Cobweb Plots for Conditional Uncertainty and Sensitivity Analysis

Description

Draw each simulated scenario as a line across parallel axes and filter on extreme outcomes to see which inputs drive them, including effects that overall correlations miss.

Global importance measures such as tornado graphs summarise each input's overall association with the output and can hide inputs that matter only in some regions. Cobweb plots, a form of parallel coordinates, show every Monte Carlo sample as a jagged line across vertical axes for all inputs and outputs. On a percentile scale the pattern of crossings between adjacent axes reveals dependence: parallel lines for rank correlation 1, a characteristic spread for independence, a central crossing for -1. The key move is conditionalisation: select the high or low tail of the output and keep only lines passing through it, then see where those lines sit on each input axis. Kurowicka and Cooke's car example shows an input with zero overall rank correlation with the output that is nonetheless high in both its extreme high and extreme low outcomes.

Attributes

Loss type
Siloed
Effort
Hours

Steps

  1. Generate a Monte Carlo sample of inputs and outputs from the uncertainty model.
  2. Draw each sample as a line across parallel vertical axes, about 200 lines for print or 1,000 on screen, coloured by quartile of one variable.
  3. Switch axes to a percentile scale and read crossing patterns between adjacent axes as dependence.
  4. Select the extreme high region of the output and note where the remaining lines fall on each input axis; repeat for the low region and for specific output bands.
  5. Compare these conditional patterns with global measures (rank correlations, tornado graphs, variance-based indices) and flag inputs that matter locally.
  6. Report the inputs that drive the outcomes of most concern, with screenshots of the conditional views.

Sources

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

Also known as

  • Cobweb plot
  • Parallel coordinate plot for uncertainty analysis
  • Conditional cobweb plot
  • Cobweb plots with interactive conditionalisation for local sensitivity

Categories

  • Analytics & Behaviour
  • Statistics & Modelling