Dirichlet Prior for Estimating Unknown Regression Error Heteroskedasticity

We propose a Bayesian procedure to estimate heteroskedastic variances of the regression error term ?O, when the form of heteroskedasticity is unknown. The prior information on ?O is based on a Dirichlet distribution, and in the Markov Chain Monte Carlo sampling, its proposal density parameters' information is elicited from the well-known Eicker-White Heteroskedasticity Consistent Variance-Covariance Matrix Estimator. We present an emprical example to show that our scheme works.