Skewed probit regression -- Identifiability, contraction and reformulation

Abstract

Skewed probit regression is but one example of a statistical model that generalizes a simpler model, like probit regression. All skew-symmetric distributions and link functions arise from symmetric distributions by incorporating a skewness parameter through some skewing mechanism. In this work we address some fundamental issues in skewed probit regression, and more genreally skew-symmetric distributions or skew-symmetric link functions. We address the issue of identifiability of the skewed probit model parameters by reformulating the intercept from first principles. A new standardization of the skew link function is given to provide and anchored interpretation of the inference. Possible skewness parameters are investigated and the penalizing complexity priors of these are derived. This prior is invariant under reparameterization of the skewness parameter and quantifies the contraction of the skewed probit model to the probit model. The proposed results are available in the R-INLA package and we illustrate the use and effects of this work using simulated data, and well-known datasets using the link as well as the likelihood.