Data-Driven Analysis Exploring the Development of Empathy in an Iranian Context

Abstract

In the present study, we explored the best regression models that explain the developmental path of dispositional empathy among Iranian participants using Bayesian Model Averaging (BMA). BMA, a data-driven analysis method, was employed to identify the most likely candidate regression models given the collected data. We reported the best regression model for each dependent variable and different components of applied questionnaires by evaluating and comparing multiple model fit indicators—Akaike information criterion, Bayesian information criterion, adjusted R², Bayes Factor model, and leave-one-out cross-validation root-mean-square error—among candidate regression models identified by BMA. We discussed the theoretical implications of the findings regarding factors associated with empathy development and the methodological implications of using data-driven analysis in the field.

1. Introduction

One of the most important questions in the field of social psychology is determining how a construct changes over the life span. Empathy is not an exception in this regard. Among all the proposed theories on empathy development (see [1]), Hoffman’s developmental model [2] is the most comprehensive conceptual framework [3,4,5,6]. This model is based on the development of self–other distinction- the prerequisite of empathic behavior- Accordingly, the empathic feeling starts from birth and changes as one grows until one acquires the skills required to physically and psychologically distinguish oneself from others.

1.1. Empathy Development in Childhood

Although Hoffman’s model has an inclusive and sequential nature in considering children’s cognitive and emotional development, this theory can only explain empathy development during infancy and early childhood. To explore empathy development beyond the period, Dymond et al. [7] conducted seminal cross-sectional and experimental studies on elementary-school children. They found a linear positive relationship between age and empathy. Similar to Hoffman [2], Dymond et al. [7] argued that higher empathy in older children occurs because of their interest in socializing and familiarity with their own internal psychological states. However, when affective and cognitive empathy [8,9] were measured separately, the pattern was reversed, meaning lower affective empathy was observed in 10–12-year-old children [10].

1.2. Empathy Development in Adolescence and Adulthood

Dymond et al.’s [7] model was also limited to a certain developmental stage and failed to explain empathy development beyond childhood. Therefore, other researchers conducted studies based on socio-cognitive [11], neuropsychological [12], dynamic integration [13], and neurodevelopmental [14] theoretical frameworks to overcome the aforementioned limitations.

According to socio-cognitive theory, there is a rapid growth of empathic understanding because of emotional maturation. People experience a broad range of emotional contexts throughout their lifespans, helping them to acquire extensive knowledge regarding emotional cues and to strengthen social bonds. Consequently, people develop an ability to respond empathically toward others’ emotional states as they get older [11,15].

The neurodevelopmental perspective of empathy development is built upon the development of the brain and self-regulation. As people get older, their prefrontal cortex, a brain region involved in cognitive behavior, such as decision-making and response inhibition [16], along with more sophisticated cognitive-based skills, such as social interaction and context evaluation, become more developed. Therefore, they acquire the knowledge/skills required to behave more altruistically. Higher emotional regulation in older people could also be related to the brain development process over time. Brain regions associated with emotion generation (e.g., amygdala) show a faster growth compared with cognition-related brain regions, such as the prefrontal cortex [17]. In support of this view, Decety and Michalska [14] recorded brain activity of 7 to 40-year-old participants and found a significant shift in generating empathic responses as age increased. Specifically, younger participants experienced stronger empathic distress and emotionally perceived and judged painful stimulus with higher intensity. On the contrary, due to a more fully developed brain, older participants were able to regulate their distress and employ cognitive processes for evaluating the observed situation, which was accompanied by empathic response.

Unlike socio-cognitive and neurodevelopmental approaches, the neuropsychological perspective proposes that empathic understanding declines over time. Accordingly, aging causes decreased brain activities, especially in the frontal and temporal lobes, leading to the loss of abilities essential for emotional understanding [18]. Moreover, older people are less empathic because they are less likely to participate in social activities compared to younger people. Social isolation can lead to fewer social interactions and make understating others’ perspectives and adjusting behaviors based on such understanding, a challenging task (e.g., [19,20]). However, decreased empathy among older people might not be due to their declined mental capacities, but might be due to their more developed wisdom based on accumulated life experiences. They might not react empathically in every situation as they try to wisely keep the balance between different virtues [21].

In support of the neuropsychological perspective, Phillips et al. [12] conducted a comparative study in which participants would report on their emotional understanding using self-reporting questionnaires and were verbally and non-verbally assessed for basic and complex emotional abilities (i.e., emotional intelligence; identifying negative and positive emotions from the face, eyes, and stories; emotional empathy). Accordingly, all aspects of emotion, except for emotional empathy and theory of mind, did not differ significantly across age groups. Older participants showed significantly lower performance in these two dimensions compared to the younger participants. They were also less accurate in decoding negative emotions (anger and sadness). However, after controlling for participants’ education there was no significant effect regarding age, suggesting that lower empathy in the elderly could be associated with their educational level rather than their age. Such trends could be due to the fact that education may help people to be more open-minded and aware of others’ perspectives, more aware of their own mental and cognitive states, and better identify the emotional states of people in situation causing suffering [12,15].

One methodological limitation in the mentioned study is that although the authors used brain aging to explain their results, their study was behavioral. Therefore, lower empathy in older people cannot be directly linked to neural-level changes due to the lack of neuroimaging evidence. Moreover, although the mentioned study motivated researchers to search for other factors (such as SES) that could potentially explain empathy development, the small sample size (N = 60) and type of data analysis (cross-sectional analysis with only two groups) threaten the generalizability of the study.

Thus, Grühn et al. [19] assessed intellectual functioning and subjective and psychological well-being in 10–87-year-old participants with different SES. The authors found that applying different analysis methods may result in different relationships between age and empathy. Specifically, cross-sectional analysis supported negative association between age and empathy, while latent-growth curve analysis showed a non-significant relationship. Unlike Phillips et al. [12], Grühn et al. [19] did not find any changes in the empathy–age relationship when educational level was controlled; thus, they argued for cohort effect rather than age or education effect on empathy. Supporting the argument regarding the effect of different analyses on the empathy–age relationship, cross-sectional studies reported a lower empathic response among the elderly, whereas the studies employing alternative approaches (e.g., longitudinal) reported different findings. Given these results, it can be concluded that that there is a lack of evidence supporting a causal relationship between age and empathy.

Although Grühn et al. [19] suggested one fundamental point—potential influences of different analysis methods on factors associated with empathy development—that should be carefully considered by other researchers, their methodology has a significant limitation. The questionnaire used for data collection, the California Personality Inventory (CPI), is not an appropriate measure for empathy. The CPI is a personality trait measurement consisting of 20 scales including emotional empathy [22], meaning that the other CPI scales could influence the participants’ responses to empathy-related items.

To overcome the measurement limitation, O’Brien et al. [13] conducted the first study exploring the developmental path of affective and cognitive empathy using the Interpersonal Reactivity Index (IRI; [23]), which is one of the most valid and reliable empathy questionnaires [24]. Beside their novelty in treating empathy as a multidimensional construct [25], their large sample size (N = 75,263) consisting of diverse Americans in late adolescence and adulthood (age range: 18 to 89) made their study exceptional. O’Brien et al. [13] found that people in middle age (50–70) reported the highest emotional and cognitive empathy compared to both younger and older adults. These results were in line with the dynamic integration theory [26], which argues for an inverse U-shaped relationship between emotional understanding and age. Accordingly, people experience maturation in both cognitive and emotional functioning until they reach the peak around middle age. Then, their cognitive abilities decline thereafter, which may cause problems in the emotional-related functions associated with empathy.

To summarize, studies focusing on empathy development during childhood are in line with Hoffman’s model [2], suggesting a shift from empathic distress (an automatic response) to empathic feeling (a thought-out emotional and cognitive reaction) as infants enter childhood. On the other hand, research on empathy development during adolescence and adulthood reported inconsistent findings. Such inconsistency in the research mostly originated from differences in empathy definitions, methodological issues (e.g., different data collection approaches—cross-sectional versus longitudinal), and negligence in differentiating between the cognitive and affective components of empathy; see [8]. Among the introduced theories regarding empathy development in adolescence and adulthood, the dynamic integration theory [26] has been the most comprehensive and convincing model, as it incorporates socio-cognitive, neuropsychological, and neurodevelopmental approaches [15].

1.3. Gender Differences in Empathy

Several studies have tried to provide empirical evidence to support the traditional belief that women are more empathic than men (e.g., [9]). For example, Van der Graaff et al. [27] conducted a 6-year longitudinal study using the IRI questionnaire and latent growth curve modeling. They observed a faster growth of cognitive empathy among adult females. Affective empathy also significantly increased in females in adolescence, while male participants showed stable affective empathy. The higher affective empathy in female participants was also found in non-Western populations [28].

Researchers explained higher empathy among women based on (1) social desirability [29], (2) parenting style [30], and (3) different brain structure and function [31]. According to first theory, women are more concerned about receiving acceptance from society, so they perhaps follow social norms requiring them to be more caring and to behave empathically [29]. Based on the parenting style theory, females are more empathic because the surrounding environment and rearing style foster emotional reactions in them. Gilligan [30] used the two factors—environment and rearing style—to argue that girls are more encouraged to be emotional and to care about others, whereas boys are raised to be more rational and suppress their emotions.

However, researchers studying brain structure and function challenged such traditional perspectives and provided evidence showing male and female adults have different brain activations in empathy-inducing situations. For instance, a higher empathy trait among women was found to be associated with increased activity in the right cerebral hemisphere [31]. In addition, Singer et al. [32] reported different brain activity between male and female participants in response to witnessing a defector receiving electric shock. Moreover, higher activity in the amygdala, which indicates higher emotional and social sensitivities, was reported for female adults [33]. In support of the latter finding, Schulte-Rüther et al. [34] also observed increased activity in the left temporoparietal junction, a brain region responsible for self–other distinction, among male participants. Given these studies reporting neural-level activity patterns associated with higher emotional understanding among women, it seems plausible that higher empathy among women is not solely due to social desirability and parenting style, but it is also rooted in biological (different brain activations) and psychological (different parenting and environment) factors.

1.4. Purpose of the Study: Examining Iranian Empathy Data from Bayesian Perspective

As suggested by multiple empathy researchers [13,19,35], it is necessary to explore the developmental trajectories of empathic understanding/feelings in Americans born after 1982 and/or in non-Western populations. The existing evidence regarding the developmental path of empathy is based on data from participants born before 1980, which would increase the possibility of cohort effects on empathic changes [19]. Furthermore, almost all studies in this field have been conducted with American population, meaning the reported findings could not be generalized to different populations. Researching empathy development in other cultures could make a substantial contribution in resolving the long-running debate regarding life-span empathy development [13].

Recently, the association between cultural background and empathic behavior has been frequently studied [34]; however, the majority of the recent studies have not paid much attention to considering developmental trajectories (see [28] for discussion). Given that cultural aspects are found to significantly influence empathy [36], the present study could make a significant contribution to the field by exploring empathy development among Iranian participants. Iranian populations possess diverse ethnic, social, cultural, and linguistic characteristics [28,37] and have been underrepresented in the psychological literature. Therefore, our study will be able to provide additional evidence on empathy development among non-European/American populations. To the best of our knowledge, the only study exploring the developmental trajectory of empathy in populations other than Westerners was conducted by Yaghoubi Jami et al. [21]. However, they employed the conventional analytical framework to the collected data, which poses the same limitation proposed by Grühn et al. [19]. As mentioned earlier, differences in the analysis method employed could lead to different findings.

Given the aforementioned reasons, we intend to examine empathy development among non-Western participants, namely Iranians, with a novel data analysis method to acquire additional evidence regarding this topic. For a better applied analysis, as an alternative approach, Bayesian analysis based on Bayesian perspective would be a likely candidate. The Bayesian perspective assumes that a specific parameter of interest, such as an effect size and an estimated coefficient in regression analysis, exists in terms of a probability distribution [38]. From a Bayesian perspective, statistical inference can be understood as an examination of a posterior probability distribution of the parameter of interest, based on the data [39]. Bayesian inference begins with setting a prior distribution of the parameter. The prior distribution is updated continuously by observing the data, and then this becomes the posterior distribution, indicating the likelihood of the parameter in a probability distribution given the acquired data [38]. Thus, with the posterior distribution actually updated by the data, it becomes possible to conduct data-driven statistical inference.

However, frequentist inference is mainly interested in whether the observed data is likely to be the case given a prior hypothesis, which is opposite to the main interest of Bayesian inference [40]. On the other hand, p-values in frequentist inference do not provide any information about whether a hypothesis is proven given the observed data; instead, the p-values merely indicate “the probability that one will observe values of the test statistic (parameters of interest) that are as extreme or more extreme than is actually observed” (p. 3) [40]. Thus, the Bayesian perspective would provide researchers who intend to conduct a data-driven applied analysis with epistemological and practical benefits [40].

Hence, we intended to employ the Bayesian analysis method for a better exploratory analysis. To examine which predictors in the collected dataset are most fundamental in predicting empathic traits, we used Bayesian Model Averaging (BMA). Unlike conventional regression analysis, which relies on the interpretation of p-values, this method allows us to search for the best models among all possible candidate models in a data-driven manner [41]. With BMA, we examined which factors explained variances in empathic traits among Iranian participants in the best regression models indicated by exploratory analyses.

2. Materials and Methods

2.1. Participants

The data were collected from 510 (123 males) Iranian participants (see

Table 1. Participant profile.

	Participants (N = 510)
	Mean	SD
Age	28.31	9.86
	n	%
Gender
Male	123	24.12%
Female	387	75.88%
Educational level
High school	105	20.59%
Undergraduate	296	58.04%
Graduate	109	21.37%
Marital Status
Single	312	61.18%
Married	184	36.08%
Divorced	14	2.75%
Have children
Yes	109	21.37%
No	401	78.63%
City size
Small	169	33.14%
Medium	69	13.53%
Large	232	45.49%
Living out of Iran	40	7.84%

for demographics). The age range of participants was 17 to 66 years old, with a mean of 28.32 (SD_age = 9.87 years). More than 50% of the participants were single and had no children. The residing cities of the participants were grouped into three categories (small, medium, and large) based on density of population in the reported city. Similarly, educational level was categorized as high school diploma, two- or four-year college degree, and postgraduate degree. The study was approved by the Institutional Review Board at a university in the southern US. and was performed in accordance with the ethical standards of the Declaration of Helsinki (1964) and its later amendments. All participants signed an online informed consent form approved by the Institutional Review Board.

2.2. Measures

Among all developed empathy questionnaires, the Interpersonal Reactivity Index (IRI; [23]), the Empathy Quotient (EQ; [42]), and the Questionnaire Measure of Emotional Empathy (QMEE; [43]) have been the most widely used (see [44] for discussion). The IRI measures both affective and cognitive facets of empathy. The EQ evaluates empathy as a one-dimensional construct. The QMEE assess the affective aspect of empathy. We employed these three questionnaires in our study to examine empathy development in a comprehensive manner. Since Farsi was the first language of our participants, the translated Farsi version of questionnaires [24,44] was distributed.

2.2.1. Interpersonal Reactivity Index

The IRI is a multidimensional assessment of empathy measuring both affective and cognitive facets [23]. It consists of 28 items in four different scales: fantasy scale (FS), empathic concern (EC), perspective taking (PT), and personal distress (PD). These subscales are scored on a 5-point Likert scale (0 = does not describe me well, to 4 = describes me well) with a final score ranging from 0 to 28 on each scale. A higher score on each scale indicates a higher tendency on that subscale (e.g., score of 28 on the empathic concern scale suggests participants reported themselves as having high affective empathy). Responses to fantasy scale items were removed due to the debate over the contribution of this scale to empathy measurement [42]. It has been reported to be a reliable instrument, and its psychometric properties have reached an acceptable level [24]. Using Cronbach’s alpha analysis, all scales showed acceptable internal consistency: α_{empathic concern} = 0.68, α_{personal distress} = 0.69, and α_{perspective taking} = 0.63.

2.2.2. Empathy Quotient

The EQ consists of 60 items (20 filler items) measuring empathy as a one-dimensional construct [42]. Participants rated statements according to their degree of agreement using a 4-point Likert scale, ranging from 1 = strongly agree to 4 = strongly disagree. The total score varies between 0 and 80, with higher score indicating higher empathic tendencies. The psychometric properties of the EQ were evaluated within English speaking contexts [45] and Farsi-speakers [44]. The internal consistency for the current study was in line with that of previous studies, supporting the reliability of the questionnaire within a Farsi-speakers context: α_EQ = 0.79. The questionnaire can be accessed through the Autism Research Center website (https://www.autismresearchcentre.com; accessed on 6 November 2022).

2.2.3. Questionnaire Measure of Emotional Empathy

The QMEE is a 33-item questionnaire measuring emotional empathy [43]. Participants rated their agreement on a 9-point scale ranging from −4 = very strongly disagree to +4 = very strongly agree. A higher score indicates the presence of higher affective empathic tendencies. The validity and reliability of the questionnaire was evaluated and supported within English-speaker populations [45] and non-English-speaker populations (e.g., Swedish; [46]). QMEE showed an acceptable internal consistency in this study: α_QMEE = 0.75.

2.3. Procedure

We followed similar procedures used in previous studies with similar questionnaires within an Iranian context [21,44]. We utilized an online platform as a method of data collection. The Farsi version of the questionnaires was distributed through Qualtrics survey-hosting provider (https://www.qualtrics.com; accessed on 6 November 2022). A flyer explaining the purpose and procedure, along with a link to the questionnaire, was posted on social media (e.g., Facebook, Instagram) to directly target the population of interest, Farsi-speakers. Interested participants was able to begin the survey using a Qualtrics link provided in the flyer, after completing the consent form.

2.4. Analysis Plans

2.4.1. Bayesian Model Averaging

We employed BMA to explore the best regression models predicting the developmental path of empathy in Iranians. Conventional frequentist regression analysis that employed p-values in interpreting the results would not be suitable for exploratory analysis. First, when frequentist regression analysis is performed, only the tested model is compared with the null model, not all other possible candidate models [47]. Second, if we are interested in examining which predictors should be included in a regression model, excluding non-significant predictors with p > 0.05 is not statistically appropriate. Statistically, p < 0.05 only means that a null hypothesis could be rejected, but not that the responsive alternative hypothesis should be accepted [48]. Thus, determining the inclusion or exclusion of predictors based on p-values would not be acceptable. Third, although there have been several variable and model selection methods developed based on the conventional approach (e.g., stepwise variable selection), such methods are likely to inflate false positives and overestimate coefficients [49]. Therefore, for data-driven analysis, the conventional regression method might be misleading.

Hence, we employed BMA that can address the aforementioned limitations of conventional regression analysis. BMA allows us to examine which models are most probable and likely to be the case given the tested data from the Bayesian perspective. BMA initiates the model selection process by assigning the same prior probability, P(M), that indicates to which extent a specific predictor should be included [50]. Through observing collected data, the prior probability of each candidate predictor is updated [51]. Then, we can acquire the posterior probability, P(M|D), indicating to which extent a specific predictor should be included, given the observed data. With the calculated P(M|D)s, it is possible to identity the regression models by including predictors with the highest P(M|D)s, the most likely models given the data, as the candidates of the best regression model.

To implement the aforementioned model exploration process, we used the BMA package in R [52]. We performed five independent BMA processes, with five dependent variables of interest, IRI-EC, IRI-PT, IRI_PD, EQ, and QMEE. For each BMA process, we entered candidate predictors as independent variables in the BMA model. The independent variables of interest included participants’ gender, age, having/not having children, educational level, and the city size. Among these, age was a continuous variable, while the rest were categorical.

Once the BMA process was completed, we identified the best candidate regression models as those with posterior odds higher than 10%. Although BMA has been primarily utilized to examine the averaged coefficient across multiple models to address model uncertainty, Hoeting et al. [51] suggested that the results from BMA can be used to search for the single best models among candidate models. To estimate the coefficients of survived predictors in each of the identified regression models, we performed multiple linear regression while using only the survived predictors in the identified model as independent variables. From the resultant multiple linear regression model, we extracted additional information for model evaluation: Akaike’s information criteria (AIC), Bayesian information criteria (BIC), adjusted R², model Bayes Factors (BF), and leave-one-out cross-validation root-mean-square deviations (LOOCV RMSE). We applied these procedures to fit and evaluate the full model that included all candidate predictors as independent variables.

2.4.2. Model Evaluation and Comparison

We evaluated and compared each model with diverse indicators mentioned in the prior section: AIC, BIC, adjusted R², model BFs, and LOOCV RMSEs. Among these, the former three indicators, AIC, BIC, and adjusted R², have been widely used in studies. AIC and BIC are calculated with the maximum likelihood and the number of free variables occurring in a regression model of interest. These can be calculated as follows:

$$\mathrm{AIC}=-2 \log \left(L\right)+2 V$$

(1)

$$\mathrm{BIC}=-2 \log \left(L\right)+V \log \left(n\right)$$

(2)

where L is the maximum likelihood, V the number of free variables, and n the number of observations [53]. These indicators tend to penalize a complex model with a large number of predictors, provided that a calculated maximum likelihood is the same when different models are compared. In the present study, we assumed that the smaller AIC or BIC indicates the superiority of a model of interest.

R² basically indicates to what extent the variance for a dependent variable interest of interest in a regression model can be explained by predictors in the regression model. One of the significant issues related to use of simple R² is that the inclusion of more predictors results in the increase in R², regardless of the complexity of the model and whether the included predictors make significant contributions to explaining the variance. Thus, for a fair model comparison, it is necessary to employ adjusted R² to address the issue related to model complexity [54]. lm function in R calculates adjusted R² as follows (type summary.lm in R for further details):

$$\mathrm{Adjusted} R^2=1-\left(1-R^2\right)\frac{n-1}{n-p-1}$$

(3)

where R² is unadjusted R², n is the number of observations, and p is the number of predictors entered into a regression model. As shown in the equation, similar to AIC and BIC, adjusted R² tend to penalize complex models with independent variables that do not explain the variance significantly. We used the greater adjusted R² value as an indication of relative model superiority.

In the present study, we employed additional indicators for better model comparison. First, we employed model BFs. In general, model BFs indicate to what extent a specific model is better supported by evidence compared with another model from a Bayesian perspective [55,56]. Let us say we are interested in whether a specific model is likely to be the case, given the collected data. We can start with setting our prior belief in the model as a prior probability, P(M), where M refers to the model to be examined. Then, the probability is updated by examining the collected data. The updated probability is a posterior probability, P(M|D), which is the probability that M is the case, given the data (D). P(M|D) can be calculated with P(M) and D following Bayes’ Theorem, as follows:

$$\mathrm{P}\left(\mathrm{M}|\mathrm{D}\right)=\frac{\mathrm{P}\left(\mathrm{M}\right)\mathrm{P}(\mathrm{D}|\mathrm{M})}{\mathrm{P}\left(\mathrm{D}\right)}$$

(4)

P(D|M)/P(D) is acquired from the observed data to update P(M|D) with P(M). When two models, Models A and B, are compared, we can refer to BF_AB, indicating to what extent D supports Model A versus model B. BF_AB is calculated with P(D|M)/P(D), as follows:

$${\mathrm{BF}}_{\mathrm{AB}}=\frac{\mathrm{P}(\mathrm{D}|{\mathrm{M}}_{\mathrm{A}})}{\mathrm{P}(\mathrm{D}|{\mathrm{M}}_{\mathrm{B}})}$$

(5)

BF_AB >> 1 indicates that Model A is more likely to be the case, given the data, compared with Model B. Several Bayesian statistical guidelines suggest that 2log(BF_AB) ≥ 2 indicates the presence of positive evidence supporting Model A versus Model B, and 2log(BF_AB) ≥ 10 indicates the presence of strong evidence [40]. In the present study, we calculated BF_A0, which indicates to what extent a specific Model A is better than the null model, only with an intercept. To compare two models of interest (e.g., A versus B), BF_AB can be calculated with BF_AB = BF_A0/BF_B0, and thus 2log(BF_AB) = 2log(BF_A0) − 2log(BF_B0), with BF_A0 and BF_B0.

Although AIC and BIC allow us to compare different models in terms of their likelihood and complexity, from a Bayesian perspective, information criteria, particularly BIC, are mere approximations of model BFs [57]. Although the calculation of BFs was computationally very demanding in the past, it has become very feasible with recently developed computers. One additional benefit of using model BFs is that we can make quantitative statistical inference (e.g., to what extent the data better supports Model A versus Model B) with the BFs [48]. Hence, we decided to employ model BFs in our study. We calculated model BFs with an R package, BayesFactor [58].

Second, in addition to the adjusted R², we calculated LOOCV RMSEs to examine whether models well explained the variance within the data that was not used for regression. Although adjusted R² can penalize complex models, it could not address the issue of overfitting because it cannot show to which extent a fitted regression model can well predict cases out of the boundary of the data used for regression. To address this issue, it is necessary to utilize cross-validation [59]. Among various cross-validation algorithms, LOOCV is one of the most prevalently utilized methods. When a total of n observations are in the data collected for regression analysis, LOOCV repeats n times regression and cross-validation, while not using one n observation for model fitting each time [60]. That one observation is used to examine the extent to which the regression model fitted without that observation (n − 1) can well predict that observation in terms of LOOCV RMSE. This LOOCV RMSE can be used to examine whether a model of interest can predict observations well and if it is robust against overfitting, given that it quantifies the extent to which an observation not used for model fitting can be well predicted by the fitted model (Xie et al., 2019). A small LOOCV RMSE can be interpreted as an indication of relatively better performance. In the present study, we used LOOCV implemented in the MuMIn package in R [61].

3. Results

As explained earlier, empathy was measured through three questionnaires: IRI, EQ, and QMEE. Among the three components of IRI, the participants’ score for empathic concern was higher (M = 19.41, SD = 4.08) than their scores for perspective taking (M = 16.43, SD = 4.03) and personal distress (M = 14.88, SD = 4.63). In other words, participants’ emotional empathy was higher than their cognitive empathy and personal distress. Moreover, on average, the participants scored 38.09 ± 8.72 on the EQ questionnaire. Finally, the participants’ scores on the QMEE exhibited more deviation (M = 41.83, SD = 22.78), showing this questionnaire to be more person-dependent compared to the other two questionnaires.

When we performed BMA for IRI-EC, IRI-PT, IRI-PD, EQ, and QMEE, we were able to identify the candidates for the best regression models with posterior probabilities ≥ 10% for all dependent variables, except IRI-PT. In the case of IRI-PT, the most probable model identified by BMA was the null model (only with an intercept), so we did not evaluate model performance. The results from BMA and model evaluation with the performance indicators AIC, BIC, model BFs, and LOOCV RMSEs are presented in

Table 2. Comparison of regression models identified by BMA (posterior probability ≥ 10.00%).

		Female	Age	No Child	Education (Reference Group: High School)			Marital Status (Reference Group: Single)			City Size (Reference Group: Small)			Posterior Probability	Adjusted R2	AIC	BIC	Log Bayes Factor (vs. Null)	LOOCV RMSE
					Associate Degree	Undergraduate	Graduate	Married	Divorced	Widow	Medium	Big	Out of Country
IRI-EC	Model 1 (most probable)		+		+	+	+							62.40%	16.41%	2798.09	2823.49	38.37	3.75
	Model 2 (+ female)	+	+		+	+	+							11.10%	16.70%	2797.30	2826.94	37.56	3.75
	Model 3 (full model)	0	+	0	+	+	+	0	0	0	0	0	0	−	16.10%	2806.85	2861.90	29.67	3.78
IRI-PD	Model 1 (most probable)	+	-											36.10%	4.55%	2991.04	3007.97	7.94	4.53
	Model 2 (+ divorced)	+	-						+					28.70%	5.43%	2987.26	3008.43	8.79	4.52
	Model 3 (full model)	+	-	0	+	0	0	0	+	0	0	0	0	−	5.58%	2994.34	3049.39	1.02	4.55
Empathy Quotient	Model 1 (most probable)		+				+							47.40%	4.97%	3635.36	3652.30	9.27	8.53
	Model 2 (- graduate)		+											21.20%	3.69%	3641.19	3653.90	7.49	8.58
	Model 3 (+ undergraduate)		+			+	+							10.50%	5.39%	3634.13	3655.30	9.00	8.51
	Model 4 (full model)	0	0	0	0	+	+	0	0	0	0	0	0	−	5.05%	3643.82	3698.87	−0.71	8.60
Emotional Empathy	Model 1 (most probable)				+	+	+				+			31.40%	14.38%	4563.48	4588.89	31.91	21.18
	Model 2 (+ female)	+			+	+	+				+			21.90%	15.13%	4559.96	4589.61	32.43	21.12
	Model 3 (+ gender, married)				+	+	+	+			+			13.60%	14.97%	4560.92	4590.56	32.33	21.13
	Model 4 (full model)	+	0	0	+	+	+	+	0	0	+	0	0	−	15.50%	4563.67	4618.72	27.17	21.20

Note. IRI-PT was not presented because the most probable model was the null model. +: coefficients > 0; −: coefficients < 0; 0: coefficients not significantly different from zero at p < 0.05. The models identified as the best models are shown in bold.

.

In the case of IRI-EC, the most probable model (Model 1) showed the best performance in terms of BIC and model BF. Simultaneously, Model 2 showed the best performance in terms of adjusted R2, AIC, and LOOCV RMSE. When the model BFs were examined, 2log(BF12) = 1.62 indicated that the difference was not positively supported by evidence. Given that the model BFs can be better indicators for model performance compared with BIC from a Bayesian perspective [57], we conclude that Model 2 is the best model predicting IRI-EC in terms of model simplicity, explained variance, and robustness against overfitting.

When the models predicting IRI-PD were evaluated, Model 1 showed the best BIC, while Model 2 showed the best AIC, model BF, and LOOCV RMSE. The full model including all candidate predictors, Model 3, reported the highest adjusted R2. In terms of model BFs, Model 2 outperformed Model 3, 2logBF(BF23) = 15.54; however, its performance was not significantly different from Model 1, 2logBF(BF21) = 1.70. Although Model 3′s adjusted R2 was the highest, its LOOVC RMSE was greater than that of Model 2, suggesting that Model 3 would be more susceptible to overfitting compared with Model 2. Hence, we assumed that Model 2 was the best model among the candidate models.

In the case of EQ, three candidate models (Models 1–3) plus the full model (Model 4) were compared. Model 1 showed the best BIC and model BF, while Model 3 showed the best AIC, adjusted R2, and LOOCV RMSE. Although Model 1 reported the greatest model BF, when it was compared with Model 3, it did not show significantly better performance, 2logBF(BF12) = 0.54. Hence, we conclude that Model 3 is the best model.

Finally, the models predicting QMEE were examined. Similar to the prior case, three most probable candidate models (Models 1–3) and the full model (Model 4) were compared. In terms of BIC, Model 1 showed the best BIC. Model 2 reported the best performance indicators in terms of adjusted R2, AIC, model BF, and LOOCV RMSE. When model BFs were compared, Model 2 did not significantly outperformed Model 1, 2log(BF21) = 1.04. Thus, we assumed that Model 2 exhibited the best performance among the candidate models.

4. Discussion

Most previous psychological studies focused on a specific group, that is “Western, Educated, Industrialized, Rich, and Democratic (WEIRD)” (p. 61) [62], or from highly academic and above-average socioeconomic American populations [63], which could limit the generalizability of the results. Such an issue can be problematic, as O’Brian et al. [13] argued, “empathy has different developmental patterns in different cultural contexts” (p. 174). Therefore, in this study, we tried to overcome this limitation by exploring developmental trajectories of empathy among Iranian participants.

In addition, the analysis method used in the present study, BMA based on Bayesian perspective, enabled us to provide useful insights into how to improve methodologies to explore the best analysis models to study empathy development in a data-driven manner. We also proposed that multiple indicators of model performance, instead of just one indicator, be employed while evaluating and comparing different candidate regression models. The following section discuss the results observed in this study, based on research questions and previous studies.

4.1. Gender

A significant gender difference was observed regarding personal distress and measurements of affective empathy: IRI-EC and QMEE. In these domains, females scored higher than their male counterparts. Our result was in line with that in previous studies [15,45] and theoretical works concerning higher empathy and personal distress among women [30]. All these studies, including the present study, suggested that females might be more emotionally aroused in response to observing or hearing others’ pain and misfortunate. This emotional arousal could either be self- or other-oriented and lead to personal distress or emotional empathy [13,20].

Higher emotional understanding in women could be linked to both nurture and nature factors. Form a biological perspective, a woman’s brain is structured and functions differently compared to a man’s, especially when it comes to evaluating emotional clues and empathizing with others [33,34]. Similarly, males and females are raised with different values, and even social expectations, especially in socio-centric cultures such as Iran [28]. Unlike men, women are thought to react emotionally and be more sensitive towards others’ pain and suffering [30]. As a result, women might develop a more diverse and complicated emotional lexicon and mindset that would lead to higher emotional empathy. However, perhaps due to a side effect of being sensitive and emotional, personal distress is also higher in women.

However, the gender difference in empathic reactions was weakened once empathy was treated as a unidimensional concept (measured by EQ). The EQ is a questionnaire that evaluates both affective arousing and cognitive arousing empathic reactions [45]. Similarly, no gender difference was observed in IRI-PT, which is a measurement of cognitive empathy. Although the same non-significant result was observed in previous studies [21,64], it shows that the results hinge upon the specific measures and definitions of empathy. It suggests that while some factors might not show any significant gender difference when a specific measure is used, the same factors might significantly differ across genders when another measure is employed. Such a point perhaps underscores the importance of treating empathy as a multi-dimensional construct [8,25,65].

4.2. Developmental Factors

Similar to gender, participants’ educational level was also significantly associated with affective empathy. Participants’ scores regarding emotional empathy and empathy in general significantly varied according to their educational level. Inconsistent with a general belief that higher education would lead to higher cognitive abilities [66], we did not find any significant association between participants’ educational level and cognitive empathy, i.e., perspective taking. Nevertheless, education was associated with emotional empathy, perhaps due to experiencing novel emotions in college. Previous studies have shown that first-hand experience of different life events, especially painful experiences, led to higher emotional understanding and hence, higher affective empathy [15,67]. Perhaps within the context of higher education, people might be able to have time to experience and deliberate upon painful emotions emerging from socio-moral issues that they had not seriously considered in the past [68,69]. Therefore, people with higher educational level could develop a more mature emotional empathy due to the fact that they undergo more diverse experiences [20].

Our findings also showed a significant association between participants’ age, emotional empathy, and general empathy. These results are in line with both socio-cognitive and neurodevelopmental theories of empathy development [11,14]. These two perspectives propose that older people are more able to regulate their emotions and deal with empathic-inducing situations in a more sophisticated way. Moreover, because of having more experiences in social interactions and involvement in society, older people are more motivated to develop a mature emotional understanding and utilize their sophisticated evaluation of emotional cues in their social interactions [28].

4.3. Methodological Implications

The use of BMA in the present study could provide researchers with useful insights about how to improve statistical analysis in empathy research, particularly in data-driven exploratory studies. Instead of using frequentist regression analysis, which is not suitable for exploration of the best regression model and likely to commit model modification based on arbitrary criteria (e.g., p-values of predictors), we employed BMA that allowed us to explore the best model with objective indicators for decision making, such as posterior probabilities [41]. If researchers are interested in exploring a dataset instead of testing a theory-driven hypothesis, without assistance from methods such as BMA that suggest candidates for best models using presenting objective indicators, the researchers are likely to experience practical difficulties in obtaining the best answers.

Hence, appropriately applying data-driven analysis methods has significant merits in data exploration, as well as hypothesis and research question generation for further research. Once we examine how scientific research is being conducted in the field, we can find that both data-driven and hypothesis-driven (or theory-driven) approaches are required for better science [70]. Jack et al., [71] suggested that the use of data-driven methods provides researchers with novel insights that could not be explored with a priori theoretical constraints and allows for diversifying psychological knowledge in the field of human psychology. Data-driven approaches will be particularly informative and constructive in studies similar to ours that examine relatively fewer cross-cultural aspects. When relevant theories or previous studies are insufficient to form hypotheses and when exploration of relatively less examined realms in a field becomes a primary purpose, data-driven approaches will become useful tools form the diversification and expansion of knowledge, as well as hypothesis generation for follow-up studies.

Moreover, we demonstrated the necessity of employing diverse model fit indicators, i.e., AIC, BIC, model BF, LOOCV RMSE, instead of a single indicator, i.e., R², in the comparison and evaluation of candidate models. As mentioned, if researchers merely rely on frequently used indicators, such as R² and even adjusted R², it is difficult to assure that the identified model is the best model in terms of its frugality and its robustness against overfitting [72]. When multiple indicators, particularly those based on cross-validation, are employed, such points can be better addressed [73]. In our study, LOOCV RMSE directly demonstrates the extent to which a tested model is susceptible to overfitting in terms of RMSE, examined with a cross-validation dataset that was not used for regression analysis or model fitting [74]. For instance, in our model exploration for emotional empathy, although Model 4, the full model, reported the highest adjusted R², its reported LOOCV RMSE was greater than that of Model 2, suggested by BMA as one of the most probable models. This suggests that reliance on traditional fit indicators is likely to result in overfitting and may eventually limit the potential generalizability of the identified model beyond the boundary of the data used for model estimation.

Given these points, our study would make significant methodological contributions to the field of empathy research by demonstrating the use of BMA. Since all R source code (with remarks and comments) files, along with the data files, are available to the public via the Open Science Framework, researchers interested in testing BMA for data exploration will be able to obtain some practical guidelines.

5. Limitations and Directions for Future Studies

The results reported in the present study need to be interpreted while considering its limitations. For example, the measure utilized for studying empathy development relied only on self-reported questionnaires. The most problematic concern of using self-reported questionnaires is associated with the accuracy of the responses. Although we excluded all the responses that were suspected to follow a pattern, such as repeating one answer across the whole questionnaire, the results are still based on participants’ perception of empathic responsiveness. Therefore, employing other means of objective and/or subjective empathy measurement, such as behavioral observation (e.g., [65]) or a mixed-method study, could provide more clarification and understanding of the developmental trajectory of empathy within an Iranian context. The other limitation that needs to be considered while interpretating the result is the use of one questionnaire for measuring cognitive empathy —IRI-PT— and two for measuring affective empathy —IRI-EC and QMEE. To the best of our knowledge, there is no other self-reported questionnaire for measuring cognitive empathy except the IRI; thus, the field would benefit from developing a reliable, fair, and valid questionnaire tapping cognitive empathy.

6. Conclusions

In the present study, we examined the best models predicting different aspects of empathic traits using Iranian samples by employing BMA. The findings from our study could provide empathy researchers with useful insights about how to better understand factors associated with empathy developed in non-Western populations, which has not been well studied by previous studies primarily focused on WEIRD populations. Moreover, we demonstrated that the methodology implemented in our study, BMA and the use of multiple fit indicators for model selection, would be potentially useful in future empathy research. It could contribute to an increase in diversity in the knowledge regarding empathy and the generation of novel hypotheses in diverse contexts, with a data-driven perspective.