Main

Social cognition plays a key role in human interaction, encompassing the mental processes involved in perceiving, interpreting and responding to the social cues of others1. The core and most studied components are emotion recognition and mentalizing2. Emotion recognition conveys the ability to identify how others feel. Mentalizing is the capacity to infer the mental states of others, such as their intentions, beliefs and desires. Aging may diminish performance in both processes3, associated with gray matter loss and functional connectivity changes across brain networks4,5. Social cognition dysfunction in aging can increase social isolation, loneliness and vulnerability6, impacting brain health7 and quality of life8. Standardized tasks of social cognition are increasingly used in research and clinical contexts to assess the performance of patients with age-related conditions such as subjective cognitive complaints (SCC), mild cognitive impairment (MCI) and dementia, in comparison to that of healthy controls (HCs)9,10,11. However, despite its relevance, several gaps persist in our understanding of the factors that influence social cognition in aging.

One critical problem is the considerable variability observed in social cognition performance among otherwise similar individuals12, especially in the older population3 and at a global scale13,14. This variability may stem from multiple factors, such as demographic characteristics (sex, age, education13,14,15,16 and socioeconomic status17,18) and individual differences in other cognitive abilities (for example, memory and processing speed4) or executive functions (for example, inhibition and working memory19). Brain reserve, defined as the accumulation during the lifespan of structural and functional brain resources that mitigate the effects of neural decline caused by aging or disease20, may also play a role in social cognition variability4,5. Crucially, when considering underrepresented aging samples from low- and upper-middle-income countries (UMICs), brain health determinants exhibit greater heterogeneity, challenging mainstream evidence from high-income countries (HICs)21,22,23,24,25. Socioeconomic disparities worsen brain health and increase the rate of dementia23. Within these heterogeneous determinants, factors related to disparities, such as social determinants of health and education can exert a stronger influence than traditional factors such as age and sex21. Thus, heterogeneity constrains standard brain–behavior and brain–phenotype associations26,27,28,29,30, and predictive models often fail to classify individuals with nonstereotypical profiles regarding demographics, clinical presentation, admixtures, cognition and brain function26,27. In the same vein, functional connectivity-based models usually fail to generalize to diverse samples and are influenced by image acquisition artifacts, particularly in-scanner head motion26,31,32. These issues limit our ability to draw generalized conclusions about social cognition in healthy and pathological aging, hampering the development of a more global agenda.

In this Article, to address these gaps, we systematically investigated combined predictors of social cognition in older individuals through a multicentric computational approach (Fig. 1a). We sought to determine whether the traditional effects of age (Fig. 1b) and patient–control differences (Fig. 1c) are indeed the primary drivers of performance variability in social cognition tasks. We assembled 1,063 participants (>50 years) from nine countries to maximize sample diversity. Our outcomes of interest were facial emotion recognition, mentalizing and a social cognition total score (that is, the combination of both measures) using a well-characterized battery, the mini-social cognition and emotional assessment (mini-SEA)33. In the mini-SEA, participants are asked to identify the emotion depicted in a subset of photos from the Ekman series and to identify unintended transgression of social rules (that is, faux pas) in short stories. The potential predictors of social cognition comprised the following factors: (1) clinical diagnosis (HCs, SCC, MCI, Alzheimer’s disease (AD) and behavioral variant frontotemporal dementia (bvFTD)); (2) demographics (sex (female or male), age (years), education (years) and country income as a proxy of socioeconomic status (HICs and UMICs)34); (3) cognition (cognitive35,36,37 and executive function38,39 screening scores); (4) brain reserve (gray matter volume derived from voxel-based morphometry (VBM) analysis40 and functional connectivity strength derived from seed analysis41 of the resting-state functional magnetic resonance imaging (fMRI) networks: salience network (SN)42, default mode network (DMN)43, executive network (EN)44, visual network (VN)45 and motor network (MN46)); and (5) in-scanner motion artifacts (average translation and rotation parameters during the resting-state sequence). The analysis consisted of three distinct model sets. The initial set focused on behavioral data, spanning clinical diagnosis, demographics and cognition. The second set integrated structural brain reserve factors (gray matter volume) with the previous behavioral predictors. Lastly, the third set incorporated functional connectivity metrics and motion artifacts, building upon the predictors from both the first and second sets.

Fig. 1: Analysis pipeline and traditional effects of age and diagnosis on social cognition performance.
figure 1

a, (i) Participants were recruited from HICs (Chile, France, Italy and the United Kingdom) and UMICs (Argentina, Brazil, Colombia, Peru and Mexico) through ReDLat, the INSCD and GERO. (ii) Diagnosis, demographics, cognition, gray matter volume (vol) and fMRI resting-state functional connectivity of brain networks, and in-scanner motion artifacts were entered into computational models as predictors of social cognition. (iii) Data were harmonized across countries (including scale transformation) and missing values were imputed. (iv) The analysis involved Bayesian optimization with k = 3 cross-validation for tuning the hyperparameters in 70:30 train and test partitioning and SVR models using a bootstrap approach. (v) Outcome variables were facial emotion recognition, mentalizing and a social cognition total score from the mini-SEA battery. Emotion recognition image was reproduced from ref. 118. b, Age significantly predicted worse performance in emotion recognition, mentalizing and the total score across the full sample (n = 998). Data were analyzed with simple linear regression analysis. Red lines and gray shadings represent the best-fit line for each simple linear regression with 95% confidence bands. **P < 0.0001 (for details, see Extended Data Table 1). c, Participants with MCI (n = 96), AD (n = 339) and bvFTD (n = 102) performed significantly worse in social cognition relative to HCs (n = 316) and the SCC group (n = 145), and participants with bvFTD also performed significantly worse than those with AD in emotion recognition. Data were analyzed with linear mixed-effects models47 controlling for sex, age, education and country of origin. The red dots and lines display the mean and s.d. P values are corrected for multiple comparisons using the Šidák method.

Full size image

We anticipate that healthy individuals, female13,14, younger in age13,14,15, highly educated13,15,16, from HICs17, with better cognitive and executive abilities16,19 and with higher brain reserve20 will exhibit higher emotion recognition, mentalizing and total scores. However, traditional factors (that is, age and diagnosis) influencing social cognition as reported in homogeneous and stereotypical samples are hypothesized to show a reduced predictive value21. Our findings have the potential to advance our understanding of social cognition in aging populations by elucidating the factors that contribute to performance variability in current assessments. This knowledge can inform the development of tailored predictive models and tools to assess and improve social cognition in brain health and age-related diseases.

Results

Traditional effects (age and diagnosis) on social cognition

Simple linear regression analyses showed that advanced age significantly predicted worse emotion recognition, mentalizing and the social cognition total score (Fig. 1b and Extended Data Table 1). Linear mixed-effects models47 controlling for sex, age, education and country of origin revealed that the diagnosis had a significant effect on emotion recognition (F = 32.88, P < 0.0001 and ηp2 = 0.12), mentalizing (F = 59.72, P < 0.0001 and ηp2 = 0.2) and the total score (F = 63.93, P < 0.0001 and ηp2 = 0.21). Šidák-corrected post hoc tests showed that HC and SCC groups outperformed MCI, AD and bvFTD groups in the three measures, and that individuals with bvFTD performed significantly worse than those with AD in emotion recognition (Fig. 1c). No other significant between-group differences were found. Diagnosis effects on mentalizing (F = 12.75, P < 0.0001 and ηp2 = 0.17) and the total score (F = 14.36, P < 0.0001 and ηp2 = 0.08) were maintained when including the participants’ performance in the mentalizing control questions of the test33 as a covariate of no interest (Extended Data Fig. 1). This analysis confirms that results were not entirely explained by a lack of attention to or understanding of the stimuli.

Combined predictors of social cognition

Support vector regression (SVR) models48 were used to predict social cognition (emotion recognition, mentalizing and the total score) from the complete set of potential predictors. Multicollinearity between predictors is assumed and addressed in our models (Methods). In any case, as multicollinearity concerns the relationships among predictors49, it does not inherently imply any circular relationship with the outcomes of our models (that is, social cognition). Data were harmonized across countries (using equivalence tables50, scale transformation and z-scores estimation), and 170 missing values were imputed using a sklearn iterative imputer with Bayesian ridge regression51. SVR models were optimized using Bayesian optimization52 with k = 3 cross-validation for tuning the hyperparameters on training (70%) and testing (30%) folds, with ten repetitions. Feature selection was performed using backward elimination53 to identify each model’s top predictors (in order of relevance). To obtain the final models, 1,000 optimized SVR regressors were trained and tested for each outcome variable using a bootstrap approach, setting aside median-stratified 30% of the data as a test partition. We report the average models’ performance and largest false discovery rate-corrected P values (statsmodels version 0.13.2) on the test partition of the data. Analyses were performed in the full sample (n = 998, after removing participants with invalid scores) and in subsamples with neuroimaging recordings, including structural MRI (n = 598) and resting-state fMRI (n = 388) sequences.

Behavioral predictors

The first set of models assessed whether behavioral data (clinical diagnosis, demographics and cognition) were able to predict social cognition (Fig. 2a). The model using emotion recognition as the outcome variable was significant (R² = 0.35, confidence interval (CI) (95%) 0.07, f2 = 0.53, F = 22.31 and P < 0.0001). The best predictors of emotion recognition were, in order of importance, cognition (β = 29.67 and P < 0.0001), executive function (β = 18.98 and P < 0.0001), education (β = 7.88 and P < 0.0001), sex (β = 7.11 and P < 0.0001), country income (β = 5.95 and P < 0.0001) and diagnosis (β = 5.51 and P < 0.0001). Age was not a significant contributor for emotion recognition (β = 3.53 and P = 0.97). Mentalizing was significantly predicted (R² = 0.34, CI (95%) 0.09, f2 = 0.52, F = 21.76 and P < 0.0001) by cognition (β = 45.31 and P < 0.0001), executive function (β = 29.87 and P < 0.0001), education (β = 16.23 and P < 0.0001), diagnosis (β = 7.17 and P < 0.0001) and country income (β = 7 and P < 0.0001). Sex and age were not significant (β = 2.51, P = 0.85 and β = 2.34, P = 0.84, respectively). Finally, the social cognition total score was successfully predicted (R² = 0.44, CI (95%) 0.06, f2 = 0.79, F = 33.07 and P < 0.0001) by cognition (β = 63.13 and P < 0.0001), executive function (β = 41.13 and P < 0.0001), education (β = 23.23 and P < 0.0001), diagnosis (β = 11.50 and P < 0.0001) and sex (β = 8.05 and P < 0.0001). Country income and age were not significant (β = 6.55, P = 0.12 and β = 3.75, P = 0.79, respectively). The results were similar when assessed without data imputation (Extended Data Table 2), and in the subsamples with structural MRI (Extended Data Table 3) or resting-state fMRI (Extended Data Table 4) recordings. A consistent pattern of behavioral predictors was also found when stratifying the sample by sex (Extended Data Table 5) and analyzing HCs separately (Extended Data Table 6). Taken together, better cognitive and executive functions and higher educational level consistently emerged as the top predictors of social cognition performance, above diagnosis and other demographic characteristics.

Fig. 2: SVR results.
figure 2

a, Models including diagnosis, demographics and cognition as predictors of social cognition performance (n = 998). b, Models including one level of brain reserve (gray matter volume) together with behavioral features as predictors of social cognition performance (n = 598). c, Models including resting-state functional connectivity features (brain networks and motion artifacts) as predictors of social cognition performance in addition to behavioral and gray matter volume predictors (n = 388). Bars plots represent the β coefficient and CI associated with each predictor in each model. Violin plots show the distribution of R² values in the test partitions of the data from the bootstrap approach (n = 1,000 optimized SVR models). Thick lines inside density plots display the IQR and whiskers show maximum and minimum values of R². The translucid panel displays a nonsignificant model. Move rot: rotation movements; move trans: translation movements. *P < 0.05 and **P < 0.01 (details are provided in the main text).

Full size image

Behavioral and structural brain reserve predictors

The second set of models included the previous behavioral predictors plus one level of brain reserve (gray matter volume) as predictors of social cognition performance (Fig. 2b). The model predicting emotion recognition was significant (R² = 0.28, CI (95%) 0.09, f2 = 0.39, F = 5.45 and P < 0.0001) and included the following features: cognition (β = 27.25 and P < 0.0001), executive function (β = 20.37 and P < 0.0001), SN volume (β = 20.19, T = 116.11 and P < 0.0001), EN volume (β = 15.44 and P < 0.0001), sex (β = 9.79 and P < 0.0001), MN volume (β = 9.01 and P < 0.0001), diagnosis (β = 8.14 and P < 0.0001) and education (β = 7.58 and P < 0.0001). DMN volume (β = 5.99 and P = 0.92), age (β = 5.63 and P = 0.23), VN volume (β = 5.04 and P = 0.69) and country income (β = 2.12 and P = 0.77) were not significant. Mentalizing was also successfully predicted (R² = 0.33, CI (95%) 0.09, f2 = 0.5, F = 6.97 and P < 0.0001) by cognition (β = 40.82 and P < 0.0001), executive function (β = 25.60 and P < 0.0001), education (β = 14.10 and P < 0.0001), country income (β = 11.49 and P < 0.0001) and diagnosis (β = 7.43 and P = 0.02). EN volume (β = 7.48 and P = 0.6), DMN volume (β = 5.87 and P = 0.83), SN volume (β = 5.18 and P = 0.78), VN volume (β = 4.46 and P = 0.95), MN volume (β = 4.31 and P = 0.44), sex (β = 3.20 and P = 0.82) and age (β = 3.15 and P = 0.22) were not significant. The social cognition total score was significantly predicted (R² = 0.43, CI (95%) 0.06, f2 = 0.73, F = 10.2 and P < 0.0001) by cognition (β = 57.93 and P < 0.0001), executive function (β = 42.49 and P < 0.0001), education (β = 23.21 and P < 0.0001), SN volume (β = 18.38 and P < 0.0001), EN volume (β = 18.13 and P < 0.0001), diagnosis (β = 18.12 and P < 0.0001) and sex (β = 10.13 and P < 0.0001). MN volume (β = 11.11 and P = 0.89), DMN volume (β = 5.95 and P = 0.73), VN volume (β = 5.88 and P = 0.95), country income (β = 3.58 and P = 0.82) and age (β = 2.88 and P = 0.72) were not significant. Models including only gray matter predictors were not significant (Extended Data Table 7). Overall, higher cognitive and executive functions and years of education remained among the top predictors of social cognition (together with diagnosis). The higher the gray matter volume of SN, EN and MN hubs, the larger the contributions to emotion recognition.

Behavioral and structural–functional brain predictors

The last set of models included the previous two set of predictors (behavior and gray matter volume) plus functional connectivity and motion artifact predictors (Fig. 2c). Emotion recognition was significantly (R² = 0.32, CI (95%) 0.11, f2 = 0.48, F = 4.13 and P < 0.01) predicted by rotation movements (β = 30.07 and P < 0.0001), cognition (β = 26.24 and P < 0.0001), translation movements (β = 17.40 and P < 0.03), SN volume (β = 14.72, P < 0.034 and P < 0.0001), executive function (β = 14.24 and P < 0.0001), education (β = 10.49 and P < 0.0001), sex (β = 8.10 and P < 0.001) and diagnosis (β = 5.22 and P < 0.036). MN volume (β = 4.95 and P = 0.46), MN (β = 4.42 and P = 0.28), age (β = 2.48 and P = 0.96) and VN volume (β = 2.41 and P = 0.93) were not significant. Mentalizing was not successfully predicted in this model (R² = 0.3, CI (95%) 0.11, f2 = 0.44, F = 5.96 and P = 1). Finally, the social cognition total score was significantly (R² = 0.4, CI (95%) 0.11, f2 = 0.69, F = 6.59 and P < 0.0001) predicted and characterized by cognition (β = 49.85 and P < 0.0001), executive function (β = 36.16 and P < 0.0001), education (β = 22.67 and P < 0.0001), SN volume (β = 17.32 and P < 0.0001), diagnosis (β = 13.73 and P < 0.0001) and MN volume (β = 9.90 and P = 0.003). EN volume (β = 15.90 and P = 0.23), translation movements (β = 14.02 and P = 0.83), DMN (β = 10.76 and P = 0.6), VN (β = 9.31 and P = 0.15) and sex (β = 7.95 and P = 0.7) did not contribute to the model. Models including functional connectivity and motion features alone (Extended Data Table 8) and functional connectivity and motion features together with gray matter predictors (that is, only brain reserve, Extended Data Table 9) were not significant. Briefly, better cognitive and executive functions, higher education and more gray matter volume of SN hubs remained among the best predictors of social cognition together with diagnosis. While brain networks did not make significant contributions to the models, higher motion artifacts were associated with social cognition.

Discussion

We investigated the top predictors of social cognition in aging. Two main strengths enabled us to address this issue systematically: (1) the use of a diverse sample comprising 1,063 older individuals from nine countries, representing a wide range of demographics and socioeconomic contexts, and (2) the development of a multicentric computational approach that thoroughly examined the combined influence of various contributing factors. The results from SVR showed that combinations of behavioral, brain reserve (gray matter volume) and motion artifact features explained between 28% and 44% of the variance in tasks involving emotion recognition and mentalizing, with large effect sizes (f2 = 0.39–0.79). Higher cognitive and executive functions consistently predicted higher social cognition across models. More years of education was also ranked among the top predictors of social cognition in most models. Such factors had a larger influence than age across models, a finding that persisted even within the group of HCs. Furthermore, a direct comparison between nested regression models unveiled that, according to different statistical criteria (R2, adjusted R2, likelihood ratio test, Akaike information criterion, Bayesian information criterion and root mean squared error), the model encompassing all potential predictors surpassed the model with age (Supplementary Information and Supplementary Table 1). Moreover, while diagnostic differences in social cognition followed the expected pattern, with MCI and dementia groups performing poorer than HCs and SCC groups9,54, and bvFTD underperforming AD only in emotion recognition55, diagnosis was not the primary determinant of performance variability. Across models, the diagnosis effect was overshadowed by other factors, particularly cognition, executive functions and education. Finally, structural (to a lesser extent) and functional brain reserve measures had small and partial effects in the models’ performance. These results challenge traditional interpretations of age-related decline, patient–control differences and brain associations of social cognition, emphasizing the importance of heterogeneous factors. This knowledge has implications for developing tailored predictive social cognition models in diverse aging populations. It also informs the development of more robust assessment and intervention tools, ultimately improving brain health and quality of life.

The strong influence of cognitive and executive functions on social cognition performance is consistent with a growing body of evidence suggesting that age-related decline in a wide range of paradigms is dependent on task demands16,17,19,56,57,58. Accurately identifying the emotions of others partially rests on attention allocation, and attentional disturbances can lead to misrecognition of emotions and the development of affective symptoms59. Mentalizing relies on the capacity to inhibit one’s own perspective in favor of adopting that of others, a process that requires executive functions (that is, working memory and set shifting)19. Thus, the well-established decrease on these general-purpose abilities in older adults60 may explain social cognition decline. Relatedly, as in previous studies16,61, higher education also consistently emerged among the top predictors of social cognition performance. Taken together, cognition and education might represent proxies of the cognitive reserve in aging, namely the ability to cope with brain pathology to maintain function62. While a previous work showed that cognitive reserve was not associated with social cognition in older adults63, such evidence came from a homogeneous HIC population, potentially failing to capture the diversity of individual differences.

Another factor that predicted better emotion recognition and mentalizing was country income (HICs). The World Bank country classification34 represents a national level measure of the socioeconomic background of an individual (that is, social and monetary wealth or power)64. Socioeconomic status is known to have robust effects in predicting brain health outcomes in older individuals65 and dementia23. However, its impact on social cognition and emotional processing has only recently been addressed, pointing to a mediator role of cognitive and executive functions17. Our results expand this emergent research by revealing a unique contribution of socioeconomic status to social cognition performance. Finally, female sex was associated with improved emotion recognition (but not mentalizing), as previously observed13,63. Women’s advantage in identifying others’ emotions may be a result of gender-role stereotypes66. However, more research is needed to determine the underlying mechanisms of such advantage. In summary, our behavioral models suggested that, in addition to cognition, executive functions, and education, socioeconomic status and sex play an important role in some social cognition domains.

Including brain reserve measures (gray matter volume) in the model architecture did not explain the additional variance in social cognition performance. Moreover, the model that solely utilized gray matter features did not yield predictive value. Cognitive and executive abilities remained the top predictors of emotion recognition, mentalizing and the total score. Consequently, cognitive reserve may be more relevant than structural brain reserve for social cognition outcomes, potentially reflecting the deployment of active mechanisms (for example, processing resources or compensation) that facilitate coping with pathology beyond brain size62. Following cognitive factors, higher gray matter volume of the main hubs of the SN (bilateral insula and anterior cingulate cortex42), the EN (bilateral middle frontal and inferior parietal cortex67) and the MN (precentral cortex67) was associated with better emotion recognition. This finding is consistent with the role of these regions in detecting and attending to salient stimuli68, as well as in the embodied processing of emotions through mirroring mechanisms69,70. Conversely, gray matter volume did not significantly contribute to mentalizing. A possible explanation for this discrepancy could be the higher cognitive demands necessary to mental state inference as opposed to facial emotion recognition, resulting in cognition capturing more variance.

The last set of models showed that fMRI brain network connectivity failed to predict social cognition when combined with behavioral features and brain volume (and also when considered independently). Moreover, mentalizing was not significantly predicted in these analyses. In contrast, translation and rotation in-scanner motion artifacts were associated with better emotion recognition (together with cognition, executive functions, SN volume, education, sex and diagnosis). Considering the existing evidence on resting-state functional connectivity associations with social cognition (particularly the SN68 and the DMN43,71), this pattern of results may appear unusual. However, it is becoming increasingly evident that clinical and demographic heterogeneity can hinder the identification of brain–behavior associations26,72. Predictive models from homogeneous samples fail to characterize nonstereotypical individuals, particularly from multisite cohorts72, with in-scanner motion parameters representing a major source of model failure26. In brief, brain networks failed to explain social cognition performance, with cognitive and motion features emerging as top predictors in the emotion model, emphasizing the need to consider disparate sources of variability in future studies.

This work reveals that social cognition components in aging are shaped by heterogeneous factors, adding to recent literature on the demographic, socioeconomic and sociocultural determinants of social cognition12,13,14,18,23. Contrary to mainstream research, our results indicate that age and clinical diagnosis are not the primary drivers of individual differences in social cognition across diverse settings. Although both factors showed the expected effects when assessed independently, such influences attenuate or vanish when other factors are considered. Previous works have failed to detect age associations with social cognition after accounting for cognitive57,58 and mood (that is, depression58) factors. Older adults might even show improved mentalizing abilities when considering education, race and ethnicity in explanatory models12. Thus, age-related normal and pathological brain mechanisms may become less influential when accounting for sample diversity.

These unforeseen findings extend beyond the social cognition field. Most studies on factors associated with brain health and pathological aging have been performed in HICs73. However, risk may differ in underrepresented regions such as Latin America where multiple social and health disadvantages converge, including poverty, limited access to formal education and healthcare, and barriers to a healthy lifestyle21,22,23,74. Indeed, recent evidence from Latin American older adults underscores a more pronounced influence of heterogeneous and disparity-related factors (that is, social determinants of health, education, mental health symptoms and physical activity) on healthy aging relative to the age and sex traditional factors21. The present work aligns with this research and contributes to current calls of increasing sample diversity26,30,72, aiming to make cognitive and behavioral science more situated75,76.

These findings also question whether social cognition can be distinguished from broader cognitive function. Existing answers in the literature are inconclusive3,77. As cognitive and executive functions mediate social cognition16,17,19,56,57,58, there are important considerations. First, even after accounting for those factors, older adults might still experience sociocognitive difficulties4,78. Second, while socioeconomic disadvantages harm cognitive and executive functions, they can paradoxically enhance social cognition in some settings18. Third, cognition and social cognition engage partially distinct neural correlates79 and are linked to somewhat different functional outcomes80,81. Our results point to a partial overlap between domains. While social cognition showed a strong association with cognition in the regression models, the expected differences in performance between patients and controls on the mini-SEA persisted after covarying for the mentalizing control questions of the task. This suggests that patients’ performance might not be entirely attributed to a failure in processing task stimuli (that is, lack of attention or understanding). However, as control of task performance does not fully capture complex cognitive/executive processes, further research is needed in older adults.

This work carries relevant implications. Our results inform the development of tailored predictive models that acknowledge the diverse characteristics of the population under study. This may lead to more accurate and ethical interpretations82, improving decision-making in region-specific approaches to brain health. The results suggest a more nuanced approach to social cognition assessment by reducing cognitive demands, accounting for potential attentional or comprehension issues (for example, by analyzing control stimuli)77, and developing norms adjusted for years of education and country. These recommendations are particularly relevant in light of recent calls to use standardized social cognition tasks in clinical settings to support patient characterization and differential diagnosis2,11,83. In the realm of interventions, the findings emphasize the need for contextual approaches when addressing social cognition impairments84,85. Demographic, socioeconomic and cognitive diversity might modulate the response to social cognition interventions and their impact on everyday function. Tailored predictive models, more sensitive assessments, heterogeneity-robust methods and situated interventions in social cognition may prove crucial to advance brain health equity.

Some limitations and additional lines of research must be acknowledged. First, although we used one of the most widely used social cognition assessment13,33,54, it has low ecological validity. Future studies should incorporate more naturalistic stimuli86. Also, other social cognition components (such as empathy and compassion87) should be investigated in older adults. Second, we included only a limited number of countries with unbalanced participants, reducing the possibilities for cross-country interpretations. Two clinical groups (SCC and MCI) were enrolled exclusively in one country and recruitment center, potentially introducing confounds. However, separate analyses of this cohort (Supplementary Table 2) and the remaining groups collectively (Supplementary Table 3) yielded results consistent with our main findings. Relatedly, assembling participants from multiple sites and using different cognitive and scanning protocols may introduce disparate sources of uncontrolled variability. However, our methods, combining data harmonization (equivalence tables for cognitive data50, site-specific z-scores for fMRI data54,88,89,90,91,92 and missing data imputation51) and machine learning algorithms (involving stratified data partition, hyperparameter tuning, cross-validation, SVR48 with Ridge regularization51, backward feature selection53 and generalization to unseen samples) prove robust to handle unbalanced and diverse samples21,50, multicollinearity between predictors49 and the identification of top contributing factors93. Taken together, our approach is suitable to leverage the intrinsic heterogeneity of our multi-setting sample. In any case, global approaches to brain health need larger and more balanced samples to perform more systematic comparisons across regions. Third, we used a country-level index of socioeconomic status, potentially lacking accuracy in reflecting the precise circumstances of each participant, particularly in Latin America, which is known by its marked inequality. Other socioeconomic indicators such as the Gini index or the human capital index might prove more sensitive to better capture the distinct inequalities inherent to each region. Moreover, global socioeconomic status indexes should be complemented with measures at the individual or family level (for example, household income and occupation prestige). Fourth, the cross-sectional nature of our study impedes causal conclusions. Further research should adopt longitudinal designs to understand the temporal dynamics between disparate factors and social cognition performance in brain health and disease. Finally, given the small sample size of our clinical groups and the disbalance across countries and sites, our design was unsuitable for exploring diagnosis stratification. Studies involving larger and balanced clinical cohorts should examine whether social cognition relies on distinct factors across different patient groups. Available evidence emphasizes a primary deficit in bvFTD and a secondary impairment in AD that would depend on memory and other cognitive functions83,94. This approach could be enriched by a more diverse range of social cognition predictors (for example, ethnicity and genetics) and diagnostic categories (for example, language variants of FTD95 and other neurodegenerative and neuropsychiatric disorders96).

Conclusions

Using a multicentric computational approach across three levels of analysis, our findings reveal that social cognition in aging is shaped by a heterogeneous array of cognitive and sociodemographic factors. The most influential predictors were cognitive and executive functions (together with education in most models), which outweighed the impact of age, clinical diagnosis and brain reserve. The results challenge traditional interpretations of age decline, patient–control differences and brain associations of social cognition. We emphasize the need to consider heterogeneous factors in further studies, with implications for predictive models, assessments and interventions, aimed at developing more global and inclusive approaches to brain health.

Methods

Participants

The study comprised 1,063 participants aged between 50 and 98 years (mean age 71.56 years, s.d. 8.42 years, 64.6% women, mean years of education 12.01, s.d. years of education 5). The recruitment was performed across 13 sites in 9 countries, 4 HICs (Chile, France, Italy and the United Kingdom, n = 476) and 5 UMICs (Argentina, Brazil, Colombia, Peru and Mexico, n = 587) as classified according to the World Bank34. The sample included HCs and individuals with different conditions associated with aging (SCC, MCI, AD and bvFTD, see below). Participants were recruited from different international consortia, including the Multi-Partner Consortium to Expand Dementia Research in Latin America (ReDLat)97, the International Network on Social Condition Disorders (INSCD)13 and the Geroscience Center for Brain Health and Metabolism (GERO)98.

All participants underwent extensive neurological, neuropsychological and neuropsychiatric examinations comprising semistructured interviews, standardized cognitive assessments and MRI scanning (when available). Clinical diagnoses were performed by multidisciplinary expert teams following established criteria as detailed below. The diagnostic process did not include the mini-SEA, ruling out a potential selection bias. HCs (n = 325) had preserved cognition and no history of neurological or psychiatric conditions. Participants with SCC (n = 145) presented cognitive complaints either self-reported or reported by a knowledgeable informant, scored 0.5 or less on the Clinical Dementia Rating scale99 and had preserved functional abilities98. The MCI group (n = 96) was composed of participants fulfilling the same criteria as those with SCC but scoring <22 in the Montreal Cognitive Assessment (MoCA)36, the most frequently used cut-off to detect MCI100. Individuals with AD (n = 389) fulfilled the National Institute of Neurological and Communicative Disorders and Stroke–AD and Related Disorders Association criteria101, were in early and middle stages of the disease, presented memory deficits and were functionally impaired. Individuals with bvFTD (n = 114) fulfilled the revised Rascovsky criteria102, were in the early and middle stages of the disease, exhibited prominent behavioral changes, lacked primary language deficits and had functional impairment. Supporting the clinical diagnosis of neurodegenerative conditions, an analysis of a subsample of participants with available structural MRI data revealed temporal and frontoparietal atrophy in the AD group101, and fronto-temporo-insular atrophy in the bvFTD group103 (Supplementary Fig. 1 and Supplementary Table 4). Demographic and cognitive information of each participant group is provided in Supplementary Table 5. The institutional review board of each recruitment site and the executive committee of the ReDLat consortium approved this study. All participants signed informed consent as approved by their respective center’s ethics committee. No compensation was provided for this study.

Social cognition assessment

Participants completed the mini-SEA, a short battery designed to assess two social cognition domains: facial emotion recognition and mentalizing33. In the facial emotion recognition subtest, participants are asked to identify the emotion being depicted by an individual in 35 different photos from the Ekman series. The following options are provided: fear, sadness, disgust, anger, happiness, surprise and neutral. Each correct item is given one point. The mentalizing subtest consists of an adaptation of the Faux Pas test. Participants are presented with ten short stories and asked to identify if the protagonist committed an unintended transgression of a social rule (that is, a faux pas). Each story also includes two control questions to assess general understanding. The maximum score for this subtest is 40 points. The scores of emotion recognition and mentalizing subtests are converted to a score of 15 each and then summed, resulting in a total score of 30, with higher scores representing better performance. From the full sample, 6.11% of participants (n = 65) were removed for lacking a valid score either in the emotion recognition or the mentalizing subtest, resulting in a final sample of 998 individuals.

Predictors of social cognition

The set of potential predictors of social cognition included:

(a) Behavioral features

(a.1) Diagnosis, HCs, SCC, MCI, AD and bvFTD.

(a.2) Demographics, sex (female, male), age (years), education (years) and country income (HICs and UMICs) following the World Bank classification34.

(a.3) Cognition

(a.3.1) Cognitive score, derived from harmonized scores in the Addenbrooke’s Cognitive Examination III (ref. 37), the Mini-Mental State Examination (MMSE)35 and the MoCA36. For details about these tools, see Supplementary Information and Supplementary Table 5 for the number of participants assessed with each tool in each group and the ‘Data harmonization’ section.

(a.3.2) Executive score, derived from harmonized scores in the INECO Frontal Screening (IFS)39 and the Frontal Assessment Battery (FAB)38 (Supplementary Information, Supplementary Table 5 and ‘Data harmonization’).

(b) Brain reserve features

(b.1) Gray matter volume, average volume of key hubs of the SN, the DMN, the EN, the MN and the VN from the Automated Anatomical Labeling atlas104 calculated using VBM analysis (see below).

(b.2) Functional connectivity, average connectivity strength of the SN, the DMN, the EN, the VN and the MN calculated via seed analysis of the fMRI resting-state series (see below).

(c) Motion artifacts, average translation and rotation movements estimated during the preprocessing of the fMRI sequence.

Neuroimage acquisition and preprocessing

This section is reported following recommendations from the Organization for Human Brain Mapping105. Whole-brain structural three-dimensional T1-weighted and resting-state sequences were obtained for 598 (195 HCs, 91 SCC, 53 MCI, 194 AD and 65 bvFTD) and 388 (125 HCs, 91 SCC, 52 MCI, 82 AD and 38 bvFTD) participants, respectively, across acquisition centers. For the resting-state sequence, participants were instructed to remain still, awake, with eyes closed and not to think about anything in particular. Demographic and cognitive information of these subsamples are provided in Supplementary Tables 6 and 7. The scanning protocols followed by each center are detailed in Supplementary Tables 8 and 9. Structural MRI scans were preprocessed using the DARTEL Toolbox following standard procedures for VBM40 through the Statistical Parametric Mapping software (SPM12 (ref. 106)). Functional images were preprocessed using the Data Processing Assistant for Resting-State fMRI toolbox (v.4.4 (ref. 107)) following published procedures41 (details in Supplementary Information). Six movement parameters (right, forward, up, pitch, roll and yaw) were estimated during realignment to calculate average translation and rotation movements per participant (group statistics are reported in Supplementary Table 10).

Data harmonization

Two procedures were applied to increase the number of participants with homogeneous cognitive and executive measures and harmonize the available data. First, cognitive screening measures were harmonized using equivalence tables following recommended methods50,108,109, validated for multicentric studies using data from Latin American underrepresented samples50. This procedure allows for MoCA and ACE scores to be estimated using MMSE scores and the MMSE scores using MoCA and ACE scores. As a result, a total of three new converted–harmonized variables were added. Then, the MMSE and MoCA scores were transformed from a 0–30 to a 0–100 scale and averaged with the ACE score to create a single cognitive score per participant (scale 0–100). All participants had a cognitive score. Finally, IFS and FAΒ scores were also transformed into a 0–100 scale and averaged to obtain a single executive score per participant. Both the IFS and the FAB have previously shown significant associations with classical executive tests across healthy individuals and patients with dementia, as well as adequate discriminatory accuracy to differentiate between those groups110. This suggests these measures have similar external validity. A correlational analysis using the subsample of participants with both tests revealed a strong correlation between the IFS and FAB transformed scores (Pearson’s r = 0.72 and P < 0.0001). This result further supports comparability (inferential equivalence111) between such measures. In total, 833 participants had an executive score. Second, we calculated z-scores for demographic (sex, age, education and country income), cognitive (cognitive score and executive score), gray matter, functional connectivity and motion artifacts variables. For neuroimaging variables, z-scores were estimated using normative data from each fMRI acquisition site according to the following equation:

$${x}_{z}=\frac{x-{\rm{\mu }}}{s}$$

where xz is the new value, x is the original raw score, μ is the mean score for HCs from the center to which the participant belongs and s is the standard deviation for HCs from the site or center to which the participant belongs.

Using site-specific z-scores is a standard procedure for harmonizing neuroimaging data in multicentric studies on neurodegeneration54,88,89,90,91,92. This procedure directly compares different sites and imaging modalities112, controlling for protocol effects (for example, various magnetic fields and scanner-related artifacts) while addressing potential neuroanatomical/neurofunctional differences between normative groups92. Site-specific standardization proves more robust than conventional covariance methods in controlling for protocol effects without losing information on diagnosis effects92.

Data imputation

A sklearn iterative imputer with Bayesian ridge regression51 (Python 3.7) was used to impute missing values for age (n = 4), education (n = 2) and executive score (n = 165). This algorithm applies a multivariate imputing strategy, modeling a column with missing values as a function of other features and using the estimate for imputation. Each feature is imputed sequentially allowing the usage of prior imputed values on the model that predicts later features. This process is repeated several times, allowing increasingly better estimates of missing values to be calculated as the missing values for each feature are estimated.

SVR models

To generate predictions of continuous variables (emotion recognition, mentalizing and total scores) from multimodal features (behavior, brain reserve and motion artifacts), we ran SVR models using the sklearn51 package in Python 3.7. SVR is a variation of support vector machine that allows linear and nonlinear regression. SVR transforms the feature space to establish a hyperplane that best fits the training data, while also minimizing the generalization error on new, unseen data48. The hyperplane is defined as the set of all points x in the feature space such that:

$$\bf{w}\cdot x+b=0$$

where w is the weight vector, b is the bias term and · denotes the dot product.

The SVR model seeks to find the weight vector w and bias term b that satisfy this constraint, while also minimizing the distance between the hyperplane and the training data. The distance is measured using a loss function, typically the ε-insensitive loss:

$$L(\,y,{\widehat{y}})= {\mathrm{max}} (|\,y-{\widehat{y}}|-{\rm{\varepsilon }},0)$$

where y is the true target value, \(\hat{y}\) is the predicted target value and ε is a small constant that defines the width of the margin around the hyperplane. The loss function penalizes errors that exceed ε, but ignores errors that fall within ε.

To find the optimal weight vector w and bias term b, SVR introduces two slack variables ξi and \(\widehat{{\xi }_{i}}\) for each training example, which allow for violations of the margin and the ε-insensitive loss, respectively. The optimization problem for SVR is then given by:

Minimize:

$$\frac{1}{2}{||{\bf{w}}||}^{2}+{{C}}\left(\mathop{\sum }\limits_{{{i}}=1}^{{{n}}}({{\rm{\xi }}}_{{{i}}}+\widehat{{{\rm{\xi }}}_{{{i}}}})\right)$$

Subject to:

$$\begin{array}{cc}{{{y}}}_{{{i}}}-\langle {\bf{w}},\phi ({{{x}}}_{{{i}}})\rangle \le {\rm{\varepsilon }}+{{{\xi }}}_{{{i}}} & {{i}}=1,\cdots ,{{n}}\end{array}$$
$$\begin{array}{cc}\langle {\bf{w}},\phi ({{{x}}}_{{{i}}})\rangle -{{{y}}}_{{{i}}}\le {\rm{\varepsilon }}+\widehat{{{{\xi }}}_{{{i}}}} & {{i}}=1,\cdots ,{{n}}\end{array}$$
$$\begin{array}{cc}{{{\xi }}}_{{{i}}}\ge 0,\,\widehat{{{{\xi }}}_{{{i}}}}\ge 0 & {{i}}=1,\cdots ,{{n}}\end{array}$$

where C is a hyperparameter that controls the trade-off between the margin width and the number of violations allowed, and n is the number of training examples. The first term in the objective function encourages a wide margin, while the second term penalizes violations of the margin and the ε-insensitive loss.

SVR can be extended to handle nonlinear regression tasks by using a kernel function to map the input data to a higher-dimensional feature space, where the problem may become linearly separable. The optimization problem then becomes:

Minimize:

$$-\frac{1}{2}\mathop{\sum }\limits_{{{i}},{{\,j}}=1}^{{{n}}}({{\rm{\alpha }}}_{{{i}}}-\widehat{{{\rm{\alpha }}}_{{{i}}}})({{\rm{\alpha }}}_{{{i}}}-\widehat{{{\rm{\alpha }}}_{{{i}}}}){{K}}\big({{{x}}}_{{{i}}},{{{x}}}_{{{j}}}\big)-\varepsilon$$

Subject to:

$$\mathop{\sum }\limits_{{{i}}=1}^{{{n}}}({{\rm{\alpha }}}_{{{i}}}-\widehat{{\alpha }_{i}})=0$$
$$0\le {{\rm{\alpha }}}_{{{i}}},\widehat{{\alpha }_{i}}\le {{C}}$$

where \(K({x}_{i},{x}_{j})\) is the kernel function that computes the inner product between the mapped feature vectors, and αi and are Lagrange multipliers that determine the importance of each training example in defining the hyperplane. The kernel function allows SVR to learn complex, nonlinear relationships between the input features and the target variable.

Sklearn’s SVR uses the L2 regularization term (Ridge regularization) in its function51. This regularization term helps control the model’s complexity, prevents overfitting113, and mitigates the impact of multicollinearity by penalizing the magnitudes of regression coefficients49. The strength of the L2 regularization is controlled by the C hyperparameter, where a smaller value of C corresponds to a stronger regularization effect.

Hyperparameter tuning

A Bayesian optimization52 with k = 3 cross-validation was applied for tuning the hyperparameters. A radial basis function kernel was used with optimized gamma value. Models with the best hyperparameters were trained on a training sample (70%) and tested in a testing set (30%), with ten repetitions (Supplementary Information).

Feature selection

We used a backward elimination approach53 to select the most significant predictors for each model. For each iteration, we dropped the predictor with the largest P value until we reached a statistically significant model, a predictor with a P value that became statistically significant or a model with two predictors. This procedure allowed us to automatically rank predictors based on their contribution to the model’s prediction accuracy without assuming a priori theoretical importance, which is required when classical statistical methods are applied93. Moreover, backward elimination mitigates the impact of multicollinearity between predictors by removing correlated features49.

Statistical analyses

VBM analysis

Using VBM preprocessed structural images, we calculated the average gray matter volume (ml, corrected by total intracranial volume) of 116 regions of the Automated Anatomical Labeling atlas104 to create gray matter volume indexes of the main hubs of the SN (average of the bilateral anterior cingulum and insula volume42), the DMN (average of the bilateral medial frontal and posterior cingulate volume43), the EN (average of the bilateral middle frontal and inferior parietal volume67), the VN (average of the bilateral occipital volume67) and the MN (average of the bilateral precentral volume67).

Functional connectivity analysis

The functional connectivity strength of the SN, the DMN, the EN, the VN and the MN was calculated using seed analysis. Two bilateral seeds were placed on cubic regions of interest (voxel size 7 × 7 × 7) for each network: the dorsal anterior cingulate cortex for the SN42, MNI coordinates 10, 34, 24 and −10, 34, 24; the posterior cingulate cortex for the DMN43, MNI coordinates 3,−54, 27 and −3, −54, 27; the middle frontal gyri for the EN44, MNI coordinates 30, −2, 62 and −30, −2, 62; the primary visual cortex for the VN45, MNI coordinates 8, −92, 8 and −8, −92, 8; and the primary motor cortex for the MN46, MNI coordinates 32, −30, 68 and −32, −30, 68. The Pearson correlation coefficient between the averaged blood-oxygen-level-dependent signal of each pair of seeds and voxels comprised in standard masks114 typically involved in each resting-state network was used to extract one feature per network for each participant. The statistical significance of the resting-state networks was tested by comparing them with null surrogate models. This approach enables robust statistical evaluations to ensure that the results observed are not obtained by chance but represent a true characteristic of the underlying system115. The surrogate data technique is based on comparing a particular property of the data (a discriminating statistic) with the distribution of the same property calculated in a set of constructed signals (surrogates) that match the original dataset but do not possess the property that is being tested. To this end, we used Fourier transform-based surrogates to recreate the brain’s complex-system dynamics, including uncorrelated and correlated noise, coupling between different brain areas, and synchronization. We found that all the computed resting-state networks were statistically significant against null connectivity (SN: P = 0.02; DMN: P = 0.02; EN: P = 0.03; VN: P = 0.02 and MN: P = 0.03), further corroborating our connectivity methods.

Age effects on social cognition

Simple linear regression analyses were used to evaluate the predictive value of age on emotion recognition, mentalizing and the social cognition total score. Analyses were performed in R software (version 4.1.3). The alpha threshold was set at P < 0.05. Effect size was evaluated with f2, following Cohen’s criteria,116 stating that 0.02 indicates a small effect, 0.15 indicates a medium effect and 0.35 indicates a large effect.

Social cognition performance across diagnostic groups

Linear mixed-effects models47 were performed in R (version 4.1.3) to examine diagnosis effects and between-group differences in emotion recognition, mentalizing and the total score. Sex, age and education were entered in the model as covariates of no interest, and the participant’s country of origin was entered as a random effect. Additional analyses included the participants’ performance in the mentalizing control questions of the mini-SEA as a covariate of no interest. Post hoc tests were corrected using the Šidák method. The alpha threshold was set at P < 0.05. Effect size was evaluated with ηp2 where 0.01 indicates a small effect, 0.06 indicates a medium effect and 0.14 indicates a large effect117.

SVR model estimation and performance assessment

We trained and tested 1,000 optimized SVR models for each outcome variable to obtain the final models using a bootstrap approach. We applied P value correction for false discovery rate using statsmodels (version 0.13.2) and set aside a median-stratified 30% of the data as a test set. To evaluate the models’ performance, we used four statistics: the coefficient of determination R², 95% CI, Cohen’s f2 (ref. 116), Fisher’s F test and the largest corrected P values. Outlier results (R² < interquartile range (IQR) − 1.5 × s.d. and R² > IQR + 1.5 × s.d.) were discarded to improve average estimates.

Inclusion and ethics statement

This work involved a collaboration between scientists in multiple countries including Argentina, Brazil, Chile, Colombia, France, Italy, Ireland, Mexico, Peru and the United Kingdom. All contributors have been listed as co-authors in acknowledgement to their work. Researchers from Latin America have led the study conceptualization, data analysis and manuscript writing. Collaborators established and agreed on their respective roles and responsibilities before the research began. Each site that participated in this study has retained ownership of all materials shared for research purposes.

All methods were performed in accordance with relevant guidelines and regulations provided by the Declaration of Helsinki. The institutional review board of each recruitment site and the executive committee of the ReDLat consortium approved this study. All participants signed an informed consent as approved by their respective center’s ethics committee. No compensation was provided for this study.

Reporting summary

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