Suppose I have the following DLM:

$x_t = \Phi x_{t-1} + w_t$

$y_t = A x_t + v_t$

$x_0 \sim N(\mu_0,\Sigma_0)$

$w_t \sim N(0,Q)$

$v_t \sim N(0,R)$

Let $\Theta = \{\mu_0,\Sigma_0,\Phi,Q,A,R\}$. I can do Gibbs sampling for the joint distribution $P(x,\Theta|y)$ by sampling from the conditionals $P(x|\Theta,y)$ and $P(\Theta|x,y)$.

As I understand it, for $P(x|\Theta,y)$, sampling is done via the Forward Filtering Backward Sampling algorithm, which translates to running a Kalman smoother, which handles partially and fully missing values in $y_t$ by more or less ignoring them in the calculations.

On the other hand, for sampling $P(\Theta|x,y)$, how should I go about handling missing values in $y_t$? Can I ignore fully missing $y_t$ when calculating the full conditionals? What about partially missing $y_t$? Or should I simulate missing $y_t$ in another Gibbs step?


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.