Dynare forums

Strange problem: Several programs I've written work for my co-author on his computer, but not on mine. In one case, parameter estimates don't converge when I run the identical program with identical data, while he gets convergence. As far as I can tell, the only difference is he's on a unix system and I'm in Windows. We both run the program from the folder that has the data file. In the attached case I get errors that prevent the program from running, specifically, this happens after the steady state is calculated, at the point where the data should load: Where my colleague gets

Loading 193 observations from autodata_ur.xls

Initial value of the log posterior (or likelihood): 573.5629


I get

Error using -
Matrix dimensions must agree.

Error in set_prior (line 270)
elseif any(bayestopt_.p5-old.bayestopt_.p5(:))

Error in dynare_estimation_init (line 123)
[xparam1,estim_params_,bayestopt_,lb,ub,M_] = set_prior(estim_params_,M_,options_);

Error in dynare_estimation_1 (line 59)
[dataset_,xparam1, M_, options_, oo_, estim_params_,bayestopt_] = dynare_estimation_init(var_list_, dname, [], M_, options_,
oo_, estim_params_, bayestopt_);

Error in dynare_estimation (line 70)
dynare_estimation_1(var_list,dname);

Error in invmod6ml_ur_base (line 226)
dynare_estimation(var_list_);

Error in dynare (line 120)
evalin('base',fname) ;

I've run programs like this before, so I'm really baffled.

The difference between Windows and Unix most probably has to do with different random numbers. Using a longer chain should still give you convergence. The second issue you report seems to come from a
previous

Code: Select all

definition.mat

file present in the prior subfolder, although I cannot trace why such an error might occur. Have you tried Dynare 4.4.3? In any case, you should be able to run the file in an empty folder.

Yes, both of these problems occurred with 4.4.3. I will try your solutions, thanks for responding.

You cannot have used Dynare 4.4.3, because there line 270 of set_prior.m is different.

I should have said they *also* occurred with 4.4.3. I had originally run them on 4.4.3, but when they didn't run on my computer, but did on my colleagues (who was running an earlier version), I then ran them on the earlier version, thinking maybe it was a problem with 4.4.3.

In any case, I was able to get the program that I had attached to run properly from an empty folder. As for the other problem with parameter estimates not converging: I should have mentioned that I was just using ML estimation, not Bayesian. Does that even use random number generation? And if so, is the chain something set in Dynare or in Matlab? Thanks again, and I was traveling for a week hence my delayed response.

The chain comment refers to Bayesian. If you do ML, the answer depends on the mode-finder. If you use

Code: Select all

mode_compute=4

that randomly perturbs the search direction, which can explain the difference (similarly for 6 and 9).