Variance-Based Global Sensitivity Analysis (Correlation Ratio and Sobol Indices)

Description

Rank uncertain inputs by the share of output variance each explains, alone and through interactions, across the whole input space.

One-at-a-time sensitivity analysis varies inputs around a single base case and misses nonlinearity and interactions. Variance-based methods ask how much output variance would disappear if an input were known. The correlation ratio, the variance of the conditional mean of the output given an input divided by the total output variance, is the first-order share for that input; it equals the squared correlation between the output and its best predictor from that input, so it covers nonlinear effects that rank correlations miss. Sobol indices decompose the model into main effects and interactions, giving first-order and total indices. From a Monte Carlo sample, the conditional mean can be estimated by binning the input, kernel regression or fitting a polynomial, with split-sample checks to avoid overfitting. The cost is computational, which is why screening often comes first.

Attributes

Loss type
Siloed
Effort
Days

Steps

  1. Specify input distributions (and dependence) and generate a Monte Carlo or Latin hypercube sample of inputs and outputs.
  2. For each input, estimate the conditional mean of the output given that input: sort and bin into m cells, use kernel regression, or fit a polynomial in the input.
  3. Guard against overfitting by fitting on half the sample and checking on the other half; increase polynomial degree or bin count only while held-out fit holds.
  4. Compute the correlation ratio: variance of the conditional means divided by total output variance.
  5. Where interactions matter, estimate Sobol first-order and total indices with a dedicated sampling design.
  6. Rank inputs by their indices and report the shares, noting inputs whose total index greatly exceeds their first-order index (interactions).

Sources

  • Uncertainty Analysis with High Dimensional Dependence Modelling (Kurowicka & Cooke, 2006)

Also known as

  • Global sensitivity analysis
  • Variance-based sensitivity analysis
  • Sobol indices
  • Sobol sensitivity indices
  • Correlation ratio
  • Global variance-based sensitivity analysis: correlation ratio and Sobol indices

Categories

  • Analytics & Behaviour
  • Statistics & Modelling