I updated to Dynare 4.4 but apparently there is something strange going on. I estimated exactly the same model and using the same data with the new version, but I get parameter estimates that differs a lot with the ones I get when I estimated with the old version. I understand that they shouldn´t be exactly the same (I´m using mode_compute = 6), but for example, a parameter related with an elasticity I previoulsy obtained an estimate much lower than 1, and now I get a number quite above 1. Is there any reason for this to happen ?
I would like to add that I constantly get the following error, whenever I change the estimation command. To make it clear, I am trying to estimate again my model, but with one less observable, but apparentely dynare stores the variable that I am trying to remove (prm_hat, in this case). I´m pretty sure that it is not a problem of the code.
Thanks,
Error using eval
Undefined function or variable ‘prm_hat’.
Error in read_variables (line 74)
dyn_tmp_01 = eval(var_names_01(dyn_i_01,:));
Error in initialize_dataset (line 32)
rawdata = read_variables(datafile,varobs,],xls.sheet,xls.range);
Error in dynare_estimation_init (line 477)
dataset_ =
initialize_dataset(options_.datafile,options_.varobs,options_.first_obs,options_.nobs,transformation,options_.prefilter,xls);
Error in dynare_estimation_1 (line 81)
[dataset_,xparam1, hh, M_, options_, oo_, estim_params_,bayestopt_] =
dynare_estimation_init(var_list_, dname, ], M_, options_, oo_, estim_params_,
bayestopt_);
Error in dynare_estimation (line 84)
dynare_estimation_1(var_list,dname);
Error in oil_estimado (line 1104)
dynare_estimation(var_list_);
Error in dynare (line 162)
evalin(‘base’,fname) ;
Hi jpfeifer ,
These are the files. I get very different results depending on which version I use to estimate, particularly the elasticities parameters. The model is Soto and Medina (07), from the database.
Thanks in advance. oil_data2.m (3.4 KB) oil.xls (29.5 KB) oil_estimado.mod (31.8 KB)
Sorry for the persistence ,but any help here ? I managed to find what I was doing wrong with the data, but I´m still struggling to find out why the parameters estimates differ so much when I estimate the model using the new version compared with the old one. Another doubt I have is about the endogenous_prior option. When I use this option in the estimation command I need to put all the priors for the parameters as in a usual bayesian estimation?
Yes, for the endogenous prior you need to specify a full prior, because the data moments are effectively used as a pre-sample to update the specified prior. This updating requires having a prior in the first place. What is your doubt about this?
I am looking into the issue, but haven’t found an answer yet. The initial likelihood is slighty different in both version, but that seems to be numerical error. I will keep you updated.
Looking into the issue, the reason seems be that the mode found by the new version has a likelihood that is higher by more than 100 log points. This suggests that your old results were wrong as they were not at the true mode and your MCMC had not converged yet.
Regarding the endogenous prior command, I tried it once, and I got an error, I will do some additional experiments to see what kind of message I get again. About the different parameter estimates, I did notice that with the new version the likelihood was higher, but I wasn´t expecting to see huge differences with the estimates, because when I estimated with the old version I got somehow similar estimates, despite the changes in the likelihood. Thank you again.