Question of Bayesian Estimation of Impulse Response Function

[dennis0125hk]

Before the Estimation of Impulse Response and Variance Decomposition , the number of times that drawn from the approximated joint distribution of parameters are initialized. I used Dynare 3 for Bayesian Estimation of Impulse Response and Variance Decomposition. I’ve tried to debug the code to find out how many times is assigned for drawing from the joint distribution to get the estimation result. I discover the number of times is of maximum 500 times, as shown in line 1444 in metropolis.m.

I have a question why maximum of 500 times is assigned to the number of draws from the approximated joint distribution? My concern is that if the approximated distribution is obtained by MH algorithm for many times, say 1000000 times, the chance of all draws are in the tail of distribution is high. If all the draws from approximated distribution are from the tail, the estimation result is no longer accurate.

[StephaneAdjemian]

Hi, We limit the number of simulations to 500 because this step can be very long... But you are right, this should be an option. We will fix this.

The 500 draws are taken randomly (uniformly) in the, say, 1000000 simulations of the MH. So I do think that the risk of having all the draws in the tails of the posterior distriution is low. Our experience on medium scaled dsge models is that if we take 5000 draws instead of 500 we get the same posterior IRFs distributions.

Best, Stéphane.