Introduction

Children’s ability to infer emotions from social cues such as facial configurations is crucial for navigating social interactions. Most extant studies have used stimuli depicting stereotypes of facial configurations in support of the view that emotional development progresses from valence-based judgment (i.e., feels good or bad) to more nuanced negative emotion differentiation during childhood, a trajectory thought to be driven primarily by increasing sensitivity to perceptual cues. However, more recent theories argue that encoding and discrimination of facial muscle movements alone are insufficient for predicting others’ feelings and behaviors1,2. Instead, children need to accumulate and integrate conceptual knowledge that includes cultural frameworks for displaying emotions and imbuing social interactions with meaning as well as linguistic input about emotions3,4,5. Even early-emerging emotion judgments based on valence are likely informed by conceptual knowledge shaped through children’s embodied and social experiences, rather than arising solely from low-level perceptual processes. Therefore, the emergence of children’s abilities to reason about others’ emotions develops through both bottom-up processes (attention to potentially important predictive cues) as well as top-down influences (refinement of conceptual understanding of behavior across individuals and situations). It is this conceptual framework that allows children to understand complexities such as the fact that the same emotion can be appropriate or inappropriate depending upon the context, that different people can experience and display emotions differently, and that the same individual can react differently cross time and situations. However, how these perceptual and conceptual processes contribute to children’s understanding of emotions remains elusive.

It is difficult to precisely measure children’s perceptual discrimination of emotion cues through behavioral measures because these processes are both spontaneous and fast. For these reasons, researchers have attempted to measure these processes through the use of EEG-derived event-related potentials (ERPs), specifically focusing on the neural correlates of facial structural encoding. Initially, it appeared that two ERP components, the N170 and N290, could prove useful in this regard. However, these methods have yielded inconsistent data across multiple ages and emotion categories, leaving any conclusions about discriminatory processes unclear6,7,8. The range of reported N170/N290 responses could reflect a lack of differentiation in children’s structural encoding of facial configurations. But it is also likely that factors such as low-level stimulus traits or constraints on young children’s memory and attention also influence these early ERP components9,10,11. Additionally, obtaining a sufficient number of clean trials for a satisfactory signal-to-noise ratio is crucial to observing consistent N170/N290 responses, but it is challenging to both expose children to a variety of stimulus conditions and ensure that each of those conditions contain many trials given that young children have limited capacity to complete long experimental sessions.

In contrast, the evidence that conceptual knowledge guides children’s reasoning about emotion has been more consistent (for review, see refs. 4,12). Children aged 2 to 4 years can map emotion words to various facial configurations and situations corresponding to various emotions, indicating early competence in understanding emotional concepts13,14,15,16,17. For example, Woodard et al.5 found that developmental changes in how emotions are represented do not simply reflect children’s responses becoming more consistent or children becoming more competent with increasing age; instead, the manner in which children prioritized and used different dimensions of emotion changed across age. They found that with greater experience, children draw more fine-grained distinctions that rely upon conceptual information. These data suggest that children shift from more perceptual, similarity-based categories to categories shaped more by rules and labels (see ref. 18). A related possibility is that growth of emotion vocabulary, including more abstract conceptions of emotion, gradually reshape children’s representations of emotion12,19,20,21,22. Most previous studies have examined perceptual discrimination and conceptual knowledge of emotions separately, leaving their interaction with age in shaping children’s emotion understanding not fully explored.

To address these gaps, the present study investigates how perceptual and conceptual processes contribute to children’s understanding of emotions across development. We focus on the period from 5 to 10 years, a developmental period marked by rapid changes in their emotional knowledge23,24. We conducted three studies to examine different aspects of emotion understanding.

In Study 1, we employed an EEG frequency tagging approach, referred to as the FPVS paradigm25, to investigate how children’s brains spontaneously discriminate between different stereotypical facial configurations. This method, adapted from steady-state visual evoked potential (ssVEP) models, provides an objective measure of automatic structural encoding and discrimination at pre-defined frequencies and achieves a superior signal-to-noise ratio within a short timeframe. It also minimizes interference from basic visual features (e.g., color, contrast, shape, and texture), attention, and conceptual influences that often confound facial recognition tasks26. FPVS has been validated in both adults27,28,29and children30,31,32, demonstrating its reliability for evaluating the neural encoding of facial configurations. In Study 2, we evaluated children’s conceptual knowledge of emotions through a conceptual similarity rating task. This task examined how children associate specific words with different emotions, and we analyzed the correlations between these ratings across different emotions to evaluate the conceptual similarity between them. In Study 3, we evaluated children’s understanding of the relationships between emotion categories using two behavioral tasks: a sorting task, which required children to label emotions into predefined categories33, and a matching task, which involved matching the emotions of two individuals24,34.

To determine how children integrate spontaneous discrimination of facial cues (Study 1) with their concurrent conceptual knowledge of emotion (Study 2) in reasoning about different emotion cues (Study 3), we adopted the representational similarity analysis (RSA) approach. This approach addresses the challenge of assessing the correspondence between different analytical modalities35 (e.g., neural responses vs. behavioral outputs) and constructed representational similarity matrices for four emotion categories (anger, fear, happiness, and sadness) based on indicators from the four tasks. We applied regression analyses to assess how the representational similarity matrices for children’s neural responses and conceptual knowledge, along with age, predict the matrices reflecting their understanding and judgments about emotion cues, while controlling for visual similarities.

Based on the outlined experiments and existing theories on emotion development, we hypothesize the following regarding emotion understanding and its two key processes (1) While we expect children to show a more refined understanding of emotional cues as they age, the influence of perceptual discrimination on this ability is anticipated to decrease, whereas the influence of conceptual knowledge will increase. If these hypotheses are supported, it would suggest that the development of emotional understanding and reasoning shifts from a reliance on perceptual features to a greater dependence on conceptual knowledge. (2) We predicted that preschool-aged children will exhibit less differentiated neural responses to stereotypical facial configurations. If this hypothesis is supported, we also expect that older children will show more distinct neural responses between configurations as compared to younger children. (3) Consistent with the observations that children develop increasingly nuanced conceptual knowledge of emotions, we predicted lower correlations between different emotional concepts with age on the rating task, which would be reflective of older children making clearer distinctions between these concepts.

Results

Study 1: discrimination of facial configurations

Study 1 used an FPVS paradigm to quantify participants’ sensitivity to changes between four types of stereotypical facial configurations. The final sample included useable data from 122 children (5–6 years group: N = 45; 7–8 years group: N = 38; 9–10 years group: N = 39) and 33 young adults (mean age = 20.91). An additional 30 children were tested but excluded from further analysis because of insufficient clean EEG data (see Methods for details). Participants were rapidly shown prototypes of facial configurations commonly associated with emotions (e.g., happiness) at 6 Hz, equating to 6 images per second. For every 5 images (at 6/5 = 1.2 Hz), a face depicting a different facial configuration appeared (e.g., fear; Fig. 1A). This “oddball” design triggers the brain to generate periodic responses at these prespecified frequencies.

Fig. 1: Overview of tasks used to measure different aspects of emotion processing.
figure 1

A The schematic of the FPVS paradigm. The images are presented at a base frequency of 6 Hz, with every fifth (e.g., 5th, 10th, 15th, etc.) image displaying an oddball facial configuration (e.g., the stereotypical configuration of fear), while the rest images expressing a base facial configuration (e.g., the stereotypical configuration of happiness). B Schematic of the conceptual similarity rating task. Participants were asked to rate each emotion on its conceptual relationship with other emotions. C Schematic of the sorting task. Participants were instructed to place the displayed configuration into the house corresponding to the respective predefined emotion. Image credit (children fight over a toy): iStock.com/Nosyrevy. D Schematic of the matching task. Participants were instructed to drag the person located at the bottom to one the top two people expresses the same emotion. Facial images used in the figure were selected from the Tsinghua Facial Expression Database61, with permission for use obtained.

Full size image

We calculated both the signal-to-noise ratio, computed as the power at oddball frequencies (i.e., 1.2 Hz and its harmonics such as 2.4 Hz, 3.6 Hz, 4.8 Hz, and so forth) divided by the average power of surrounding frequency bins, and the baseline-corrected amplitudes, computed by substracting the average power of the surrounding bins from the power at oddball frequencies28,29,36. We averaged responses across all conditions and age groups to identify the oddball frequencies used for analysis, selecting those with z-scores exceeding 1.64 (p value < 0.05). A one-tailed test was used because negative z-scores indicate that the power at the oddball frequency is lower than at surrounding frequency bins, and therefore reflect a lack of signal rather than a meaningful inverse effect. Only harmonics significant across all age groups were included in the analysis (i.e., 3.6 and 4.8 Hz) to minimize bias and ensure fair and reliable comparisons of neural responses across development. Over two-thirds of participants in each age group exhibited significant oddball discrimination responses at the individual level. Details on the calculation of signal-to-noise ratio and baseline-corrected amplitudes can be found in the Methods section. Together, these two measures reflect the spontaneous encoding of brief changes in facial configurations (i.e., perceptual differences between two configurations); the magnitude of this response is proportional to the representational disparities between the two alternating configurations. Although the two measures were correlated (see Supplementary Methods for analyses), we used the signal-to-noise ratio to control for the differences in EEG power among ages due to age-related confounding factors such as changes in skull thickness with age (Supplementary Fig. 2), as it normalizes data relative to adjacent frequency bins, mitigating the decline in EEG power with frequency and across ages25,37,38,39.

An orthogonal behavioral task that required the participants to detect a randomly presented black-and-white image was employed during the experiment. During the presentation, one or two images were randomly presented in black and white, and participants were instructed to press the “space bar” whenever they detected these black-and-white images (Fig. 1A). This task facilitated sustained attention throughout the FPVS session. The results from this task indicate that the participants across ages were thoroughly engaged in the FPVS presentation (Mean accuracy: 0.903, Mean reaction time = 0.64 s).

There was a clear oddball response in the EEG power spectrum at 1.2 Hz and its harmonics (2.4, 3.6, 4.8 and 7.2 Hz, Z-scores > 1.64, p value < .05, one-tailed) across all conditions and age groups (Fig. 2A, B). We calculated an average EEG spectrum across all conditions, separately for each age group. Even the 5–6 years group exhibited a signal-to-noise ratio greater than 1 at 1.2 Hz and its harmonics, indicating that their responses at the odd frequencies were stronger than at the surrounding 16 frequency bins (8 on each side). Furthermore, when examining the oddball responses for each condition separately, the 5–6 years group also exhibited above-threshold neural discrimination (signal-to-noise ratio > 1) across different emotion pairs, including those involving negative facial configurations (i.e., fear–anger, fear–sadness, anger–sadness, Supplementary 7). This demonstrates that even 5–6 year-olds are capable of discriminating between alternating facial configurations not only across valence boundaries but also within the domain of negative emotions.

Fig. 2: EEG oddball responses to changes of facial configurations (i.e., emotion pairs) in the FPVS experiment.
figure 2

A The signal-to-noise ratio spectra averaged across all emotion pair conditions, separately for age groups. The responses at oddball and base frequencies are marked in blue triangles and orange triangles, respectively. B Topographical maps of the averaged oddball responses for different emotion pairs presented separately for different age groups (5–6 years, 7–8 years, 9−10 years, and young adults). Channels with significant responses are marked with a star (Z-scores > 1.64, p-value < .05, one-tailed). C Box plots illustrate the averaged oddball responses at 3.6 Hz and 4.8 Hz for six emotion pairs, presented separately for four age groups. Each box shows the median (center line) and the 25th and 75th percentiles (box bounds), with whiskers extending to the most extreme values within 1.5× the interquartile range, following Tukey’s method. Dots represent individual outliers. Responses were obtained from biologically independent participants (n = 122 children: 5–6 years, n = 45; 7–8 years, n = 38; 9–10 years, n = 39; and n = 33 adults), with no technical replicates. Colors denote different emotion pairs, as indicated in the legend. A repeated-measures ANOVA was conducted to assess the interaction between age group and emotion pair, which revealed no statistically significant effect (F(12.964, 652.521) = 0.702, p = .784, ηp2 = 0.014, 95% CI [0, 0.013]).

Full size image

We next conducted a repeated measures ANOVA to examine the signal-to-noise ratio at the selected oddball harmonics as a function of “emotion pair” (6 levels: happiness-fear, happiness-anger, happiness-sadness, fear-anger, fear-sadness, and anger-sadness), age group (4 levels: 5–6 years, 7–8 years, 9–10 years, and 18–25 years), and ROI (3 levels: left temporal-occipital, right temporal-occipital, middle occipital). For all repeated measures ANOVAs reported, Mauchly’s test was used to assess the assumption of sphericity, and Greenhouse–Geisser corrections were applied where violations occurred. Visual inspection of residuals suggested no substantial deviations from normality. Unless otherwise specified, all p-values value reported are two-tailed and evaluated at a significance level of 0.05. Bonferroni corrections for multiple comparisons were applied to each analysis. There was an interaction between ROI and age, F(4.909, 247.104) = 6.411, p < .001, ηp2 = 0.113, 95% CI [0.037, 0.175], as the age-related differences in oddball responses primarily shown in the central and right regions (Supplementary Fig. 3B). Moreover, significant differences in EEG activity were found across different ROIs (F(1.636, 247.104) = 34.640, p < .001, ηp2 = 0.187, 95% CI [0.105, 0.268]), with the stronger activity also in the central and right temporal cortical regions. The main effect of emotion was also significant (F(4.321, 652.521) = 27.549, p < .001, ηp2 = 0.154, 95% CI [0.103, 0.200]), with heightened neural discrimination responses observed between happiness and most negative emotion categories (Supplementary Fig. 2C). However, the interaction between emotion pair and age was not statistically significant (F(12.964, 652.521) = 0.702, p = 0.784, ηp2 = 0.014, 95% CI [0, 0.013]). The main effect of age (F(3, 151) = 2.329, p = 0.077, ηp2 = 0.044, 95% CI [0, 0.108]) was not statistically significant either. These findings suggest that the ability to discriminate facial configurations has already emerged by 5 to 6 years of age. Further details on the analyses and post hoc results can be found in the Supplementary Methods (Table 8-12).

To determine if the observed oddball responses reflect participants’ processing of changes in facial features (i.e., individual elements such as the shape of the eyes or mouth and the presence of wrinkles), configurations (i.e., the spatial arrangement and relative positioning of facial features, such as the relative positions of eyes, nose and mouth), or both, we tested a new cohort of children (N = 30, Mage = 7.87, SD = 1.71, 14 females) and adults (N = 15, Mage = 21.00, SD = 1.36, 6 females) with both upright and inverted face configurations. This approach leverages the well-established face inversion effect, where inverted faces are known to disrupt high-level processing of facial configurations but leave low-level, image-based cues largely unaffected40. Note that these participants did not take part in other experiments. This analysis revealed that inverted faces elicited much lower oddball responses in discrimination of facial configurations (F(1,41) = 71.806, p < 0.001, ηp2 = 0.637, 95% CI [0.436, 0.742]), especially in the right temporal-occipital region (Supplementary Fig. 4). Thus, the oddball responses observed in Study 1 were primarily driven by the encoding of structural changes in the configurations rather than individual facial features. The detailed results for this section, including orthogonal behavioral task performance, contrasts between discrimination responses of different emotion categories, comparisons of upright versus inverted conditions across age groups and ROIs can be found in the Supplementary Methods.

Study 2: conceptual knowledge of emotions

To evaluate how children’s conceptual understanding of emotions develops, we adapted a conceptual similarity rating task34. We asked the parents to select the top 20 out of 40 words from Books and Freeman34 they believed their child could read or understand. From these selections, we compiled a final list of 20 terms representing thoughts, bodily sensations, and associated behaviors, and verified that all stimulus words had equivalent frequency of use in Chinese contexts (Supplementary Table 2). Children rated how related each word (e.g., “crying”) was to various emotion categories (e.g., “How related is ‘crying’ to anger, fear, happiness, or sadness?”) on a 7-point scale (from 1 = not to all to 7 = extremely). Of note, this task does not assess how well children’s responses align with cultural norms but rather measures the distinctiveness of their conceptual patterns based on their individual experiences within the broader context of Chinese cultural influences. The final sample of this task includes 144 children (5–6 years group: N = 61; 7–8 years group: N = 39; 9–10 years group: N = 44) and 39 young adults (mean age = 20.77).

To evaluate children’s views of the similarity structure between various emotion categories, we calculated Pearson correlations between each pair of emotions using the ratings (20 for each emotion) from the conceptual similarity rating task41,42 (Fig. 3). We conducted a repeated-measures ANOVA to examine these correlations as a function of emotion pair and age. As expected, the main effect of emotion pairs was significant (F(2.187, 391.481) = 402.854, p < 0.001, ηp2 = 0.692, 95% CI [0.645, 0.728]), with participants viewing happiness as most distinct from the negative emotion categories. The ANOVA also revealed a notable interaction between emotion pair and age (F(6.561, 391.481) = 6.925, p < 0.001, ηp2 = 0.104, 95% CI [0.042, 0.150]). Specifically, happiness and negative emotions became conceptually more distinct with age, while negative emotions came to be viewed as more similar. For instance, correlations between happiness and negative emotions were higher in the 5–6 years group compared to young adults (Fig. 3). Conversely, correlations among negative emotions were lower in younger groups (5–6 years and 7–8 years) compared to older groups (9–10 years and young adults; Supplementary Tables 23–25). We examined these age-related changes using regression models (see Supplementary Methods for details). To further support the observed similarity structure in conceptual knowledge of emotions and its changes with age, we employed a multidimensional scaling (MDS) analysis that extracts the lower-dimensional representation of different emotions using Euclidean distances. The MDS results also revealed that with age, the conceptual similarity between negative emotions gradually increased, providing a complementary perspective on the developmental changes in children’s conceptual understanding of emotions (Supplementary Fig. 6).

Fig. 3: Scores of conceptual similarities between emotion pairs for different age groups in Study 2.
figure 3

The Pearson correlation coefficient was employed as the calculation method, where the correlation values span from −1 to 1. Lower numerical values signify a reduced conceptual similarity between two emotions. Each box represents the median (center line) and the 25th and 75th percentiles (box bounds), with whiskers extending to the most extreme values within 1.5× the interquartile range, according to Tukey’s method. Dots represent individual outliers. Data were derived from biologically independent participants, with no technical replicates (n = 144 children: 5–6 years, n = 61; 7–8 years, n = 39; 9–10 years, n = 44; and n = 39 adults), with no technical replicates. A repeated-measures ANOVA revealed a significant interaction between emotion pair and age group (F(6.561, 391.481) = 6.925, p < 0.001, ηp2 = 0.104, 95% CI [0.042, 0.150]).

Full size image

Additionally, we performed a regression analysis to examine whether conceptual knowledge could predict perceptual discrimination, based on evidence from adult studies suggesting that perceptual discrimination can be influenced by higher-level conceptual processes, even in the absence of conscious awareness43,44. This analysis revealed no statistical significant association between perceptual discrimination and conceptual knowledge in children (Supplementary Fig. 13A).

We also utilized a parent-reported word comprehension questionnaire to investigate whether the above developmental changes in conceptual knowledge are associated with children’s emotional vocabulary. Correlation analyses revealed no statistical significant associations (see Supplementary Methods for details).

Study 3: assessment of emotion judgments

Study 3 employed a sorting task33 and a matching task24,34, to respectively measure children’s understanding of emotion across contexts. In the sorting task, children were presented with emotion-related scenes (referred to as “houses”) and asked to categorize stereotypical facial configurations into different scenes (Fig. 1C). In the matching task, children were instructed to match emotions conveyed by different configurations in a forced-choice paradigm (i.e., “Please match the two people who have the same feelings”). Mouse trajectories towards the selected image was recorded to capture the dynamic decision-making process (Fig. 1D). The same four categories of facial configurations from Study 1 and 2 were used in Study 3. The final sample for these two methods includes 145 and 144 children, respectively, evenly distributed across ages, and 39 young adults.

For the sorting task, we calculated the number of facial images that were sorted into “incorrect” emotion categories (e.g., facial configurations that should be categorized as “happiness” but were sorted by a child as representing “sadness”). For the matching task, we calculated a maximum deviation (MD) parameter, a well-validated index of the co-activation of unselected category responses during perception34,45,46. MD reflects the maximum perpendicular deviation from an ideal straight line between the mouse starting and ending points. Greater MD values between two emotion categories in the matching task indicate greater similarity in judgment and understanding between those categories.

The results from these two behavioral tasks revealed that children exhibited consistent patterns across the two tasks, with younger children initially demonstrating the ability to distinguish between positive and negative emotions before progressing to be able to understand and categorize within the negative categories. A repeated measures ANOVA in the sorting task revealed the main effect of age (F(3,180) = 11.512, p < 0.001, ηp2 = 0.161, 95% CI [0.066, 0.247]) and emotion pair (F(3.301, 594.137) = 260.936, p < 0.001, ηp2 = 0.592, 95% CI [0.544, 0.629]). Younger children (5–6 age group) showed greater confusion in judging emotions compared to older groups, and confusion was consistently higher among negative emotion pairs than happiness-negative pairs across all age groups (Supplementary Fig. 7A). In the matching task, an interaction between age and emotion pair (F(13.866, 827.327) = 2.847, p < 0.001, ηp2 = 0.046, 95% CI [0.009, 0.060]) indicated that younger children exhibited greater MD values between negative emotions and happiness-negative pairs. However, these differences diminished with age- except for anger-sadness, which consistently elicited the highest values (Supplementary Fig. 7B). Detailed statistical analyses are provided in the Supplementary Methods (Supplementary Tables 29–35).

RSA: relative contributions of discrimination and conceptual knowledge to emotion judgments

We conducted RSA to investigate how the ability to spontaneously discriminate facial configurations (measured through the FPVS task, Study 1) and conceptual knowledge about emotion (assessed using the conceptual similarity rating task, Study 2) predict children’s emotion judgment performance in the sorting and matching tasks. To achieve this, we constructed representational similarity matrices for each experiment using the following indicators: (1) the reciprocal of the average signal-to-noise ratio at oddball frequencies (3.6 Hz and 4.8 Hz) across three ROIs (left temporal-occipital, right temporal-occipital, middle occipital) from the FPVS paradigm, reflecting similarity in perceptual responses to facial configurations; (2) pearson correlation coefficients between children’s ratings for emotion categories, indicating the relationship between their conceptual knowledge of those emotions; (3) confusion scores from the sorting task and MD values from the matching task, representing the similarity in children’s judgment and understanding of emotion categories (Fig. 4A; Supplementary Fig. 9). We subsequently conducted two generalized estimating equations (GEEs) regression analyses34 to predict the similarity matrices for confusion scores and MD values, respectively, using the similarity matrices for perceptual discrimination and conceptual knowledge. In these models, age was included as an additional predictor to explore potential interactions with the independent variables. A visual similarity matrix for the stereotypical facial configurations, created using the Facial Action Coding System (FACS), was also included to control for any physical resemblances among the configurations. This coding also ensured that the study measured the contributions of perceptual discrimination and conceptual knowledge to emotion judgments, rather than effects driven by low-level physical similarities47,48 (Fig. 4A and Supplementary Fig. 10). All GEE models were estimated using an exchangeable working correlation structure with robust standard errors, which provides reliable inference even when the assumed correlation structure is misspecified. Visual inspection of residual plots did not reveal major violations of normality or homoscedasticity. Unless otherwise specified, all reported p-values value are two-tailed and evaluated at a significance threshold of 0.05. FDR correction was applied to all analyses. The GEE model predicting children’s emotion judgments in the sorting task revealed a significant interaction between age and perceptual discrimination, B = –0.020, SE = 0.008, Z = –2.678, p = 0.017, 95% CI [–0.035, –0.005]. To explore this interaction, we conducted simple slope tests by dividing participants into three groups: the older (Mean +1 SD), middle-aged (Mean −1 SD - Mean +1 SD), and younger children (Mean −1 SD). Simple slope tests revealed that perceptual discrimination predicted emotion judgment in the younger and middle-age groups (younger children: B = 0.183, SE = 0.047, Z = 3.894, p < 0.001, 95% CI [0.091, 0.276]; middle-aged: B = 0.143, SE = 0.027, Z = 5.206, p < .001, 95% CI [0.089, 0.196]). However, the effect was not statistically significant in the older children (B = 0.030, SE = 0.026, Z = 1.170, p = 0.242, 95% CI [–0.020, 0.080]). The estimated values demonstrated a declining trend with age, indicating that the prediction of the ablity to discriminate facial configurations on emotion judgment decreases with age (Fig. 4B, C). The main effect of perceptual discrimination (B = 0.254, SE = 0.066, Z = 3.837, p < .001, 95% CI [0.124, 0.383], Supplementary Fig. 11A) and conceptual knowledge (B = 0.085, SE = 0.036, Z = 2.331, p = .035, 95% CI [0.014, 0.156], Supplementary Fig. 11B) was significant, but conceptual knowledge does not statistically interact with age (B = 0.001, SE = 0.004, Z = 0.121, p = 0.904, 95% CI [–0.008, 0.009]).

Fig. 4: The schematic of the RSA approach.
figure 4

Panel A and results for the predictive modeling of the emotion judgment performance in the sorting B, C and matching D, E tasks. A Illustrating the RSA GEE model of how perceptual discrimination and conceptual knowledge predict children’s emotion judgment with age,while controlling for FACS. Grand averages matrices are used for illustration here, whereas individual matrices were emplyed for analysis; (B). The interaction between age and perceptual discrimination in the prediction of emotion judgment. Statistical analyses were performed using GEE, with two-sided tests and p-values value adjusted for multiple comparisons using the FDR correction; (C). Linear prediction slope of perceptual discrimination by age. Each point represents the group-level mean, with error bars indicating ±1 SEM based on biologically independent participants; (D). The interaction between age and conceptual knowledge in the prediction of emotion judgment. Statistical analyses were performed using GEE, with two-sided tests and FDR-corrected p values; (E). Linear prediction slope of conceptual knowledge by age. Each point represents the group-level mean, with error bars indicating ±1 SEM based on biologically independent participants.

Full size image

The GEE model predicting children’s emotion judgments in the matching task revealed a significant interaction between age and conceptual knowledge (B = 0.026, SE = 0.010, Z = 2.572, p = 0.024, 95% CI [0.006, 0.046]). Conceptual knowledge predicts emotion judgment in all three groups, with the predicted values progressively increasing (younger children: B = 0.081, SE = 0.039, Z = 2.076, p = 0.038, 95% CI [0.005, 0.157]; middle-aged: B = 0.172, SE = 0.024, Z = 7.098, p < 0.001, 95% CI [0.125, 0.220]; older children: B = 0.239, SE = 0.042, Z = 5.667, p < .001, 95% CI [0.156, 0.322]). These findings suggest that the association between conceptual knowledge and emotion judgment strengthens with age (Fig. 4D, E). Additionally, there was a main effect of perceptual discrimination (B = 0.491, SE = 0.177, Z = 2.777, p = 0.019, 95% CI [0.144, 0.837], Supplementary Fig. 11C). However, the main effect of rating was not statistically significant (B = -0.048, SE = 0.081, Z = -0.591, p = .554, 95% CI [-0.206, 0.110], Supplementary Fig. 11D), nor were the interaction between perceptual discrimination and age (B = -0.039, SE = 0.021, Z = –1.858, p = 0.089, 95% CI [–0.081, 0.002]).

These regression models were applied to children’s data due to the continuous nature of age as a variable and the substantial age gap between children and young adults (18–25 years). For reference, we also analyzed young adults’ data. The GEE models for both emotion judgment tasks (sorting and matching) consistently revealed that only conceptual knowledge predicted adults’ emotion judgment (sorting: B = 0.061, SE = 0.006, Z = 9.482, p < 0.001, 95% CI [0.049, 0.074]; matching: B = 0.114, SE = 0.022, Z = 5.177, p < 0.001, 95% CI [0.071, 0.157]). This finding also underscores the increasing importance of conceptual knowledge in emotion judgments, which becomes the dominant factor as individuals progress from childhood to adulthood. Additional details about the results of these GEE models for children and adults can be found in the SI.

Discussion

The present study aims to explore how the ability to discriminate facial configurations and conceptual knowledge about emotions contributes to emotional development. Findings from Studies 1 and 2 reveal that both perceptual and conceptual processes undergo developmental changes from preschool age to early adolescence. However, consistent with our primary hypothesis, the RSA demonstrates a shift in the contributions of these respective processes to children’s understanding of emotions: The influence of perceptual discrimination on understanding emotion cues decreases with age, while the influence of conceptual knowledge gradually increases. This shift underscores the idea that emotional understanding is not static or uniformly continuous. Rather, it appears to be constructed through the accumulation and integration of conceptual knowledge combined with a shift in underlying mechanisms, shaped by children’s cultural and social experiences.

The finding from Study 1 that children demonstrate proficient discrimination of stereotypical facial configurations from preschool age is contradictory to our hypothesis made based on extant pediatric EEG/ERP literature. Previous studies on child N290 and N170 components did not observe a consistent pattern for perceptual discrimination between facial configurations before early adolescence6,7,8. The key finding in the current study is likely attributable to the FPVS approach employed, which offers a more objective, higher signal-to-noise ratio, and less interference-prone measure of structural encoding, potentially making it a more sensitive metric for evaluating children’s spontaneous discrimination of facial configurations25,26. Additionally, the observation that smiling and “negative” facial configurations were better discriminated than distinctions among negative configurations, along with increased discrimination response with age in the right temporal-occipital region, suggests that children’s perceptual processing of these configurations improves with experience and exposure to them. This developmental pattern aligns with previous research on children’s behavioral outputs, which shows a broad-to-differentiated progression in their ability to distinguish and categorize stereotypical facial configurations5,48,49.

The changes in conceptual relationship between emotions observed in Study 2 suggest a growing comprehension of the complexity of emotions across different contexts. While happiness and negative emotions becomes more conceptually dinstint in children’s mind, they conceptualize various negative emotions as more similar to each other with age. In adults, the likelihood of experiencing emotions with the same valence simultaneously in complex situations is indeed higher, with a mix of emotions within the same valence being considered a more cognitively complex experience50,51. In this regard, the current findings imply that as children’s emotional concepts strengthen, they increasingly recognize the complexity and interrelatedness of negative emotions across different contexts, acknowledging that multiple emotions can coexist within a single context52. The acquisition of emotional concepts is likely driven by children’s increasing exposure to diverse emotional situations and predictive patterns as they grow older, which makes them more attuned to cultural nuances20,53,54. As children grow older, their language skills also improve drastically, allowing them to learn labels and refine their conceptual frameworks for emotions4,55. While differences in cognitive and linguistic abilities could potentially influence task performance in younger children, we consider it unlikely that such limitations were the primary drivers of the observed developmental pattern, given the low rate of skipped items and the consistency in conceptual structure across age groups.

In contrast, in Study 3, we observed an increasing ability to distinguish between negative emotional categories during emotion judgments. This finding is consistent with literature24,33,56 but may seem contradictory to the findings in Study 2. We interpret the observed differences as reflecting different facets of emotional development across tasks. As discussed, findings from Study 2 reflect higher-order categorization, where children recognize shared features and contexts across negative emotions (e.g. anger and fear can co-occur in situations involving perceived threat), indicating a more abstract and interconnected conceptual framework. At the same time, the improved behavioral differentiation in the emotion judgment tasks indicates that with age, children become better at differentiating the emotional nuances of these configurations in real-world contexts – for instance, recognizing that fear typically involves withdrawal and uncertainty, whereas anger is often associated with approach and blame.

The acquisition of emotional concepts is crucial for developing a sophisticated understanding of emotions. The data in RSA revealed a shift from relying on the ability to discriminate facial movements to depending on conceptual knowledge in children’s understanding of emotions. This developmental change aligns with the theory of constructed emotion, which posits that early in life, children rely on broad affective cues such as valence, arousal, and physical signals to infer others’ feelings. While perceptually salient, these cues are increasingly interpreted through conceptual knowledge as children accumulate more emotion-evoking experiences and undergo cognitive development1,18,57. The current finding that only conceptual knowledge contributes to adult participants’ emotion judgments in behavioral tasks further supports the notion that the reliance on conceptual knowledge increases with age. This may be because perceptual discrimination reaches a ceiling in adults, resulting in efficient and consistent neural responses that no longer predict behavioral performance. Alternatively, adults’ emotion understanding relies more heavily on top-down conceptual processes34,58. In sum, children’s evolving conceptual frameworks of emotions enable them to interpret emotional cues in increasingly flexible and context-sensitive ways, reflecting a more nuanced understanding of how emotions are perceived and inferred within specific cultural and contextual influences.

It’s worth noting that we observed that the results from the RSA models predicting children’s emotion judgments varied between the sorting and matching tasks. Specifically, the interaction of conceptual similarity and age was significant in the matching but not in the sorting task, whereas the interaction of perceptual discrimination and age was significant in the sorting but not in the matching task. This discrepancy may arise from the distinct cognitive demands of the two tasks59. The sorting task requires participants to identify and categorize facial configurations into appropriate senarios described with scripts and labels, which necessitates a high level of semantic understanding of the information60. This high demand for conceptual knowledge may overshadow developmental differences in conceptual knowledge, as we found a main effect of conceptual knowledge on emotion judgment in the sorting task. Considering that the production and comprehension of emotion labels might pose challenges for younger children, future research may incorporate more non-verbal tasks that allow us to better assess children’s conceptual knowledge about emotions (also see ref. 5). On the other hand, the matching task may depend on perceptual processing as a prerequisite for emotion judgment, explaining the significant main effect of perceptual discrimination that does not vary with age. Thus, an alternative explanation is that the two tasks engage distinct underlying mechanisms for emotion judgment, differentially tapping into the perceptual and conceptual processes. This suggests that no single cohesive mechanism operates across all contexts, and emotion judgment may depend on distinct and complex processes as needed.

We also recognize that the current study has certain limitations, which point to promising directions for future research. The EEG approach used in the current study lacks the spatial resolution to provide detailed information about underlying brain networks. Therefore, we cannot determine how the brain regions are involved in face structural encoding versus emotion conceptual processing. Subsequent analyses could consider using fMRI to pinpoint specific brain networks involved in emotion understanding. However, it is worth noting that the current EEG approach has its advantages in probing the temporally early structural encoding of faces and is more child-friendly, especially considering the additional behavioral tasks participants need to complete. Additionally, future research could explore whether signal-to-noise ratio or baseline-corrected amplitude in EEG FVPS designs is more suitable for different contexts and age groups.

Because this study is cross-sectional, we could not assess longitudinal trajectories associated with individual experiences, but this is an approach that is ripe for future research. Finally, to maintain consistency with most previous research with children and to ensure robust data collection and maintain the attention of young children throughout the experiment, we used static images of posed facial configurations from a small set of emotion categories. But a next phase of research needs to expand to more naturalistic stimuli that includes contextual information. In the current study, we included only one positive emotion category in order to reduce task complexity and maintain feasibility for younger participants. Future studies should incorporate a broader range of both positive and negative emotion categories to conduct a more comprehensive analysis. We anticipate that using more naturalistic stimuli that represent a broader array of possible instances of emotional experiences will amplify the influence of children’s conceptual knowledge on emotion reasoning4.

In conclusion, understanding others’ emotions is crucial for human interaction and adaptation, yet the intricacies of how individuals develop an ability to make these types of inferences remains to be fully understood. A developmental shift from making perceptual discriminations to reliance on learned conceptual knowledge makes intuitive sense and fits with existing data about the emergence of expertise in other domains of behavior. This developmental change may be shaped in part by children’s emotional experiences and their socialization into ways of expressing and interpreting emotion. The findings suggest that fostering conceptual knowledge about emotions, rather than solely focusing on perceptual cues, may help children better navigate social interactions, as emotion recognition becomes increasingly reliant on the social and cultural nuances of emotions.

Methods

All studies were approved by the Research Ethics Committee of the School of Psychological and Cognitive Sciences at Peking University, and conducted in accordance with relevant ethical regulations. Written informed consent was obtained from all adult participants and from the guardians of child participants, who also provided financial compensation post-experiment.

Participants were recruited through poster promotion and word-of-mouth. The same cohort took part in all three studies. Data were collected from 152 children (mean age = 7.73, SD = 1.85, age range = 5.00–11.75, 71 females) and 41 adults (mean age = 20.78, SD = 1.82, age range = 18–25, 22 females). Gender was determined by parental report for child participants and by self-report for adult participants. The sample included approximately equal numbers of males and females across age groups; however, gender was not considered in the study design or included in the data analysis. Power analysis for repeated measures ANOVA conducted using G*Power indicated that a minimum of 24 participants per age group would be sufficient to detect medium effects. To ensure adequate power and account for potential data loss, we aimed to recruit approximately 40 participants per group. Participants were matched for gender across all age groups and none of the participants had a history of psychiatric or neurological disorders. Both the children and adults participated in the FPVS, conceptual similarity rating, matching, and sorting tasks. The final sample size for each task is reported in the following sections. Additionally, the guardians of the children were asked to complete a questionnaire concerning their children’s comprehension of emotional words.

Study 1. Discrimination of facial configurations

Participants

The final sample of the FPVS task included 122 children (61 female, Mage = 7.89, SD = 1.84, range 5.00–11.75) and 33 adults (15 female, Mage = 20.91, SD = 1.89, range 18–25) were included. Additionally, 15 adults (6 female, Mage = 21.00, SD = 1.36, range 18 – 22) and 30 children (14 females, Mage = 7.87, SD = 1.71 range 5.00–11.75) completed an upright-inverted facial configurations comparison task as a complementary experiment of study 1. Data from 30 children and 8 adults were excluded from the analysis due to the following reasons: 8 children exhibited excessive noise in their raw data which cannot be deleted in the further analysis, 6 children had incomplete data for at least one condition, and 16 children, along with 8 adults, failed to complete FPVS task.

Stimuli

Face images of eight females and eight males embodying stereotypical facial configurations (happiness, fear, anger, sadness) were extracted from the “Tsinghua facial expression database”61. The selection of the four emotion categories was driven by both developmental and methodological considerations. Developmentally, these emotions undergo significant improvement in discrimination and understanding from preschool through early adolescence56. Methodologically, limiting the emotion categories helped maintain young children’s attention and ensured robust data collection. The images were equalized for luminance and contrast using the shinetoolbox (https://osf.io/auzjy/). The images were presented at the center of a 27-inch LED screen (60 Hz refresh rate, resolution: 1920×1080 pixels) against a white background (255 / 255,255 / 255,255 / 255) to ensure prominent contrast at an approximate viewing distance of 60 cm, subtending a large part of the visual field of participants (12 × 13.61° of visual angle).

Procedure

The stimuli were presented at a base rate of 6 Hz (six images per second). At this rate, each stimulus lasts about 167 ms (1 s /6). One basic facial configuration (e.g. stereotype of happiness) was presented at the base rate (6 Hz) and a change of configuration (e.g. stereotype of fear) was introduced every 5 stimuli, resulting in an oddball configuration frequency of 1.2 Hz. To minimize low-level repetition effects, the size of faces varied randomly between 85% and 115% at every stimulus-onset. The implementation of square wave stimulation was undertaken, wherein the faces were fully unveiled (i.e. 100% contrast) for 60% of their exposure duration and subsequently became invisible for the remaining 40% of the duration28 (Fig. 1A). It’s important to highlight that our use of square-wave contrast in this task marks a departure from previous applications of this paradigm, which typically employed sinusoidal contrast modulation28,29. The square-wave contrast is used here for optimizing the measurement of the temporal dynamics in the facial discrimination response, particularly enhancing the clarity of stimulus onset62 and is not expected to change the magnitude and nature of the response63. Each stimulation sequence presented the face of one individual. The association between facial configurations and identities has been counterbalanced between participants.

The task included four distinct emotions: happiness, fear, anger, sadness, culminating in a total of 12 blocks, wherein each emotion pair (e.g., happiness-sadness) was presented for 2 blocks, with one emotion (e.g., happiness) as the basic facial configuration and the other one (e.g., sadness) as the oddball configuration for one block, and vice versa for the other block, resulting in 6 conditions for different emotion pairs.

Participants, seated in a dimly lit room, completed two blocks of presentation for each condition. Considering the tolerance of the preschoolers to the duration of EEG experiments, each block consisted of two trials, with each trial of stimuli was presented continuously for 20 s for children aged 5–6 years and 30 s for the older children and young adults. In the analysis of FPVS data, we extracted the first 20 s of data from both the children aged 7–10 years and the young adults to ensure consistency in data length across participants. Additionally, we conducted a separate analysis using 30 s of data for the same groups, and the outcomes to be consistent with those derived from the 20-s duration (see Supplementary Methods for details; Supplementary Fig. 14). To ensure the participants were engaged in the presentation, they were instructed to press the spacebar when noticing a color change of the facial stimuli. We additionally included a control condition for an external cohort of 15 adults, in which we presented both upright and inverted faces.

Data preprocessing

EEG was recorded from a 128-channel HydroCel Geodesic Sensor Net (HGSN) that was connected to a NetAmps 400 amplifier (Electrical Geodesic Inc., Eugene, OR). Channel impedances were kept at or below 50 kΩ and signals were sampled at 500 Hz. All EEG processing steps were carried out using Matlab R2020a (The Mathworks). EEG data were first notch filtered to eliminate 50 Hz and 100 Hz power line interference and then bandpass filtered at 0.1–40 Hz (Butterworth filter, second order). Noisy channels identified by FASTER64 were interpolated using a spherical spline interpolation if there were fewer than 18 (15%) electrodes marked8,65. EEG data were then rereferenced to a common average reference before further analyses.

The 20-s EEG data were further segmented into 10 s epochs with 50% overlap for each trial. The segments within each trial were averaged to generate a single time-domain epoch. A Fast Fourier Transform (FFT) was applied to the time-down average epoch to extract the amplitude (power) spectrum for all electrodes with a frequency resolution of .1 Hz (1 / 10 s). Z-scores were computed for each frequency bin as in previous studies, using the mean and standard deviation of the 16 surrounding bins28,66 (10 bins on each side, excluding the 2 immediately neighboring bins and the 2 most extreme values). We used the Z-scores at the base frequency and its harmonics (6, 12, 18, 24 Hz, etc.) to further remove artificial epochs, as the amplitude at these base frequencies is associated with the engagement (looking) of the participants and data quality (e.g., limited body and eye movements, fewer artifacts, etc.). The epochs were removed if the averaged Z-score on the base frequencies was smaller than 2.33 (ps < .010, one-tailed). The “clean” epochs survived from artifacts rejection were averaged for each condition for further analysis (5–6 age group: 12 epochs deleted, 7–8 age group: 3 epochs deleted, 9–10 age group: 4 epochs deleted, young adult: 0 epochs deleted).

Frequency analysis

The individual data were grand-averaged across facial configurations conditions, separately for different age groups. A Fast Fourier Transform (FFT) was then applied to the averaged epoch for each condition to extract the amplitude (power) spectrum with a frequency resolution of 0.1 Hz. The amplitude spectrum was subsequently baseline-corrected by subtracting the mean amplitude of the surrounding 16 bins from the amplitude of each frequency bin. Z-scores were calculated in the same way as how they were calculated for individual epoch. Harmonics were considered as significant when the Z-scores were above 1.64 (ps < 0.05, one-tailed). For analyses investigating the response to different emotions, we calculated the average signal-to-noise ratio on significant harmonics across age groups and conditions. Signal-to-noise ratio is expressed as the amplitude value divided by the average amplitude of the 16 surrounding frequency bins. Signal-to-noise ratios of significant harmonics were then averaged together, as they all corresponded to the same underlying phenomenon66,67,68. Based on previous studies using a similar paradigm28,36,66, subsequent analysis focused on 3 regions of interest: left temporal-occipital region(TO_L), middle occipital region(M_O), and the right temporal-occipital region (TO_R) (Supplementary Fig. 1).

Study 2. Conceptual knowledge of emotions

Participants

A total of 144 children (Mage = 7.76, SD = 1.86, range 5.00–11.75, female 71) and 39 young adults (Mage = 20.77, SD = 1.87, range 18–25, female 20) completed the conceptual similarity rating task with usable data.

Stimuli

The conceptual similarity rating task involved 20 terms on thoughts, bodily sensations, and associated behaviors (e.g., “crying”, “turning away”) related to the four emotions (happiness, fear, anger, sadness). These terms were centrally presented on the center of a 13.6-inch LED screen (resolution: 2560 × 1664 pixels) against a gray background (128 / 255,128 / 255,128 / 255) to ensure prominent contrast at an approximate viewing distance of 60 cm, subtending a large part of the visual field of participants (12×13.61° of visual angle).

Procedure

The stimulus presentation and response collection were programmed using the Psychopy v2021.2.3 toolbox while the participants were seated in a quiet and child-friendly behavioral testing room. In the conceptual similarity rating task, participants were instructed to rate each of the 20 items connected to the four emotions in a randomized order. Each was rated for its relevance to the related category on a seven-point scale (e.g., “On a scale from 1 = not to all to 7 = extremely, e.g., how related is ‘crying’ to the emotion Fear?”). This process included 80 trials in total, i.e., four trials for each item. If a child was struggled to understand an item, they could request clarification. The experimenter would skip a particular item if the child still found an item incomprehensible after the explanation. On average, 5.33% of the items were skipped per child in the 5–6 years group, and 2.50% of the items were skipped in the 7–8 years group. No items were skipped for 9−10 years group or young adults.

Analytic approach

For each term, participants provided ratings reflecting its relationship with each of the four emotion categories. Based on these ratings, we computed Pearson correlation coefficients between emotion pairs (e.g., anger and sadness) across the 20 terms for each participant. These correlations served as an index of the conceptual similarity between emotions.

Study 3. Assessment of emotion judgments

Participants

The final sample for the matching task included 144 children (Mage = 7.76, SD = 1.86, range 5.00–11.75, female 71) and 39 young adults (Mage = 20.77, SD = 1.87, range 18–25, female 20). The final sample for the sorting task included 145 children (Mage = 7.73, SD = 1.84, range 5.00–11.75, female 70) and 39 young adults (Mage = 20.77, SD = 1.87, range 18–25, female 20).

Stimuli

We utilized the identical pictorial stimuli as those adopted in Study 1, while maintaining consistency in presentation settings with Study 2. The emotional intensity of the images used in the sorting task was calibrated between 60% and 100% using the “Morph Age” software to avoid ceiling effect in behavioral performance.

Procedure

We adopted the sorting task used in a previous study33. Specifically, four “cartoon houses” were sequentially presented to the participant on the monitor, with a scenario picture (ice cream, wolf, broken toy, or quarrel) drawn the roof of each house to identify the houses (respectively happy, fearful, sad or angry emotional scenarios). The presentation of each house was associated with a uniquely emotional narrative in their native language (translated as follows):“The people in this house are feeling really (1) scared after looking out the front window and seeing a wolf; (2) happy because their parents said they can eat ice cream; (3) sad because their toy is broken; (4) angry because other kids are taking their toys away.” The narratives were pre-recorded and audibly conveyed to the participants via the computer system. After reviewing the four scenarios, the experimenter proceeded to introduce the task to the participants: “I have a whole lot of people right here and I need you to look at each person and think about how they’re feeling, and put them in the house that matches.” Participants were asked to use the touchpad of the computer to sort different facial emotion images into corresponding houses. The task proceeded until all 64 facial stimuli were sorted.

In the matching task, participants were asked to manipulate the computer touchpad to pinpoint one of the two emotions (i.e., response options), located at the left or right top corner respectively, that matched the emotion of the individual depicted at the bottom of the screen. The three faces used in each trial possessed distinct identities.

Analytic approach—sorting task

To measure the confusion scores between two emotions, the number of times each emotion was misclassified as another emotion was calculated (e.g. the number of times facial configurations images representing “sadness” were misclassified as ‘fear’ and vice versa was counted and divided by the total number of images of these two emotions to represent the confusion of these two emotions). Additionally, participants’ ability to judge different emotions was reflected by calculating the accuracy of their categorization of emotions.

Analytic approach—matching task

The coordinates of the mouse along the x and y axes over time were recorded for each trial. After excluding incorrect choice trials, trajectories were normalized into 101 time bins using linear interpolation and rescaled into a coordinate space with [−1.0,1.5] at the top left and [1,0] at the bottom right, leaving [0,0] at the start location. MD of each mouse trajectory towards the unselected response option on the opposite side of the screen was calculated as the maximum perpendicular deviation from an idealized straight line between its start and end point. MD has been validated in previous research as an index of coactivation between competing response options during decision-making, rather than a measure of raw motor accuracy or speed46. The MD of the participants in the incorrect response trials was considered as 1, and a weighted average with the MD in the correct trials was used as an index of perceptual similarity between the two facial emotions. Additionally, participants’ ability to judge each type of emotion was calculated.

RSA: relative contributions of discrimination and conceptual knowledge to emotion judgments

Participants

The final sample for the RSA analysis consisted of 122 children (61 female, Mage = 7.89, SD = 1.84, range 5.00–11.75) and 33 adults (15 female, Mage = 20.91, SD = 1.89, range 18–25). These participants were selected because they had valid FPVS and behavioral data, ensuring that the RSA analysis was based on reliable and comprehensive datasets.

RSA approach

We applied representational similarity analysis (RSA) techniques to integrate the relationship between various experimental paradigms. This approach was utilized to evaluate the predictive impacts of structural encoding and conceptual knowledge on emotion judgment. The 10 unique similarity values under the diagonal of the 4 × 4 structural, conceptual, visual, and representational similarity matrices of emotion judgment were submitted to multiple regression. To account for repeated measurements within subjects, we used a multilevel regression approach with generalized estimating equations (GEEs).

Visual controls

The Facial Action Coding System (FACS) coding was conducted to control the physical resemblances among facial configurations in the RSA34. The FACS is an anatomically based technique for the quantitative analysis of facial muscle organization, allowing for the decomposition of facial configurations into individual components (AUs)47. An independent certified FACS coder, unaware of the research hypotheses, coded the facial configuration images stimuli used in the study. This process resulted in a FACS representational similarity matrix for all kinds of expressions, which was used as a covariate in the RSA.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.