Hello,
I’m exploring the application of Sobol sensitivity analysis over fragmented or conditional regions of the input parameter space and would appreciate any insights on how best to approach this using (or adapting) SALib.
As a simplified case, consider the analytical model:
y = a * b²,
with a and b sampled independently from the unit hypercube [0, 1]. Computing the standard first-order and total Sobol indices using SALib is straightforward in this case.
However, I’m interested in quantifying sensitivity not across the full input domain, but only in a subregion defined by an output condition, e.g., y > 0.5. That is, how can we estimate Sobol indices conditionally—restricted to samples where the model response exceeds a certain threshold?
Also, does anyone know of alternative libraries, frameworks, or SA methodologies beyond SALib that are better suited to analyzing conditional subspaces of the input domain?
Thank you! I appreciate the help!