The Meta-analyses Of Observational Studies in Epidemiology (MOOSE) Checklist was followed throughout the meta-analysis. Moreover, most recent recommendations for a neuroimaging meta-analysis were followed through the process (when applicable) (Müller et al., 2018). The literature search was conducted using PsycINFO, PubMed, and Web of Science (17 January 2022). The search was directed to the fields of title and abstract, and we set no restrictions regarding publication date, language, number of citations, or publication status (see Supplementary methods for the search terms).

A PRISMA flowchart on the article selection process can be found in Figure 1. After removing duplicates, all identified studies were screened on the basis of title and abstract and classified as eligible/ineligible for this meta-analysis. Thereafter, the eligible full-text articles were screened more precisely on the basis of the exclusion and inclusion criteria. Besides of original studies, the reference lists of all meta-analyses and reviews (identified by the search terms) were manually checked for any additional eligible studies. In case some necessary information was missing, the authors of the original studies were contacted (Supplementary Table 4 describes the studies about which we received additional details from the authors). Included studies are listed and described in Supplementary Tables 1, 2. Literature search was conducted by AS (Ph.D in psychology; Ph.D in medicine; and Ph.D in educational sciences). In order to increase transparency of the literature search, we have provided additional descriptive information of the included studies in Supplementary Table 3; the statistical contrasts selected from each study in Supplementary Table 4; and the primary reasons for excluding articles in Supplementary Table 5 (exclusion on the basis of title and abstract) and in Supplementary Table 6 (exclusion on the basis of full-text version). Further, the full statistical data (e.g., all the coordinates and statistical estimates) can be requested from the corresponding author.

We had the following inclusion criteria: an original peer-reviewed fMRI article; n ≥ 10; adult sample (mean age > 18 years); a non-clinical sample (the subjects did not have any reported diseases or medications); a study design including exposure to ethnic in-group and out-group; a stimulus material including visual exposure to ethnic in-group and out-group members, including their faces (so that skin color could be approximately obtained); study design included face visual material of ethnic in-group and out-group (regardless whether the study topic was in/out-group favoritism/degoration or other ethnic perceptual or cognitive processing); the coordinates had been reported in the Talairach Atlas (Tal) or the Montreal Neurological Institute (MNI) space; T or Z statistics or p values of the observed BOLD response toward ethnic in-group vs. out-group members were available; the study reported results of whole-brain analyses or analyses with very limited masking. More precisely, we did not include small-volume corrected results or region-of-interest (ROI) based results because a previous review recommended to only include studies with whole-brain results available (Radua and Mataix-Cols, 2012). We allowed masking if the study had proved that masking did not limit regions where significant in-group vs. out-group differences could be obtained; if masking was limited to regions (e.g., the brain stem or occipital lobe) that are not theoretically relevant in the context of inter-group bias; or if masking was limited to such brain regions that, in preliminary analyses, had been shown to correlate with the mental process under investigation (e.g., in-group vs. out-group contrast during facial processing was investigated in regions that were significant in face vs. non-face contrast). The exclusion criteria (mostly consisting of the opposite criteria to the inclusion criteria) can be found in Supplementary methods.

Data extraction is described in Supplementary methods.

The meta-analyses were conducted using the Seed-based d Mapping (SDM) software (version 6.21). Further meta-analytical details of the SDM software are available elsewhere (Radua et al., 2012, 2014; Albajes-Eizagirre et al., 2019).

During the data pre-processing, the lower and upper bounds of the possible effect-size values of the studies were estimated (full anisotropy = 1.0, full width at half maximum of Gaussian kernel = 20 mm, voxel size = 2 mm). Next, a mean analysis was conducted, representing the weighted mean difference in the fMRI activity (i.e., BOLD responses) toward in-group vs. out-group. There is evidence that sufficient statistical stability can be reached with 20 imputations (Radua et al., 2012). To be on the safe side, we used 50 imputations. The statistical threshold consisted of an uncorrected voxel p value <0.005, cluster extent ≥10 voxels, and SDM-Z < 1, in accordance with previous recommendations (Radua et al., 2012). Regarding other analytical details, we used the default settings of the SDM software that are described in other sources more precisely (Radua et al., 2012, 2014; Albajes-Eizagirre et al., 2019). All the meta-analyses were adjusted for age and gender.

The literature search process is illustrated in Figure 1. The literature search resulted in a total of 50 eligible original datasets that had been published between 2008–2021 and included altogether 1,211 subjects (M_age = 24.5 years, 50.1% female). There were 39 datasets with samples including majority members (n = 945, M_age = 24.1, 48.9% female, a total of 411 brain loci with significant differences in BOLD responses toward in- vs. out-group in the original studies); and 11 datasets with samples including minority members (n = 246, M_age = 26.2, 54.6% female, 51 loci). Additionally, we found altogether 21 datasets with White target individuals (n = 477, M_age = 23.8, 49.0% female, 89 loci); 21 datasets with Black target individuals (n = 557, M_age = 26.4, 48.0% female, 297 loci); and 8 datasets with Asian target individuals (n = 204, M_age = 32.8, 57.2% female, 89 loci). Most minority members were newly moved immigrants: they had been living only a short time in the country, and many of them were students of an abroad study program or had not had close contacts with the ethnic out-group before the study (see Supplementary Table 5).

Further details about the included studies can be found in the Supplementary material. Supplementary Table 1 presents studies categorized on the basis of subjects’ ethnic majority vs. minority status; Supplementary Table 2 presents studies categorized on the basis of target individuals’ ethnicity; and Supplementary Table 3 includes additional descriptive information on subjects’ ethnic background.

First, we conducted a meta-analysis including all eligible studies. There was higher activity toward ethnic in-group in the left superior parietal gyrus (x = −22, y = −78, z = 46, SDM-Z = 3.263, p = 0.00055) and left median network (x = −6, y = −48, z = 28, SDM-Z = 2.787, p = 0.00266) when compared to ethnic out-group. No brain region exhibited the opposite activity pattern. There was not significant publication bias, as indicated by a visual inspection of the funnel plots (see Supplementary Figure 1) and the results of metabias tests (p = 0.998–0.994 for the peak voxels). Heterogeneity was minor in the left median network (I² = 0.256) and low in the left superior parietal gyrus (I² = 10.034).

This meta-analysis provided a novel perspective to the current literature by investigating the neural basis of ethnic inter-group bias (a) separately among ethnic minority and majority members (toward visually presented ethnic minority vs. majority members) and (b) separately toward visually presented Black, White, and Asian individuals. The pooled evidence from the original studies showed that, first, neural inter-group bias was non-evident in ethnic minority members but evident in ethnic majority members. Second, we found neural inter-group bias toward Black and White individuals but not toward Asian individuals in subjects not belonging to the respective ethnic group. Finally, heterogeneity was mostly minor or low in the brain regions related to inter-group bias, and the metabias tests did not identify any significant publication bias. In summary, this study showed that neural inter-group bias is differently expressed among ethnic minority vs. majority members, and toward different ethnic groups (White, Black, and Asian individuals).

We did not find any neural inter-group bias in ethnic minority members: ethnic minority members did not show different responses to ethnic minority vs. majority members. Hence, in minority members, we obtained neither signs of out-group derogation (such as deactivation of perspective-taking-related brain regions when observing majority members); nor signs of anxiety or threat responses toward ethnic majority members. In the original studies, most minority members were newly moved immigrants: they had been living only a short time in the country, and many of them had not had close contacts with the ethnic out-group beforehand (see Supplementary Table 5). It is possible that inter-group bias could be evident after spending a longer time period in a country as an ethnic minority member. Finally, it is necessary to consider that we had a smaller number of minority (than majority) members in this meta-analysis, providing a further possible explanation for the null results.

The fusiform gyrus showed more activity toward ethnic minority (vs. majority) members; and also toward Black (vs. non-Black) individuals. The fusiform gyrus constitutes a face-specialized region (Haxby et al., 2000), shows activity in response to visual objects of expertise (Wong et al., 2009), and responds sensitively to face color (Nakajima et al., 2014). There is evidence from experimental studies and computational models that facial features, including ethnicity and race, are stored through previous experience in a multidimensional memory space that guides face perception (Valentine, 1991; Caldara and Hervé, 2006): ethnic majority members may have developed “a field of expertise” of processing face of their own ethnic group. Related to this, we found that the left lingual gyrus showed more activity toward Black (vs. non-Black) individuals. The left lingual gyrus is involved in working memory for schematic faces (Kozlovskiy et al., 2014), responds more strongly to faces than non-faces (Patriquin et al., 2016), and exhibits higher reactivity to submissive than neutral faces (Chiao et al., 2008). This provides further evidence of biased responses to Black faces, and also arouses a cautious question whether Black faces are more easily interpreted as submissive than non-Black faces.

Our meta-analysis showed higher activity toward ethnic minority (vs. majority) members in the right inferior frontal gyrus; and also higher activity toward Black (vs. non-Black) individuals in the left median cingulate cortex. The right inferior frontal gyrus is a region of the emotional empathy network (Schulte-Ruther et al., 2007; Jabbi and Keysers, 2008; Seitz et al., 2008; Perry et al., 2012; Peled-Avron et al., 2019; Li et al., 2021), and is also involved in executive functioning such as inhibitory control and attentional monitoring according to original studies (Aron et al., 2003, 2014; Hampshire et al., 2010; Gavazzi et al., 2017) and meta-analyses (Simmonds et al., 2008; Gavazzi et al., 2021). The left median cingulate cortex is a part of the mirror neuron system and involved in prosocial emotions (Amodio, 2014; Molenberghs and Louis, 2018). Consequently, these findings would support an interpretation that empathic responses might be stronger toward minority (vs. majority) members and toward Black (vs. non-Black) individuals.

Our results also showed, however, that observing Black (vs. non-Black) individuals correlates with higher activity of the right insula, which is commonly interpreted as a stronger disgust reaction (Amodio, 2014). Additionally, we found higher activity toward minority members (vs. majority) members and toward Black (vs. non-Black) individuals in the right inferior temporal gyrus, which is found to show activation when one overestimates his/her capacity to express empathy for others (Sollberger et al., 2014). This implies that majority members and non-Black (White or Asian) individuals may be prone to overestimate their empathic expressions toward minority members and Black individuals. Taken together, a very cautious interpretation might be that, toward Black and minority members, there may be a kind of spontaneous disgust reaction that may be compensated by activating the frontal empathy-related regions and an attempt to express empathy toward them.

Finally, we did not find any inter-group bias in the subcortical regions (e.g., the amygdala), indicating that there may not be any threat-related reactions when observing ethnic majority or minority members. We found, however, that ethnic majority members showed higher activity toward ethnic majority (than minority) members in the left superior parietal gyrus. It is a part of the dorsal frontoparietal network that is thought to maintain a “salience map” and direct attention to such external stimuli that are salient on the basis of semantic knowledge (Corbetta and Shulman, 2002). Thus, a cautious interpretation might be that, when ethnic majority members have directed attention to majority members (compared to minority members), they may be more actively retaining semantic knowledge from their memory.

The left superior frontal gyrus showed lower activity toward White (vs. non-White) individuals; and also higher activity toward Black (vs. non-Black) individuals. Thus, the left superior frontal gyrus seemed to respond in opposite directions depending on the target individuals’ ethnicity. In general, this brain region is thought to be involved in working memory, especially during high cognitive load and spatial processing (du Boisgueheneuc et al., 2006). Interestingly, higher activity of the left superior frontal gyrus also associates with social semantics (Wood et al., 2003), stronger feelings of social punishment (Zinchenko, 2019), and higher social dominance in response to disgusting facial expressions (Aleman and Swart, 2008). Thus, our finding possibly implies that a non-Black observer may have stronger feelings of social punishment or disgust-related dominance when observing a Black individual, while those feelings may be less evident when observing a White individual.

We found biased responses toward Black (vs. non-Black) individuals in the middle occipital gyrus that is involved in perception of visuospatial objects (Yang et al., 2015), and in the postcentral gyrus is known to include the primary somatosensory cortex. Thus, both regions contribute to primary sensory processing. Interestingly, we also found biased reactivity to Black (vs. non-Black) individuals in the left anterior thalamic projections that constitute a link between the primary sensory regions and frontal cortex (where we also found biased responses) (Mitchell, 2015). The thalamocortical projections play a role in sensory selection during divided attention (Wimmer et al., 2015) and in maintaining frontal activity during working memory performance (Bolkan et al., 2017). Hence, there seems to be biases in the neural mechanisms regulating attentional resources toward different ethnic groups. This attention-related interpretation was supported but that we also found lower activity toward Black (vs. non-Black) individuals in the left posterior cingulate gyrus. This region typically deactivates when concentrating on an external task and when sustaining a vigilant attentional state (Leech and Sharp, 2014).

There was lower activity toward White (vs. non-White) individuals in the left supplementary motor area (SMA). The SMA is involved in the preparation and execution of voluntary movements (Weilke et al., 2001). Thus, our finding may reflect biased motor responses, as most fMRI tasks of the original studies included pressing a button in response to ethnic groups. Additionally, the SMA is also in self-other distinction and empathic perspective-taking (Schulte-Ruther et al., 2007). Thus, an alternative interpretation of our finding might be that White (vs. non-White) target individuals may arouse lower perspective-taking responses.

We found no signs of neural inter-group bias toward Asian (vs. non-Asian) individuals in subjects with non-Asian ethnicities. This is in accordance with some behavioral studies showing that Asian individuals encounter less racial discrimination in the UK labour market when compared to e.g., African individuals (Heath and Di Stasio, 2019); and Asian-Indians are more likely to encounter benevolent than hostile prejudice in the US (Ramasubramanian and Oliver, 2007). Finally, we had fewer original datasets available when examining inter-group bias toward Asian individuals (than toward White or Black individuals). Overall, we conclude that the pooled results of the currently published original studies did not find evidence for a neural inter-group bias toward Asian individuals.

Due to a lack of a sufficient number of original studies, this meta-analysis could not investigate neural inter-group biases between different ethnic minority groups, or inter-group biases toward Arabs, Middle Eastern individuals, or Latinos. Some pieces of evidence suggest that ethnic discrimination is more evident toward Arabs and Middle Eastern than Asian individuals (Zschirnt and Ruedin, 2016). Further, some single findings have supported the presence of inter-group biases between ethnic minorities, such as between Arab and Israeli groups in the US (Bruneau et al., 2012). Thus, inter-group bias toward these ethnic groups remains an intriguing topic for future studies.

It is necessary to keep in mind that, the brain responses toward particular ethnicities (Black, White, Asian) may differ between individuals from different ethnic groups. In our analyses, it was not possible to examine brain responses toward White/Black/Asian individuals separately among White/Black/Asian individuals (due to limited number of studies in each category). Therefore, results per ethnicity might theoretically reflect the combination of activation clusters linked to the different ethnic out-groups. This remains a research question for future studies.

In all included studies, the fMRI tasks consisted of visual (including facial) processing of ethnic groups. Nevertheless, there was some degree of variation in the task content between single original studies: for example, whether the stimulus material included video clips or photographs; whether the stimulus material presented only faces (with neutral or painful expression) or also a body; and whether subjects were instructed to press a button on the basis of group membership or some other indicator (e.g., target individual’s gender). As a result, each sub-group meta-analysis included a slightly different combination of different fMRI tasks. The fMRI task of each original study is described in Supplementary Tables 3, 4. In order to increase comparability between the original studies, we included only studies with visual processing (excluded studies with merely auditory processing) and studies where the target individuals’ skin color could be seen (excluded studies where target individuals’ ethnic group was described only in a written statement).

Inter-group bias can encompass either in-group favoritism or out-group derogation. In this meta-analysis, however, it was not possible to certainly define whether the biased BOLD responses reflected either in-group favoritism or out-group derogation. This is because, first, most brain regions (e.g., the amygdala, frontal regions) are involved in both positive and negative emotions and may also process complex sociocognitive information. Second, when finding a significant difference (such as higher activity toward out- than in-group), it is not possible to deduce whether it refers to elevated activity to out-group or reduced activity to in-group. This is because the brain responses cannot be directly interpreted as reflecting changes in activity level (e.g., increased activity toward in-group vs. decreased activity toward out-group) but merely a difference in activity level.

Finally, this meta-analysis was limited to non-clinical study samples with adult-aged subjects. The original datasets were approximately balanced by sex: ca. 48.0–57.2% were female in different analyses of this manuscript. A large proportion of the original samples consisted of young adults, with the mean age of 24.1–32.8 years in our meta-analyses. Consequently, our results cannot be directly generalized to low-educated groups, clinical populations, elderly, adolescents, or children. Inter-group bias may be less evident in children (Abrams et al., 2003), and more evident in low-educated groups as low educational level associates with a stronger need for ethnic distance (Hello et al., 2006).