I have recently been learning about Sobol indices and have found them quite informative. However, as I have explored them I have encoutnered situations where I have found them counter intuitive and in fact more misleading than helpful.

I am lead to assume that this must be a misconception on my part, so perhaps somebody could shed some light on this for me?

For our considerations let $ X,Y \sim \mathcal N(0,1) $ be standard normally distribution and consider a model $ Z = f(X,Y) $.

Scenario (1): Let $ f(X,Y) = aX+bY $, then the Sobol indices are $ S_X = \frac{a^2}{a^2+b^2} $, $ S_Y = \frac{b^2}{a^2+b^2} $ and $ S_{XY} = 0 $, which makes a lot of sense each of them contributing a fraction of the variance proportional to their own variance.

Scenario (2): Let $ f(X,Y) = XY $, then the Sobol indices are $ S_X = S_Y = 0 $ and $ S_{XY} = 1 $, which again makes sense. All variation is contributed by their combination.

Scenario (3): Let $ f(X,Y) = X^2Y $. Now it gets counter-intuitive to me. Here the Sobol indices are $ S_X = 0 $, $ S_Y = \frac13 $ and $ S_{XY} = \frac23 $. I can compute the indices, no problem, but the way I see it $ X $ is much more significant to the variance than $ Y $ here, which is not clear from the indices at all. In fact, they kind of seem to tell the opposite story.

Of course, $ S_X = 0 $ in (3) is due to it being proportional to $ \mathbb E_Y[f(X,Y)] = 0 $, so $ f(X,Y) = X^kY $ for all $ k > 0 $ has $ S_X = 0 $.

What exactly am I supposed to take away from scenario (3)? And if this is kind of an artifact of the setup, how am I supposed to identify such an issue in another setting where the model $ f $ is way more obscure? or straight up black box?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.