Variance Decomposition

[jpfeifer]

That's fine, but if you specify a filter in a subsequent stoch_simul-command, that would explain the differences.

[HouseC]

That's fine, but if you specify a filter in a subsequent stoch_simul-command, that would explain the differences.

Thanks Johannes.
But I use the following default setting.

Code: Select all

stoch_simul(loglinear,order=1,irf=0)

In this case, I do not call HP filter. is it possible to explained the difference? In my previous example, I says that the estimation assign large weight on productivity shock while the simulation returns smaller weight. When I plot the IRF for this productivity shock and other shocks using posterior mean/median. It seems that variance decomposition from simulated version is more reasonable and easier to interpret. I just try to figure out what can result in such inconsistent result.

[jpfeifer]

It seems you are using

Code: Select all

loglinear

only for

Code: Select all

stoch_simul

That might explain the difference.

[HouseC]

It seems you are using

Code: Select all

loglinear

only for

Code: Select all

stoch_simul

That might explain the difference.

Dear Johannes,

I'm a bit confused. I guess the simulation command and estimation command is similar with your replication on Garcia-Cicco, Pancrazi and Uribe (2010). I put my model into the dynare similar to you. Where I speficy the original model in "model" block and variables are the original variable. In the simulation part, I put the 'loglinear' option so that the simulation will be done for loglinearized model. In the estimation block, I specify the 'logdata' because the input data is logged. All intereted variables are log-diffrence variables (the growth rate), same as the GPU paper.
Could you elaborate more about the difference?
By checking the variable decomposition from stoch_simul command in your shared file , the results are pretty closed to the Table 5 in the paper. Do it mean that the paper report the variance decomposition based on the simulation?

[jpfeifer]

In my mod-file, there is a

Code: Select all

stoch_simul(loglinear)

before

Code: Select all

estimation

This implies that the loglinear option is already set when estimation is run. If you don't do this, the estimation will not be using the

Code: Select all

loglinear

option. I updated the Garcia-Cicco, Pancrazi and Uribe (2010) mod-file to make explicit that you would need to set that option explicitly.

[HouseC]

In my mod-file, there is a

Code: Select all

stoch_simul(loglinear)

before

Code: Select all

estimation

This implies that the loglinear option is already set when estimation is run. If you don't do this, the estimation will not be using the

Code: Select all

loglinear

option. I updated the Garcia-Cicco, Pancrazi and Uribe (2010) mod-file to make explicit that you would need to set that option explicitly.

Thanks, Johannes.
I guess I try to clarify that I follow the estimation and simulation exactly in your way. After following this estimation command (simulation with log linear is specified), I got the variance decomposition from this estimation. After the estimation, I also put the poseterior mean into simulation with log linear command and find that two variance decomposition are pretty inconsistent. Again,my question is can this inconsistency be explained?

[jpfeifer]

Does your posterior look approximately normal? Because estimation is going to provide the mean decomposition, not the decomposition at the mean. Those two objects can be very different if the distribution looks funny.

[HouseC]

Does your posterior look approximately normal? Because estimation is going to provide the mean decomposition, not the decomposition at the mean. Those two objects can be very different if the distribution looks funny.

Thanks Johannes. I understand your suggestion!