Bayesian mode regression using mixtures of triangular densities

Bayesian semiparametric models for mean and median regressions abound, but a void for mode regressions exists. We fill this gap by nonparametrically modeling the error distribution in such regressions that entails constructing prior distributions on densities which exhibit flexibility, while fixing the mode at 0. Such priors exist when constraining the mean and median but, to our knowledge, there is none for the mode. Our solution with mixtures of triangular distributions results in a conditionally conjugate prior on the space of unimodal, untruncated, convex densities. Consistency properties of the resulting modal estimators are studied, followed by simulated and real data illustrations.