I have a discrete time Markov Chain $\{X_n: n \in \mathbb{N}_0\}$ with unknown transition matrix $P \in \mathbb{R}^{M \times M}$ on the state space $\mathcal{S}_X = \{1,2, \dots, M\}$, with $M \geq 2$. At each time step, the random variable $Y_n$ is defined by $Y_n = g(X_n)$, where $g: \mathcal{S}_X \to \{0,1\}$ is a deterministic function.

I can work out the emission probabilities $\mathbb{P}(Y_n = 0|X_n)$ and $ \mathbb{P}(Y_n=1|X_n)$, given $P$, i.e. these probabilities are known and computable functions of the entries of $P$.

Let's say I now observe $Z_n = \sum_{i =1}^k Y_n^i = \sum_{i =1}^k g(X_n^i)$, where $k \geq 1$ is unknown, where the sum is taken over $k$ iid Markov Chains $X_n^i$. Note that while I can write down the likelihood for $Z_1, Z_2, \dots$, the function is intractable for even moderately small values of $k$.

I am interested in estimating both $P$ and $k$ from my data $Z_n$. I know that I can use some sort of simulation based particle filter to estimate $P$ from the data in $Z$, however, I am wondering how I can estimate $k$. The value for $k$ is typically large (>1000). I am not sure that this problem requires advanced/transdimensional MCMC methods (e.g. reversible jump), since $P$ is independent of $k$, but I am not sure how to proceed as $k$ is large. It seems inefficient to fix $k$ and fit $P$ to each corresponding model, and then performing Bayesian model selection.

Does anyone know of any quick but accurate methods to perform model selection?



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.