2.4.5 | RF-ABC model-choice for the admixture history
of ACB and ASW populations
For the ACB and ASW observed data separately, we performed model-choice
prediction and estimation of posterior probabilities of the winning
model using the predict.abcrf function in the abcrfpackage, using the complete simulated reference table for training the
Random-Forest algorithm (100,000 SNPs, 50 individuals in population H,
90 and 89 individuals in the African and European sources respectively)
(Figure 3 , Supplementary Table S1 ).
2.4.6 | Posterior
parameter estimation with Neural-Network ABC
It is difficult to estimate jointly the posterior distribution of all
model parameters with RF-ABC (Raynal et
al., 2019). Furthermore, although RF-ABC performs satisfactorily well
with an overall limited number of simulations under each model
(Pudlo et al., 2016), posterior parameter
estimation with other ABC approaches, such as simple rejection
(Pritchard et al., 1999), regression
(Beaumont et al., 2002;
Blum & François, 2010) or Neural-Network
(NN) (Csilléry et al., 2012), require
substantially more simulations a priori . Therefore, we performed,
for posterior parameter estimations, 90,000 additional simulations, for
a total of 100,000 simulations under the best scenarios identified with
RF-ABC for the ACB and ASW separately. For comparison purposes, we
performed 100,000 simulations under the loosing scenario Afr2P-Eur2P
(see Results ), and conducted anew the below parameter
estimation and error evaluation procedures for this scenario.