Statistics Glossary → Q² predictive relevance (Stone-Geisser, PRESS)
Q² predictive relevance (Stone-Geisser, PRESS)concept
An R²-like measure based on cross-validation that shows how well a model predicts data points not used in its estimation.
It is computed as Q² = 1 - PRESS / SS, where PRESS is the sum of squared errors of predictions made with each observation (or block of observations) left out of the model, and SS is the sum of squared deviations of the outcome from its mean. R² measures fit to the data the model was fitted on and rises with every extra variable; Q² measures prediction on unseen data and falls when the model overfits, and can even be negative. Q² > 0 means the model predicts better than simply using the mean.
In partial least squares regression, the point where Q² stops rising and starts falling as components are added is used to choose the number of components. In PLS-SEM, the Stone-Geisser Q² is computed by 'blindfolding': every D-th value of the data matrix is omitted in turn, the model is estimated on the remaining data, and the omitted values are predicted. D (the omission distance) is usually chosen between 5 and 10, and the number of observations should not be an exact multiple of D. Blindfolding-based Q² is not a full out-of-sample prediction test, which is why k-fold approaches such as PLSpredict are also recommended in PLS-SEM.
An everyday example
A café builds a model predicting daily sales from the weather and type of day; on its own data the model gives R² = 0.85. When each day is left out in turn and predicted from a model built on the remaining days, Q² = 0.62. The gap shows the model looks better at explaining the past than at predicting the future, and that it overfits somewhat.
In MerQur
Available in PLS-SEM as the 'Q2 (blindfolding)' check box and the 'Omission distance (D, default 7)' option; the Stone-Geisser Q2 is computed with the cross-validated redundancy approach for reflective endogenous constructs. In partial least squares regression the 'Select number of components by CV' option gives PRESS, RMSECV and Q2 = 1 - PRESS/SST for each number of components.
Analyses where this concept appears
Partial Least Squares Regression (PLS)PLS-SEM (Partial Least Squares Path Modeling)