About Variance Decomposition: Some shocks have zero impacts

Stochastic simulations
#1

Dear Colleagues

I am Ha

This questions is not related to Dynare Code but how the results make sense.
I carefully checked the code and the mathematical derivations and nothing goes wrong.

However when I run the Bayesian Estimation to obtain the Variance Decomposition, the results come out quite weird since the cost-push shock (eps_theta) and the trend inflation shock (eps_Pibar) have zero impact on all variables. I checked parameters relating to these shock to make sure that these shocks will play some roles in the model but the results remain the same.

Please take a look at my code and let me know how to fix the problem.

Another question is relating to the data used in the model. I used the non-linear model and the raw data. For example, regarding inflation, I used gdp delator , thus the gross inflation will be (gdp deflator (t)/ gdp deflator (t-1) and use this as the data estimation. ex, if gdp deflator (t+1)= 108.606; gdp deflator (t)=108.009, the gross inflation is 108.99/108.606=1.0055 and use it as the data for inflation. For the non-linear model, the way I did is right?

This is my codeModel_Uncertainty_Est_Base.mod (11.5 KB)
here is datadata_est.m (18.5 KB)

Any comment and suggestion are greatly appreciated

Thank you so much

#2

Already in your calibrated model, all shocks except for one play hardly any role. Thus. you need to check the intuition behind that result. Maybe there is still a mistake how the shocks enter the model. Also, check your prior for the shocks. The initial value of the likelihood is too big, suggesting there is still a problem with the initial values.

Regarding data treatment, it seems correct.

#3

Dear Professor

Thank you so much for your promp reply.

Regarding initial value of posterior that suggest that the prior might be incorrect. Could you suggest me some way to find out which one that is incorrect? For example, looking the mode_check or prior plots, we can have some evidence to make conclusion on that prior distribution?

If we can, please suggest me some others ways to solve it? Like using other distributions?

Thank you Professors so so much

#4

You should try something like a prior predictive. If you calibrate the model to your prior mean, do the IRFs and second moments make sense? If not, you might want to modify your prior.