Repository navigation
Wrong second-order effects for lognormal inputs #43
Description
Activity
Are we seeing the same thing on the scatter plots? Because the bottom one does not show a lognormal distribution.
Exactly, when there are two uniforms, the indices are adequate, but when one is uniform and another lognormal -> weird negative second-order, eventhough inputs are independent.
For the same problem in matlab the second-order is about 1% for both cases, which is a desired reasonable result.
Here is a test dataset, first column is the output, the other two are inputs, one uniform and one lognormal
test_data_43.csvand the correct indices for it (computed with Matlab package)
- added a commit that references this issue
on Feb 26, 2026 - linked a pull request that will close this issueFix SOE with lognormal distribution #44
on Feb 26, 2026
Python package (and dashboard) returns wildly negative second-order effects for independent inputs, if one of them is lognormal and another uniform (I haven't tested other combinations). In the Matlab package, the second-order is adequate.
Source: The difference is that in MATLAB we have a custom binning function, while in Python we use a ready-made one from some library.
Solution: So I guess there is no better solution than adopt the Matlab binning algorithm to Python as well.
Example.
(Matlab for the same problem with lognormal interest_rate returns a meaningful SOE=0.01)