1
$\begingroup$

I have used one at a time (OAT) approach in my model for sensitivity analysis. Which gives me elementry effects and variances for the input parameters, helping me to eliminate the less influencial parameters. OAT is a local sensitivity analysis approach.

I now want to use Sobol analysis (global method) in my model on the parameters which I have selected from OAT. Sobol method is a variance based method which will give me the interaction effects in terms of sensitivity indices.

But I am unsure what exactly I will get from Sobol method or how will it be different from OAT method. Hence I want to have a reasoning before using the method. Can we compare both the methods? Can anyone please come up with

  1. some comparing points between these 2 methods and
  2. advantages/uses of Sobol method in general?
$\endgroup$
2
$\begingroup$

OAT is a local method, meaning that you consider one point of your input space, and you then change its components/features one at a time. It gives you information on the sensitivity of your model on that specific point not on the entire space.

On the contrary, Sobol analysis is global method. You can have

  • the global importance of the input features on the compact space you consider (with 1st order indices)
  • the importance of the interactions of input features between each others (with higher order indices and total indices), which you cannot have at all with OAT.

There is a particular type of OAT method called Morris which consists in repeating OAT on multiple points of the input space, so that from these multiple local sensitivity analyses, you can get a global analysis. This method is however not as precise as Sobol indices (no way to differentiate between interaction and non-linear effect! for example) and is most of the time only advised if you want to do screening, i.e. removing the really non-important features as a first step.

$\endgroup$
0

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.