0
$\begingroup$

I have a regressive model of this form \begin{equation} y=f(\theta)+\varepsilon \end{equation} to describe observations $y$, with noise $\varepsilon$ and a parametric function $f$ with parameters $\theta$. In reality the regressive model is much more complicated than what I've shown but that's not relevant for my question.

Now I am interested in the scenario where the parameters $\theta$ vary throughout the experiments, i.e. they have an inherent variability, and I assume that their distribution is multivariate normal $\theta\sim\mathcal{N}(\mu,\Sigma)$. Hence, I decided to build a hierarchical model in order to identify the hyperparameters $\mu,\Sigma$, having chosen a multivariate normal $\mathcal{N}(\mu_0,\Sigma_0)$ and an inverse Wishart $\mathcal{W}^{-1}(\Psi_0,\nu_0)$, respectively, as their hyperpriors.

Then, I have implemented a Metropolis-within-Gibbs routine in MATLAB, to sample these conditional posteriors: \begin{equation} \label{eq:20} \small \mu|\theta,\Sigma,y \sim \mathcal{N}\left((\Sigma_0^{-1} + n\Sigma^{-1})^{-1}( \Sigma_0^{-1}\mu_0 + n \Sigma^{-1} \theta ),(\Sigma_0^{-1} + n \Sigma^{-1})^{-1}\right) \end{equation} for the mean vector, and \begin{equation} \label{eq:21} \small \Sigma|\theta,\mu, y\sim\mathcal{W}^{-1}(\Psi_0 +\sum_{i=1}^n (\theta_i - \mu) (\theta_i - \mu)^T,n+\nu_0) \end{equation} for the covariance matrix, where $n=1$. The above choices are motivated by conjugate priors. During the iterative process of the Gibbs sampler, the Metropolis-Hastings is used to take one sample from the posterior distribution $\theta|\mu,\Sigma,y$ (which is also multivariate normal).

My current results show a very wide posterior for $\Sigma$ and I cannot understand why that is. I am still unsure of:

  • how many degrees of freedom $\nu_0$ I should consider for the inverse Wishart;
  • if it is correct to assume that $n=1$ since I am taking only one sample from the MH sampler;
  • and if during the Gibbs routine I need to multiply the inverse Wishart sample by a factor $(\nu_0-k)$ as shown here ($k$ is the number of parameters $\theta$)?

I took 10000 samples to generate the below plots (one of them is a zoom of the variance of the parameter).

Histogram of parameter (left) and its mean (right) Histogram of Variance of one parameter Histogram of Variance of one parameter (zoom)

UPDATE According to this question my formulation might be fundamentally wrong, in the sense that I should be using a Normal Inverse Wishart instead of a separate MVN and inverse Wishart. I would like to confirm this, and gain further insight into how the Gibbs sampler needs to consider the conditional posteriors. Any answer to this specific question is very welcome.

$\endgroup$

Your Answer

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

Browse other questions tagged or ask your own question.