Reversible jump Markov chain Monte Carlo
Posted on
The note is for Green, P.J. (1995). “Reversible Jump Markov Chain Monte Carlo Computation and Bayesian Model Determination”. Biometrika. 82 (4): 711–732.
Abstract
MCMC methods for Bayesian computation have until recently been restricted to problems where the joint distribution of all variables has a density with respect to some fixed standard underlying measure. Not available for application to Bayesian model determination, where the dimensionality of the parameter vector is typically not fixed.
A new framework for the construction of reversible Markov chain samplers that jumps between parameter subspaces of differing dimensionality, which is flexible and entirely constructive.
Illustrated with applications to multiple change-point analysis in one and two dimensions, and to a Bayesian comparison of binomial experiments.
A discrete choice between a set of models, a parameter vector with an interpretation depending on the models in question, and data, influenced by the model and parameter values
Model
Suppose that we have a countable collection of candidate models $\{\cM_k,k\in\cK\}$. Model $\cM_k$ has a vector $\theta^{(k)}$ of unknown parameters, assumed to lie in $\IR^{n_k}$, where the dimension $n_k$ may vary from model to model. We observe data $y$. There is a natural hierarchical structure expressed by modelling the joint distribution of $(k,\theta^{(k)},y)$ as
\[p(k,\theta^{(k)},y) = p(k)p(\theta^{(k)}\mid k)p(y\mid k, \theta^{(k)})\,,\]that is, the product of model probability, prior and likelihood. We can abbreviate the pair $(k,\theta^{(k)})$ by $x$. For given $k$, $x$ lies in $\cC_k = \{k\}\times \IR^{n_k}$.
Consider a concrete example, consider a change-point problem in which there is an unknown number of change-points in a piecewise constant regression function on the interval $[0,L]$. For $k\in \cK=\{0,1,2,\ldots\}$, model $\cM_k$ says that there are exactly $k$ change-points. To parametrise the resulting step function, we need to specify the position of each chang-point, and the value of the function on each of the $(k+1)$ subintervals into which $[0,L]$ is divided. Thus $\theta^{(k)}$ is a vector of length $n_k=2k+1$\,.
Bayesian inference about $k$ and $\theta^{(k)}$ will be based on the joint posterior $p(k,\theta^{(k)}\mid y)$, which is the target of the Markov chain Monte Carlo computations described below. We often factorise this as
\[p(k,\theta^{(k)}\mid y)=p(k\mid y)p(\theta^{(k)}\mid k,y)\]Consider the Bayes factor for one model relative to another,
\[\frac{p(y\mid k_1)}{p(y\mid k_0)} = \frac{p(k_1\mid y)}{p(k_0\mid y)}\div \frac{p(k_1)}{p(k_0)}\]Some work on MCMC computation with application to aspects of Bayesian model determination:
- jump-diffusion samplers
- embedding method in which the ${\cC_k}$ are mapped onto subsets of a single parameter space.
Let $\pi(dx)$ be the target distribution of interest. In MCMC, construct a Markov transition kernel $P(x,dx’)$ that is aperiodic and irreducible, and satisfies detailed balance,
\[\int_A\int_B\pi(dx)P(x,dx') = \int_B\int_A\pi(dx')P(x',dx)\]for all appropriate $A,B$.
In a typical application with multiple parameter subspaces $\{\cC_k\}$ of different dimensionality, it will be necessary to devise different types of move between the subspaces.
When the current state is $x$, we propose a move of type $m$, that would take the state to $dx’$, with probability $q_m(x,dx’)$. For the moment, this is an arbitrary sub-probability measure on $m$ and $x’$. Thus $\sum_mq_m(x,\cC)\le$, and with probability $1-\sum_mq_m(x,\cC)$, no change to the present state is proposed.
The transition kernel can be written as
\[P(x,B) = \sum_m\int_B q_m(x,dx')\alpha_m(x,x') + s(x)I(x\in B)\]for Borel sets $B$ in $\cC$, and the probability of not moving from $x$, either through a proposed move being rejected, or because no move is attempted is
\[s(x)=\sum_m\int_\cC q_m(x,dx')\{1-\alpha_m(x,x')\}+1-\sum_mq_m(x,\cC)\]The detailed balance relation requires the equilibrium probability of moving from $A$ to $B$ to equal that from $B$ to $A$, for all Borel sets $A,B$ in $\cC$. It is sufficient that
\[\int_A\pi(dx)\int_Bq_m(x,dx')\alpha_m(x,x') = \int_B\pi(dx')\int_A q_m(x',dx)\alpha_m(x',x)\]By some discussion, we can take
\[\alpha_m(x,x') = \min\Big\{1,\frac{\pi(dx')q_m(x',dx)}{\pi(dx)q_m(x,dx')}\Big\}\]Example: Coal mining disasters
For data points ${y_i,i=1,2,\ldots,n}\in [0,L]$ from a Poisson process with rate given by the function $x(t)$, the log-likelihood is
\[\sum_{i=1}^n\log\{x(y_i)\} - \int_0^Lx(t)dt\]Assuming that the rate function $x(\cdot)$ on $[0,L]$ is a step function. Suppose that there are $k$ steps, at positions $0 < s_1 < s_2 < \ldots < s_k < L$, and that the step function takes the value $h_j$, which we call its height, on the subinterval $[s_j,s_{j+1})$ for $j=0,1,\ldots,k$, writing $s_0=0, s_{k+1}=L$ for convenience. The prior model is specified by supposing that $k$ is drawn from the Poisson distribution
\[p(k) = e^{-\lambda}\frac{\lambda^k}{k!}\]the heights $h_0,h_1,\ldots,h_k$ are independently drawn from $\Gamma(\alpha,\beta)$.
If $x$ is a step function on $[0,L]$, some possible transitions are:
- a change to the height of a randomly chosen step
- a change to the position of a randomly chosen step
- “birth” of a new step at a randomly chose location in $[0,L]$
- “death” of a randomly chosen step
the last two involve changing the dimension of $x$, the countable set of moves is denoted as ${H,P,0,1,2,\ldots}$
At each transition, an independent random choice is made between attempting each of the at most four available move types $(H,P,k,k-1)$. These have probabilities $\eta_k$ for $H$, $\pi_k$ for $P$, $b_k$ for $k$, and $d_k$ for $k-1$.