We show analytically that RVM selects predictors when the absolute z-ratio (|least squares estimate|/standard error) exceeds 1 in the case of orthogonal predictors and, for M = 2, that this still holds true for correlated predictors when the other z-ratio is large. RVM selects the stronger of two highly correlated predictors. In experiments with real and simulated data, RVM is outcompeted by other popular regularization methods (LASSO and/or PLS) in terms of the prediction performance. We conclude that type II ML is not the general answer in high dimensional prediction problems.
In extensions of RVM to obtain stronger selection, improper priors (based on the inverse gamma family) have been assigned to the inverse precisions (variances) with parameters estimated by penalized marginal likelihood. We critically assess this approach and suggest a proper variance prior related to the Beta distribution which gives similar selection and shrinkage properties and allows a fully Bayesian treatment.