A data-driven reversible jump for estimating a finite mixture of regression models

We propose a data-driven reversible jump (DDRJ) method for selecting and estimating a mixture of regression models in a single run, which can also be applied as a robust regression model to outliers. We compare the clustering and estimation performance of the proposed method with Expectation–Maximization and Gibbs sampler algorithms combined with model selection criteria in synthetic data sets. Under tested conditions, DDRJ outperforms these traditional methods in identifying the number of groups, classification and precision of estimates. When compared with traditional reversible jump algorithms, the data-driven procedure simplifies the calculations and implementation and shows a better mixing and faster convergence. Finally, we apply the proposed method to analyze two well-studied data sets: tone perception (a simple data set) and baseball salaries (a more complex data set with a larger number of covariates).