An Integrated Strategy for Prediction Uncertainty Analysis (Supplement)

As mentioned in the main paper, we ran Profile Likelihoods for each mode detected with the Monte Carlo Multiple Minimisation. The reason for this is that the method by which these profiles are computed are local (reoptimising from the current point at each step). The result is that one runs the risk that the profile likelihood fails to leave a mode before the error reaches the threshold (see for example the green profile for parameter 4 in the top row). It is therefore important to validate that all modes have been reached after running the Profile Likelihoods. Mode switches can also be observed (for an example, see the blue profile for parameter 1 in the bottom row). The profiles can subsequently be merged afterwards. Note that some of the parameters were structurally non-identifiable and showed clear relationships between the parameters when plotted in a scatter plot (see Figure 1). Shown in Figure 2 are the three separate Profile Likelihoods performed in order to obtain the merged version in the paper. The starting value for each profile is denoted with a cross.