Hi Ryan,

• Empirical Jacobian can be calculated from your par/obs stacks; see here for example. My understading is that this is pretty (very) dirty though - I don't know that you could get a useful sensitivty measure from it using traditional methods for example.
• We tend to use the mean parameter shift (per par or group of pars), and/or normalised variance (or range) reduction from prior --> posterior as indicators of senstivity. For example, maps of kh variance reduction can be really informative.

Sorry for the delay @RyanConway91

@cnicol-gwlogic has it right - that empirical jco is really sloppy (and algorithmically, ies never actually forms an nobs X npar matrix for computational reasons). More conceptually, ensemble methods lack the "causality" that a sensitivity implies. So if you need/want to know what the primary/important input-output relations are, I think you'll have to do something more deliberate like sensitivity analysis...#nofreelunch ...

Thank you, I will look into mean parameter shift/variance and maybe use those results to guide a true sensitivity analysis.