Fig. 4.

Violin plots displaying change in selected fixed effects parameter values between the respective method and default estimation, relative to true values, delta values in percent, for the non-identifiable models E (top) and F (bottom), at their respective lowest − 2log(likelihood). The four methods compared are, in order from the left, perturbation and re-estimation, one saddle-reset step, three saddle-reset steps, and one saddle-reset step with computed R-matrix. A wider distribution and separation from zero indicates better performance in exposing the non-identifiability. Using a computed R-matrix produces parameter values that are vastly different from the default estimation, clearly indicating non-identifiability. Some parameters remain identifiable, such as baseline in model E and proportional error in model F. The total number of estimations that reached the lowest known OFV (n_tot), and the number of estimations that produced the same parameter estimates in default estimation and in the respective method (n_θ̃=θ̃^new), is shown in the bottom panel for each method in each model. A lower n_tot indicates that estimations crashed or did not reach the lowest OFV. A lower n_θ̃=θ̃^new means that more estimations unveiled non-identifiability