I am now studying the Metropolis-Hastings algorithm and I want to apply it in order to made a Bayesian Inference of a function $y=f(x)$ to a dataset $D=\{x_i,y_i\}$. Five parameters of the function need to be determined $\theta=\{\theta_1,...,\theta_5\}$.

Here I want to first introduce how to develop the M-H algorithm.

I start from an initial guess of the model parameters $\theta^{(0)}$ and, using a candidate distribution (in my case a Multivariate Normal Distribution centred in $\theta^{(0)}$ with a certain variance $\delta$ that has to be tuned), select a candidate $\theta^{(*)}$. Since the candidate distribution is symmetrical, the acceptance probability is the ratio:

\begin{equation} \alpha=min\{1, \frac{\pi(\theta^{(*)})}{\pi(\theta^{(0)})} \} \end{equation} Where $\pi(\bullet)$ is the target function, proportional to the Posterior Distribution.

$\theta^{(*)}$ is accepted ($\theta^{(1)}=\theta^{(*)}$) if $\alpha$ is bigger than $u=U(0,1)$, otherwise it is rejected ($\theta^{(1)}=\theta^{(0)}$). A new candidate can be drawn and the process repeated.

In case I want to use the M-H algorithm in a inference problem, the target function $\pi(\theta)$ can be substituted by the Likelihood function $L(\theta)$.

However, if I want to made a Bayesian inference, it is not really clear to me the role of the prior distributions associated to the elements $\theta_i$ of the vector parameter. How do they figure in the acceptance probability?

Thanks for the answers and the comments

  • $\begingroup$ I am not sure if I understand your question. If I guess, you can use improper prior (pr(parameter)=1). MH is well developed in literature and there are many packages that use MH, like MHadaptive, BDWreg ... in R. Just take a look at the source of the packages. $\endgroup$ – TPArrow Jul 21 '16 at 14:41

The posterior distribution can be expressed as $$ p(\theta|y) = \frac{p(y|\theta)\,p(\theta)}{p(y)} $$ where $$ p(y) = \int p(y|\theta)\,p(\theta)\,d\theta. $$ The likelihood is given by $L(\theta) := p(y|\theta)$ and the prior is $p(\theta)$. The target function can be expressed as $$ \pi(\theta) = c\,L(\theta)\,p(\theta), $$ where $c^{-1} = p(y)$. The (possibly unknown) constant of proportionality $c$ cancels out in the ratio. The algorithm then proceeds as outlined.

  • $\begingroup$ Thank you. So, the target function is given by the product between the likelihood and the prior distribution. Thank you! $\endgroup$ – D.Leo Jul 21 '16 at 16:33

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.