Variance Decomposition

[jakegunner]

Dear all,

My question is regarding the unconditional variance decomposition post estimation. In the command window, dynare reports the following table:

VARIANCE DECOMPOSITION (in percent)

I am unsure as to whether I should use the values reported here or those reported in oo_.PosteriorTheoreticalMoments.dsge.VarianceDecomposition.hpdinf.<NAME_OF_THE_ENDOGENOUS_VARIABLE>.<NAME_OF_THE_EXOGENOUS_VARIABLE>. They are wildly different. Any help on the matter would be really appreciated.

Thanks

[jpfeifer]

Your description is insufficient to provide a conclusive answer. But is seems you are comparing apples and oranges.
The output of the variance decomposition comes from the calibrated model, I guess. The fields you describe seem to come from the model estimation. Those are two completely different things.
On top of that, the HPDI field refers to the upper bound of the credible set (Bayesian confidence interval). Of course the upper bound is different from the mean or median.

[jakegunner]

Many thanks Prof,

I was under the impression that dynare, in the command window, would report variance decomposition for those parameter values evaluated after estimation. However, it is indeed reported for those in the calibration block.
J

[costa]

I always thought that conditional_variance_decomposition after estimation would automatically considers the posteriors means obtained from estimation. Is this correct ?

[jpfeifer]

If you run stoch_simul after estimation, you will obtain the conditional_variance_decomposition at the posterior mean and the output will be displayed in the command window. But this post referred to someone using stoch_simul BEFORE estimation, while also requesting conditional_variance_decomposition in the estimation command. In this case, there will only be output for the conditional_variance_decomposition from the calibrated model. There will be no output in the command window from the estimation command for the conditional_variance_decomposition as this command, as documented in the the manual, only stores the results in oo_.

[sbehpour]

Hi Mr Pfeifer
I am writing a paper by using "dsge models" and I have a problem.
could you help me how can I do variance decomposition in dynare? I do not know what code is required and where I should put that..
thank you

[jpfeifer]

Are you doing estimation or simulations?

[sbehpour]

thank you Mr Pfeifer
I want to do both of them (estimation and simulation)..

[jpfeifer]

See the manual on

Code: Select all

conditional_variance_decomposition

For estimation, see also the manual on

Code: Select all

moments_varendo

[ZBCPA]

If you run stoch_simul after estimation, you will obtain the conditional_variance_decomposition at the posterior mean and the output will be displayed in the command window. But this post referred to someone using stoch_simul BEFORE estimation, while also requesting conditional_variance_decomposition in the estimation command. In this case, there will only be output for the conditional_variance_decomposition from the calibrated model. There will be no output in the command window from the estimation command for the conditional_variance_decomposition as this command, as documented in the the manual, only stores the results in oo_.

Dear Johhanes,

If I use such command:

Code: Select all

estimation(datafile=...,moments_varendo,....);
stoch_simul(order=1,irf=0);

Then I will get two types of unconditional variance decomposition results.
The first one displayed in the command window is the unconditional variance decomposition at the posterior mean ;

My quesiton is , how to explain the second type of unconditional variance decomposition stored in oo_.PosteriorTheoreticalMoments.dsge.VarianceDcomposition ? Even though the two types are very close in value.

In published paper, people generally report which type of unconditional variance decomposition?

Thanks in advance.

Best regards,
Huan

[jpfeifer]

The main difference is that the stoch_simul variance decomposition is computed at the calibrated parameter combination (here the posterior mean), while the Bayesian variance decomposition is the mean variance decomposition (not to be confused with the variance decomposition at the mean). That is, the latter is the average of the variance decomposition over the parameter draws while the former is the variance decomposition at the average over the parameters.

You can use both. In the context of Bayesian estimation, the ones from the estimation command are theoretically preferable, but most people use the ones from stoch_simul, because they are easier to interpret. For example, Christiano/Motto/Rostagno (2014) in their Risk Shocks paper in the AER do this.

[ZBCPA]

The main difference is that the stoch_simul variance decomposition is computed at the calibrated parameter combination (here the posterior mean), while the Bayesian variance decomposition is the mean variance decomposition (not to be confused with the variance decomposition at the mean). That is, the latter is the average of the variance decomposition over the parameter draws while the former is the variance decomposition at the average over the parameters.

You can use both. In the context of Bayesian estimation, the ones from the estimation command are theoretically preferable, but most people use the ones from stoch_simul, because they are easier to interpret. For example, Christiano/Motto/Rostagno (2014) in their Risk Shocks paper in the AER do this.

Dear Johhanes,

Thank you for your time and patience and professional clear explanation.

Best regards,
Huan

[HouseC]

Dear all,

Many thanks for your discussion. It's really helpful. My estimation exercise return significantly difference between variance decomposition using posterior mean and the mean variance decomposition. For instance, I want to look at TFP shock's explainary power on output and consumption. The bayesian estimation exercise reveals that this shock captures around 50% of output fluctuation. Yet when I check the variance decomposition for calibrated model using posterior mean. the explainary power is reduced to 10%. Is it possible to get such result? And how to understand ?

Best regard

[jpfeifer]

Are you using a different filter option?

[HouseC]

Are you using a different filter option?

Dear Johannes,

Thanks for your reply. My command in estimation is pretty standard and I guess I don't use another filter. The estimtaion command is the following

Code: Select all

estimation(datafile=xxxx,
        logdata,
        plot_priors=0,
        mode_check,
        mode_compute=6,
        optim=('AcceptanceRateTarget',0.25),
        moments_varendo,
        mh_nblocks=1,
        mh_replic=20000
               );