Sensory Informed Incomplete Block Designs
This technical presentation from University of Guelph and Compusense details the statistical methodology underlying sensory-informed incomplete block designs for consumer segmentation. The core challenge: traditional clustering requires complete data, but products causing sensory fatigue (wine, alcohol, intense flavors) necessitate incomplete block designs. Simple mean imputation creates artifacts that distort cluster structures. The solution is an incremental (column-wise) EM algorithm that efficiently estimates covariance matrices and imputes missing values while preserving the underlying data structure. The method uses Gaussian mixture models with factor analyzer covariance structures, with model selection via BIC. Validation on the Iris dataset (with simulated missing data) showed near-perfect recovery of true clusters. Application to a white bread study (420 consumers, 12 products, 12-present-6 design with nested 12:3 and 12:4 designs) successfully identified six stable consumer segments using a two-factor model—demonstrating that smaller nested designs yield comparable segmentation to larger designs.
Browne, R.P, Findlay, C. J., McNicholas, P. D., & Castura, J. C. (2012). Sensory Informed Incomplete Block Designs. In: 11th Sensometric Meeting. 11-13 July. Rennes, France.