ABSTRACT
Durbin and Koopman (1997) consider a Gaussian state equation and a non-Gaussian observation process. They approximate the density p(y1lc1) by a Gaussian conditional density g(y1 lc-1) to arrive at a Gaussian approximation to the likelihood. Simulation techniques are then used to compute the adjustment from the approximate likelihood to the correct likelihood. The conditional mode and expected value of £ 1 are also calculated efficiently using their method. An interesting future research question is to compare the performance of their methods with that proposed here and with the use of full simulation of the exact likelihood based on the Markov chain Monte Carlo ideas described above.
