We used ARGweaver to infer the ancestral recombination graph.36 ARGweaver is based on the standard coalescent model and is sensitive to balancing selection, such that regions under balancing selection have older times to the most recent common ancestor than comparable neutral regions. To account for uncertainty in both the mutation and recombination rates, we used a grid approach. We tested mutation rates of 0.72×10−8, 0.97×10−8, and 1.44×10−8 mutations per generation per site.34 We tested recombination rates of 5×10−9, 1×10−8, 1.5×10−8 (the default value for human data), 2×10−8, and 2×10−7 recombinations per generation per site.34 We ran the sampler for 3,000 iterations and saved the ancestral recombination graph from every tenth iteration. We used the functions heidel.diag and geweke.diag in the coda library of R, version 3.2.3, to assess convergence diagnostics on the basis of the posterior distribution of the number of recombination events.37 To convert generations into years, we assumed a generation interval of 28 years.38, 39