Robustness of the three methods under various demographic and selective scenarios (see materials and methods for full details)

Growth−0.531.15 (1.00–1.35)1.10 (1.00–1.25)1.30 (1.20–1.42)
Bottleneck0.250.60 (0.50–0.70)0.65 (0.55–0.70)0.56 (0.50–0.64)
Sweep–1.251.02 (0.78–1.32)0.95 (0.80–1.10)0.94 (0.84–1.06)
Four-island even0.140.190.85 (0.75–0.95)0.80 (0.70–0.90)0.88 (0.82–0.96)
Four-island uneven0.250.210.60 (0.50–0.70)0.65 (0.55–0.75)0.62 (0.56–0.68)
Two-island even0.190.220.82 (0.72–0.92)0.85 (0.75–1.00)0.82 (0.74–0.88)
Two-island uneven0.210.250.76 (0.68–0.86)0.80 (0.80–0.95)0.76 (0.70–0.84)
  • For each scenario and method the average log-likelihood curve across 100 10-kb data sets was calculated; the value given is the ratio of the value of ρ that maximizes this curve over the true value of ρ0. This ratio gives the factor by which the estimator will tend to under- or overestimate the true recombination rate. A value of 1 shows that the method is robust; values of 2 or 0.5 would suggest that for large data sets the estimates will tend to be respectively twice or half the truth. In parentheses are ∼95% confidence intervals based on the curvature of the expected log-likelihood curves. For each scenario the average values of Tajima's D and (for the island models) FST are given.