Quantifying cultural sharing

We used the D-PLACE database (Gray, Reference Gray1999; Kirby et al., Reference Kirby, Gray, Greenhill, Jordan, Gomes-Ng, Bibiko and Gavin2016; Murdock, Reference Murdock1967) to collect data on the cultural makeup of 1094 mainland traditional societies (Figure S2), defined as groups of people at a focal location whose language and cultural traditions set them apart from neighbouring groups (Kirby et al., Reference Kirby, Gray, Greenhill, Jordan, Gomes-Ng, Bibiko and Gavin2016). Out of the total 1291 societies represented in the Ethnographic Atlas dataset within D-PLACE, we excluded societies located on islands, based on the geographic positions of societies in Glottolog (Hammarström et al., Reference Hammarström, Forkel and Haspelmath2019) and using the island polygons described in Weigelt et al. (Reference Weigelt, Jetz and Kreft2013) as reference for continental and island terrain. We considered 67 cultural traits related to subsistence, house construction, property rules, community organisation, marriage and kinship, politics, class, labour, and ritual practices (Data S1). For each cultural trait, we scored whether a pair of societies, our unit of analysis, exhibited the same or different trait values. For example, each society exhibits a dominant mode of subsistence (dependence predominantly on gathering, hunting, fishing, pastoralism or agriculture), and two societies are identified as sharing the dominant subsistence mode if they both rely predominantly on, for example, the gathering of wild plants and small land fauna relative to other subsistence modes. In practice, this means that a pair of societies shares a cultural trait if societies share the codes allocated for that trait in the D-PLACE database. For some traits, we modified the original D-PLACE coding to better reflect the question of interest. For example, we considered societies that rely on casual, extensive, intensive agriculture (or any combination of these) as sharing an agricultural mode of subsistence. Further, we did not consider societies in which a cultural trait of interest is absent, e.g. when looking at whether societies share the pattern for the largest matrilinear group, we excluded societies in which matrilinear groups are absent. We also did not consider societies with missing cultural data. Lastly, we excluded cultural traits with only two discrete code categories. Data S1 contains a list with all traits for which we scored cultural similarity (a total of 67) alongside their code values, number of categories, and a notes tab justifying the codes used. All the traits we used are discrete.

Quantifying environmental barriers to cultural transmission

We built pairs of societies by considering each society and its closest 100 neighbours based on geodesic distances. To determine geodesic distances, we rely on the coordinates of societies as they are listed in Glottolog (Hammarström et al., Reference Hammarström, Forkel and Haspelmath2019). From the initial 100 pairs for each society, we select the pairs for which we have cultural data. We then evaluated whether the odds of sharing the same values for a given cultural trait are impacted by environmental parameters associated with latitude. Following Diamond's hypothesis, we focus on dissimilarities in environmental conditions at the locations of societies in pairs, as well as on the magnitude of ecological barriers that the use of a cultural trait may have encountered as it expanded its range. If Diamond's ecological premise is supported by the data, we expect that the odds of cultural similarity will be negatively impacted by high environmental dissimilarities between the home ranges of the societies in the pair, and by high environmental heterogeneity along the paths that connect them. We follow Diamond's original framework and primarily focus on ecological drivers of cultural transmission; we do, however, extend our analyses to include topographic factors, i.e. barriers associated with the movement of humans.

Environmental costs to cultural transmission between pairs of societies were estimated from data on 100+ year time series on local mean, variability and predictability of temperature and precipitation, as available in the EcoClimate database (Lima-Ribeiro et al., Reference Lima-Ribeiro, Varela, González-Hernández, de Oliveira, Diniz-Filho and Terribile2015). These climatic variables exhibit pronounced latitudinal gradients and are heavily featured in Diamond's (Reference Diamond1997) book. Given that the patterns of geographic variation in these parameters are highly correlated, we reduced them to two composite variables through a global principal component analysis (PCA) at a 0.5 × 0.5° resolution (Botero et al., Reference Botero, Dor, McCain and Safran2014a; Table S1). Only land cells were included. The resulting components captured global gradients in the variation of ‘temperature harshness’ (PC1) and ‘aridity index’ (PC2). PC1 values reflect a continuum from cold, unpredictable and variable temperatures to warm, predictable and invariable conditions. PC2 values reflect a continuum from arid, unpredictable and variable precipitation to humid, predictable and invariable regimes (Figure S3).

Environmental dissimilarity between societies in pairs was estimated as the absolute difference in temperature harshness and aridity index, respectively, between the geographic positions of societies (Figure S4). We estimated environmental heterogeneity in the intervening space between societies in pairs using least-cost path algorithms. Specifically, starting from each society in a pair, we computed the path toward the other society that minimised aridity index differentials with the starting location (Alberti, Reference Alberti2019; van Etten, Reference van Etten2017; Figure 2, see Figure S4 for details about the least-cost path algorithm). A high accumulated cost along this path indicates high dissimilarity in aridity regime from the starting location to the destination society. Similarly, longer paths indicate that longer, hence costlier, diffusion (and travel) routes are needed to achieve a minimal aridity regime difference between the starting and destination societies. For each society pair ‘AB’, we computed the paths of least resistance starting from the coordinates of A towards the coordinates of B and vice-versa. We used the mean accumulated cost values on the paths A to B and B to A, and the mean length of paths A to B and B to A as estimates of environmental heterogeneity in the intervening space between A and B (Figure S4). We then repeated this algorithm to estimate the path of minimal temperature harshness differences. Lastly, we measured the length and cost of least-cost topographic paths (as measured through changes in slope of the terrain), alongside geodesic distances as indicators of barriers to cultural similarity related to human travel costs (see Figure S4 for details). All analyses included geographic correction to account for the spatial distortion of latitude/longitude map projections in the WGS 84 coordinate system (van Etten, Reference van Etten2017).

Figure 2. Example of (a) topographic, (b) temperature, and (c) aridity least-cost paths between the locations of two societies (starting point is marked by a star). Dark colours on the cost maps indicate cells of high cost – i.e. high elevation (a) or high differences in temperature harshness (b) and aridity index (c) as compared with the starting point. Environmental heterogeneity along the least-cost path connecting two societies in a pair is estimated as the log value of accumulated cost and log value of path length.

Common ancestry and potential of indirect transmission from neighbouring cultural groups

Patterns of cultural similarity between societies are known to be shaped by common ancestry and horizontal diffusion (Mace & Holden, Reference Mace and Holden2005; Mace & Jordan, Reference Mace and Jordan2011). We thus accounted for the possibility that societies in a pair share the same cultural trait values owing to shared ancestry by considering relatedness as a predictor in our models. Language phylogenies are commonly used to account for cultural similarity owing to shared ancestry (Evans et al., Reference Evans, Greenhill, Watts, List, Botero, Gray and Kirby2021). This is because languages are a good representation of the history of expansions and splits of cultural groups and thus reasonably estimate population history. We estimated relatedness using a recent Bayesian phylogeny of all extant languages (Bouckaert et al., Reference Bouckaert, Redding, Sheehan, Kyritsis, Gray, Jones and Atkinson2022, preprint). This super-tree incorporates information on established language classification within families, language location, historical, linguistic, and archaeological evidence for the deeper splits, as well as genetic and archaeological evidence about the colonisation of the world by humans. The Bayesian approach incorporates uncertainty about the relationships among languages in the form of a posterior distribution of possible trees. For the analyses in the main text, we used TreeAnnotator to build a maximum clade credibility tree, setting node heights to common ancestor (Drummond et al., Reference Drummond, Suchard, Xie and Rambaut2012). We prune this tree for the societies in our dataset (Figure S5), and then compute our relatedness metric as the phylogenetic cophenetic distance (based on branch lengths) to the nearest common ancestor of the two societies in a pair9) (we used the R package {ape}, Paradis & Schilep, Reference Paradis and Schliep2019). To test the robustness of our results to phylogenetic uncertainty, we also run the analyses using a randomly drawn sample of 50 trees from the posterior of alternative trees (Data S3).

Our analyses are focused on cultural similarity between pairs of societies. When two societies share the same cultural trait value, they could be doing so because of cultural transmission between them. Alternatively, cultural transmission could have happened between each of the two societies in a pair and other close-by neighbours. We note here that Diamond's thesis is not reliant on the assumption of direct transmission between pairs of societies. However, we believe that comparing society pairs facilitates a principled definition of cultural influences and enables a more straightforward interpretation of results, particularly when compared with difficult to define cultural ‘neighbours’. Thus, our models ask: do geographical and environmental barriers promote cultural similarity between pairs of societies above and beyond what is expected from shared ancestry or transmission from other close-by societies? In practice, for each society, we calculated the proportion of the five closest neighbours (based on geodesic distances) for which cultural data exist in the D-PLACE database that share the trait of its pair member (following Botero et al., Reference Botero, Gardner, Kirby, Bulbulia, Gavin and Gray2014b). If a high percentage of nearby neighbours of society ‘A’ express the trait of society ‘B’ (pair member of ‘A’), there is a high chance that ‘A’ has the same trait as ‘B’ not because of cultural transmission between ‘B’ and ‘A’, but because of cultural transmission from the close neighbours of ‘A’ to ‘A’. That is, ‘A’ got the trait value of ‘B’ because of transmission within its close neighbourhood (Figure S6). For each pair ‘AB’, we compute the potential for indirect transmission from other cultural groups for society ‘A’, then for ‘B’, and then average the two values. We assume that higher values of this predictor indicate that a society could have picked a shared cultural trait from its immediate neighbours rather than from the other society in the pair (Botero et al., Reference Botero, Gardner, Kirby, Bulbulia, Gavin and Gray2014b). This is a conservative assumption, as it equates a pattern of spatial clustering with the mechanism of horizontal diffusion, even though in reality we cannot know whether cultural transmission actually happened or not (Evans et al., Reference Evans, Greenhill, Watts, List, Botero, Gray and Kirby2021). As a sensitivity analysis, we also computed our metric of cultural transmission from other groups using each society's 10 closest neighbours, regardless of whether these neighbouring societies had data in the D-PLACE dataset (i.e. what proportion of the closest 10 neighbours of ‘A’ have the trait value of ‘B’?, and vice-versa). In this way, the selection of which societies to consider as the ‘close neighbourhood’ was not affected by sampling in the D-PLACE dataset. Pairs in which neighbourhood structure could not be estimated were discarded (if cultural data were available for at least three out of the 10 nearest neighbours for both societies in the pair, the pair was kept in the analyses).

Modelling cultural sharing as a function of environmental dissimilarities

We used logistic regressions to correlate the odds of societies sharing cultural traits with our initial 11 candidate drivers of cultural sharing: temperature and aridity dissimilarity, the length and accumulated cost along the least-cost path for human travel, the geodesic distance, the length and accumulated cost along the temperature and aridity least-cost paths, shared ancestry, and cultural transmission from other nearby societies (Figure 1a, Figure S1a, the green ‘Raw variables’ box). To recapitulate, in our models, cultural sharing is a binary variable, with ‘1’ indicating the two societies are culturally similar (i.e. the pair shares the codes allocated for that trait in the D-PLACE database), and ‘0’ indicating the reverse. We set the error family as Bernoulli. However, most such models suffered from strong multicollinearity (variance inflation factors > 5; Zuur et al., Reference Zuur, Ieno and Elphick2010). Thus, we reduced the initial set of predictors to five composite variables via a varimax-rotated PCA (Figure 1a and Figure S1a, the burgundy ‘Variables reduced by PCA’ box). The PCAs were run separately for each cultural trait because the potential for transmission from nearby neighbours is based on patterns that are specific to each trait (Table S2 shows how each raw variable connects to a component in each of these PCAs). For each trait, we used five composite variables as predictors for cultural sharing in mixed-effects models that included the identities of societies in pairs as random effects (modelled as random intercepts; Figure 1a and Figure S1a, blue box). Thus, we asked whether the odds of cultural sharing in pairs (i.e. whether or not societies in a pair share the same focal cultural trait value) were predicted by five candidate drivers: (i) temperature turnover (component capturing variation in temperature harshness heterogeneity at the location of societies in pairs and along the temperature-based least-cost paths); (ii) aridity turnover (capturing aridity index heterogeneity at the location of societies in pairs and along the aridity-based least-cost paths); (iii) costs to human movement (capturing topographic costs and climate-related path lengths); (iv) common ancestry; and (v) transmission from other close-by neighbours (Figure 1a and Figure S1a, blue box). In our models, higher principal component values show higher environmental and travel barriers. Thus, if Diamond's ecological assumption is supported, we expect to see negative associations (negative estimated effects) between cultural similarity and temperature turnover, aridity turnover, and travel costs. We used the function glmer() to run the models in R (Bates et al., Reference Bates, Mächler, Bolker and Walker2015). Spatially neighbouring data points could show similarity in predictors, response variables, and/or the relationship between the two (i.e. the assumption of independent residuals is breached). Models that suffer from spatial autocorrelation can show inflated Type I errors. We thus used Moran's I spatial autocorrelograms to determine whether model residuals showed spatial dependence. We used the midpoint between the locations of pair members as coordinate references in spatial autocorrelograms (Vilela & Villalobos, Reference Vilela and Villalobos2015). The autocorrelograms show that our models do not suffer from spatial dependence in the residuals (Figure S7).

In the main text, we focus on the results in which the potential for similarity owing to transmission from other close-by societies is computed using, for each society, its five closest neighbours with D-PLACE data. We excluded imbalanced models (the percentage of pairs that show cultural sharing was more than 90% or less than 10%), as well as models that did not converge, or had fewer than 50 datapoints. In total, this analysis included 54 cultural traits (listed in Table S4a). Significance levels were adjusted for false discovery rates to account for multiple independent comparisons (Benjamini & Yekutieli, Reference Benjamini and Yekutieli2001), using the function p.adjust() in R. Ninety per cent of models show areas under the curve (AUC) higher than 0.8 (the lowest AUC value was 0.73, and five models had AUC values between 0.7 and 0.8, Table S4a), as estimated by the function roc() in the R package {pROC} (Robin et al., Reference Robin, Turck, Hainard, Tiberti, Lisacek, Sanchez and Müller2011).

Quantifying the strength of environmental barriers within corridors of agriculture transmission

To formally test the geographic premise that Eurasia's East–West orientation translated into weaker environmental barriers to cultural transmission, we use a critical test case and arguably the most prominent example in Diamond's (Reference Diamond1977) book: the spread of agriculture. Specifically, the axis of orientation hypothesis predicts that the magnitude of ecological barriers to cultural spread should be smaller along potential routes of transmission of agriculture in Eurasia than in other regions of the planet. We used the geographic ranges of 17 known areas of independent origin of domestication of plants and animals as standardised points of reference (Larson et al., Reference Larson, Piperno, Allaby, Purugganan, Andersson, Arroyo-Kalin and Fuller2014). Out of the 19 origin areas proposed in Larson et al. (Reference Larson, Piperno, Allaby, Purugganan, Andersson, Arroyo-Kalin and Fuller2014) we excluded New Guinea and Japan, as we are focused here on cultural transmission among continental societies. Following Kavanagh et al. (Reference Kavanagh, Vilela, Haynie, Tuff, Lima-Ribeiro, Gray and Gavin2018), we considered the Northwestern and Northern Lowlands of South America as one area of domestication origin. We selected societies that are located within these areas based on their geographic positions in Glottolog (Hammarström et al., Reference Hammarström, Forkel and Haspelmath2019). As in our prior analyses, we quantify the ecological costs to cultural transmission for pairs of societies. That is, we evaluated environmental barriers between societies located within the areas of independent agricultural origin and each of their closest 100 neighbours. To investigate the potential for long-range transmission, we paired societies within the agricultural points of origin with their closest 100 neighbours located at least 2500 km away (Figure S8 shows the location of societies in each area of agricultural origin, as well as the location of their closest 100 neighbours and closest 100 neighbours located at least 2500 km away). This long-range threshold was chosen in order to investigate the broad-scale cultural transmission examples mentioned in Diamond's thesis, such as from the Fertile Crescent to Southern and Central Europe or to North and East Africa, or from Mesoamerica to Southern North America and Equatorial South America; Figure S8). Therefore, we estimated environmental barriers at both small and large spatial scales within corridors of early agriculture transmission across the world.

For each pair of societies in these analyses we compute the dissimilarity in temperature harshness and aridity index at the locations of the two societies in a pair, accumulated cost and length of the paths of least environmental resistance, accumulated cost and length of the path minimising topographic costs and geodesic distance (Figure 1b and Figure S1b, the green ‘Raw variables’ box). Then, for each focal society (i.e. a society in an agricultural point of origin), we average each of these metrics among its closest 100 neighbours (for short-distance analyses), and among its closes 100 neighbours located at least 2500 km away (for long-distance analyses). Thus, each starting location contributes a single data point in the short- and long-distance analyses, respectively (detailed in Figure S9).

Comparing continental differences in ecologically induced costs to agricultural spread

For each society in an agricultural point of origin we have thus quantified the potential costs to cultural transmission from the society's location to its neighbours: dissimilarity in temperature harshness and aridity regimes, cost accumulated along the temperature harshness and aridity index least-cost paths, the lengths of temperature harshness and aridity index least-cost paths, and costs related to human movement (based on elevation and geodesic distances; Figure 1b and Figure S1b, ‘Raw variables’ green box). As we have noted in our ecological test for Diamond's hypothesis, these metrics are highly correlated (Figure S10). We thus follow the same approach and reduce dimensionality in the data by running two separate varimax-rotated PCA analyses: one for close and one for long-range nodes, where a society located in a domestication origin area is the unit of analysis. The close-range PCA gave us three estimates of environmental barriers to cultural spread: (i) a component capturing variation in the topographic travel costs, geodesic distance and climate-related least-cost paths (close-range PC1, or crPC1); (ii) aridity turnover (crPC2); and (iii) temperature turnover (crPC3). As with the first set of analyses, ‘turnover’ here captures environmental dissimilarity at societies’ locations and the accumulated cost across the least-cost path. The long-range PCA gave us four estimates of environmental barriers: (i) temperature turnover (long-range PC1, or lrPC1); (ii) a component capturing the lengths of climate-related least-cost paths (lrPC2); (iii) aridity turnover (lrPC3); and (iv) a component capturing topographic travel costs and geodesic distance (lrPC4, Figure 1b and Figure S1b, the burgundy ‘Variables reduced by PCA’ box). Table S3 shows how each raw variable respectively associated with each close-range and long-range component. We used these principal components in separate linear models: we regressed each PC against a categorical variable with levels represented by 16 areas of domestication origin around the globe (we exclude South India from the analyses, as we only have coordinates for one society in this area of agricultural origin; Figure 1b and Figure S1b, the blue box). We thus ask: are there substantial differences in the mean of each environmental barrier metric (denoted by the values of each principal component) between the areas of agricultural origin? We ran variance-weighted linear models to account for heterogeneity in variance across the levels of the categorical variable. Moran's I tests revealed that models can suffer from spatial autocorrelation in the residuals (Figure S11). We hence accounted for the spatial dependence by using the function SpatialFiltering() in the R package {spatialreg} to select eigenvectors in a semi-parametric spatial filtering approach (Tiefelsdorf & Griffith, Reference Tiefelsdorf and Griffith2007). Eigenvectors were introduced as covariates in the models, and accordingly, Moran's I test on the new models showed few spatial dependencies (Figure S11). If Diamond's geographic premise is supported by the data, we expect to consistently see smaller means of environmental barriers (i.e. principal component values) within Eurasian corridors of agricultural spread compared with other regions of our world.

Comparing continental differences in ecologically induced costs to agricultural spread

using reconstructed climatic conditions at 12, 8 and 4 kya

Because the exact timeline of the spread of agriculture is relatively unknown in many regions of the world, we repeated the analyses in this section using climate reconstructions from different time periods. We obtained values for the reconstructed annual mean and variation of temperature, as well as mean and coefficient of variation of precipitation conditions around the globe at 12, 8 and 4 kya (thousand years ago) from Kavanagh et al. (Reference Kavanagh, Vilela, Haynie, Tuff, Lima-Ribeiro, Gray and Gavin2018). We choose these timestamps to cover the entire timeframe in which agriculture presumably originated and spread in different regions of the planet (Diamond & Bellwood, Reference Diamond and Bellwood2003; Larson et al., Reference Larson, Piperno, Allaby, Purugganan, Andersson, Arroyo-Kalin and Fuller2014). For each of these time points, we (i) ran a global PCA at a 0.5 × 0.5° resolution, to obtain two axes that capture global variation in temperature harshness and aridity index and (ii) used the two environmental axes computed in (i) to estimate the following environmental barriers to cultural spread within the close- and long-range neighbouring societies located within areas of agricultural origin: dissimilarity in temperature and aridity conditions at the locations of societies, the accumulated cost and length of the least-cost path given a landscape of costs determined by temperature and aridity heterogeneity. The values for the geodesic distance and travel costs (determined by elevation) between societies were not recomputed, as they do not depend on reconstructed temperature harshness and aridity index values. We then repeated the analyses quantifying and comparing differences in the strength of environmental barriers between areas of agricultural origin in order to identify if – at any of the time points considered (12, 8 and 4kya) – Eurasian areas showed consistent ecological advantages over other areas of the world (Table S7). Thus, this section of our analyses broadly asks whether Eurasia exhibits ecological advantages for the transmission of culture either presently or in earlier time points that are relevant to the spread of agriculture.