Morris Elementary-Effects Screening

Description

Screen many model inputs cheaply with randomised one-at-a-time trajectories, separating negligible inputs from influential ones and flagging nonlinearity or interaction.

When a model has many inputs and each run is costly, a full sensitivity analysis is unaffordable. Morris screening scales each input to [0, 1] on a grid of p levels and builds r randomised trajectories through the input space, each changing one input at a time by a fixed step. Each change yields an elementary effect: the output change divided by the step. A trajectory costs n + 1 runs for n inputs, so the design needs (n + 1)r runs rather than 2rn. For each input the mean of its absolute elementary effects measures overall influence and their standard deviation signals nonlinearity or interaction with other inputs. Plotting standard deviation against mean sorts inputs into negligible, linear and nonlinear or interacting groups. Unlike one-way sensitivity analysis from a single base case, the trajectories cover the whole input space.

Attributes

Loss type
Siloed
Effort
Hours

Steps

  1. List the n inputs and their plausible ranges, and scale each to [0, 1].
  2. Choose the number of grid levels p (often 4 or 6) and the step delta (with p even, delta = p / (2(p - 1))).
  3. Generate r random trajectories (often 10 to 20), each from a random base point, changing one input at a time in random order and direction.
  4. Run the model at every trajectory point, (n + 1)r runs in total.
  5. Compute each input's elementary effects and their mean absolute value and standard deviation.
  6. Plot standard deviation against mean; drop inputs near the origin from further analysis and investigate those with large means or large standard deviations with variance-based methods.

Sources

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

Also known as

  • Morris method
  • Elementary effects method
  • Morris screening

Categories

  • Analytics & Behaviour
  • Statistics & Modelling