Effect of Emotional Valence on Emotion Recognition in Adolescents with Autism Spectrum Disorder

We first determined the emotional valence of each eye region stimulus from the RMET-C and RMET-A. On the RMET-C, 36% of items were classified as negative valence with the rest evenly split between positive and neutral valence. On the RMET-A, 39% were classified as positive valence with the rest evenly split between negative and neutral valence.

We next compared performance of ASD and TD groups on both the RMET-A and RMET-C and quantified how emotional valence impacted this performance. Results showed that adolescents with ASD had higher error rates compared to TD on both the RMET-C and RMET-A. While adolescents in the ASD group consistently demonstrated higher error rates than TD with all stimuli regardless of valence and in both versions of the RMET, differences between groups were greater with positive and negative valence items than with neutral items. Both groups had the lowest error rate when performing the RMET-C test with negative valence stimuli. Both groups performed better overall on the RMET-C compared to the RMET-A; however, this difference was more pronounced in the performance of individuals with ASD.

Multiple approaches to emotional valence classification of the RMET stimuli have been used. Hudson and colleagues (2020) outline these methodological variations and show how such inconsistencies, together with varying sample sizes, make it challenging to draw conclusions about how valence may affect RMET performance. Of these previous studies, we compared our emotional valence classification system to the four most commonly used (Koizumi & Takagishi 2014, Baribeau et al. 2015, Harkness et al. 2005, Scott et al. 2011). The first two of these studies applied emotional valence classification to the RMET-C while the second two applied emotional valence classification to the RMET-A. We compared how consistent our valence classification was to those published in the existing literature and discuss the impact of methods used to establish classifications.

Our RMET-C valence classifications have a 57% agreement with those of Koizumi & Takagishi (2014) and a 54% agreement with Baribeau and colleagues (2015) (Appendix 1). Koizumi & Takagishi (2014) state that items were classified by their emotional valence but do not indicate the method used for their classification. They cite Fertuck and colleagues (2009), who in turn used the classifications of Harkness and colleagues (2005) on the 36-item RMET-A, which may not be directly applicable to the 28-item RMET-C. Also, there is only 56% agreement between the classification of overlapping items in Koizumi & Takagishi (2014) and Harkness et al. (2005). One possible explanation for these discrepancies is that similar to Hudson et al. (2020), both studies included the “correct” emotion label with the photograph during the classification task and, therefore, may have imparted bias to the subjects. Since the correct labels differ between the two RMET versions, the potential for bias can be seen in Fig. 3, which displays an item that was classified by Koizumi & Takagishi as “neutral” (correct label is “serious” in the RMET-C), but that same item was classified by Harkness and colleagues as being “negative” (correct label is “accusing” in the RMET-A). By removing the labels and having subjects classify the valence using only the image of the eye region, our study addresses the potential bias associated with the label.

Baribeau and colleagues (2015) derived their own RMET-C valence classifications from Koizumi & Takagishi (2014) and two differing RMET-A classifications (Harkness et al.,2005; Scott et al., 2011)). They assigned the valence by consensus of the other three study assignments, choosing a neutral valence for the two items that were assigned to three differing categories by the three prior studies (see Appendix 1 – RMET-C Items 17 and 26, RMET-A Items 19 and 34). Thus, their classification scheme suffered from a combination of the weaknesses of the three prior studies.

Our RMET-A valence classifications have a 44% agreement with the study by Harkness and colleagues (2005) and a 64% agreement with the study by Scott and colleagues (2011) (Appendix 1). While the valences used in the study Harkness and colleagues were determined by raters who were shown each face together with its correct label, valences in the study by Scott and colleagues were determined independently of the labels, likely explaining the greater agreement with the present study. The participants who classified the stimuli also differed between the studies: Harkness and colleagues surveyed 12 undergraduate women, while Scott and colleagues surveyed 40 undergraduate students (males and females), and in the present study we surveyed 113 medical school students (45 male, 68 female). Due to the differences in sample sizes, it is possible that the study with the lower number of individuals surveyed led to some stimuli being classified as neutral due to not meeting the significance threshold (Type 2 error). Most of the differences (9/13 items; 69%) in classifications between our study and the study by Scott and colleagues (2011) could be accounted for by decreased power to detect significant positive or negative valence in the earlier study. It is possible that the age, clinical experience and training of the medical students used for valence classification in our study contributed to the different classifications of the remaining 4 items between these two otherwise similar studies. Figure 4 shows the items with disagreement, not explained by insufficient power in the previous study, along with the assigned valence in our study and in the study by Scott and colleague. Further clinical assessments were not performed on these convenience samples, so it is unclear if differences in social cognition between these groups contributed to the differing classification results.

Individuals with ASD had a higher error rate on both versions of the RMET compared to TD, which is consistent with previous ASD RMET studies (Baribeau et al., 2015; Demurie et al., 2011; Kaland et al., 2008) and a substantial body of research suggesting impaired ToM in ASD (Capps et al., 2000; Happe et al., 1996; Harms et al., 2010; Lartseva et al., 2015). Individuals with ASD performed significantly worse on positive items compared to neutral and negative items in a prior study by Baribeau and colleagues (2015). In contrast, we found individuals with ASD performed significantly worse on both positive and neutral items compared to negative on the RMET-C, and they performed slightly worse on the RMET-A overall. This was likely a result of our differing valence classifications, given that almost half of the eye regions were classified differently in our present study. Given the methodological limitations of the previously used classification schemas outlined above, the present study may be a more accurate reflection of emotion recognition accuracy by valence.

Our findings differ from those of non-RMET valence studies that suggest subjects with ASD have greatest difficulty identifying emotions of negative valence (Ashwin et al., 2006; Howard et al., 2000; Pelphrey et al., 2002; Wallace et al., 2011). Much of this difference likely stems from these studies’ use of different experimental paradigms, consisting of far fewer total emotions than either RMET version with a greater proportion being negative (fear, sadness, disgust, and anger) (Howard et al., 2000; Pelphrey et al., 2002; Wallace et al., 2011).

Similar to the present study, Kaland and colleagues (2008) studied performance on both the RMET-C and RMET-A in TD adolescents and in individuals with ASD. Both groups showed a higher accuracy on the child version than the adult version; however, the authors did not report on the significance of the differences between versions in comparing each group’s performance.

In our classification of RMET stimuli, we found a relatively narrow valence range across all images. In a study of facial emotion recognition, Wallace and colleagues (2011) demonstrated that adolescents with ASD require more intense facial expressions for accurate emotion identification compared to TD controls. Since our stimuli did not have a high intensity of valence, this supports the ability of the RMET to detect more subtle difficulties with emotion recognition in ASD. However, it is necessary to establish validated valence classifications for stimuli across both RMET versions, as the classifications to date show limited concordance. We believe that the strengths of our valence classification method compared to previous reports, including increased sample size and identification of valence without biasing results by including “correct” labels, make our classification results more reliable. However, future validation of the results by replication is warranted.

Interestingly, while individuals with ASD consistently performed worse compared to TD, RMET-A performance did not significantly differ by valence within either group. By contrast, in the RMET-C both groups performed better at identifying negative valence emotions. Negative emotional stimuli tend to capture more attentional resources compared to positive emotional stimuli, which may explain this trend (Balconi et al., 2012; Charles et al., 2003; Johansson et al., 2004; Öhman et al., 2001; Rumpfa et al., 2012). Our study demonstrates that impairment in the ability of individuals with ASD to identify emotions is primarily seen in identification of positive or negative valence emotions, with no difference between groups on identification of neutral items. The increased language complexity of the RMET-A, meanwhile, further compounded the increase in error rate when individuals with ASD were asked to identify positive or negative emotions. This suggests that a task involving processing of complex language interacts with the emotional demands required to identify extremes of emotional valence, exacerbating the challenge that individuals with ASD experience when interpreting nonverbal emotional cues. This finding has clinical implications, as individuals with ASD may be more significantly impacted by their social communication deficits when combined with demands that stretch their verbal language processing capacity.

The present study should be interpreted in light of the following limitations. Due to the low number of subjects, we may have lacked the power needed to identify significant differences in RMET performance, especially in the TD group where differing performance among valences and between RMET versions was more subtle. Given the results in the TD group on RMET-A, we estimate that a prospective study would need to include nearly 3 times as many TD participants (n = 90) in order to have 80% power to detect a difference in error rate at least as large as observed here between the valences. Another limitation was the IQ difference between groups, which may contribute to the higher discrepancy between groups in the RMET-A. However, both groups had above average IQ, and the difference between groups was not statistically significant at the α = 0.05 level. Further study with more subjects that are matched more closely for IQ will help to clarify these results.