Dear Dynare users, this is my first post on this forum. I would like to thank you all for the help you give in the use of this software.
I have a common problem with bayesian estimation, mode_compute=4 does not work and gives me the classic error
??? Error using ==> chol
Matrix must be positive definite.
Most of you suggest of using the calibrated initial values for the estimation or to run an other routine like mode_compute=6. In my experience the only mode compute that gives proper convergence is 4 (also 10). Since I have no idea of what means “calibrated initial values” or at least I am not sure of it, i would like to ask you if would make sense to run the mode_compute=6 (that works always) and use the posterior mode or the final posterior mean as initial values for a new estimation with mode_compute=4.
I read somewhere to use a “combination of the two mode_compute” but i don’t know if you mean this.
Sometimes I do this (estimate with mode compute 6 ) and put the results as the inputs for another estimation, and I manage to get better results. But as an user I´m not sure if that is theoretically right.
I think you are referring to the estimated_params_init option, where you can put the calibration option. As far as I know it should be placed before the estimation command, before the priors.
I hope I have not confused you even more.
Yes i was referring to the part, actually i use to set this value after the estimated_params, I giess is the same:
x, init_value, upper bound,lower bound, prior shape,prior mean,...
Sequentially running different mode_compute makes a lot of sense. In case you did not find the true mode, you increase the likelihood of finding it. In case you already found it, running mode_compute a second time will just cost you time, but will not do any other harm as the optimizer will stay at the same mode you found before.
Using the calibrated value as starting values is recommended in case your mode_computation fails because of invalid starting values. If that is no problem, you don’t need to use it. Using the full estimated_params-block with initial value, upper and lower bound makes the estimated_params_init-redundant.
Thank you for the kind reply.
By “sequentially using” do you mean taking the posterior mode values from a computation and use it as initial parameter values for an other? Or there is an automatic way of running sequential routines?
By calibrated values do you mean posterior mean values after the metropolis hastings?
Sorry for the trivial questions.
mode_file option of estimation to load the results from the previous mode-computation. That part you have to do manually.
With calibrated values I mean the values set before the model block if you would run stoch_simul. That is, the values before you even would run estimation with mode-computation.