Latin Hypercube Sampling with Iman-Conover Rank Correlation

Description

Propagate uncertain and possibly correlated inputs through a model with a stratified sample that covers each input's range evenly, needing fewer runs than plain random sampling.

Plain Monte Carlo draws inputs at random, so with few runs some regions are over-sampled and others missed. Latin hypercube sampling splits each input's distribution into N equal-probability intervals, draws one value from each, and pairs values across inputs by random permutation, so every input's range is covered evenly with only N runs. It gives unbiased estimates of means and suits models that are slow or costly to run, such as large spreadsheets, simulations or calls to external services. Inputs are rarely independent; the Iman-Conover method induces target rank correlations by reordering each column of the Latin hypercube sample to match the ranks of a correlated normal sample, leaving the marginal distributions untouched. The model is then run on the N input vectors and the output distribution summarised.

Attributes

Loss type
Siloed
Effort
Days

Steps

  1. Specify a distribution for each uncertain input from data or structured expert judgement.
  2. Choose the number of runs N and divide each input's distribution into N equal-probability intervals.
  3. Draw one value at random within each interval for each input and map it through the inverse cumulative distribution.
  4. Pair values across inputs by independent random permutations.
  5. To induce dependence, generate N correlated normal vectors with the target rank correlation matrix, rank them, and reorder each input column to match those ranks (Iman-Conover).
  6. Run the model on the N input vectors and summarise the output distribution, percentiles and contribution of inputs.

Sources

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

Also known as

  • Latin hypercube sampling
  • LHS
  • Iman-Conover method
  • Probabilistic uncertainty propagation
  • Monte Carlo uncertainty propagation with Latin hypercube sampling and Iman-Conover rank correlation

Categories

  • Analytics & Behaviour
  • Statistics & Modelling