Two major hypotheses have been suggested as potential mechanisms for the global spread of plant and animal domestication (farming) among human societies. The diffusion hypothesis suggests that farming technology was passively shared between neighboring groups, while the takeover hypothesis suggests that farming spread through the forceful eviction and replacement of non-farming competitors. Investigating the role and relative importance of these two processes based on historical texts, local archeological surveys, and genetic analyses alone has proven both difficult and uninformative. Here we use the tools of spatial ecology, phylogenetic methods, niche reconstruction, and machine learning to develop a rigorous quantitative framework to better distinguish between these alternatives.
We modeled the spatial and phylogenetic distribution of farming under four different modes of transmission: inheritance, inheritance plus diffusion, inheritance plus takeover, and inheritance plus diffusion and takeover. These models were replicated extensively (ca. 100k replicates per mode) and the resulting phylogenetic and spatial patterns were collectively used to train a random forest machine-learning algorithm to identify the most likely mode of transmission of any given simulation output. Once trained, the algorithm was used to determine which transmission mode most likely led to current phylogenetic and spatial distribution of farming among human cultures.
Results/Conclusions
The trained random forest algorithm was able to properly classify the transmission mode of any provided simulation with 59% accuracy, but demonstrated a particularly strong ability to distinguish between the diffusion and takeover models. In the relatively few instances where the algorithm misclassified transmission modes, it most frequently confused strict diffusion or takeover with the model including both processes or with the baseline model including only inheritance.
When applied to the current spatial and phylogenetic distribution of farming across human societies, our random forest algorithm unambiguously suggests that the most likely mechanism behind the spread of farming was strict cultural diffusion. We interpret this result as indicative of the different timescales at which the two hypotheses for the spread of agriculture operate. Diffusion predominates because it can potentially occur immediately after the origins of domestication. In contrast, takeover is likely to require some time for domesticating societies to accumulate enough resources and organizational skills as to gain a competitive advantage over others. More broadly, our results indicate that a combination of process-based modeling and summary statistics can help elucidate the mechanisms that drove current patterns of trait diversity even under high levels of uncertainty in model parameters.