Why does MAP vastly outperform sample in bayesian clustering?

Is there any work around where I could have the non-mixing property of MvNormal while using censored univariate normals? If that is not straight forward I will open another topic illustrating as to in what context I need such a model.