The Bayesian Multivariate Linear Regression wikipedia article states that posterior parameters of a multivariate linear regression model with a Inverse-Wishart prior can be obtained from the following equations:

$$ \mathbf{V_n}=\mathbf{V_0}+(\mathbf{Y}-\mathbf{XB_n})^{\rm T}(\mathbf{Y}-\mathbf{XB_n})+(\mathbf{B_n}-\mathbf{B_0})^{\rm T}\boldsymbol\Lambda_0(\mathbf{B_n}-\mathbf{B_0}) $$ $$ \boldsymbol\nu_n=\boldsymbol\nu_0+n $$ $$ {B_n}=(\mathbf{X}^{\rm T}\mathbf{X}+\boldsymbol\Lambda_0)^{-1}(\mathbf{X}^{\rm T}\mathbf{Y}+\boldsymbol\Lambda_0\mathbf{B_0}) $$ $$ \boldsymbol\Lambda_n=\mathbf{X}^{\rm T}\mathbf{X}+\boldsymbol\Lambda_0 $$

However, from this equations we get that the dimensions of the parameter $V_n$ would be 1x1, meaning that the scale matrix of the Inverse-Wishart distribution, and therefore the sampled covariance matrix is a scalar.

Is this correct? And if it is, what is the difference of using the Inverse-Wishart prior instead of the Inverse-Gamma distribution?


In the article you link to, the dimensions of $\mathbf{Y}$ are given as $n \times m$ (where $m$ is the number of outcome variables). From this (and consistent with the other matrix dimensions listed in the article), it follows that the $(\mathbf{Y}-\mathbf{XB_n})^{\rm T}(\mathbf{Y}-\mathbf{XB_n})$ term must have dimensions $m\times m$, and so those must also be the dimensions of $\mathbf{V_n}$.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.