Convergence and Adaptive phases

Estimation
1

Dear all,

I am attempting to improve my Bayesian estimation of a large-scale multisectoral model. The idea is to implement an adaptive phases methodology to improve the starting point of the MCMC process.

I need to verify whether there was a draw from a higher mode and update the calibration of the variance-covariance matrix of the jump distribution. If I use mode_compute=6 initially and then proceed with

mode_compute = 0, mode_file = FILE, and load_mh_file,

it appears that I will simply continue the same chain. However, according to the algorithm, I need to change the initial draw point and the var-cov matrix. Am I missing something? And if that’s the case, is there any way to implement this directly in Dynare?

Thanks!

2

In that case, you don’t want to use load_mh_file. Rather, you would provide the initial draw and covariance matrix within the mode-file you load.

1 Like
3

Thanks for your reply professor! If you could kindly help me with two another questions:

  1. If I only provide the mode_file, is that sufficient to carry out the procedure? Since in the second estimation, I will be using mode_compute = 0, I haven’t found a way to provide the updated matrix to initialize the estimation; It seemed possible only if I estimate with mode_compute = 6 because of the 'InitialCovarianceMatrix' option.

  2. Would it be better to maximize the mode in the subsequent estimations as well?

4
  1. The mode-file contains a vector xparam1 with the parameter values and a hh matrix that is the covariance. That should allow you to provide all your need.
  2. Often, it does not hurt to do an additional optimization step. But it’s not necessary.
1 Like