This page is published in English.
Sensitivity
Morris Screening Method - Elementary effects for global sensitivity analysis
Global sensitivity - elementary effects screening
Morris, M.D.1991doi:10.1080/00401706.1991.10484804 ↗
Overview
Rank factors by μ* descending. Factors with μ* near zero are negligible. High σ/μ* ratio signals non-linearity or interactions - follow up with Sobol' indices for those factors.
- Output
- sensitivity index, higher is better
- Data
- Crisp, complete numeric matrix
- Size
- 2+ alternatives, 3-20 criteria works best
- Used for
- Sensitivity analysis, uncertainty quantification, MCDM robustness
Fits when / Look elsewhere when
Fits when
- •r(k+1) evaluations - very cheap
- •No distributional assumption required
- •Detects non-linearity and interactions via σ
Look elsewhere when
- •Full variance decomposition required - use Sobol'
- •Only 2-3 factors - Sobol' is not significantly more expensive
Assumptions to verify
- Factors can be perturbed independently
- Model evaluation is deterministic given inputs
Limitations
- •Does not quantify exact sensitivity indices (only ranking)
- •Results depend on grid resolution p and number of trajectories r
Edge cases and pitfalls
Using μ (signed) instead of μ* - cancellation in μ can mask important non-monotone factors.
Too few trajectories (r<10) gives unreliable estimates.
Morris screens importance but not interaction structure - Sobol' needed for full decomposition.
Works with
Its derived weights can feed
How to cite
Morris, M.D. (1991). Factorial sampling plans for preliminary computational experiments. Technometrics. https://doi.org/10.1080/00401706.1991.10484804
System ID, as it appears in reports and the API
SENSITIVITY-MORRIS