Table 4. Approximate Bayesian Computation mean posterior
parameter errors under the winning Scenario AfrDE-EurDE, for the ACB and
ASW populations separately, using four different methods: NN estimation
of the parameters taken jointly as a vector, NN estimation of the
parameters taken separately, Random Forest (parameters taken
separately), and Rejection (parameters taken separately). For each
target population separately and for each method, we conducted an
out-of-bag cross validation by considering in turn 1,000 separate
parameter inferences each using one of the 1,000 closest simulation to
the observed ACB (or ASW) data as the target pseudo-observed dataset.
All posterior parameter estimations were conducted using the other
99,999 simulations under the AfrDE-EurDE scenario (Figure 1 ,Table 1 ), a 1% tolerance rate (i.e. 1,000 simulations), 24
summary statistics, logit transformation of all parameters, four neurons
in the hidden layer per neural network and 500 trees per random forest.
Median was considered as the point posterior parameter estimation for
all parameters. First column provides the average absolute error; second
column shows the mean-squared error; third column shows the mean-squared
error scaled by the parameter’s observed variance (see Materials
and Methods for error formulas).