Links between language diversity and species richness can be confounded by spatial autocorrelation.
autocorrelation and congruence in the distribution of language and mammal richness: a reply to Cardillo et al. Our paper on the relationship between the spatial distribution of language richness and mammal species richness recently argued that little congruence exists between the distribution of threatened languages and threatened mammals in New Guinea, despite high overlap between areas of high language richness and high mammal species richness across this island . In reply to this, Cardillo et al. re-analysed the data at one of the spatial resolutions that we considered in our paper, claiming that the original analyses were statistically flawed . In short, Cardillo et al. argue that the inclusion in our analyses of coastal pixels containing both land and sea, as well as a lack of consideration of spatial autocorrelation, has led us to wrongly derive conclusions about patterns of distribution in language and species richness. We believe such conclusions are misleading for various reasons. First, the inclusion of coastal pixels in our analyses has no impact on the results reported. Re-running the analyses at the 50 km resolution without the inclusion of these mixed pixels, we show that the direction and significance of the Pearson correlation coefficients between language richness and mammal richness (r ¼ 0.23, p ¼ 0.004) and between threatened language richness and threatened mammal richness (r ¼ 20.15, p ¼ 0.03) both still hold. Although the correlation between threatened language richness and threatened mammal richness becomes non-significant when only coastal pixels are considered (r ¼ 20.03, p ¼ 0.75), we also show that both the direction and significance of the relationship between overall language richness and mammal richness still hold for analysis of coastal cells alone (r ¼ 0.41, p , 0.001). Furthermore, a high number of languages, primarily in the recent Austronesian language radiation, show a narrow distribution along the northern coast of New Guinea  and are thus largely restricted to coastal pixels; excluding such an important component of New Guinea linguistic diversity from analysis may therefore be expected to bias conclusions about island-wide patterns and correlates of regional language richness. Second, there is a fundamental difference between correlation and regression , which the authors fail to acknowledge. The use of regression implies a search for assumed causality, while the use of correlation simply looks for similarity (or dissimilarity) in patterns between associated variables. In this study, we did not expect a decrease in …