Mixture Experiments (Simplex-Lattice and Simplex-Centroid Designs)

Description

Designs and models for factors that are shares of a fixed whole summing to 100 percent, so the response depends on the blend rather than absolute amounts.

In a mixture experiment the factors are proportions that must sum to one, so changing one forces another to change and ordinary factorial designs cannot be used. The design space is a simplex. Simplex-lattice designs place points at equally spaced proportions; simplex-centroid designs use the pure components, every equal blend of subsets and the overall centroid (2^k - 1 points); axial designs add points along component axes so all components are present. Responses are modelled with Scheffe canonical polynomials, which have no intercept, and results are shown as contours on triangular coordinates. Lower and upper bounds on components, and process variables, can be added. Coefficients describe blending behaviour rather than separate component effects, which is the main interpretive trap.

Attributes

Loss type
Unmigrated
Effort
Weeks

Steps

  1. Define the k components and confirm their proportions must sum to 1, adding any lower or upper bounds.
  2. Choose a simplex-lattice, simplex-centroid or axial design with at least as many distinct blends as model parameters, plus replicates of some points for an error estimate.
  3. Randomise the run order.
  4. Fit a linear or quadratic Scheffe canonical polynomial without intercept.
  5. Plot the fitted response on simplex coordinates and identify the best blend, then confirm it with a follow-up run.
  6. Add process variables or constrained-region designs if the blend interacts with other settings.

Sources

  • Design and Analysis of Experiments, Vol. 1 (Hinkelmann & Kempthorne, 2008)

Also known as

  • Mixture designs
  • Simplex-lattice design
  • Simplex-centroid design
  • Scheffe canonical polynomials
  • Experiments with mixtures

Categories

  • Analytics & Behaviour
  • Experimentation & Causal Inference