This study is positioned within the perspective of empathy online, namely that it is possible to express “traditional empathic characteristics such as concern and caring for others … through computer-mediated communications” (Terry and Cain, 2016, p. 1). In fact, this study focuses specifically on empathy in virtual contexts, because empathy itself is not online, but rather, occurs within individuals as they establish interpersonal relations in virtual contexts. To date, few studies have considered this specificity and have assessed empathy with adapted instruments. That is, few studies have considered the online characteristics of empathy, when studying cyberbullying. Nonetheless, some studies have already taken empathy in virtual contexts into account. For example, Carrier et al. (2015) and Manasia and Chicioreanu (2017) found that virtual empathy was positively related with empathy in face-to-face interactions, however, virtual empathy was lower for both sexes. Complementarily, Marín-López et al. (2020) found no differences between the different cyberbullying roles with respect to online empathy. Considering the scarce literature with respect to empathy in virtual contexts and cyberbullying (Marín-López et al., 2019, 2020), it is crucial to develop further research in this area of knowledge.

Assessing empathy is important to explain bystanders’ role in cyberbullying situations. For instance, Macaula and Boulton (2017) found that when comparing positive bystanders’ responses in bullying and cyberbullying, the rate of responses tended to be higher in cyberbullying. Moreover, this type of responses in both bullying and cyberbullying was positively and moderately correlated with empathy. Also, positive bystander responses tended to increase, as a result of cyberbullying severity. Another study (Schultze-Krumbholz et al., 2018) found that higher levels of both cognitive and affective empathy were associated with prosocial defending, when compared to passive bystander behavior. Notwithstanding, the research presented above considered measures of empathy without accounting for the online context.

From a phenomenological perspective, Fuchs (2014) proposed that it is not possible for empathy to occur in online contexts, since we lose our perceptual access to other individuals’ physical presence, and thus, we lose our direct empathic access to others. Accordingly, for empathy to occur, we need to perceive other individuals’ “lived body” (see Osler, 2021), and this is not possible in online “disembodied communication” (Fuchs, 2014, p. 167). Moreover, the temporal delay and the loss of perceptual queues (i.e., the perception we have is not apprehended by all our sensory capabilities) that occurs in technological mediated communication prevents us from perceiving someone’s physical and emotional experience. This was not a concern in face-to-face interactions, but do come into play in online interactions (Osler, 2021). Despite these perspectives, we believe it is possible to feel empathy in online contexts, even if individuals do not see others in person. We consider this to be true because empathic skills can be developed through the use of virtual reality (e.g., Bertrand et al., 2018), which is also different from face-to-face interactions. Moreover, although there are differences between online and offline communication, individuals tend to use other cue systems at their disposal, with the objective of promoting and detecting these cues, as well as developing relationships (Walther, 1995). Therefore, if relationships can be developed, empathy can also be possible in online interactions. In fact, through interpersonal communication online, individuals are able to infer what others might be thinking/feeling in a certain situation (Carrier et al., 2015). Nonetheless, the specificities of online contexts, may make it difficult for empathic reactions (Terry and Cain, 2016). Despite the fact that few studies have investigated empathy in virtual contexts and its specificities, it has been already proven that empathy can be experienced online. For example, Preece (1999) found that empathy online was quite common in support groups, which corroborates our position. This author discussed that the difference between synchronous and asynchronous systems impacts communication. Firstly, the pace of interaction is very different between these systems, that is, in one it is almost immediate, whereas in the other, it can take much more time (i.e., hours, days, or weeks differing from the platform). Moreover, another important difference is regarding the mode of expression, and other features that allowed nonverbal expression, whereas in the asynchronous system the primary mode is written text. It is important to highlight that this investigation is from the 1990’s, and several features of online communication have changed. However, more recent studies have found that text-type emoticons and graphic emojis are processed in a similar way to in-person facial expressions (Gantiva et al., 2019), and participants who viewed text-type emoticons exhibited face imitation mirroring (O’Neil, 2013). Therefore, we can argue that it is possible to feel empathy when interacting in virtual contexts.

Considering the importance of accounting for online features in measuring empathy, we sought new instruments on empathy that were developed according to the online context. To date, we found three instruments directly adapted from the Basic Empathy Scale (Jolliffe and Farrington, 2006), that is, the Virtual Empathy Scale (Carrier et al., 2015), the Online Empathy Questionnaire (Marín-López et al., 2019) and the Virtual Basic Empathy Scale (Manasia and Chicioreanu, 2017). Also, another instrument was adapted by García-Pérez et al. (2016) based on the Basque version (Gorostiaga et al., 2014) of the Test de Empatía Cognitiva y Afectiva (TECA) from López-Pérez et al. (2008). Additionally, Happ and Pfetsch (2015) developed the Media-Based Empathy (MBE) Scale (original name Skalazumedienbasierter Empathie) based on a pool of items according to the Interpersonal Reactivity Index (Davis, 1980) and an instrument to assess media empathy by Früh and Wünsch (2009), which included media concern, affective media empathy, cognitive media empathy, and immersion in video games, with items related to different types of media, as well as fictional and real people. Of all these instruments, only the Online Empathy Questionnaire (Marín-López et al., 2019) was used in relation to cyberbullying behavior.

Despite the valuable contributions in terms of the aforementioned instrument development and validity studies, and after a detailed analysis of the respective items, we found that the Empathy Quotient (EQ) by Baron-Cohen and Wheelwright (2004) would be appropriate to reach our objectives. Specifically, these authors defined empathy as “The drive or ability to attribute mental states to another person/animal and entails an appropriate affective response in the observer to the other person’s mental state” (Baron-Cohen and Wheelwright, 2004, p. 168). The term “quocient” derives from the Latin word “quotiens” which means “how much” or “how many” (Baron-Cohen and Wheelwright, 2004, p. 166). According to this perspective (Baron-Cohen, 2011), that if individuals only focus on their own problems or interests, they are likely to feel less empathy. In fact, when individuals feel empathy, they are able to identify what others are thinking or feeling and are able to provide an adaptive emotional response. Thus, this view of empathy entails two fundamental stages: recognition and response. Accordingly, empathy occurs when there is recognition and an adaptive response, which helps avoid hurting others and fosters prosociality.

Some studies have provided evidence that the Empathy Quotient was the third most used instrument (e.g., Ilgunaite et al., 2017) and a recent a meta-analysis by Hall and Schwartz (2019) determined that it was the second most used instrument in research. For this investigation the aim was to choose an instrument that had been widely used and already validated for several countries (e.g., Redondo and Herrero-Fernández, 2018), but that also included items assessing accurate interpersonal perception (Hall and Schwartz, 2019), since it is an important feature when assessing empathy, specifically in the virtual contexts, as is the case with this study. Moreover, we preferred to adapt the short form of this questionnaire, which had already been developed by Wakabayashi et al. (2006), and adapted for the Portuguese population (Rodrigues et al., 2011). Our study provides an important contribution, since it proposes to adapt this last version of the instrument to a younger population and for online contexts.

Considering the literature reviewed, one of the main purposes of this study is to present and evaluate a new version of the Portuguese short form of the EQ for adolescents communicating online, entitled Empathy Quotient in Virtual Contexts (EQVC).

According to some of the literature, empathy can be developed over time (Gerdes et al., 2010) and may be considered a capacity (or ability), suggesting that individuals have the potential to empathize or not (Hall and Schwartz, 2019). In fact, in some circumstances, feeling empathy requires effort and cognitive costs, and therefore, individuals may avoid feeling empathy (Cameron et al., 2019). Thus, considering the specificities of the online environment and its consequences in interpersonal relationships, we felt the need to assess empathy that occurs specifically in virtual contexts. Moreover, empathy can be situation and context specific (Cameron et al., 2019) such as in cyberbullying situations. Nonetheless, despite the widespread consensus that empathy is predetermined by circumstances (Barlińska et al., 2013), none of the empathy definitions clearly state that empathy can decrease in some situations. That is, for example, in a bullying situation, an individual might feel empathy, however, if a similar situation occurs online, the same individual might not feel the same degree of empathy. This is one of the reasons we opted to adapt an empathy instrument for online contexts, as it may be more difficult for individuals to feel empathy toward others in these digital environments (Pfetsch, 2017).

For all the studies presented, authorization to complete the questionnaires in the online context was granted by the Ministry of Education of Portugal, the Portuguese National Commission of Data Protection, the Deontology Committee of the researchers’ institution, the schools’ boards of directors, the teachers, the parents and the adolescents themselves. Before the completion of the questionnaires, students were informed that psychological assistance was available if needed, considering the sensitivity of the subject in study. Additionally, students were informed that all information collected was anonymous and confidential and that they could quit at any time if they were not comfortable. This study was not preregistered. Further information regarding the initial adaptation and construction of the instruments, all items (Portuguese version), and additional information are available in the Supplementary material.

Thirty-four 9th grade students (Mage = 14.29, SD = 0.72, 53% female) participated in an in-depth semi-structured interview with fictitious scenarios.

A qualitative study was conducted to explore adolescents’ MD in cyberbullying situations. In-depth semi-structured interviews with scenarios were conducted and verbatim transcribed. Later, we performed a content analysis with a mixed approach (deductive/inductive), based on the Social Cognitive Theory (Bandura, 2002). The coding units we established were adolescents’ written verbalizations with meaning (Amado et al., 2014), summing a total of 396 verbalizations, which were analyzed. We performed an initial phase, where categories were created, and a re-checking phase, where a set of verbalizations were analyzed by two other researchers and adjustments were made to the operational definition of the categories. Finally, two independent coders rated the data. Inter-rater reliability was excellent, as mentioned in the literature (McGraw and Wong, 1996), with an ICC = 0.99, with a 95% confident interval = 0.99–0.99. From this analysis, the categorization process went beyond the Social Cognitive Theory. That is, several categories of MD mechanisms emerged from the analysis, as well as other attributions (Figure 1), both regarding aggressors’ and bystanders’ behavior from the scenarios (see Francisco et al., 2022 for a detailed description).

It is important to highlight that we considered MD as a process, since several mechanisms tend to be used before the aggression, during the behavior and after as consequents of the behavior, as presented in Figure 1. Thus, considering this novel approach, the qualitative data was the starting point of the development of the PMDCI because we sought to develop an instrument that could capture adolescents’ beliefs and perspective of this phenomenon as accurately as possible. Hence, from the categories that emerged from the content analysis, we created the items for the PMDCI. All the procedures regarding scale development can be found in the Supplementary Appendix A.2.

A total of 234 students participated in the exploratory factor analysis (EFA) study (M_age = 13.24; SD = 1.18; 51.7% girls), 35.9% of whom were in the 7th grade, 25.6% were in the 8th grade and 38.5% were in the 9th grade (Supplementary Appendix A.4). All 234 participated in the EFA of the EQVC and 230 participated in the EFA of the PMDCI.

The new created version of EQ (EQVC) and the new developed instrument (PMDCI) were administered on-line in a classroom context, individually with the guidance of an educational psychologist. Students took approximately 40 minutes to complete both questionnaires. After the data gathering, EFA was conducted with FACTOR 10.10.02 (Ferrando and Lorenzo-Seva, 2017) to understand the factorial structure of both instruments. Specifically, we intended to explore if the EQVC yielded the same structure of the EQ-short form (Portuguese version), or if considering the new context and different population, the structure of the instrument would change. Regarding the PMDCI, since it was developed considering the four loci (i.e., Behavior, Agency, Outcome and Recipient) and the respective MD mechanisms, we intended to evaluate the best way to validate the instrument. That is, we were interested in understanding if the instrument should be considered as a single scale, or if it should be regarded as a questionnaire with different scales (i.e., one scale for each locus) involving the distinct locus of the MD.

In order to uncover the underlying structure of the EQVC, we performed an EFA (see Supplementary Appendix A.5 for more details). We present the correlations and descriptive statistics of all items, including skewness and kurtosis (Supplementary Appendix Table A.7). Regarding univariate normality, all variables were approximately normally distributed according to the literature, with skewness absolute values less than 2 and kurtosis absolute values less than 2 (George and Mallery, 2016). We also analyzed multivariate normality accordingly to Bollen and Long (1993), where multivariate normality is accepted if Mardia’s coefficient is lower than P (P + 2), considering P the number of observed variables. Considering that the EQVC presented 22 observed variables, Mardia’s coefficient for skewness of 78.41 < 22(22 + 2) = 528 and for kurtosis is 605.06 > 22(22 + 2) = 528. Moreover, as for the correlation matrix, we used polychoric correlations (Muthén and Kaplan, 1985; Brown, 2006) (Supplementary Appendix Table A.7). Furthermore, before proceeding to the EFA results, Kaiser–Meyer–Olkin (KMO) and Bartlett Sphericity were assessed. As for KMO it was 0.89 revealing sampling adequacy, and Bartlett Sphericity test was χ²(₂₃₁) = 2543.4 (p < 0.001), which indicated that we could proceed with factor analysis. In order to retain the appropriate number of factors we used Horn Parallel analyses (O’Connor, 2000). In the FACTOR program (Ferrando and Lorenzo-Seva, 2017) the Optimal Implementation of Parallel Analysis (Timmerman and Lorenzo-Seva, 2011) suggested that two factors should be extracted. We used the Unweighted Least Squares (ULS) method for factor extraction. Specifically, Robust Factor Analysis based on the Robust Unweighted Least Squares (RULS) was used to fit the factor solution. Robust Promin Rotation was used to achieve factor simplicity (Lorenzo-Seva and Ferrando, 2019). As according to the literature (Bandalos and Finney, 2010), we took into account all items with structure coefficients superior to 0.30, and no items revealed loadings greater than 0.40 on the two factors (Supplementary Appendix Table A.2). According to the literature (McDonald, 1999), goodness-of-fit values (GFI = 0.98) and (AGFI = 0.98), residuals statistics (RMSR = 0.06) were good. The EQVC presented 48% of the explained variance. We then compared the bi-factorial model to the unifactorial model (Supplementary Appendix A.6 and Supplementary Appendix Table A.9). Considering the results, we decided to keep the bi-factorial model since the percentage of explained variance was higher. Regarding reliability, McDonald’s Omega (Hayes and Coutts, 2020) was also assessed for both factors: factor 1 presented ω = 0.68, 95% CI [0.58, 0.74], showing acceptable reliability, and factor 2 presented ω = 0.91, 95% CI [0.88, 0.93], with excellent reliability (Supplementary Appendix A.6).

Later, we conducted a Multidimensional Normal-ogive Graded Response Model (Reckase, 1985), whose parameters can be seen in Supplementary Appendix Table A.8, as well as the item loadings. This model presents a discrimination parameter (a), which is important in the preliminary adjustment of questionnaires and item selection (Matteucci and Stracqualursi, 2006). Most items revealed moderate item discrimination, however, items 1, 4, and 5 revealed low item discrimination, presenting values between 0.424 and 0.586, as indicated in the literature (Baker, 2001). Item discrimination reveals how well an item differentiates individuals scoring high and low on the latent ability being measured (Depaoli et al., 2018). Then, we performed the analysis again without items 1, 4, and 5 to see how the model change. Lastly, we had some participants with Weighted Mean-Squared Index larger than 2.0 (Ferrando et al., 2016), thus, these participants were removed and the analysis was performed again. Table 1 shows a comparison between 4 proposed EFA models: (1) with all participants and all items, (2) with all participants and without items 1, 4 and 5, (3) without infit/outfit participants and all items and (4) without infit/outfit participants and without items 1, 4, and 5.

The elimination of participants improved the % of explained variance (from 48 to 51%); the RMSEA and the RMSR were the fit indices that had better improvement. Moreover, the elimination of the 3 items improved the model essentially in terms of % explained variance (from 48 to 55%), and also the same indices as described above. Considering these improvements, we conducted Confirmatory Factor Analysis (CFA) with this structure.

With the aim of assessing the structure of the PMDCI, we performed an EFA with data from 230 participants to the 5 scales included in the questionnaire (4 scales regarding Loci of MD and 1 scale regarding Attributions for the cyberbullying behavior), considering the Bystanders’ perspective (i.e., Bystander scale). We present the correlations and descriptive statistics of all items, including skewness and kurtosis (Supplementary Appendix Table A.10).

Regarding univariate normality, most of the variables were normally distributed, with skewness absolute values less than 2 (Bollen and Long, 1993), with the exception of the items from the Attribution Scale. Regarding kurtosis, all variables had less than 5 in absolute value. With respect to multivariate normality, according to Bollen and Long (1993), it is accepted if Mardia’s coefficient is lower than P(P + 2), considering P the number of observed variables. Moreover, as for the correlation matrix, we used polychoric correlations (Muthén and Kaplan, 1985; Brown, 2006). Furthermore, before proceeding to the EFA, Kaiser–Meyer–Olkin (KMO) and Bartlett Sphericity were assessed (Supplementary Appendix Table A.11). All scales had high KMO which revealed sampling adequacy, as well as a significant Bartlett Sphericity test, which indicates that we could proceed with factor analysis.

In order to retain the appropriate number of factors, we followed the same procedures used for the EQVC. Our EFA suggested that a single factor should be extracted of each scale of the PMDCI. As for the factor structure (Supplementary Appendix Table A.12), we took into account all items with structure coefficients superior than 0.30 (Bandalos and Finney, 2010). Regarding reliability, all scales reveal good internal consistency values (Supplementary Appendix Table A.11).

Regarding Explained Variance, all scales were above the minimum range, as according to the literature (Hair et al., 2014). As for the model fit indices, all scales presented satisfactory values of goodness-of-fit values and residuals statistics (Supplementary Appendix Table A.11), according to the literature (McDonald, 1999).

Later, we conducted a Multidimensional Normal-ogive Graded Response Model for unifactorial models (Samejima, 1969), whose parameters can be seen in Supplementary Appendix Table A.12, as well as the item loadings, for all the 5 scales. Considering the discrimination parameter values, it was concluded that all items from all scales revealed good discrimination (Baker, 2001), indicating that there was no need to remove items. Thus, we conducted CFA with the original structure of all 5 scales.

For the CFA, our sample consisted of 345 students (Mage = 13.13; SD = 1.27; 51% boys), 40.5% of whom were in the 7th grade, 27.1% in the 8th grade and 32.4% in the 9th grade. Most students were Portuguese (85.8%). All 345 participated in the CFA of the EQVC and 342 participated in the CFA of the PMDCI, as well as in the correlational study.

Before proceeding to the CFA, univariate and multivariate normality of all scales were evaluated and the distributions were considered non-normal. This is consistent with the literature (Yuan and Bentler, 1998), since non-normality is prevalent in real data (Blanca et al., 2013) and it would dictate the possibilities in the data analysis, because structural equation modeling assumes the normality of latent variables (Bollen, 1989). Thus, several estimation methods were investigated and analyzed considering the nature of our data (for a detailed description see Supplementary Appendix A.7).

With this in mind, we attempted to analyze several estimation methods that could be applied to our data. As a way of summarizing our results, we only mentioned the ULS parameters in the text, as advised by Bollen (1989) because it does not make distributional assumptions regarding the observed variables. Moreover, the other estimation procedures are presented in the Supplementary Material and referred to when they are considered relevant.

For the CFA of the EQVC and PMDCI we used IBM, SPSS AMOS 24.0 (Arbuckle, 2019) and the lavaan package (Rosseel, 2012) in R Project (R Core Team, 2020). ULS and ML with Bollen-Stine Bootstrapping were conducted in AMOS, and ML with Satorra-Bentler correction and WLSMV were conducted using the lavaan package in R software. Several Fit Indices will be presented according to the different estimation methods (Supplementary Appendix A.7), and organized by their main classification. Considering that the covariance matrix might not be as asymptotically distributed as chi-square with the ULS method (Bollen, 1989), several statistics are not reported, such as the chi-square test and other fit indexes based on this statistic. Instead, we used the following fit indexes to ascertain the tested models: GFI, AGFI and PGFI (more information regarding Fit Indices are in the Supplementary Appendix A.9).

As for the correlational study, Spearman correlation coefficients were used to examine the relationship between the variables.

We examined the multivariate normality and considering that the critical ratio for both skewness and kurtosis was outside the interval of [−1.96, +1.96] (Byrne, 2010), some procedures were made to account for the non-normal distribution of the data. Thus, first several multivariate outliers were removed, and multivariate normality was assessed again. However, the distribution was still non-normal.

We tested various possible models so as to confirm the initial structure of the EQVC suggested by the EFA with confirmatory factor analysis. We attempted to test a model with all participants and no covariances (model 1), a model without outliers and no covariances (model 2) and a model without outliers and with covariances (Supplementary Appendix Table A.13 and Supplementary Appendix Figure A.1) between the error terms (model 3). From the results presented, we chose model 3, which according to the literature (Jöreskog and Sörbom, 1984; Cole, 1987; Blunch, 2008) presented good reference values [χ²(149) = 151.626, χ²/df = 0.793, GFI = 0.969, AGFI = 0.961, SRMR = 0.054, NFI = 0.930, PGFI = 0.759, PNFI = 0.810].

Despite the good fit of the model, several relationships between each factor and corresponding items were lower than the cut-off value of 0.5, as suggested in the literature (e.g., Bandalos and Finney, 2010). All unstandardized path coefficients¹ were significant at p < 0.05, with the exception of item 3, which was equal to 0.05 (Supplementary Appendix Figure A.1). Moreover, the construct reliability scores were low for the Difficulties in Empathizing and higher than 0.80 (Hair et al., 2014) for the Self-efficacy regarding Empathy (Table 2). Thus, the second factor presented good construct reliability; however, the first, which only has 4 items, revealed low reliability. Convergent validity was low for both factors since the Average Variance Extracted (AVE) scores were lower than 0.50 (Henseler et al., 2009). Nonetheless, the Average Shared Variance scores below the AVE scores (Hair et al., 2014) indicated good discriminant validity of both factors. Additionally, the simplified model also presented lower Modified Expected Cross-Validation Index (MECVI), indicating that it has better validity in the population we are studying (Marôco, 2014).

The bi-factorial structure that we found could be the result of reverse coding (Woods, 2006). Even though the factor Difficulties in Empathizing revealed low construct reliability, we decided to keep the bi-factorial structure, since this is a pilot study of an adapted instrument to online contexts, which is quite different from the offline environment. Nonetheless, further studies are required to better assess the EQVC, and to better understand if the bi-factorial structure results from reverse coding, or from the characteristics of online contexts.

In order to confirm the initial structure suggested by the EFA of the scales from the PMDCI, various possible models were tested for the 5 scales (Supplementary Appendix Tables A.14–A.18). Hence, we attempted to test a model with all participants and no covariances (model 1), a model without outliers and no covariances (model 2) and a model without outliers and with covariances between the error terms (model 3).

Considering the Locus Behavior scale, the best model (model 3) presents several covariances between items (Supplementary Appendix Table A.14 and Supplementary Appendix Figure A.2). According to the literature (Jöreskog and Sörbom, 1984; Cole, 1987; Blunch, 2008), the factor model we opted for presented good reference values [χ²(25) = 9.638, χ²/df = 0.386, GFI = 0.991, AGFI =0.983, SRMR = 0.051, NFI = 0.975, PGFI =0.550, PNFI = 0.677].

As for the Locus Agency scale, model 3 which presents the covariances between two error terms of items (Supplementary Appendix Table A.15 and Supplementary Appendix Figure A.3) presented good reference values [χ²(8) = 1.233, χ²/df = 0.154, GFI = 0.997, AGFI =0.992, SRMR = 0.032, NFI = 0.987, PGFI =0.380, PNFI = 0.526], as according to the literature (Jöreskog and Sörbom, 1984; Cole, 1987; Blunch, 2008).

As for the Locus Outcome scale, we only assessed 2 models, since the Modification Indices did not indicate the need to covariate error terms of items (Supplementary Appendix Table A.16 and Supplementary Appendix Figure A.4), thus we only had model 1 with all participants, and model 2 without outliers. Model 2 presented good values [χ²(9) = 0.904, χ²/df = 0.100, GFI = 0.997, AGFI =0.993, SRMR = 0.028, NFI = 0.993, PGFI =0.427, PNFI = 0.596], as according to the literature (Jöreskog and Sörbom, 1984; Cole, 1987; Blunch, 2008). Nonetheless, Model 1 presented better validity in the population of study, since it has lower MECVI (Marôco, 2014).

Considering the Locus Recipient scale, model 3 presented the covariances between four error terms (Supplementary Appendix Table A.17 and Supplementary Appendix Figure A.5). According to the literature (Jöreskog and Sörbom, 1984; Cole, 1987; Blunch, 2008), the factor model we opted for presented good reference values [χ²(7) = 6.366, χ²/df = 0.909, GFI = 0.993, AGFI =0.979, SRMR = 0.042, NFI = 0.979, PGFI =0.331, PNFI = 0.457].

Finally, for the Attribution scale, model 3 presented the covariances between two error terms (Supplementary Appendix Table A.18 and Supplementary Appendix Figure A.6) revealed good reference values [χ²(26) = 1.198, χ²/df = 0.046, GFI = 0.992, AGFI =0.987, SRMR = 0.05, NFI =0.987, PGFI =0.573, PNFI = 0.713], according to the literature (Jöreskog and Sörbom, 1984; Cole, 1987; Blunch, 2008).

Despite the good fit of the selected models, PGFI did not present good values for all scales. It was below the cutoff of 0.6 (Blunch, 2008) in the Locus Agency, Outcome and Recipient and near the cutoff in the Locus Behavior and Attribution scale. Nonetheless, the other estimation procedures revealed good fit indices, supporting our model choice, as can be seen by comparing RMSEA and AIC. Also, all models chosen presented lower MECVI (Marôco, 2014), indicating better validity in the population of study, except for the Locus Outcome scale.

As can be seen in Supplementary Appendix Figures A.2–A.6, several relationships between each factor and corresponding items were lower than the cut-off value of 0.5 (Bandalos and Finney, 2010). Nevertheless, all unstandardized path coefficients were significant at p < 0.05. Moreover, the composite reliability scores ranged from 0.62 to 0.88, revealing medium to high construct reliability (Hair et al., 2014), as can be seen in Table 3. However, the AVE was low for Locus Behavior, Agency and Recipient and approximate of the 0.50 as indicated in the literature (Henseler et al., 2009) for Locus Outcome and Attributions. Thus, for the former scales, convergent validity was low, and for the later, convergent validity was almost adequate. Nonetheless, the Average Shared Variance (ASV) scores below the AVE scores (Hair et al., 2014) indicated good discriminant validity for all scales, except for Locus Outcome, of which the ASV could not be calculated, since this scale did not have correlation between error terms.

In this investigation, we found that empathy in online contexts appeared to be divided in two factors (i.e., Difficulties in Empathizing and Self-efficacy regarding Empathy), and that Moral Disengagement with respect to cyberbullying situations was composed of 4 different loci (i.e., Behavior, Agency, Outcome and Recipient) and Attributions (for the definition of each scale/variable see Supplementary Appendix A.10). Thus, regarding the first research question, Difficulties in Empathizing was negatively and significantly correlated with Attributions (r = −0.135, p < 0.05) and 3 Locus of MD [Agent (r = −0.169, p < 0.01), Outcome (r = −0.218, p < 0.01), and Recipient (r = −0.142, p < 0.01)]. That is, the more difficulty participants had in empathizing, the less attributions and the three different Loci were used. However, with respect to self-efficacy in empathizing, it was not statistically significantly correlated with any variable. Considering the second research question, difficulties in empathizing was negatively and significantly correlated with gender (r = −0.114, p < 0.05), meaning that girls tended to have more difficulties in empathizing, and boys tended to have less. Additionally, gender was positively and significantly correlated with Attributions (r = 0.223, p < 0.01), Locus of Behavior (r = 0.174, p < 0.01), Locus of Agency (r = 0.226, p < 0.01), Locus of Outcome (r = 0.136, p < 0.05) and Locus of Recipient (r = 0.196, p < 0.01). This means that boys tended to use more attributions and MD Loci with regards to cyberbullying. Correlations can be found in Table 4.