Nature Human Behaviour

Published by Springer Nature
Online ISSN: 2397-3374
Learn more about this page
Recent publications
When speaking to infants, adults often produce speech that differs systematically from that directed to other adults. To quantify the acoustic properties of this speech style across a wide variety of languages and cultures, we extracted results from empirical studies on the acoustic features of infant-directed speech. We analysed data from 88 unique studies (734 effect sizes) on the following five acoustic parameters that have been systematically examined in the literature: fundamental frequency (f0), f0 variability, vowel space area, articulation rate and vowel duration. Moderator analyses were conducted in hierarchical Bayesian robust regression models to examine how these features change with infant age and differ across languages, experimental tasks and recording environments. The moderator analyses indicated that f0, articulation rate and vowel duration became more similar to adult-directed speech over time, whereas f0 variability and vowel space area exhibited stability throughout development. These results point the way for future research to disentangle different accounts of the functions and learnability of infant-directed speech by conducting theory-driven comparisons among different languages and using computational models to formulate testable predictions.
Study designs
a, On each trial, participants chose between two slot machines and received feedback on their choices. The participants could see the labels of the slot machines before making a choice and were explicitly instructed about the differences between a stable and a fluctuating machine before the task. b, The prediction task was completed at the end of each block in Experiment 2, where participants reported their reward predictions for the options they encountered in this block and rated their confidence. c, An example reward structure for a stable and a fluctuating arm. In the experiment, the means of both options were resampled from the zero-mean Gaussian at the beginning of each block.
Predictions of choice probability function change across conditions and probit regression results
a, Directed exploration predicts a preference for the uncertain option, which manifests as a shift in intercept in opposite directions for SF and FS trials. P (choose option 1): probability of choosing the option on the left. b, Random exploration predicts more choice stochasticity when total uncertainty is high, equivalent to a steeper curve for FF trials than for SS trials. c–f, Across two experiments: Experiment 1, N = 501 (c,d) and Experiment 2, N = 484 (e,f), the intercept of FS trials was larger than that of SF trials (Experiment 1: F(1, 150,292) = 6.20, P = 0.013, ΔM = 0.061; Experiment 2: F(1, 144,892) = 10.17, P = 0.001, ΔM = 0.062), while the intercepts of FF and SS trials did not differ (Experiment 1: F(1, 150,292) = 0.06, P = 0.814, ΔM = 0.004; Experiment 2: F(1, 144,892) = 0.04, P = 0.834, ΔM = 0.003). The slope of FF trials was larger than that of SS trials (Experiment 1: F(1, 150,292) = 81.07, P < 0.001, ΔM = 0.423; Experiment 2: F(1, 144,892) = 31.23, P < 0.001, ΔM = 0.175), while the slopes of SF and FS trials did not differ (Experiment 1: F(1, 150,292) = 2.77, P = 0.096, ΔM = 0.067; Experiment 2: F(1, 144,892) = 0.03, P = 0.867, ΔM = 0.005). Points indicate the fixed effect coefficients of intercept and slope for each condition using maximum likelihood estimation. F-tests (two-sided) were used to compare intercepts and slopes across conditions. No multiple comparisons correction was applied. Error bars are 95% confidence intervals. *P < 0.05, **P < 0.01, ***P < 0.001.
Exploratory factor analysis results (N = 501)
a, Scree plot of eigenvalues. b, Factor loadings of items on trait anxiety factors. Items from different subscales are distinguished by their colours. STICSA-T somatic: somatic subscale of STICSA-T; STICSA-T cognitive: cognitive subscale of STICSA-T; STAI-T absent: STAI-T anxiety absent items; STAI-T present: STAI-T anxiety present items.
Effects of trait anxiety factors on exploration strategies
a, Data from Experiment 1 (N = 501). b, Data from Experiment 2 (N = 484). Factor scores were obtained using EFA results in Experiment 1. All factors were standardized and entered into the same model together with age and gender. Points indicate the fixed effect coefficients fit for each predictor using maximum likelihood estimation. Error bars represent 95% confidence intervals. A positive coefficient for RU:Factor (V/TU:Factor) indicates increased directed (random) exploration. A positive coefficient for V:Factor indicates decreased undirected exploration. Significance of the coefficients was assessed using t-test (two-sided). Multiple comparisons correction was not applied. Across the two experiments, Somatic anxiety factor negatively correlated with RU (Experiment 1: t(150,273) = −2.12, P = 0.034, β = −0.070, 95% CI (−0.134, −0.006); Experiment 2: t(145,173) = −2.14, P = 0.032, β = −0.050, 95% CI (−0.096, −0.004)) and V (Experiment 1: t(150,273) = 3.32, P < 0.001, β = 0.217, 95% CI (0.089, 0.345); Experiment 2: t(145,173) = 4.05, P < 0.001, β = 0.194, 95% CI (0.100, 0.289)). *P < 0.05, **P < 0.01, ***P < 0.001.
Anxiety has been related to decreased physical exploration, but past findings on the interaction between anxiety and exploration during decision making were inconclusive. Here we examined how latent factors of trait anxiety relate to different exploration strategies when facing volatility-induced uncertainty. Across two studies (total N = 985), we demonstrated that people used a hybrid of directed, random and undirected exploration strategies, which were respectively sensitive to relative uncertainty, total uncertainty and value difference. Trait somatic anxiety, that is, the propensity to experience physical symptoms of anxiety, was inversely correlated with directed exploration and undirected exploration, manifesting as a lesser likelihood for choosing the uncertain option and reducing choice stochasticity regardless of uncertainty. Somatic anxiety is also associated with underestimation of relative uncertainty. Together, these results reveal the selective role of trait somatic anxiety in modulating both uncertainty-driven and value-driven exploration strategies.
  • Frederik V. SeersholmFrederik V. Seersholm
  • Hans HarmsenHans Harmsen
  • Anne Birgitte GotfredsenAnne Birgitte Gotfredsen
  • [...]
  • Anders J. HansenAnders J. Hansen
The success and failure of past cultures across the Arctic was tightly coupled to the ability of past peoples to exploit the full range of resources available to them. There is substantial evidence for the hunting of birds, caribou and seals in prehistoric Greenland. However, the extent to which these communities relied on fish and cetaceans is understudied because of taphonomic processes that affect how these taxa are presented in the archaeological record. To address this, we analyse DNA from bulk bone samples from 12 archaeological middens across Greenland covering the Palaeo-Inuit, Norse and Neo-Inuit culture. We identify an assemblage of 42 species, including nine fish species and five whale species, of which the bowhead whale (Balaena mysticetus) was the most commonly detected. Furthermore, we identify a new haplotype in caribou (Rangifer tarandus), suggesting the presence of a distinct lineage of (now extinct) dwarfed caribou in Greenland 3,000 years ago. Seersholm et al. analysed permafrozen middens from Inuit and Viking settlements to uncover evidence of diet in prehistoric Greenland. Using ancient DNA, they identified 42 different species and found that whales were surprisingly common.
| The structure of the human social world. Personal social networks have a hierarchically inclusive layered structure, with the layers having distinctive sizes that are determined by the frequency of contact and perceived emotional closeness 86 . The indicated values are robust population averages. In each case, there is interindividual variation due to gender, age, personality and circumstances. These values always have a fractal structure with a scaling ratio of ~3. The darker circle at 150 denotes the normal limit for personal social networks in which relationships are reciprocated, are relatively stable and have a personal history; beyond this, the outer layers consist of individuals with whom relationships are casual, unreciprocated and more fluid. Note that all layers include both friends and extended family, generically referred to as 'friends'. Most work colleagues would be placed in the 'Acquaintances' layer, except for the few that have graduated into being formal friends.
| Experienced social isolation shows a brain signature implicating especially the higher-order association circuits. A Bayesian hierarchical model was applied to ~40,000 UK Biobank participants to distinguish lonely (target group, encoded as 1) from non-lonely participants (control group, encoded as 0), by quantifying the degree of structural differences in brain region volume measurements in 100 cortical regions (Schaefer-Yeo atlas; for the details, see ref. 37 ). Yellow and green show positive and negative volume associations, respectively, indicating (for example) bigger volume effects in yellow areas in lonely participants. CO, central operculum; ITG, inferior temporal gyrus; pSTS, posterior superior temporal sulcus; TPJ, temporoparietal junction; IPL, inferior parietal lobe; dACC, dorsal anterior cingulate cortex; dlPFC, dorsolateral prefrontal cortex; RSP, retrosplenial cortex; FG, fusiform gyrus; IVG, inferior visual gyrus; L/R, left/right . m hemisphere. Figure reproduced with permission from ref. 37 .
Intense sociality has been a catalyser for human culture and civilization. Within this context, our relationships at a personal level play a pivotal role in our health and well-being. These relationships are, however, sensitive to the time we invest in them. To understand how and why this should be, we rst outline the evolutionary background in primate sociality from which our human social world has emerged. We then review de ning features of that human sociality, putting forward a framework within which one can understand new evidence on the consequences of mass social isolation during the COVID-19 pandemic, including mental health deterioration, stress, sleep disturbance and substance misuse. We outline recent research insights on the neural basis of prolonged social isolation, highlighting especially higher-order neural circuits such as the default network. Our survey of studies witnesses the various negative e ects of prolonged social deprivation and the multifaceted drivers of day-today pandemic experiences. Humans, like all anthropoid primates, are intensely social. We now have considerable evidence that interindividual differences in social embed-dedness affect a variety of health and fitness indicators. In humans, the single best predictor of physical health and well-being, as well as future longevity, is the number and quality of close friendships, with the more conventional suspects (such as diet, obesity, alcohol consumption and air quality) ranking a distant second 1,2. Indeed, the frequency of social engagement predicts psychological health and well-being 3 , self-rated feelings of happiness, satisfaction with life, and trust in one's local community 4. . m. m. m. m. m. m The COVID-19 lockdowns of the past two years were a global stress test-large-scale social deprivation in a more dramatic extent and form than ever before in recorded history. At the peak of public health restrictions, >3.6 billion people worldwide were subject to government-imposed stay-at-home orders. On the individual scale, we know that we respond poorly to isolation. However, existing psychological and neuroscience research had little to say about the possible consequences of mass isolation. In contrast, there have been many large-scale epidemiological studies of the effects of social deprivation in the elderly 5. Almost all of these investigations yielded strong signals for detrimental effects on cognitive capacity, psychological and physical well-being, and even longevity. It became clear that the chronic experience of social isolation escalated the risk of depression and dementias, as well as cardiovascular disease and certain types of cancer 6-8. The present review provides a frame of reference that can help situate current and future findings on the ongoing mass social isolation by incorporating established knowledge on human sociality and its underlying neurobiological mechanisms. To this end, we first place human sociality within the broader context of anthropoid primate sociality, with its behavioural and neurobiological determinants. Our aim is to provide a more grounded explanation as to why the neurobiol-ogy of human sociality takes the form it does. We then survey some of the unfolding evidence with direct relevance to the neurobiological and psychological consequences of the large-scale lockdown during COVID-19 and subsequent social rehabilitation.
Statistical inference is the optimal process for forming and maintaining accurate beliefs about uncertain environments. However, human inference comes with costs due to its associated biases and limited precision. Indeed, biased or imprecise inference can trigger variable beliefs and unwarranted changes in behaviour. Here, by studying decisions in a sequential categorization task based on noisy visual stimuli, we obtained converging evidence that humans reduce the variability of their beliefs by updating them only when the reliability of incoming sensory information is judged as sufficiently strong. Instead of integrating the evidence provided by all stimuli, participants actively discarded as much as a third of stimuli. This conditional belief updating strategy shows good test–retest reliability, correlates with perceptual confidence and explains human behaviour better than previously described strategies. This seemingly suboptimal strategy not only reduces the costs of imprecise computations but also, counterintuitively, increases the accuracy of resulting decisions.
An overview of the study design
We randomly divided the BBJ mainland samples into ten subsets to apply the LOGO method. We conducted GWAS using training samples and withholding the target subset using GCTA-fastGWA. We derived PGS for even-/odd-numbered chromosomes (PGSodd/PGSeven) in the target subset using the PRS-CS method and estimated GPD for even-/odd-numbered chromosomes (θeven_to_odd and θodd_to_even). We then meta-analysed the GPD estimates across the ten subsets. For the six independent Japanese or EAS cohorts, we derived PGSodd/PGSeven based on fastGWA results from the whole mainland sample in BBJ. Finally, we performed a meta-analysis of the GPD estimates across all EAS datasets (n = 172,270). We adopted the same LOGO method to estimate GPD in the UKB data (n = 337,139).
Estimates of GPD for 81 complex traits in the Japanese population
For 81 human complex traits, we quantified GPD as the correlation between trait-specific PGSs for odd-/even-numbered chromosomes. We selected the meta-analysed GPD estimate (θ) with the larger variance between θeven_to_odd and θodd_to_even in all Japanese or EAS cohorts (n = 172,270). P values were determined by two-sided Wald test. We set a study-wide significance threshold at P < 6.2 × 10⁻⁴ (=0.05/81) by applying Bonferroni’s correction for multiple comparison. Statistically significant traits are marked with an asterisk and in bold. Detailed results are presented in Supplementary Table 6. The bar plots represent the point estimates, and error bars represent the s.e. Freq., frequency.
Correlations between GPD estimates from even to odd chromosomes and GPD estimates from odd to even chromosomes for 81 traits in BBJ
a, The correlation plot of 81 traits between θeven_to_odd and θodd_to_even in the Japanese population. b, The enlarged plot of a around the trait of ever versus never drinking. The x axis indicates the meta-analysed GPD estimated from even to odd chromosomes (θeven_to_odd), and the y axis indicates that from odd to even chromosomes (θodd_to_even). The error bars represent s.e. The dashed line represents θodd_to_even = θeven_to_odd.
GPD estimates for six complex traits in the UKB
For six complex traits, we quantified GPD as the correlation between trait-specific PGSs for odd-/even-numbered chromosomes. GPD estimate (θ) with the larger variance between θeven_to_odd and θodd_to_even was selected in white British individuals from UKB data (n = 337,139). The bar plots represent the point estimates, and error bars represent the s.e.
Forest plots of GPD estimates of five traits related to AM
Forest plots of five significant traits, T2D, CAD, light-PA, natto consumption and yoghurt consumption. For all plots, a given label and number along vertical axes represent the name and the sample size of the cohort, respectively. The points indicate the point estimates, and error bars indicate 95% confidence intervals. Osaka Univ., Osaka University healthy cohort; Meta, meta-analysis.
Assortative mating (AM) is a pattern characterized by phenotypic similarities between mating partners. Detecting the evidence of AM has been challenging due to the lack of large-scale datasets that include phenotypic data on both partners, especially in populations of non-European ancestries. Gametic phase disequilibrium between trait-associated alleles is a signature of parental AM on a polygenic trait, which can be detected even without partner data. Here, using polygenic scores for 81 traits in the Japanese population using BioBank Japan Project genome-wide association studies data (n = 172,270), we found evidence of AM on the liability to type 2 diabetes and coronary artery disease, as well as on dietary habits. In cross-population comparison using United Kingdom Biobank data (n = 337,139) we found shared but heterogeneous impacts of AM between populations.
The two images used to label the buttons in the identity treatment
Instead of clicking on a button labelled with @ or #, as in the neutral treatment, participants in the identity treatment had to choose by clicking on one of two buttons with these images embedded in the buttons themselves (Supplementary Figs. 2 and 3).
Distributions of normalized spillovers by treatment
a, The distribution of spillovers in the neutral treatment. b, The distribution of spillovers in the identity treatment. The spillover²² is a normalized measure of how common the alternative becomes in an experimental group (Methods), and it can take any value in [−1, 1]. Negative values occur when the final proportion choosing the alternative behaviour is less than the proportional size of the intervention. Positive values occur when the final proportion choosing the alternative behaviour is greater than the proportional size of the intervention. The difference in spillovers by treatment is large and highly significant (Table 2).
Choice dynamics by treatment
a,b, The status quo behaviour was the choice associated with the norm that emerged in the pre-intervention phase of a session. With a status quo established, the alternative behaviour was simply the other choice option, which was always favoured by the intervention (Table 1). Here we show the proportion of choices, over all relevant sessions, in which the participants coordinated on the status quo (orange), coordinated on the alternative (green) or miscoordinated (blue) for each period. ‘Time’ refers to the period of play, which is centred on the intervention period (0). In the neutral sessions, the participants were relatively slow to converge on the status quo before intervention and relatively fast to converge on the alternative after intervention (a). In the identity sessions, the participants converged quickly before intervention, requiring fewer overall trials to meet the intervention criteria, but persisted in a state of chronic disagreement after intervention (b). For reference, under random matching, the maximum possible expected rate of miscoordination is 0.5.
Choice of alternative behaviour by treatment
The effect sizes and 95% confidence intervals (CIs) are from Model 1 in Table 3, N = 1,546 observations. The P values are based on two-sided z-tests. No adjustments were made for multiple comparisons. The omitted category consists of the neutral treatment, non-targeted participants, pre-intervention (Neutral, NT, Pre-int), and all other effects are relative to this benchmark. The curly brackets show the results from various linear combinations discussed in the main text, all of which are based on the cluster-robust standard errors in Table 3. The red dashed vertical line represents no effect. To present these linear combinations graphically, we show the effects in a different order than in Table 3.
Pay-off dynamics
a, Mean pay-offs by treatment and period. The pay-offs are measured in experimental currency points (100 points = US$1). ‘Time’ refers to the period of play, which is centred on the intervention period (0). The dashed lines are 95% confidence intervals from a bootstrapping algorithm clustered at the level of the experimental group. Compared with the neutral treatment, political labels in the identity treatment provided a ready focal point⁶⁶ that allowed the participants to converge on a norm quickly before intervention. After intervention, however, chronic disagreement (Fig. 3) prevented the participants in the identity treatment from transitioning to new norms in the same way the participants did in the neutral treatment. b, The accumulated difference in mean pay-offs, identity minus neutral, shows the monetary opportunity costs that the participants in the identity sessions ultimately paid.
Social tipping can accelerate behaviour change consistent with policy objectives in diverse domains from social justice to climate change. Hypothetically, however, group identities might undermine tipping in ways that policymakers do not anticipate. To examine this, we implemented an experiment around the 2020 US federal elections. The participants faced consistent incentives to coordinate their choices. Once the participants had established a coordination norm, an intervention created pressure to tip to a new norm. Our control treatment used neutral labels for choices. Our identity treatment used partisan political images. This simple pay-off-irrelevant relabelling generated extreme differences. The control groups developed norms slowly before intervention but transitioned to new norms rapidly after intervention. The identity groups developed norms rapidly before intervention but persisted in a state of costly disagreement after intervention. Tipping was powerful but unreliable. It supported striking cultural changes when choice and identity were unlinked, but even a trivial link destroyed tipping entirely.
There has been increasing interest in using neuroimaging measures to predict psychiatric disorders. However, predictions usually rely on large brain networks and large disorder heterogeneity. Thus, they lack both anatomical and behavioural specificity, preventing the advancement of targeted interventions. Here we address both challenges. First, using resting-state functional magnetic resonance imaging, we parcellated the amygdala, a region implicated in mood disorders, into seven nuclei. Next, a questionnaire factor analysis provided subclinical mental health dimensions frequently altered in anxious-depressive individuals, such as negative emotions and sleep problems. Finally, for each behavioural dimension, we identified the most predictive resting-state functional connectivity between individual amygdala nuclei and highly specific regions of interest, such as the dorsal raphe nucleus in the brainstem or medial frontal cortical regions. Connectivity in circumscribed amygdala networks predicted behaviours in an independent dataset. Our results reveal specific relations between mental health dimensions and connectivity in precise subcortical networks.
Rising partisan animosity is associated with a reduction in support for democracy and an increase in support for political violence. Here we provide a multi-level review of interventions designed to reduce partisan animosity, which we define as negative thoughts, feelings and behaviours towards a political outgroup. We introduce the TRI framework to capture three levels of intervention—thoughts (correcting misconceptions and highlighting commonalities), relationships (building dialogue skills and fostering positive contact) and institutions (changing public discourse and transforming political structures)—and connect these levels by highlighting the importance of motivation and mobilization. Our review encompasses both interventions conducted as part of academic research projects and real-world interventions led by practitioners in non-profit organizations. We also explore the challenges of durability and scalability, examine self-fulfilling polarization and interventions that backfire, and discuss future directions for reducing partisan animosity. Rachel Hartman and colleagues review interventions designed to reduce partisan animosity in the United States and introduce a framework to categorize interventions across three levels: thoughts, relationships and institutions.
Decades of research indicate that some of the epistemic practices that support scientific enquiry emerge as part of intuitive reasoning in early childhood. Here, we ask whether adults and young children can use intuitive statistical reasoning and metacognitive strategies to estimate how much information they might need to solve different discrimination problems, suggesting that they have some of the foundations for ‘intuitive power analyses’. Across five experiments, both adults (N = 290) and children (N = 48, 6–8 years) were able to precisely represent the relative difficulty of discriminating populations and recognized that larger samples were required for populations with greater overlap. Participants were sensitive to the cost of sampling, as well as the perceptual nature of the stimuli. These findings indicate that both young children and adults metacognitively represent their own ability to make discriminations even in the absence of data, and can use this to guide efficient and effective exploration. Adults and children can represent the relative difficulty of discriminating two populations and recognize that larger samples are required for populations with greater overlap. This suggests that they have foundations for ‘intuitive power analyses’.
Non-random mating affects the genetic makeup of populations and challenges the validity of popular genetics methods. A new study explores the unique patterns of non-random mating in the Japanese population and underscores the importance of large-scale genetic studies outside European-descended groups.
The low representation of academics with disabilities is a longstanding problem on which progress has been slow. Drawing on my research on disability-related barriers and my experiences of disability, I make six practical suggestions for how academic staff and people with disabilities can help make academia more disability inclusive.
Map of Philadelphia County (where the Philly Vax Sweepstakes occurred)
Treatment zip codes in Philadelphia are shown in dark green, control zip codes are shown in medium green and all other Philadelphia zip codes are shown in light green. This map was created using border data from Google Earth.
Daily manual registrations for the Philly Vax Sweepstakes at
Daily manual registrations are plotted as a function of the total city population over the six weeks following the sweepstakes’ launch (on 7 June 2021) and up to (and including) the day before the final drawing (on 19 July 2021).
Manual registrations for the Philly Vax Sweepstakes at by region and treatment period
a–c, The percentage of manual registrations for the Philly Vax Sweepstakes is calculated based on the total city population for each treated zip code (19126 in a, 19133 in b and 19142 in c) and for the control zip codes during each of the three treatment periods, highlighting the relevant treatment period for each treated zip code.
Vaccinations in the treatment and control zip codes
a–f, The panels on the left present the weekly number of first-dose vaccinations per 100,000 adult Philadelphians in each of the treated zip codes (19126 in a, 19133 in c and 19142 in e) versus the pooled 17 control zip codes and a synthetic control group. The panels on the right present the difference in the raw number of weekly first-dose vaccinations per 100,000 adult Philadelphians between the treated and control zip codes (19126 in b, 19133 in d and 19142 in f). The weekly data are plotted on the last day of a given week (for example, the data for the week 31 May–6 June is plotted on 6 June).
Lotteries have been shown to motivate behaviour change in many settings, but their value as a policy tool is relatively untested. We implemented a pre-registered, citywide experiment to test the effects of three high-pay-off, geographically targeted lotteries designed to motivate adult Philadelphians to get their COVID-19 vaccine. In each drawing, the residents of a randomly selected ‘treatment’ zip code received half the lottery prizes, boosting their chances of winning to 50×–100× those of other Philadelphians. The first treated zip code, which drew considerable media attention, may have experienced a small bump in vaccinations compared with the control zip codes: average weekly vaccinations rose by an estimated 61 per 100,000 people per week (+11%). After pooling the results from all three zip codes treated during our six-week experiment, however, we do not detect evidence of any overall benefits. Furthermore, our 95% confidence interval provides a 9% upper bound on the net benefits of treatment in our study. A citywide experiment tested the effects of three high-pay-off, geo-targeted lotteries to motivate adults to get a COVID-19 vaccine. Zip-code-targeted lotteries in which residents were given 50–100× boosts in their chances of a win did not result in higher vaccine uptake.
Despite the special role of tenure-track faculty in society, training future researchers and producing scholarship that drives scientific and technological innovation, the sociodemographic characteristics of the professoriate have never been representative of the general population. Here we systematically investigate the indicators of faculty childhood socioeconomic status and consider how they may limit efforts to diversify the professoriate. Combining national-level data on education, income and university rankings with a 2017–2020 survey of 7,204 US-based tenure-track faculty across eight disciplines in STEM, social science and the humanities, we show that faculty are up to 25 times more likely to have a parent with a Ph.D. Moreover, this rate nearly doubles at prestigious universities and is stable across the past 50 years. Our results suggest that the professoriate is, and has remained, accessible disproportionately to the socioeconomically privileged, which is likely to deeply shape their scholarship and their reproduction.
Plotting the distributions of income for Putnam County, Ohio, and Chambers County, Texas
a, Income bucket representation: the proportion of earners per income bucket is shown for two counties that have approximately the same Gini coefficient (0.46). b, Lorenz curve representation: the same income distributions are plotted as Lorenz curves, which reveals that while overall levels of inequality are the same for both distributions (that is, the same area under the curve), where inequality is concentrated differs between the counties.
The strength of evidence in favour of the two-parameter Ortega model
a, The histogram plots the AICc differences (Δi,j) between the one-parameter lognormal model (i) and the two-parameter Ortega (j). To categorize the strength of evidence, we define the following ranges: Δi,j > 10 implies decisive evidence that model j is superior to model i; Δi,j ∈ [4, 10] implies some evidence; Δi,j ∈ [−4, 4) implies inconclusive evidence; and Δi,j < −4 implies counter-evidence (that is, evidence in favour of model i over j). b, An example to illustrate the goodness of fit of one-parameter versus two-parameter models of Lorenz curves to empirical data. For the two-parameter model, we fitted the Ortega Lorenz curve model using the empirical data points and MLE, plotted next to the empirically best-fitting one-parameter model (the lognormal Lorenz curve model).
Using simulations to systematically vary the two Ortega parameters to identify their impacts on the shape of the income distribution
a, The disproportionate change exhibited by the Lorenz curve when the Ortega parameter α varies within the range of 0.01 to 1.5 leads to a more pronounced change for lower income percentiles. (The dashed off-diagonal line facilitates the recognition that the Lorenz curve is stretched more intensely in lower income percentiles.) b, Conversely, when the Ortega parameter γ varies within the range 0.01 to 0.99, the Lorenz curve exhibits a disproportionate change in the top income percentiles. For comparison, the empirical estimates across counties for α range from 0.12 to 1.23; for γ, they range from 0.3 to 0.93.
Different representations of inequality across counties in the United States
a, The Gini coefficient. b, The first Ortega parameter, α (a measure of more bottom-concentrated inequality). c, The second Ortega parameter, γ (a measure of more top-concentrated inequality). An interactive version of this figure is available at Sources: Esri, HERE, Garmin, © OpenStreetMap contributors, and the GIS User Community.
A two-parameter Ortega approach reveals significant correlations between inequality and policy outcomes across N = 3,049 US counties that the Gini coefficient misses in our dataset
Point estimates of the Pearson correlations (Gini coefficient) and partial Pearson correlations (Ortega parameters) with policy outcomes are visualized with the bounds of the 0.9995 confidence interval, using a Bonferroni correction. The figure shows the subsample of covariates (33 of 100) for which the Pearson correlations with the Gini coefficient were not significant but that exhibited at least one statistically significant partial correlation with the Ortega parameters. M, male; F, female; Q, income quartile; frac., fraction; raceadj., race adjusted.
Prior research has found mixed results on how economic inequality is related to various outcomes. These contradicting findings may in part stem from a predominant focus on the Gini coefficient, which only narrowly captures inequality. Here, we conceptualize the measurement of inequality as a data reduction task of income distributions. Using a uniquely fine-grained dataset of N = 3,056 US county-level income distributions, we estimate the fit of 17 previously proposed models and find that multi-parameter models consistently outperform single-parameter models (i.e., models that represent single-parameter measures like the Gini coefficient). Subsequent simulations reveal that the best-fitting model—the two-parameter Ortega model—distinguishes between inequality concentrated at lower- versus top-income percentiles. When applied to 100 policy outcomes from a range of fields (including health, crime and social mobility), the two Ortega parameters frequently provide directionally and magnitudinally different correlations than the Gini coefficient. Our findings highlight the importance of multi-parameter models and data-driven methods to study inequality. Moving beyond the Gini coefficient in studying inequality, Blesch et al. identify two parameters that capture inequality concentrated at the top and bottom. The results challenge mixed associations between inequality and policy outcomes.
Drawing on her personal experience as an autistic scientist–practitioner, Eloise Stark explores how we can empower neurodivergent populations in academia.
Behaviour- and demography-informed epidemic modelling (BD model)
a, Overview of our BD model, where each CBG maintains its specific SEIR model and connects with other CBGs via mobility flows. b, Capability in fitting representative curves of daily deaths. Whether daily deaths grow sub-linearly or almost linearly, our BD model (orange) fits more accurately to ground truth (green) than the SEIR (grey) and metapopulation (blue) models do. The shaded regions show the results of parameter sets that achieve an RMSE within 150% of the best result. c, Distribution of NRMSE in daily death prediction (across 100 bootstrap samples). The width of the violin indicates the probability density, and the line within the violin indicates the median value. Our BD model (orange) reduces the error by 51.9% and 35.7% compared with the SEIR model (dashed lines) and the metapopulation model (blue), respectively. d, Predicted uneven fatality rates among communities with different demographic features (across 30 independent experiments). The error bars indicate the 25th and 75th percentile values. Our BD model (orange) successfully captures the high mortality risks faced by communities with high older adult ratios, low household income, high essential worker ratios and high minority ratios, while two baseline models fail. e, Joint distributions of demographic features and mobility. The older adult ratio and average household income negatively correlate with per capita mobility (r = −0.29 and r = −0.45, respectively), while the essential worker ratio and minority ratio positively correlate with per capita mobility (r = 0.39 and r = 0.35).
Social utility and equity under different vaccine distribution strategies
a, Changes in social utility and equity, compared with the Homogeneous baseline. The red and blue points represent the strategies prioritizing the most and least disadvantaged communities, respectively. In each plot, the first to the fourth quadrants represent (1) simultaneously improving utility and equity, (2) improving equity but damaging utility, (3) simultaneously damaging utility and equity, and (4) improving utility but damaging equity. b, Change in social utility under different scenarios of vaccine hesitancy (nine MSAs). The bottom and top of each box indicate the 25th and 75th percentile values. The whiskers indicate 1.5× the interquartile range below and above the 25th and 75th percentile values. The line inside the box represents the median value of the results. When vaccine hesitancy in low-income communities is stronger, the benefit to social utility brought by prioritizing disadvantaged communities diminishes and is eventually erased, making it inferior to the baseline.
Design and justification of community risk and societal risk
a, Illustration of community risk (CR) and societal risk (SR). Each node represents a community, the node size reflects the community’s vulnerability and the colour tint reflects the number of deaths in the community, quantified by the value of D. Each edge represents inter-community mobility connections, with thickness reflecting mobility intensity. For each community, CR equals the community’s own mortality risk (green boxes), and SR equals the sum of its own mortality risk and the mortality risk it potentially presents to others (red boxes). As two representative cases, community A of transmission chain I has large CR but small SR, while community B of transmission chain II has small CR but large SR. b, OLS regression of changes in social utility with and without societal risk (across 20 bootstrap samples). The bottom and top of each box indicate the 25th and 75th percentile values. The whiskers indicate 1.5× the interquartile range below and above the 25th and 75th percentile values. Regressions with only demographic features explain on average 38.7% of the variance, measured by adjusted R² (grey boxes). The incorporation of societal risk raises the explained variance to an average of 67.5% (red boxes), greatly improving the goodness of fit of the regression model. c, OLS regression of changes in equity with and without community risk (across 20 bootstrap samples). The width of the violin indicates the probability density, and the line within the violin indicates the median value. Regressions with only demographic features explain on average 62.9%, 46.0%, 41.4% and 48.5% of the variance, respectively (grey shapes). The incorporation of community risk raises the explained variance to an average of 70.4%, 57.7%, 52.1% and 57.9%, respectively (green shapes), greatly improving the goodness of fit of the regression model. d, Joint probability distribution of community risk and societal risk, where brighter colours indicate larger probability density. There is a non-negligible positive correlation (r = 0.29) between community risk and societal risk.
Performance of the Comprehensive distribution strategy under various vaccination rates and timings
a, Changes in social utility and four dimensions of equity under eight vaccine distribution strategies. Values are normalized by the result of the Comprehensive strategy. The Comprehensive strategy (red) surpasses or is comparable to all other strategies in the five metrics, indicating its well-rounded effectiveness. In contrast, SVI-Informed (grey) and Comprehensive-Ablation (violet) result in degradation in certain dimensions of equity. b, Changes in social utility in each MSA. The bottom and top of each box indicate the 25th and 75th percentile values. Whiskers indicate 1.5× the interquartile range below and above the 25th and 75th percentile values. c, Changes in equity by age, income, occupation and race/ethnicity in each MSA. d, Overall performance of strategies under different vaccination rates. Overall performance is the sum of relative improvements in social utility and the four dimensions of equity compared with the Homogeneous baseline. The star shows overall performance if a vaccine is distributed proportionally to its real-world distribution, with a vaccination rate of 56% (that is, close to the current rate in the United States). e, Overall performance of strategies under different vaccination timings.
Balancing social utility and equity in distributing limited vaccines is a critical policy concern for protecting against the prolonged COVID-19 pandemic and future health emergencies. What is the nature of the trade-off between maximizing collective welfare and minimizing disparities between more and less privileged communities? To evaluate vaccination strategies, we propose an epidemic model that explicitly accounts for both demographic and mobility differences among communities and their associations with heterogeneous COVID-19 risks, then calibrate it with large-scale data. Using this model, we find that social utility and equity can be simultaneously improved when vaccine access is prioritized for the most disadvantaged communities, which holds even when such communities manifest considerable vaccine reluctance. Nevertheless, equity among distinct demographic features may conflict; for example, low-income neighbourhoods might have fewer elder citizens. We design two behaviour-and-demography-aware indices, community risk and societal risk, which capture the risks communities face and those they impose on society from not being vaccinated, to inform the design of comprehensive vaccine distribution strategies. Our study provides a framework for uniting utility and equity-based considerations in vaccine distribution and sheds light on how to balance multiple ethical values in complex settings for epidemic control. The authors use data-informed computational modelling and show that prioritizing vaccination efforts for the most disadvantaged communities can simultaneously improve equity and prevent the spread of disease.
The COVID-19 pandemic has increased negative emotions and decreased positive emotions globally. Left unchecked, these emotional changes might have a wide array of adverse impacts. To reduce negative emotions and increase positive emotions, we tested the effectiveness of reappraisal, an emotion-regulation strategy that modifies how one thinks about a situation. Participants from 87 countries and regions (n = 21,644) were randomly assigned to one of two brief reappraisal interventions (reconstrual or repurposing) or one of two control conditions (active or passive). Results revealed that both reappraisal interventions (vesus both control conditions) consistently reduced negative emotions and increased positive emotions across different measures. Reconstrual and repurposing interventions had similar effects. Importantly, planned exploratory analyses indicated that reappraisal interventions did not reduce intentions to practice preventive health behaviours. The findings demonstrate the viability of creating scalable, low-cost interventions for use around the world.
Schematic illustration of the scope and structure of this systematic literature review
Aiming to give an overview of VH, we recognize three types of operationalizations: conceptualizations (blue), identification of subpopulations (orange) and measurements (green). Conceptualizations of VH are analysed at three levels: (1) common themes, (2) closely related concepts and (3) potential variation in conceptualization between research field and vaccine type. Each type of operationalization and its levels are discussed in separate sections.
PRISMA flow diagram
Visualization of the process involving identification of records from databases, screening of records, assessing reports for eligibility, inclusion of eligible studies and exclusion of non-eligible reports with reasons for exclusion. The number of records or reports in each step of the process is shown in brackets.
Vaccine hesitancy (VH) is considered a top-10 global health threat. The concept of VH has been described and applied inconsistently. This systematic review aims to clarify VH by analysing how it is operationalized. We searched PubMed, Embase and PsycINFO databases on 14 January 2022. We selected 422 studies containing operationalizations of VH for inclusion. One limitation is that studies of lower quality were not excluded. Our qualitative analysis reveals that VH is conceptualized as involving (1) cognitions or affect, (2) behaviour and (3) decision making. A wide variety of methods have been used to measure VH. Our findings indicate the varied and confusing use of the term VH, leading to an impracticable concept. We propose that VH should be defined as a state of indecisiveness regarding a vaccination decision. This systematic review of 422 studies of vaccine hesitancy finds that the term is used inconsistently. Vaccine hesitancy should be defined as a psychological state of indecisiveness that people may experience when making a vaccination decision.
Marginalized scholars are often excluded from key scientific conferences owing to visa and travel restrictions, which increases inequity among academics. Every year, many scholars are prevented from presenting their accepted research at scientific conventions for reasons related to structural barriers surrounding conference attendance. In such contexts it is partly financial disparities that divide attendees from those invisibly absent, but there are additional issues that foster nonattendance, including travel bans pertaining to ethnic origin and acquisition of visas. These barriers produce synthetic selection effects that, over time, lead to systematic discrepancies between academics based on unjust sources of variation. Open and accessible link to article in Nature Human Behaviour:
Retrospective duration estimation is affected by lockdown, lockdown stringency and mobility
a, Retrospective duration estimates (minutes) as a function of veridical clock duration (minutes) during lockdown (S1; pink) and outside of it (SC; grey). Each dot represents a single participant. The regression lines were estimated from the linear mixed effect model; their 95% CIs are shown with grey shading. b, Relative retrospective duration estimates (unitless) as a function of the stringency index (a.u. between 0 and 100) for all sessions (coloured). The coloured dots are individual data points per participant and per session. The regression line was estimated from the linear model; the 95% CI is shown with grey shading. The more stringent governmental rules were, the shorter retrospective durations were estimated to be. c, Relative retrospective duration estimates (unitless) as a function of the mobility index (percent change relative to baseline, prior to lockdown; see the main text) for all sessions (coloured). Each dot is an individual data point per participant and per session. The black line is a regression line estimated from the linear model; the 95% CI is shown with grey shading. The closer to baseline mobility, the shorter retrospective durations were estimated to be.
Passage of time and subjective confinement
a, Distribution of VAS rating (0 to 100) counts for passage-of-time judgements as a function of session (colour coded). b, Passage-of-time ratings as a function of subjective confinement (5 to 20). The grey dots are individual data points (per participant, per session, per run). The black dots are the mean passage-of-time ratings binned by subjective confinement. Their size scales with the underlying number of individual data points. The black line is a regression line estimated from the linear mixed effect model; the 95% CI is shown with grey shading. The less lonely the participants felt, the faster the passage of time felt.
Subjective temporal distances
a, Subjective temporal distances from the first day of lockdown as a function of the mobility index. The black dots are the mean subjective temporal distances binned by mobility. Their size scales with the underlying number of individual grey data points. The black line is a regression line estimated from the linear regression model; the 95% CI is shown with grey shading. b, Subjective temporal distances from the first day of lockdown as a function of the index of subjective confinement. c, The distribution of future subjective temporal distances obtained for ‘next week’ (light green) significantly differed from those obtained for ‘next month’ (dark green; F(1, 3,169) = 1,171.9, P = 2.2 × 10⁻¹⁶). d,e, Subjective temporal distances to ‘next week’ (d) and ‘next month’ (e) as a function of age. The light and dark green dots are the mean subjective temporal distances binned by age. Their size scales with the underlying number of individual grey data points. f,g, Subjective temporal distances to ‘next week’ (f) and ‘next month’ (g) as a function of the subjective confinement index. h,i, Subjective temporal distances to ‘next week’ (h) and ‘next month’ (i) as a function of stringency. The black lines are regression lines estimated from the linear model; the 95% CIs are shown with grey shading.
The COVID-19 pandemic and associated lockdowns triggered worldwide changes in the daily routines of human experience. The Blursday database provides repeated measures of subjective time and related processes from participants in nine countries tested on 14 questionnaires and 15 behavioural tasks during the COVID-19 pandemic. A total of 2,840 participants completed at least one task, and 439 participants completed all tasks in the first session. The database and all data collection tools are accessible to researchers for studying the effects of social isolation on temporal information processing, time perspective, decision-making, sleep, metacognition, attention, memory, self-perception and mindfulness. Blursday includes quantitative statistics such as sleep patterns, personality traits, psychological well-being and lockdown indices. The database provides quantitative insights on the effects of lockdown (stringency and mobility) and subjective confinement on time perception (duration, passage of time and temporal distances). Perceived isolation affects time perception, and we report an inter-individual central tendency effect in retrospective duration estimation.
Hate crimes by group rank, conditional on group size
The figure displays binned scatter plots of the county-level correlation between hate crimes committed by white offenders against each group (per 100,000 inhabitants) and group size (as a share of the total county population). The blue diamonds and red dots denote bins of county-year cells where the group is, respectively, the first- and second-largest minority. The differences between first and second rank are as follows: for Black victims, t(2,564) = −3.79; P < 0.001; β = −0.965; 95% CI, (−1.464, −0.466); for Hispanic victims, t(3,084) = −9.10; P < 0.001; β = −0.660; 95% CI, (−0.802, −0.518); and for Asian victims, t(1,330) = −3.04; P = 0.002; β = −0.304; 95% CI, (−0.501, −0.108). All P values are two-tailed.
Source data
Change in hate crimes around the rank change threshold
a,b, Binned scatter plots of the county-level correlation between hate crimes committed by white offenders against each group (per 100,000 inhabitants) and the difference in size of each group from the largest among the remaining three groups. The data are restricted to county-decades where the group is first or second in rank. Linear regression lines are fitted on each side of the rank change threshold. Panel a displays the raw data, and panel b displays the data after a 90% winsorization.
Source data
Effect of rank on hate crimes
The red circles and blue squares are the point estimates from ordinary least squares regressions, and the error bars are 95% CIs of βn, the marginal effect of size rank on hate crimes per 100,000 county residents committed by white offenders. ‘Baseline’ refers to estimates from equation (1) in Supplementary Section C, and ‘Rank changes’ refers to estimates from equation (8) in Supplementary Section F. The baseline effects of rank relative to the fourth-ranked reference group (from column 1 of Supplementary Table C.1) are as follows: for the first, β = 0.249; robust s.e. = 0.034, P < 0.001; for the second, β = 0.126, robust s.e. = 0.026, P < 0.001; and for the third, β = 0.120, robust s.e. = 0.019, P < 0.001. The rank change effects of rank relative to the fourth-ranked reference group (from column 1 of Supplementary Table F.1) are as follows: for the first, β = 0.143, robust s.e. = 0.046, P = 0.002; for the second, β = 0.075, robust s.e. = 0.033, P = 0.023; and for the third, β = 0.020, robust s.e. = 0.022, P = 0.352.
Source data
People are on the move in unprecedented numbers within and between countries. How does demographic change affect local intergroup dynamics? Complementing accounts that emphasize stereotypical features of groups as determinants of their treatment, we propose the group reference dependence hypothesis: violence and negative attitudes towards each minoritized group will depend on the number and size of other minoritized groups in a community. Specifically, as groups increase or decrease in rank in terms of their size (for example, to the largest minority within a community), discriminatory behaviour and attitudes towards them should change accordingly. We test this hypothesis for hate crimes in US counties between 1990 and 2010 and attitudes in the United States and United Kingdom over the past two decades. Consistent with this prediction, we find that as Black, Hispanic/Latinx, Asian and Arab populations increase in rank relative to one another, they become more likely to be targeted with hate crimes and more negative attitudes. The rank effect holds above and beyond group size/proportion, growth rate and many other alternative explanations. This framework makes predictions about how demographic shifts may affect coalitional structures in the coming years and helps explain previous findings in the literature. Our results also indicate that attitudes and behaviours towards social categories are not intransigent or driven only by features associated with those groups, such as stereotypes. Cikara et al. propose and test the group reference dependence hypothesis, stating that violence and negative attitudes towards minoritized groups depend on the number and size of other minoritized groups in a community. Using data on hate crimes in US counties between 1990 and 2010, they show that as groups increase in rank in terms of their size, hate crimes against them become more likely.
SNP-based heritability estimates in the European ancestry sample for each of the six MAF/LD bins, and sums across bins
The error bars represent standard errors. ‘Rare’ is the sum of the MAF 0.1–1% and MAF 0.01–0.1% bins. ‘Common’ is the sum of the other MAF bins. ‘Total’ is the sum of ‘Rare’ and ‘Common’. All estimates were adjusted for demographic variables and 20 PCs (half of them from rare variants) as fixed effects along with a random effect of cohort, except for CigDay, which was adjusted for five common PCs to allow model convergence.
SNP-based heritability estimates in the European ancestry sample from sensitivity analyses
The error bars represent standard errors. The figure shows SNP-based heritability estimates from different sensitivity conditions. Heritability was estimated after adjusting for 20 common and 20 rare variant PCs (‘40 PCs’) or 50 common and 50 rare variant PCs (‘100 PCs’), after removing individuals who share IBD segments more than 2.5% of the total genome length (‘Long IBD’), after adjusting for the top 20 PCs from the IBD-based GRM (‘IBD PCs’), and after adjusting for recruitment site as a random effect (‘Site’).
Comparison of heritability estimates between current and published studies
The figure shows SNP heritability estimates across different studies. The error bars denote standard errors. ‘Pedigree’ and ‘Pedigree-FHS’ refer to ĥped2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat h^2_{{{{\mathrm{ped}}}}}$$\end{document} from whole TOPMed pedigree samples and from the FHS only. ‘WGS_EUR’ and ‘WGS_AFR’ refer to WGS-based SNP heritability estimates in individuals of European and African ancestries. Note that WGS_AFR is based on common variants only for all phenotypes except for SmkInit, which includes the contribution from MAF 0.1–1% variants. ‘Evans_imputed’ and ‘Liu_LDSC’ refer to SNP heritability estimates from Evans et al. (MAF, 1–50%; relatedness threshold, 0.02; ref. ⁷⁸) and LDSC analysis from a recent meta-analysis of tobacco use¹⁸. The red dashed line indicates the heritability estimate of smoking from a recent large meta-analysis of twin studies⁷⁹.
Common genetic variants explain less variation in complex phenotypes than inferred from family-based studies, and there is a debate on the source of this ‘missing heritability’. We investigated the contribution of rare genetic variants to tobacco use with whole-genome sequences from up to 26,257 unrelated individuals of European ancestries and 11,743 individuals of African ancestries. Across four smoking traits, single-nucleotide-polymorphism-based heritability (hSNP2) was estimated from 0.13 to 0.28 (s.e., 0.10–0.13) in European ancestries, with 35–74% of it attributable to rare variants with minor allele frequencies between 0.01% and 1%. These heritability estimates are 1.5–4 times higher than past estimates based on common variants alone and accounted for 60% to 100% of our pedigree-based estimates of narrow-sense heritability (hped2, 0.18–0.34). In the African ancestry samples, hSNP2 was estimated from 0.03 to 0.33 (s.e., 0.09–0.14) across the four smoking traits. These results suggest that rare variants are important contributors to the heritability of smoking. The team of authors led by Seon-Kyeong Jang use whole-genome sequencing data and show that rare genetic variants explain much of the ‘missing heritability’ in smoking behaviours. These results help address a long-standing mystery in behavioural genetics.
Mental health, neuroscience and neuroethics researchers must engage local African communities to enable discourses on cultural understandings of mental illness. To ensure that these engagements are both ethical and innovative, they must be facilitated with cultural competence and humility, because serious consideration of different contextual and local factors is critical.
Data has tremendous potential to build resilience in government. To realize this potential, we need a new, human-centred, distinctly public sector approach to data science and AI, in which these technologies do not just automate or turbocharge what humans can already do well, but rather do things that people cannot.
When academics support refugee scholars, everyone benefits. Scholars who are refugees face complex challenges, including bureaucratic, cultural, linguistic and academic barriers. Ahmad Al Ajlan discusses key steps that academic communities can take to support and integrate their refugee colleagues.
Policing efforts to thwart crime typically rely on criminal infraction reports, which implicitly manifest a complex relationship between crime, policing and society. As a result, crime prediction and predictive policing have stirred controversy, with the latest artificial intelligence-based algorithms producing limited insight into the social system of crime. Here we show that, while predictive models may enhance state power through criminal surveillance, they also enable surveillance of the state by tracing systemic biases in crime enforcement. We introduce a stochastic inference algorithm that forecasts crime by learning spatio-temporal dependencies from event reports, with a mean area under the receiver operating characteristic curve of ~90% in Chicago for crimes predicted per week within ~1,000 ft. Such predictions enable us to study perturbations of crime patterns that suggest that the response to increased crime is biased by neighbourhood socio-economic status, draining policy resources from socio-economically disadvantaged areas, as demonstrated in eight major US cities. Rotaru et al. introduce a transparent crime forecasting algorithm that reveals inequities in police enforcement and suggests an enforcement bias in eight US cities.
Schematic of the experimental procedure
fNIRS data were recorded at onset (T0, baseline), 5 h later (T1) and another 2 h later (T2). Training involved exposure to forward and backward stimuli in blocks, and test sessions involved random presentations of a specific set of vowels pronounced naturally or played backwards. In the consolidation phase, neonate participants were at rest and received no stimulation.
[HbO] mean amplitude results
a, Plot of β estimates for the BLUPs of the three-way interaction between group contrast (active control vs experimental), stimulus type (forward vs backward) and test phase contrast (T1 vs T2) on [HbO] mean amplitude. β values are plotted per channel on a neonate brain model (37 weeks) elaborated by ref. ⁹⁴, using the BrainNet Viewer toolbox⁹⁵. b, Violin plots of observed [HbO] values in response to forward and backward vowels in 5 of the channels listed in Table 1 (results for channel 10 (not pictured) closely resembled those illustrated for channel 7). Experimental group n = 22, active control group n = 23 and passive control group n = 21. Black dots depict means and error bars display 95% confidence intervals. Brain regions are labelled according to the abbreviations used in the main text. c, Representative examples of [HbO] and [Hb] variation over time in each of the three groups and test sessions in channel 7 set over the left ST region. Waves depict mean concentration evolution over time averaged across individual data, bounded by s.e.m. in the corresponding transparent shade.
[HbO] peak latency analysis results
a, Plot of β estimates for the BLUPs of the three-way interaction between the active control vs experimental group contrast, the stimulus type contrast (forward vs backward) and the T0 vs mean (T1, T1) phase contrast on [HbO] mean peak latency. β values are plotted per channel on a neonate brain model (37 weeks) elaborated in ref. ⁹⁴, using the BrainNet toolbox⁹⁵. b, Violin plots of observed [HbO] peak latencies in response to forward and backward vowels for channels listed in Table 2. Experimental group n = 22, active control group n = 23 and passive control group n = 21. Black dots depict means and error bars display 95% confidence intervals. c, Representative examples of [HbO] and [Hb] variation over time in each of the three groups and test sessions in channel 6 set over the left IF region. Waves depict mean concentration evolution over time averaged across individual data, bounded by s.e.m. in the corresponding transparent shade.
Functional connectivity results
Dots represent channel locations reconstructed by computing the midpoint between optodes and sensors on the basis of 3D coordinates registered for each neonate participant (Supplementary Table 3). Seed channels are highlighted with a white circle and blue halo. Dots are coloured on the basis of the average Z-score observed for the corresponding channel across the entire correlation matrix (all channels included). Lines represents correlation z-scores between pairs of channels over a threshold of 0.413, which is the absolute value of the most negative correlation observed in the experimental group at rest (see Methods and density plots in lower left quadrant). Functional connectivity intensity as measured by z-scores is depicted by hue (see colour scale), thickness (the greater the thicker) and transparency (the weaker the more transparent). The neonate brain model is from ref. ⁹⁴ and visualization was implemented using the BrainNet Viewer toolbox⁹⁵.
Human neonates can discriminate phonemes, but the neural mechanism underlying this ability is poorly understood. Here we show that the neonatal brain can learn to discriminate natural vowels from backward vowels, a contrast unlikely to have been learnt in the womb. Using functional near-infrared spectroscopy, we examined the neuroplastic changes caused by 5 h of postnatal exposure to random sequences of natural and reversed (backward) vowels (T1), and again 2 h later (T2). Neonates in the experimental group were trained with the same stimuli as those used at T1 and T2. Compared with controls, infants in the experimental group showed shorter haemodynamic response latencies for forward vs backward vowels at T1, maximally over the inferior frontal region. At T2, neural activity differentially increased, maximally over superior temporal regions and the left inferior parietal region. Neonates thus exhibit ultra-fast tuning to natural phonemes in the first hours after birth.
We must often infer latent properties of the world from noisy and changing observations. Complex, probabilistic approaches to this challenge such as Bayesian inference are accurate but cognitively demanding, relying on extensive working memory and adaptive processing. Simple heuristics are easy to implement but may be less accurate. What is the appropriate balance between complexity and accuracy? Here we model a hierarchy of strategies of variable complexity and find a power law of diminishing returns: increasing complexity gives progressively smaller gains in accuracy. The rate of diminishing returns depends systematically on the statistical uncertainty in the world, such that complex strategies do not provide substantial benefits over simple ones when uncertainty is either too high or too low. In between, there is a complexity dividend. In two psychophysical experiments, we confirm specific model predictions about how working memory and adaptivity should be modulated by uncertainty. Tavoni et al. show that complex inference strategies are worth the cognitive effort only in environments of moderate statistical complexity.
Cultural evolution of baby names
a–g, Frequency trajectories of names in the United States exhibit rich empirical dynamics (a–c, name popularity plotted as frequency divided by average frequency among the top 50 names). Some names, such as Elizabeth or Daniel (a), are perennially popular whereas others show boom–bust patterns (b). Currently trending names Emma and Evelyn (c) are actually part of century-long, intergenerational cycles (1850–1880 data from US decadal census85–100). d, Names can also be driven by popular culture: Maverick as a first name began with the TV broadcast Maverick (1957), but the movie Top Gun (1986) initiated a meteoric rise that expanded even to include female Mavericks. e, Using all name trajectories, we infer how name fitness depends on frequency. The average frequency of the top 50 names has declined over time (red line) as the United States has diversified. f, An example curve s(p) in which fitness is frequency-dependent. The relative growth rate of a type depends on the type’s frequency, p, in the population. Types with log fitness >s¯\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bar{s}$$\end{document} tend to increase in abundance while those below this replacement fitness tend to decrease. g, We approximate the curve s(p) using piecewise-constant parameterization, in which we infer the selection coefficient si for each frequency bin i (Supplementary Information 1.2).
Frequency-dependent selection on first names
a, We infer that selection tends to favour rare names and disfavour common ones in the United States, the Netherlands, France and Norway, even though any particular name (for example, Sarah, François, Bjørn) might occupy different frequencies and experience different growth rates from year to year. b, All inferred fitness curves (repeated from a) overlap and intercept zero growth at frequencies between 8 and 32 births per 10,000 (vertical grey bars). These frequencies form the threshold between rare favoured names and disfavoured common names. c, We plot abundance distributions as 1 minus the empirical cumulative distribution of names at frequency >4 × 10⁻⁶ (or >7.8 × 10⁻⁵ for Norway). These abundance distributions feature cusps near the frequencies that divide favoured and disfavoured names (grey bar), reflecting a deficit of names considered too common. Aggregation of uncensored data from the Netherlands (a) into either 1- or 5-year time steps allows us to probe lower frequencies than inverse annual birth cohort size. In all panels, shaded boxes indicate frequency bin boundaries and bootstrap 95% CIs (Supplementary Information 1.9). Bins with fewer than five names, or bias possibly exceeding 0.8% yr–1, are omitted.
Female, male and biblical names
a–c, Male versus female (a) and biblical versus non-biblical names (b) differ in inferred frequency-dependent fitness, with higher growth rates among male and biblical names. The depressed growth rates of extant female names reflects dilution caused by higher rates of innovation among female names. Biblical names, by contrast, preclude innovation: they genuinely enjoy higher fitness across all frequencies, which counteracts their dilution and produces enrichment of biblical names among top names (c). The inferred frequency-dependent fitness for the entire population (a,b, black; coinciding with Fig. 2a) is the weighted average fitness of its subpopulations.
Frequency dependence and novelty preference in dog breeds
a,b, We inferred negative frequency-dependent selection (a) in dog breed preference from an 80-year time series of >50 million dogs registered with the AKC (b). The rarest breeds increased by >10% yr–1 whereas the most common slowly declined. c, Popular dog breeds are subject to large boom–bust fads that do not occur in a model where selection is determined by frequency alone. d,e, A novelty selection model (Supplementary Information 5), in which each new type has fitness Δs greater than the previous newest type, exhibits boom–bust cycles (d) and creates an emergent association between selection and frequency (e). We fit parameters μ and Δs so that ŝ(p)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{s}(p)$$\end{document} inferred from novelty simulations matches ŝ(p)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{s}(p)$$\end{document} inferred from the AKC data (e). f, The mean effective frequency dependence inferred from these simulations converges to the true association between frequency and selection in the novelty model (Supplementary Information 1.6).
The frequency of a cultural trait can influence its tendency to be copied. We develop a maximum-likelihood method to measure such frequency-dependent selection from time series data, and we apply it to baby names and purebred dog preferences over the past century. The form of negative frequency dependence we infer among names explains their diversity patterns, and it replicates across the United States, France, Norway and the Netherlands. We find different growth rates for male versus female names, attributable to different rates of innovation, whereas biblical names enjoy a genuine selective advantage at all frequencies, which explains their predominance among top names. We show how frequency dependence emerges from a host of underlying selective mechanisms, including a preference for novelty that recapitulates boom–bust fads among dog owners. Our analysis of cultural evolution through frequency-dependent selection provides a quantitative account of social pressures to conform or to be different. Newberry and Plotkin show that the frequency of a cultural trait can influence its tendency to be copied. They develop a method to measure frequency-dependent selection and describe how it relates to the dynamics and diversity of first names and dog breed preferences, in different countries and cultures.
Illustration of the task and participants’ choice behaviour
a, In each trial, ten numbers are sequentially presented, alternating in colour between red and green. The participant then chooses a colour, without a time constraint, by pressing the corresponding key. Feedback is given, which reports the two averages and emphasizes the chosen one. b, In each trial, both red and green numbers are drawn from the same prior distribution over the range [10.00, 99.99]. In different blocks of consecutive trials, the prior is Downward (a triangular distribution with the peak at 10.00, orange line), Uniform (blue line) or Upward (a triangular distribution with the peak at 99.99, green line). PDF, probability density function. c, Choice probability: the fraction of trials in which participants choose the ‘red’ average, as a function of the true difference in the averages of the two series of numbers. d, Probability P(red|x) of choosing ‘red’ conditional on a red number x being presented, as a function of x, for the participants (thick lines) and the ideal participant (thin lines). e, Decision weight, defined as |P(red|x) − 0.5|, for the participants (thick lines) and the ideal participant (thin lines), as a function of the sum of the prior mean, x¯\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bar{x}$$\end{document}, and the absolute difference between the number and the prior mean, ∣x−x¯∣\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$| x-\bar{x}|$$\end{document}. The participants’ decision weights for numbers below the mean (dashed lines) and above the mean (solid lines) are appreciably different, whereas for the ideal participant there is only a small difference due to sampling error. In c, d and e, each point of the curves is obtained by taking the average, combining all participants’ data, of the quantity of interest (ordinate) over a sliding window of length 10 of the quantity on the abscissa, incremented by steps of length 1. In c and d, the shaded areas represent the standard error of the mean.
Models with both nonlinear transformation and varying noise best capture participants’ behaviour
a, Choice probability under the unbiased, constant-noise model (N(x, s²)) as a function of the difference in the averages of the presented numbers, for the three prior conditions. b, Model comparison statistics for the four one-stage models of noisy estimation and for the efficient-coding, Bayesian-decoding model, both with homogeneous parameters and with prior-specific parameters. For each model class, the difference in BIC from the least restrictive model class (general transformation, variable noise, prior-specific) is reported. The lower BIC for the efficient-coding, Bayesian-decoding model (a special case of the general one-stage model) indicates that it captures the participants’ behaviour more parsimoniously. c, Decision weights, |P(red|x) − 0.5|, for the unbiased, constant-noise model (top left); the unbiased, varying-noise model (top right); the model with transformation of the number and constant noise (bottom left); and the model with transformation of the number and varying noise (bottom right), for numbers below the mean (dashed lines) and above the mean (solid lines). Only the two models with transformation of the number reproduce the behaviour of the participants shown in Fig. 1e. All models are fit to the participants’ combined data.
Best-fitting noise and bias: participants encode less frequent numbers with greater noise
a, The best-fitting noise function s(x) in the N(m(x), s²(x)) model (solid lines) and the prior distribution (dashed lines) in the Downward (left), Uniform (middle) and Upward (right) conditions. The scale refers to the noise (the scale of the prior PDF is not shown). b, The bias (solid lines) and the derivative of the noise variance (multiplied by 1−p2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{1-p}{2}$$\end{document}, dashed lines) for the best-fitting N(m(x), s²(x)) model, as a function of x, in the three prior conditions. The efficient-coding, Bayesian-decoding model requires these two functions to be the same (equation (4)). The model was fit to the participants’ combined data. In b, the ranges of the vertical axes are adapted to the values taken by the function.
Efficient-coding, Bayesian-decoding model: the Fisher information, fitted to the participants’ data, is adapted to the prior
The noise function, s(x), equal to the inverse square root of the Fisher information, fitted to the participants’ data (solid lines); and the prior π(x) raised to the power −1/(p + 1), with the best-fitting value p = 0.7 (dashed lines), in the Downward (left), Uniform (middle) and Upward (right) conditions. The ordinate scale refers to the noise function (the scale for the prior is not shown). The model is fit to the participants’ combined data.
The efficient-coding, Bayesian-decoding model reproduces the participants’ behaviour, and the Fisher information functions fitted to the participants’ data improve the performance ratio in the Upward and Downward conditions
a, Noise s(x) implied by the Fisher information in the three prior conditions. b, Bias m(x) − x in the three prior conditions, also implied by the Fisher information. c, Decision weights |P(red|x) − 0.5| predicted by the efficient-coding, Bayesian-decoding model. d, Probability of choosing ‘red’ conditional on a red number x being presented, in the model (solid lines) and our approximation of the model prediction (dashed lines; Methods). e, Choice probabilities implied by the model as a function of the difference between the averages of the numbers presented. f, Performance ratios over numbers sampled from the Downward (left), Uniform (middle) and Upward (right) priors, for the participants (black dots) and implied by the estimated encoding rules (bars) for each of the three prior conditions. With numbers sampled from the Downward prior, the Downward encoding rule, IDownward, yields the best performance (left orange bar), whereas with numbers sampled from the Upward prior, the Upward encoding rule, IUpward, results in the best performance (right green bar). The black vertical lines represent the standard deviations of the performance ratios (for the model, each standard deviation is obtained from 30,000 computations of the performance ratio of the model simulated with all the trials faced by the participants; for the participants, this quantity is bootstrap-estimated from the data, with 30,000 bootstrap samples in each case). All analyses use the participants’ combined data.
Humans differentially weight different stimuli in averaging tasks, which has been interpreted as reflecting encoding bias. We examine the alternative hypothesis that stimuli are encoded with noise and then optimally decoded. Under a model of efficient coding, the amount of noise should vary across stimuli and depend on statistics of the stimuli. We investigate these predictions through a task in which the participants are asked to compare the averages of two series of numbers, each sampled from a prior distribution that varies across blocks of trials. The participants encode numbers with a bias and a noise that both depend on the number. Infrequently occurring numbers are encoded with more noise. We show how an efficient-coding, Bayesian-decoding model accounts for these patterns and best captures the participants’ behaviour. Finally, our results suggest that Wei and Stocker’s “law of human perception”, which relates the bias and variability of sensory estimates, also applies to number cognition. Prat-Carrabin and Woodford show that the bias and variance in participants’ estimates of numbers both depend on the numbers and on the prior, suggesting an optimal use of limited representational capacities through efficient coding and Bayesian decoding.
Heterogeneous treatment effects of remote learning on dropout risk and standardized test scores by grade
a,b, Effect sizes (bars) estimated through grade-specific OLS regressions using the differences-in-differences model, with 95% confidence intervals (error bars) based on standard errors clustered at the school level, where the dependent variable is high dropout risk (=1 if the student had no maths or Portuguese grades on record for that school quarter, and 0 otherwise, N = 8,543,586) (a) or scores from quarterly standardized tests (AAPs), averaging maths and Portuguese scores for that school quarter (N = 7,097,042) (b). All regressions follow the specification in column 5 of Table 1, only restricting observations to each grade. We normalize each effect size by its baseline mean, to express them as percentage changes. In a, the estimates are divided by the variation in the percentage of students with dropout risk = 1 between Q1 and Q4 of 2019 within each grade. In b, the estimates are divided by the variation in standardized test scores between Q1 and Q4 of 2019 within each grade. All columns include an indicator variable equal to 1 for municipalities that authorized schools to reopen from September 2020 onwards, and 0 otherwise (allowing its effects to vary at Q4), and a third-degree polynomial of propensity scores, and re-weight observations by the inverse of their propensity score.
The transition to remote learning in the context of coronavirus disease 2019 (COVID-19) might have led to dramatic setbacks in education. Taking advantage of the fact that São Paulo State featured in-person classes for most of the first school quarter of 2020 but not thereafter, we estimate the effects of remote learning in secondary education using a differences-in-differences strategy that contrasts variation in students’ outcomes across different school quarters, before and during the pandemic. We also estimate intention-to-treat effects of reopening schools in the pandemic through a triple-differences strategy, contrasting changes in educational outcomes across municipalities and grades that resumed in-person classes or not over the last school quarter in 2020. We find that, under remote learning, dropout risk increased by 365% while test scores decreased by 0.32 s.d., as if students had only learned 27.5% of the in-person equivalent. Partially resuming in-person classes increased test scores by 20% relative to the control group.
Study paradigm and different neural responsivity to semantic and syntactic violation tasks
a, Examples of stimuli in the sentence processing task. The words marked as their corresponding colour are the keywords for analyses unless otherwise specified. Correct sentences (CORR, red); syntactic violation of local phrase (SYN-P, orange); violation of syntactic category (SYN-C, green); semantic violation (SEM, blue); the black line indicates word stimuli presenting time (400 ms) and the purple line indicates fixation time (100 ms). b, Classification accuracy from an array of frequency bands using the support vector clustering, indicated that 70–150 Hz is the best frequency band to interpret the difference in processing each of the corresponding linguistic components. The x axis indicates the lower frequency boundary and the y axis indicates the upper-frequency boundary. c, MRI reconstruction of participant 3’s brain with high-density grid electrodes (grey). Electrodes that responded to linguistic stimuli were labelled in red (Methods). d, Cortical activity for representative electrodes from the first presentation word to the last word of one participant; the coloured arrow indicates the onset of the keyword (mean ± s.e.m.). e, Cortical activity for representative electrodes after the keyword (mean ± s.e.m.). The shaded area corresponds to the discriminating period (P < 0.05, ANOVA, two tailed, Bonferroni correction for the number of electrodes).
Source data
Spatial representation of syntax and semantics
a, Clustered time course from all responding electrodes. b, F value of responding electrodes to each violation condition and Venn diagram. The F value was from the comparison of respective violation sentences with the correct sentences for each responding electrode. c, Euclidean distance of the response to each linguistic condition; the y axis indicates the relative distance. d, KDE analysis for the centre of each linguistic feature processing. e, Permutation test for the epicentre of each violation condition. f, Simulation of the diameter of each functional circuit.
Source data
Temporal representation of syntax and semantics
a, Example electrodes showing temporal dissociations of different conditions (the box centre defined as the mean of the data; the box extends from the first quartile to the third quartile of the data, with a line at the median; the whiskers extend from 10% to 90% of ranked data). b, Time course of each condition with top 30 responding electrodes ranked by responding amplitude. Black-dashed lines indicate averaged onset and offset. c, Onset, and peak delay of all responding electrodes. d, Quantification of responding onset and peak (left) and comparison among SYN-P, SYN-C, and SEM. ***P < 0.001, one-way ANOVA, two tailed, Bonferroni corrected, n = 106, electrodes in pars opercularis, pars triangularis and precentral. e, Peak latency along anterior–posterior and superior–inferior axis. The x axis is the x or y coordinate of electrodes and the y values are their responding peaks and onset time (solid line, linear regression model for the location and latency; shaded colour area, 95% confidence interval of the linear regression model). f, Summary of the temporal dissociation. Filled circles denote response peak; hollow squares denote response onset.
Source data
Human languages are based on syntax, a set of rules which allow an infinite number of meaningful sentences to be constructed from a finite set of words. A theory associated with Chomsky and others holds that syntax is a mind-internal, universal structure independent of semantics. This theory, however, has been challenged by studies of the Chinese language showing that syntax is processed under the semantic umbrella, and is secondary and not independent. Here, using intracranial high-density electrocorticography, we find distinct spatiotemporal patterns of neural activity in the left inferior frontal gyrus that are specifically associated with syntactic and semantic processing of Chinese sentences. These results suggest that syntactic processing may occur before semantic processing. Our findings are consistent with the view that the human brain implements syntactic structures in a manner that is independent of semantics.
Temporal behaviour of the fractions of Searches and News from All Sources in Italy
The Searches (red, left y axis) and News from All Sources (blue, right y axis) for the keyword ‘coronavirus’ were recorded from 6 December 2019 to 31 August 2020. Searches are reported as a percentage of the maximum observed in the monitored period. News from All Sources is represented by the daily fraction of articles containing at least three keyword occurrences (see Methods). The improved model (black line) leverages the past News from All Sources and Searches, together with present Searches, to infer the dynamics of News from All Sources.
The ranked components of Stot, representing ‘coronavirus’ subdomains sorted by total news demand over the observed time
Middle: Stot, red text. On the sides of each keyword, a tag indicates the rank in NAStot\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{N}}}}_{{{{{\mathrm{AS}}}}}_{\mathrm{tot}}}$$\end{document} for News from All Sources (left), and in NQStot\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{N}}}}_{{{{{\mathrm{QS}}}}}_{\mathrm{tot}}}$$\end{document} for News from Questionable Sources (right). Tags are distanced from the centre by the amount of rank mismatch to Searches ranks. Tags are coloured to highlight the rank closest to the Searches rank: blue for News from All Sources and green for News from Questionable Sources.
The combined index and the normalized time series of news whose sources were annotated as questionable (NQS)
The time series of NQS was normalized using the total number of ‘coronavirus’-related News compared with the combined index. The combined index is defined as a linear combination of the weighted modelling error for the local fitting of NAS within the improved Vector Auto-Regression model and the cosine distance between the semantic vectors of Searches and News from All Sources. The parameters of the combination were fitted in the training set and then tested in the validation set. Green background represents the training set while pink background represents the validation set.
Misinformation threatens our societies, but little is known about how the production of news by unreliable sources relates to supply and demand dynamics. We exploit the burst of news production triggered by the COVID-19 outbreak through an Italian database partially annotated for questionable sources. We compare news supply with news demand, as captured by Google Trends data. We identify the Granger causal relationships between supply and demand for the most searched keywords, quantifying the inertial behaviour of the news supply. Focusing on COVID-19 news, we find that questionable sources are more sensitive than general news production to people’s interests, especially when news supply and demand mismatched. We introduce an index assessing the level of questionable news production solely based on the available volumes of news and searches. We contend that these results can be a powerful asset in informing campaigns against disinformation and providing news outlets and institutions with potentially relevant strategies. Studying news supply and demand amidst the COVID-19 pandemic in Italy, Gravino and coauthors show that news production by unreliable sources is more sensitive to the public interest than reliable news.
Behavioural task
a, The simplified two-step task was presented on a computer screen with four circles visible on a grey background: two central circles (upper and lower) and two side circles (right and left). Each circle was coloured yellow when available for selection and black when unavailable. The circles could be selected by pressing the corresponding arrow key on the computer keyboard. Each trial started with the central circles turning yellow, prompting the first-step choice between either the upper or the lower circle (1). Following this, one of the side circles (left or right) turned yellow (2), with differing probabilities (b). The participant then selected the yellow side circle, resulting in a probabilistic monetary reward, indicated by the circle changing to the image of a coin (3, left). No reward was indicated by the circle changing back to black (3, right). b, Transition probabilities linking the first-step choice (up or down) to the second-step state (left or right). Each first-step option commonly (in 80% of trials) led to one second-step state and rarely (in 20% of trials) to the other. In the Fixed version of the task, the transition probabilities were counterbalanced across participants, with half experiencing the type A probabilities (top) and half the type B (bottom). In the Changing version of the task, the transition probabilities alternated between types A and B in blocks. c, Reward probability blocks. The reward probabilities for the side circles changed in blocks that were higher on one side or the other (p = 0.8 versus p = 0.2, non-neutral blocks) or neutral (p = 0.4 for both sides). Non-neutral blocks ended when participants consistently chose the first-step option that most frequently led to the high-reward-probability side. Neutral blocks ended probabilistically, independent of the participants’ behaviour (Methods). To maximize the reward rate, the participants had to choose the first-step action that commonly led to the second-step state with a higher reward probability, tracking the best option across reward-probability reversals. Credit: a, Coin image, United States Mint.
Uninstructed behaviour is predominantly model-free
a, Session 1 stay probability analysis showing the probability of repeating the first-step choice on the next trial as a function of trial outcome (rewarded (1) or not rewarded (0)) and state transition (common (C) or rare (R)). The error bars indicate cross-participant s.e.m. b, Stay probability for rewarded and non-rewarded trials as a function of trial number in session 1. The shaded areas show across-participant standard error. c, Stay probabilities for session 3. d, Logistic regression analysis of how the outcome (rewarded or not), the transition (common or rare) and their interaction predict the probability of repeating the same choice on the subsequent trial. The dots indicate maximum a posteriori parameter values for individual participants, the horizontal bars indicate the population mean and the vertical bars indicate the 95% CI of the mean. In this and other panels, blue indicates session 1, while red indicates session 3. The influence of both state transition (null 95% CI, (−0.18, 0.18); coefficient change, 0.27; P = 0.003; permutation test) and the transition–outcome interaction (null 95% CI, (−0.25, 0.24); coefficient change, 0.39; P < 0.001) increased between sessions 1 and 3. e, Reaction times after common and rare transitions in sessions 1 and 3. Key-press reaction times at the second step became faster overall between sessions 1 and 3 (main effect of session, F1,66 = 21.1, P < 0.0001, ηp² = 0.24), but this was more pronounced following common than rare transitions (session–transition interaction, F1,66 = 21.1, P = 0.008, ηp² = 0.1, repeated measures ANOVA). f, Comparison of mixture model fits between session 1 and session 3. The dots and bars are represented as in d. The value learning rate increased significantly between sessions 1 and 3 (null 95% CI, (−0.17, 0.17); parameter change, 0.18; P = 0.03). All panels show the results of analysis of uninstructed behaviour in 67 healthy volunteers. RL model parameters: MF, model-free strength; MB, model-based strength; αQ, value learning rate; λ, eligibility trace; αT, transition probability learning rate; bias, choice bias; pers., choice perseveration. Asterisks indicate significant differences: *P < 0.05; **P < 0.01; ***P < 0.001.
Impaired learning of model-based control from experience in OCD
a–h, Participants with OCD (n = 46) are represented in a–d, and those with mood and anxiety disorders (n = 49) are represented in e–h. a,e, Stay probability analysis for session 1 (left, blue) and session 3 (right, red), as in Fig. 2a. b,f, Logistic regression analysis of stay probabilities, as in Fig. 2d. In the OCD group, the influence of trial outcome on stay probability increased between sessions 1 and 3 (null 95% CI, (−0.35, 0.36); coefficient change, 0.58; P < 0.001; permutation test). In the group with mood and anxiety disorders, the influence of outcome (null 95% CI, (−0.33, 0.35); coefficient change, 0.63; P < 0.001) and transition (null 95% CI, (−0.25, 0.25); coefficient change, 0.33; P = 0.011) increased. c,g, Second-step reaction times after common and rare transitions in sessions 1 and 3. In the OCD group, reaction times were faster following common than following rare transitions (main effect of transition, F1,45 = 51.3, P < 0.0001, ηp² = 0.53, repeated measures ANOVA) and faster in session 3 than session 1 (main effect of session, F1,45 = 10, P = 0.003, ηp² = 0.18). In the group with mood and anxiety disorders, second-step reaction times were faster following common than following rare transitions (main effect of transition, F1,48 = 34.2, P < 0.0001, ηp² = 0.42, repeated measures ANOVA) and faster in session 3 than session 1 (main effect of session, F1,48 = 30.5, P < 0.0001, ηp² = 0.39). d,h, Comparison of RL mixture model fits, as in Fig. 2f. In the OCD group, the influence of model-free action values on choice increased between sessions 1 and 3 (null 95% CI, (−1.36, 1.32); parameter change, 1.71; P = 0.012; permutation test). In the mood and anxiety disorders group, the value learning rate increased between sessions 1 and 3 (null 95% CI, (−0.21, 0.21); coefficient change, 0.26; P = 0.011).
Explicit knowledge increases model-based control
a–j, To avoid ceiling effects, analysis of debriefing effects was conducted in healthy volunteers who were model-free at session 3, as assessed by a likelihood ratio test (debriefing group, n = 41, a–e; no-debriefing group, n = 16, f–j). a,f, Per-participant likelihood ratio test for the use of a model-based strategy at session 3 (left) and session 4 (right). The colours indicate whether each participant’s data were better explained by a mixture of model-free and model-based RL (green) or by model-free RL only (blue), using a P < 0.05 threshold for rejecting the simpler model. The y axis shows the difference in log likelihood between the models. b,g, Stay probability analysis showing the probability of repeating the first-step choice on the next trial as a function of trial outcome and state transition. In these and the remaining panels, red indicates session 3, while yellow indicates session 4. The error bars show cross-participant s.e.m. c,h, Logistic regression analysis of how the outcome, the transition and their interaction predict the probability of repeating the same choice on the subsequent trial. The dots indicate maximum a posteriori values for individual participants, the horizontal bars indicate the population mean and the vertical bars indicate the 95% CI of the mean. Following debriefing, the influence of state transition (null 95% CI, (−0.42, 0.42); coefficient change, 0.75; P < 0.001; permutation test) and the transition–outcome interaction (95% CI, (−0.51, 0.50); coefficient change, 1.07; P < 0.001) increased. d,i, Second-step reaction times following common and rare transitions. Following debriefing, the influence of transition on reaction time increased (session–transition interaction, F1,40 = 59.6, P < 0.0001, ηp² = 0.59, repeated measures ANOVA). e,j, Comparison of mixture model fits. The dots and bars are as in c. Following debriefing, the influence of model-based action values on choice increased (null 95% CI, (−0.70, 0.70); parameter change, 1.17; P < 0.001), the influence of model-free action values on choice decreased (null 95% CI, (−0.79, 0.79); parameter change, −1.04; P = 0.006), value learning rate increased (null 95% CI, (−0.18, 0.18); parameter change, 0.29; P < 0.001), the eligibility trace parameter decreased (null 95% CI, (−0.16, 0.17); parameter change, −0.23; P = 0.006) and the perseveration parameter increased (null 95% CI, (−0.75, 0.76); parameter change, 1.63; P < 0.001). RL model parameters are as in Fig. 2f.
Explicit knowledge increases model-based control in OCD
a–j, Effects of debriefing in 41 individuals with OCD (a–e) and 37 individuals with mood and anxiety disorders (f–j) who were model-free at session 3 as assessed by likelihood ratio tests. a,f, Per-participant likelihood ratio test for the use of a model-based strategy at session 3 (left) and session 4 (right), as in Fig. 4a. b,g, Stay probability analysis, as in Fig. 4b. c,h, Logistic regression analysis of stay probabilities, as in Fig. 4c. In the OCD group, debriefing increased the influence of both the transition (null 95% CI, (−0.35, 0.37); coefficient change, 0.56; P < 0.001; permutation test) and the transition–outcome interaction (null 95% CI, (−0.50, 0.51); parameter change, 0.88; P < 0.001). In the group with mood and anxiety disorders, debriefing increased the influence of both transition (null 95% CI, (−0.40, 0.41); coefficient change, 0.64; P < 0.001; permutation tests) and the transition–outcome interaction (null 95% CI, (−0.57, 0.59); coefficient change, 1.20; P < 0.001). d,i, Second-step reaction times following common and rare transitions. In both the group with OCD and those with mood and anxiety disorders, debriefing increased differences in second-step reaction times between common and rare transition trials (session–transition interaction, OCD group, F1,40 = 30.8, P < 0.0001, ηp² = 0.43; mood and anxiety group, F1,36 = 26.2, P < 0.0001, ηp² = 0.42; repeated measures ANOVA). e,j, Comparison of mixture model fits, as in Fig. 4e. In the OCD group, following debriefing, the influence of model-based action values on choice increased (null 95% CI, (−0.63, 0.64); parameter change, 1.22; P < 0.001), the eligibility parameter decreased (null 95% CI, (−0.19, 0.19); parameter change, −0.24; P = 0.017), the transition learning rate decreased (null 95% CI, (−0.21, 0.21); parameter change, −0.24; P = 0.019) and the perseveration parameter increased (null 95% CI, (−0.67, 0.66); parameter change, 0.77; P = 0.023). In individuals with mood and anxiety disorders, following debriefing, the influence of model-based action values on choice increased (null 95% CI, (−0.81, 0.81); parameter change, 1.63; P < 0.001), the influence of model-free action values decreased (null 95% CI, (−0.69, 0.71); parameter change, −0.82; P = 0.019), the value learning rate increased (null 95% CI, (−0.17, 0.18); parameter change, 0.21; P = 0.015), the eligibility parameter decreased (null 95% CI, (−0.15, 0.15); parameter change, −0.15; P = 0.043) and the transition learning rate decreased (null 95% CI, (−0.34, 0.33); parameter change, −0.38; P = 0.024).
Explicit information obtained through instruction profoundly shapes human choice behaviour. However, this has been studied in computationally simple tasks, and it is unknown how model-based and model-free systems, respectively generating goal-directed and habitual actions, are affected by the absence or presence of instructions. We assessed behaviour in a variant of a computationally more complex decision-making task, before and after providing information about task structure, both in healthy volunteers and in individuals suffering from obsessive-compulsive or other disorders. Initial behaviour was model-free, with rewards directly reinforcing preceding actions. Model-based control, employing predictions of states resulting from each action, emerged with experience in a minority of participants, and less in those with obsessive-compulsive disorder. Providing task structure information strongly increased model-based control, similarly across all groups. Thus, in humans, explicit task structural knowledge is a primary determinant of model-based reinforcement learning and is most readily acquired from instruction rather than experience.
Detecting and learning temporal regularities is essential to accurately predict the future. A long-standing debate in cognitive science concerns the existence in humans of a dissociation between two systems, one for handling statistical regularities governing the probabilities of individual items and their transitions, and another for handling deterministic rules. Here, to address this issue, we used finger tracking to continuously monitor the online build-up of evidence, confidence, false alarms and changes-of-mind during sequence processing. All these aspects of behaviour conformed tightly to a hierarchical Bayesian inference model with distinct hypothesis spaces for statistics and rules, yet linked by a single probabilistic currency. Alternative models based either on a single statistical mechanism or on two non-commensurable systems were rejected. Our results indicate that a hierarchical Bayesian inference mechanism, capable of operating over distinct hypothesis spaces for statistics and rules, underlies the human capability for sequence processing. Maheu et al. show that human probabilistic and deterministic sequence processing can be modelled under a hierarchical Bayesian inference model, with distinct hypothesis spaces for statistics and rules, linked by a single probabilistic currency.
Making good decisions requires thinking ahead, but the huge number of actions and outcomes one could consider makes exhaustive planning infeasible for computationally constrained agents, such as humans. How people are nevertheless able to solve novel problems when their actions have long-reaching consequences is thus a long-standing question in cognitive science. To address this question, we propose a model of resource-constrained planning that allows us to derive optimal planning strategies. We find that previously proposed heuristics such as best-first search are near optimal under some circumstances but not others. In a mouse-tracking paradigm, we show that people adapt their planning strategies accordingly, planning in a manner that is broadly consistent with the optimal model but not with any single heuristic model. We also find systematic deviations from the optimal model that might result from additional cognitive constraints that are yet to be uncovered.
Cross-cultural regularities in infant-directed vocalizations
a, We recorded examples of speech and song from 21 urban, rural or small-scale societies, in many languages. The map indicates the approximate location of each society and is colour-coded by the language family or subgroup represented by the society. b, Machine-learning classification demonstrates the stereotyped acoustics of infant-directed speech and song. We trained two LASSO models, one for speech and one for song, to classify whether recordings were infant- or adult-directed on the basis of their acoustic features. These predictors were regularized using fieldsite-wise cross-validation, such that the model optimally classified infant-directedness across all 21 societies studied. The vertical bars represent the mean classification performance across societies (n = 21 societies for both speech and song; quantified via receiver operating characteristic/AUC); the error bars represent 95% CI of the mean; the points represent the performance estimate for each fieldsite; and the horizontal dashed lines indicate chance level of 50% AUC. The horizontal bars show the six acoustic features with the largest influence in each classifier; the green and red triangles indicate the direction of the effect, for example, with median pitch having a large, positive effect on classification of infant-directed speech. The full results of the variable selection procedure are in Supplementary Table 2, with further details in Methods.
How people alter their voices when vocalizing to infants
Eleven acoustic features had a statistically significant difference between infant- and adult-directed vocalizations, within-voices, in speech, song or both. Consistent with the LASSO results (Fig. 1b and Supplementary Table 2), the acoustic features operated differently across speech and song. For example, median pitch was far higher in infant-directed speech than in adult-directed speech, whereas median pitch was comparable across both forms of song. Some features were highly consistent across fieldsites (for example, lower inharmonicity in infant-directed speech than adult-directed speech), whereas others were more variable (for example, lower roughness in infant-directed speech than in adult-directed speech). The boxplots, which are ordered approximately from largest to smallest differences between effects across speech and song, represent each acoustic feature’s median (vertical black lines) and IQR (boxes); the whiskers indicate 1.5× IQR; the notches represent the 95% CI of the medians; and the doughnut plots represent the proportion of fieldsites where the main effect repeated, based on estimates of fieldsite-wise random effects. Only comparisons that survived an exploratory–confirmatory analysis procedure are plotted; the faded boxplots denote comparisons that did not reach statistical significance in confirmatory analyses. Significance values are computed via linear combinations with two-sided tests, following multilevel mixed-effects models (n = 1,570 recordings); *P < 0.05, **P < 0.01, ***P < 0.001; no adjustments were made for multiple comparisons, given the exploratory–confirmatory approach taken. Regression results are in Supplementary Table 3 and a full report of fieldsite-level estimates is in Supplementary Table 5. Note: the model estimates are normalized jointly on speech and song data so as to enable comparisons across speech and song for each feature; as such, the absolute distance from 0 for a given feature is not directly interpretable but estimates are directly comparable across speech and song. ID, infant-directed; AD, adult-directed.
Naive listeners distinguish infant-directed vocalizations from adult-directed vocalizations across cultures
Participants listened to vocalizations drawn at random from the corpus, viewing the prompt ‘Someone is speaking or singing. Who do you think they are singing or speaking to?’ They could respond with either ‘adult’ or ‘baby’ (Extended Data Fig. 3). From these ratings (after exclusion n = 473 song recordings; n = 394 speech recordings), we computed listener sensitivity (d′\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d^{\prime}$$\end{document}). a, Listeners reliably detected infant-directedness in both speech and song, overall (indicated by the diamonds, with 95% CI indicated by the horizontal lines) and across many fieldsites (indicated by the black dots), although the strength of the fieldsite-wise effects varied substantially (see the distance between the vertical dashed line and the black dots; the shaded regions represent 50%, 80% and 95% CI, in increasing order of lightness). Note that one fieldsite-wise d′\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d^{\prime}$$\end{document} could not be estimated for song; complete statistical reporting is in Supplementary Table 5. b, The participants in the citizen-science experiment hailed from many countries; the gradients indicate the total number of vocalization ratings gathered from each country. c, The main effects held across different combinations of the linguistic backgrounds of vocalizer and listener. We split all trials from the main experiment into three groups: those where a language the listener spoke fluently was the same as the language of the vocalization (n = 82,094), those where a language the listener spoke fluently was in the same major language family as the language of the vocalization (n = 110,664) and those with neither type of relation (n = 285,378). The plot shows the estimated marginal effects of a mixed-effects model predicting d′\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d^{\prime}$$\end{document} values across language and music examples, after adjusting for fieldsite-level effects. The error bars represent 95% CI of the mean. In all three cases, the main effects replicated; increases in linguistic relatedness corresponded with increases in sensitivity.
Human inferences about infant-directedness are predictable from acoustic features of vocalizations
To examine the degree to which human inferences were linked to the acoustic forms of the vocalizations, we trained two LASSO models to predict the proportion of ‘baby’ responses for each non-confounded recording from the human listeners. While both models explained substantial variability in human responses, the model for speech was more accurate than the model for song, in part because the human listeners erroneously relied on acoustic features for their predictions in song that less reliably characterized infant-directed song across cultures (Figs. 1b and 2). Each point represents a recorded vocalization (after exclusions n = 528 speech recordings; n = 587 song recordings), plotted in terms of the model’s estimated infant-directedness of the model and the average ‘infant-directed’ rating from the naive listeners; the barplots depict the relative explanatory power of the top eight acoustical features in each LASSO model, showing which features were most strongly associated with human inferences (the green or red triangles indicate the directions of effects, with green higher in infant-directed vocalizations and red lower); the dashed diagonal lines represent a hypothetical perfect match between model predictions and human guesses; the solid black lines depict linear regressions (speech: F(1,526) = 773, R² = 0.59; song: F(1, 585) = 126, R² = 0.18; both P < 0.0001; P values computed using robust standard errors); and the grey ribbons represent the standard errors of the mean, from the regressions.
When interacting with infants, humans often alter their speech and song in ways thought to support communication. Theories of human child-rearing, informed by data on vocal signalling across species, predict that such alterations should appear globally. Here, we show acoustic differences between infant-directed and adult-directed vocalizations across cultures. We collected 1,615 recordings of infant- and adult-directed speech and song produced by 410 people in 21 urban, rural and small-scale societies. Infant-directedness was reliably classified from acoustic features only, with acoustic profiles of infant-directedness differing across language and music but in consistent fashions. We then studied listener sensitivity to these acoustic features. We played the recordings to 51,065 people from 187 countries, recruited via an English-language website, who guessed whether each vocalization was infant-directed. Their intuitions were more accurate than chance, predictable in part by common sets of acoustic features and robust to the effects of linguistic relatedness between vocalizer and listener. These findings inform hypotheses of the psychological functions and evolution of human communication.
Causal framework
Proposed causal framework with a directed acyclic graph outlining potential relationships between large-scale cash transfer programmes and HIV-related outcomes, mediated through an anti-poverty effect. The green box is the exposure of interest (cash transfer programmes). The blue boxes are the HIV-related outcomes of interest, two more proximal (sex behaviours and HIV treatment cascade) and two more distal (HIV incidence and AIDS-related deaths). The orange boxes are ancestors of both the exposure and the outcomes (that is, confounders). Underneath each box are the covariates used to measure the constructs within the boxes. The covariates from the DHS are individual-level, and all other covariates are country-level. DHS, Demographic and Health Survey; GDP, gross domestic product; PEPFAR, The US President’s Plan for AIDS Relief; UNAIDS, The Joint United Nations Programme on HIV/AIDS.
Timeline of the included countries. The green lines indicate the cash transfer periods, and the coloured dots indicate years during which a DHS was conducted.
Temporal trends in country-level outcomes
Adjusted IRRs of new HIV infections and AIDS-related deaths (N = 976 country-years), and adjusted change in the proportion of people with HIV receiving antiretroviral therapy (N = 796 country-years), as a function of year of the cash transfer period. The data are presented as IRRs with 95% CIs.
Interaction analyses
Interaction analyses of baseline HIV prevalence at the start of the cash transfer period (above or below the median, 3.7%) and impoverished population coverage of the cash transfer programme(s) (above or below the median, 23%) for individual-level (stratified by sex) and country-level outcomes, with adjusted ORs for the individual-level outcomes and adjusted IRRs for the country-level outcomes, and P values for the interactions (F statistic, two-sided comparisons, no adjustment for multiple comparisons). The data are presented as IRRs with 95% CIs, with sample sizes as follows: country-level N = 976 country-years; individual-level female N = 1,295,177; individual-level male N = 590,556. STI, sexually transmitted infection.
Many countries have introduced cash transfer programmes as part of their poverty reduction and social protection strategies. These programmes have the potential to overcome drivers of HIV risk behaviours and usage of HIV services, but their overall effects on HIV-related outcomes remain unknown. Here we evaluate the effects of cash transfer programmes covering >5% of the impoverished population on country- and individual-level HIV-related outcomes in 42 countries with generalized epidemics. Cash transfer programmes were associated with a lower probability of sexually transmitted infections among females (odds ratio, 0.67; 95% confidence interval (CI), 0.50–0.91; P = 0.01), a higher probability of recent HIV testing among females (odds ratio, 2.61; 95% CI, 1.15–5.88; P = 0.02) and among males (odds ratio, 3.19; 95% CI, 2.45–4.15; P < 0.001), a reduction in new HIV infections (incidence rate ratio, 0.94; 95% CI, 0.89–0.99; P = 0.03) and delayed improvements in antiretroviral coverage (3%; 95% CI, 0.3–5.7 at year 2; P = 0.03) and AIDS-related deaths (incidence rate ratio, 0.91; 95% CI, 0.83–0.99 at year 2; P = 0.03). Anti-poverty programmes can play a greater role in achieving global targets for HIV prevention and treatment. Cash transfers are a popular anti-poverty strategy worldwide. In this study of 42 countries over 24 years, Richterman and Thirumurthy find that large cash transfer programmes were associated with improvements in a variety of HIV-related outcomes.
MR plots for the relationship of relative carbohydrate intake (N = 268,922) with MDD (N = 143,265)
a, Scatter plot of SNP effects on relative carbohydrate intake versus MDD, with the slope of each line corresponding to the estimated MR effect per method. The data are expressed as raw β values with 95% CIs. b, Forest plot of individual and combined SNP MR-estimated effect sizes. The effect estimates represent the log odds for MDD per one-s.d. increase in mean relative carbohydrate intake, and the error bars represent 95% CIs.
Mediation analysis of the effect of relative carbohydrate intake on MDD via potential mediators
a, Two-step MR analysis framework. Step 1 estimated the causal effect of the exposure on the potential mediators, and step 2 assessed the causal effect of the mediators on MDD risk. ‘Direct effect’ indicates the effect of relative carbohydrate intake on MDD risk after adjusting for the mediator. ‘Indirect effect’ indicates the effect of relative carbohydrate intake on MDD risk through the mediator. IVs, instrumental variables. b, Summary MR estimates derived from the IVW, weighted median, weighted mode, MR-RAPS and MR-Egger methods for the effect of relative carbohydrate intake (N = 268,922) on BMI (N = 322,154) (left) and the effect of BMI (N = 681,275) on MDD (N = 143,265) (right). The error bars represent 95% CIs. All statistical tests were two-sided. P < 0.05 was considered significant.
Growing evidence suggests that relative carbohydrate intake affects depression; however, the association between carbohydrates and depression remains controversial. To test this, we performed a two-sample bidirectional Mendelian randomization (MR) analysis using genetic variants associated with relative carbohydrate intake (N = 268,922) and major depressive disorder (N = 143,265) from the largest available genome-wide association studies. MR evidence suggested a causal relationship between higher relative carbohydrate intake and lower depression risk (odds ratio, 0.42 for depression per one-standard-deviation increment in relative carbohydrate intake; 95% confidence interval, 0.28 to 0.62; P = 1.49 × 10−5). Multivariable MR indicated that the protective effect of relative carbohydrate intake on depression persisted after conditioning on other diet compositions. The mediation analysis via two-step MR showed that this effect was partly mediated by body mass index, with a mediated proportion of 15.4% (95% confidence interval, 6.7% to 24.1%). These findings may inform prevention strategies and interventions directed towards relative carbohydrate intake and depression. Using genomic data and Mendelian randomization techniques, Yao and coauthors show that higher relative carbohydrate intake may have a protective effect, lowering depression risk.
‘Intuitive physics’ enables our pragmatic engagement with the physical world and forms a key component of ‘common sense’ aspects of thought. Current artificial intelligence systems pale in their understanding of intuitive physics, in comparison to even very young children. Here we address this gap between humans and machines by drawing on the field of developmental psychology. First, we introduce and open-source a machine-learning dataset designed to evaluate conceptual understanding of intuitive physics, adopting the violation-of-expectation (VoE) paradigm from developmental psychology. Second, we build a deep-learning system that learns intuitive physics directly from visual data, inspired by studies of visual cognition in children. We demonstrate that our model can learn a diverse set of physical concepts, which depends critically on object-level representations, consistent with findings from developmental psychology. We consider the implications of these results both for AI and for research on human cognition. Piloto et al. introduce a deep-learning system which is able to learn basic rules of the physical world, such as object solidity and persistence.
Spending timelines after receiving the COVID-19 relief stimulus payment
Spending before and after receiving a 2020 CARES Act stimulus payment for lower-income (earning under US$28,001 per year), middle-income (US$28,001–US$68,000) and higher-income (above US$68,000) individuals. The baseline average (light blue line) is the amount spent 60 days prior to receiving the payment. The left plot presents proportional spending compared with a standard baseline. The right plot presents the same information but uses actual spending values. Apart from the days immediately following receipt, the base-standardized spending patterns are almost identical for all three groups.
Source data
Global indications of intertemporal choice
a–f, Maps of choice preferences in aggregate and by individual anomaly indicate heterogeneity in intertemporal choice patterns. While some subtle patterns emerge, particularly stronger preferences for delayed gains in higher-income regions, choice preferences are broadly consistent across 61 countries in the sense that all anomalies appear in all locations. No location consistently presents extremes (high or low) of each anomaly. The results are based on the models specified in Supplementary Table 13. g,h, Conditional smooth effects (black) and 95% confidence intervals (light blue). Map from Natural Earth (
Source data
Baseline temporal discounting and GDP
a–c, There is a clear trend of lower GDP³⁶ being associated with higher preferences for immediate gains and later payments. However, all locations indicate some preference for immediate over delayed. Taken together, this provides support for the hypothesis that baseline temporal discounting is observed globally and that the economic environment may shape its contours. The results are based on the models specified in Supplementary Table 14. Smooth terms and 95% confidence intervals are presented in black and grey, respectively.
Source data
Anomalies and temporal discounting scores by country
a,b, Proportions (solid bars are overall means) of participants that demonstrated inconsistent choice preferences (a) and the proportion of each country sample that aligned with the five anomalies of interest (b). Apart from absolute magnitude and present bias, no consistent rate was based on wealth, and all countries indicate some presence of each anomaly. c–h, Each plot presents the distribution of values ordered by mean or proportion value. Plot c presents the distribution of discounting scores for each country, including means, prediction intervals (coloured) and standard deviations (grey). Plots d–h show the proportions of participants that presented each anomaly. While the difference from lowest to highest for each is noteworthy, similar variabilities exist across all. See Supplementary Figs. 3–8 for the full values and sample sizes for each point.
Source data
Wealth, debt, inequality and temporal discounting
a–f, Plots using standardized scores for temporal discounting indicate an overall trend that greater wealth and income at the individual and national levels are associated with lower overall temporal discounting, and greater economic inequality and individual debt are associated with lower overall temporal discounting. Inflation has a modest relationship with discounting, which becomes much stronger at substantially high levels of inflation. The results for each variable by score are from models specified in Supplementary Table 16. Smooth terms and 95% confidence intervals are presented in black and grey, respectively.
Source data
Economic inequality is associated with preferences for smaller, immediate gains over larger, delayed ones. Such temporal discounting may feed into rising global inequality, yet it is unclear whether it is a function of choice preferences or norms, or rather the absence of sufficient resources for immediate needs. It is also not clear whether these reflect true differences in choice patterns between income groups. We tested temporal discounting and five intertemporal choice anomalies using local currencies and value standards in 61 countries (N = 13,629). Across a diverse sample, we found consistent, robust rates of choice anomalies. Lower-income groups were not significantly different, but economic inequality and broader financial circumstances were clearly correlated with population choice patterns.
Top-cited authors
Brian Nosek
  • University of Virginia
C. Sunstein
  • Harvard University
Jay Van Bavel
  • New York University
Michele Gelfand
  • University of Maryland, College Park
Sander van der Linden
  • University of Cambridge