Description
Regresses an outcome on many correlated predictors by projecting them onto a few latent components that maximise covariance with the outcome, working when cases are few.
PLS regression handles the classic small-sample UX problem: dozens of correlated predictors (interface metrics, attitude items, behavioural counts) but only 20-60 participants. Ordinary least squares breaks down under collinearity or when predictors exceed cases; PLS extracts a small number of orthogonal components that capture predictor variance most relevant to the outcome, then regresses on those. Developed for chemometrics, where spectra have hundreds of collinear wavelengths and few samples, it is rigorous when the number of components is chosen by cross-validation and predictor importance is read through VIP scores rather than raw weights. It predicts well but its components are not causal constructs.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Goal
- Predict performance
- Stage
- Evaluation
- Data
- Quantitative
- Effort
- Days
Steps
- Assemble a predictor matrix X (e.g. 30 interface metrics) and outcome y for each participant or screen; standardise all variables.
- Fit PLS models with 1 to 10 components and choose the number minimising cross-validated prediction error (leave-one-out or 10-fold).
- Report Q-squared (cross-validated R-squared); values above 0.5 are generally considered good predictive relevance.
- Compute Variable Importance in Projection scores and treat VIP above 1.0 as influential, below 0.8 as candidates for removal.
- Refit with the reduced predictor set and confirm predictive performance does not degrade.
- Interpret the leading components by their loadings and validate any design implication with a targeted study.
Sources
- Applied Predictive Modeling (Max Kuhn & Kjell Johnson)
- R for Marketing Research and Analytics (Chris Chapman & Elea McDonnell Feit)
- Wold, Sjöström & Eriksson (2001) PLS-regression: a basic tool of chemometrics
- Applied Predictive Modeling (Kuhn & Johnson)
- R for Marketing Research (Chapman & Feit)
- The Inquiry Series, Ch 43 — Regression, Moderation & Mediation for UX Researchers
Also known as
- Partial least squares regression
- Projection to latent structures
- PLS-R
Categories
- Analytics & Behaviour
- Statistics & Modelling