3.5 | Comparing NN, RF, and Rejection ABC posterior
parameter estimation accuracy
The three types of posterior parameter estimation errors (scaled
mean-square error, mean-square error, absolute error) were
systematically lower for the two NN methods (joint or independent
posterior parameter estimations) than for the RF and Rejection
independent posterior parameter estimations (Table 4 ).
Altogether, these results showed that considering the NN estimation for
parameters taken jointly as a vector is overall preferable, since it
further allowed the joint interpretation of parameter values estimateda posteriori , with little accuracy loss.
The lengths of 95% CI estimated with NN joint parameter estimation
were, across all parameters, more accurate than those obtained with all
other methods with, on average, 95.1% and 95.2% of true parameter
values falling within the estimated 95% CI, for the ACB and ASW
respectively (Supplementary Table S3 ). Furthermore, the lengths
of 95% CI estimated with NN and RF independent posterior parameter
estimations were systematically under-estimated, with less than 94% of
the true parameter values falling into the estimated 95% CI. Finally,
the lengths of 95% CI estimated with the Rejection method were also
rather accurately estimated, although on average slightly over-estimated
compared to the NN joint estimation with, on average, 95.5% of the
1,000 cross-validation true parameter values within the estimated 95%
CI for the ACB, and 95.8% for the ASW.