2.4.11 | Comparing the accuracy of posterior parameters
estimations using NN, RF, or Rejection ABC
We compared four ABC posterior parameter estimation methods: NN-ABC
estimation of the parameters taken jointly as a vector (similarly as in
the above procedures), NN-ABC estimation of the parameters taken in turn
separately, RF-ABC estimation of the parameters which also considers
parameters in turn and separately (Raynal
et al., 2019), and simple Rejection ABC estimation for each parameter
separately (Pritchard et al., 1999). For
each method, we used in turn the 1,000 simulations closest to the real
data as pseudo-observed data and the 99,999 remaining simulations as
reference table. We consider the same parameters for the NN, and we 500
decision trees for the RF to limit the computational cost at little
accuracy cost a priori (Supplementary Figure S3 ). We
computed the three types of errors and the accuracies of the 95% CI for
each ABC method similarly as previously (Table 4 ).