifically derived, and psychometrically independent, from other measures of personality (Mayer & Salovey, 1997; Salovey & Mayer, 1990). On the other hand, advocates of the mixed model argue that their approach is highly correlated with desired organisational outcomes, and is of significant value to organisational and leadership development (Emmerling & Goleman, 2003; Goleman, 1995, 1998).

In addition to the debate surrounding the appropriateness of various models of emotional intelligence, a second debate has focused on the operationalisation (i.e. measurement) of the construct itself. Various authors have argued that it is better to use performance-based measures of emotional intelligence when it is conceptualised as an ability (Ashkanasy & Daus, 2005; Joseph & Newman, 2010; O'Boyle et al., 2011). In contrast, self-report measures of emotional intelligence, usually associated with the mixed model, seem to be contextualised for organisational research (Smollan & Parry, 2011, p. 438). In addition, such measures are easier and faster when collecting data in comparison to measures of ability. More specifically, it has been found that ability-based measures of emotional intelligence seem to be more theoretically sound, yet suffer from poor predictive validity with regard to job performance. In contrast, mixed-model measures of emotional intelligence show strong relationships with job performance, but may suffer from questionable theoretical value (Joseph & Newman, 2010, p. 71). The following quote summarises the current state of the debate: ‘although EI is a wildly popular tool in organizations, organizational science has yet to answer many theoretical, measurement, and validity questions surrounding the construct’ (Joseph & Newman, 2010, p. 54).

Table 1 present the goodness-of-fit statistics based on Sample 1 (n = 470) associated with the 40-item and 22-item versions of the Rahim EQI.

From Table 1 it is evident that, when using the 22-item version of the Rahim EQI, the original five-factor structure fits the data best (especially when examining the values associated with the CFI, RMSEA, and SRMR). It is also clear that this model has the lowest value associated with AIC.

The goodness-of-fit statistics based on Sample 2 (n = 308) associated with the 30-item and 22-item versions of the Rahim EQI are reported in Table 2.

On the basis of the information reported in Table 2, it appears the original five-factor structure fits the data best (especially when examining the values associated with the AIC, CFI, RMSEA, and SRMR) when using the 22-item version of the Rahim EQI. It is also clear that the bifactor model has acceptable fit. This may suggest that the Rahim EQI has a strong general factor (emotional intelligence), with five group factors (self-awareness, self-regulation, motivation, empathy and social skills).

Given the fact that the original structure of the Rahim EQI, using the 22-item version, provided the best fit to the data, the intercorrelations amongst the five dimensions (of this version of the questionnaire) are reported in Table 3.

From Table 3 it is clear that all the dimensions have high correlations (strong relationships). These high correlations may also lend support to the possibility that the Rahim EQI only measures a strong general factor. Additional support for this will be provided in Table 7.

The reliability estimates for each of the three versions of the Rahim EQI are presented in Table 4 and Table 5.

On the basis of the results reported in Table 4 and Table 5, it is evident that the five dimensions measured by the Rahim EQI have acceptable reliabilities, all exceeding 0.80. It should be noted that the reliabilities obtained from Sample 2 (n = 308) are slightly higher than those obtained from Sample 1 (n = 470).

Table 6 provides the standardised loadings and standard errors associated with the 22-item version of the Rahim EQI.

It is clear that all the items have significant loadings on their respective dimensions.

Table 7 provides the standardised loadings of the bifactor model associated with the 22-item version of the Rahim EQI.

From Table 7 it is evident that all of items have higher loadings on the general factor than on the group factors (i.e. dimensions of the Rahim EQI). This is indicative of a strong general factor (Reise, Moore & Haviland, 2010). Cronbach's alpha may be misleading when used to interpret how well a measure reflects a single construct, should multidimensionality be assumed (Cortina, as cited by Reise et al., 2010, p. 555). However, the omega hierarchical coefficient is a statistic that is based on a bifactor model representation (Reise et al., 2010). The omega hierarchical coefficient was 0.90.

When comparing the goodness-of-fit statistics associated with each of the three versions (22-item, 30-item and 40-item) of the Rahim EQI, the following conclusions can be drawn.

The only measurement model that pointed to acceptable levels of fit was the 22-item version of the Rahim EQI. When looking at the goodness-of-fit statistics of the current measurement model, it outperformed both the 30-item and 40-item versions of the Rahim EQI. The results obtained in this study were also better than those reported by Rahim et al. (2002). The measurement model associated with the 22 items produced a poor-fitting model in the South African sample (n = 84). The reliabilities associated with the five dimensions, in both samples (n = 470 and n = 308), were acceptable, ranging between 0.84 and 0.94. These reliabilities were better than those reported for the South African sample (Rahim et al., 2002). Given the findings of the current study, it seems that the 22-item version of the Rahim EQI represents the most parsimonious operationalisation of the construct requiring fewer parameters to be estimated resulting in a less complex measurement model. With fewer variables, such models have the potential for better fit (Hair et al., 2006, p. 826). Hence, it can be concluded that the 22-item version of the Rahim EQI has acceptable psychometric properties.

Both the 30-item and 40-item versions of the Rahim EQI produced acceptable levels of fit when taking into consideration both the RMSEA (0.066 and 0.062) and SRMR (0.037 and 0.052). In contrast, the CFI (0.912 and 0.873) indicated poor fit. The reliabilities associated with the five dimensions in both samples (n = 470 and n = 308) all exceeded 0.90. In a study conducted by Schlechter (2005), the psychometric properties of the 40-item version of the Rahim EQI were evaluated. Schlechter (2005, p. 197) reported very high reliabilities associated with each of the five dimensions: self-awareness (α = 0.92), self-regulation (α = 0.95), motivation (α = 0.84), empathy (α = 0.93) and social skills (α = 0.91). It is clear that these results are comparable to the reliabilities reported in the current study. In addition, the goodness-of-fit statistics associated with the original five-factor structure are comparable to those obtained in the current study. Specifically, the current study found slightly better fit with regard to both the RMSEA (0.062 versus 0.1085) and the SRMR (0.052 versus 0.069), than that reported by Schlechter (2005, p. 199). In contrast, the latter study reported better fit when considering CFI (0.9678 versus 0.873). The current study concurs with Schlechter's (2005, pp. 200–201) conclusion that the original five-dimensional structure, using the 40-item version of the Rahim EQI, fitted the data poorly. Although the 30-item version of the Rahim EQI exhibited better fit than the 40-item version, there is still room for improvement.

When comparing the psychometric properties of the Rahim EQI (22 items) with other similar instruments, the results are encouraging. Schutte et al. (1998) claim that their instrument measures a general EI factor. Petrides and Furnham (2000) criticise this assumption and continued to fit a unidimensional model. The latter model did not fit their data well (RMSEA = 0.105, GFI = 0.69, CFI = 0.51). They continued with exploratory factor analysis that suggested a four-dimensional model. Unfortunately, Petrides and Furnham did not provide any fit statistics. In a South African sample, Jonker and Vosloo (2008) reported a four-dimensional structure associated with the instrument developed by Schutte et al. (1998). However, no fit statistics were calculated for the observed measurement model.

More emotionally intelligent employees yield higher work performance (O'Boyle, Humphrey, Pollack, Hawver & Story, 2011), as well as other desired workplace outcomes such as work engagement and flourishing (Schutte & Loi, 2014). Various studies have indicated that emotional intelligence can be increased through training (Schutte & Maouff, 2013). The Rahim EQI can therefore be used as a pre-measure and post measure to the EI training intervention, in order to determine to what extent the training had an impact on individuals’ levels of self-perceived EI.

The current study agrees with Rahim et al. (2002, p. 317) that ‘it is evident that continued research is necessary to improve the EQI’. Although the current study provides some insight into the structure of the Rahim EQI, there are some suggestions to improve future investigations.

The current study used data collected from two samples that were from different industries, with Sample 1 (excluding medical) requiring physical attributes to be honed and Sample 2 (excluding medical) requiring social attributes to be honed. Also, the managerial levels within the organisations were not considered in the measurement. Both these sampling issues might have had an influence on the results. Samples that are statistically similar in terms of the attributes mentioned (e.g. managerial levels and industry) are advised for future researchers when planning similar investigations.

It is also advisable to investigate the factorial invariance of the Rahim EQI in different South African sub-samples. This will enable researchers to determine whether culture plays a role in emotional intelligence. Finally, there are more advanced statistical techniques, such as Rasch analysis, that should be used to supplement the results obtained in the current study. Rasch analysis is the preferred analysis technique to investigate unidimensionality of a measuring instrument (Bond & Fox, 2007). More specifically, ‘the unidimensionality assumption states that the observations on the manifest variables (e.g. the items) are solely a function of a single continuous latent person variable’ (De Ayala, 2009, p. 20). The Rasch model will allow researchers to identify those items that can be used to measure an individual's standing on the latent variable (emotional intelligence). Those items that may lead to misfit can be removed to improve the accuracy of the Rahim EQI as a measure of emotional intelligence. Therefore, the unidimensionality assumption should be tested using the Rasch model in order to investigate the validity of the suggestion that emotional intelligence can be treated as a unidimensional construct.

The psychometric properties of the 22-item version of the Rahim EQI seem to be acceptable. The original five-factor structure of the emotional intelligence construct fitted the data well, as evidenced by the overall goodness-of-fit statistics. The reliabilities associated with all the five dimensions (self-awareness, self-regulation, motivation, empathy and social skills) were high. However, the results obtained from the bifactor model cast doubt about the usefulness of these subscales. The bifactor model suggested that there was a strong general factor and that the five subscales added very little understanding to the construct over and above the general factor. Researchers should therefore be cautious when using these subscales, as the overlap amongst them is so large that treating them as unique and separate dimensions may be overly optimistic.