Guidance for Bayesian Estimation

Estimation ML/Bayesian estimation
1

Hi all,
does dynare provides any guidance video and codes for Bayesian estimation or MLE? The summer school 2021 video does not have many example. No further guidance for the RBC_tax code or fs2000 code? What about Bayesian replication code for An and Schorfheide (2007) paper?

Moreover, for Bayesian estimate, do we have requirement for the number of observed data (varobs)? Like SMM, we should have moments no fewer than estimated parameters. Any requirement for Bayesian? And suppose the exogenous shock variable is government spending, then must I include that data?

2
  1. No, we don’t provide much more material online beyond what you saw at Dynare Summer School 2021 · Wiki · Dynare / dynare · GitLab and e.g. DSGE_mod/RBC_baseline at master · JohannesPfeifer/DSGE_mod · GitHub
  2. The fs2000.mod at examples/fs2000.mod · master · Dynare / dynare · GitLab is taken directly from the cited paper.
  3. No, we don’t replicate the An/Schorfheide paper, but there is a mod-file at tests/identification/as2007/as2007.mod · master · Dynare / dynare · GitLab
  4. For Bayesian estimation, you need at least as many shocks as observables. There are not a lot more restrictions.
  5. No, you data should ideally be informative about what you are trying to study. But most data series should give you identification, even if not directly observing some important data.
3

Thank you professor, for the point 4, is that means if we have one shock in the model, then we just need at least one observed data?

4

No, if you have one shock in the model, you can only have one observable.

5

Hi Professor, but you said I need “at least” as many shocks as observables? In the RBC_tax question you used in the 2021 summer school, the model has 4 shocks but you include 5 observed variables.

In addition, I think that the data in RBC_tax case is also logged data, but why don’t you use the logdata option (“loglinear” and “logdata” in estimation command) but you did in the fs2000?

Meanwhile, it seems that the RBC_tax_estimation code use “moment_varendo” instead of “bayesian_irf” in default. Are there any difference?

Thank you again!

6
  1. That file also has measurement error on output, which is the fifth shock.
  2. The loglinear and logdata options are to automate things that may you have to otherwise do manually. In the present example, I implemented something similar manually.
  3. moment_varendo is for moments, bayesian_irf is for IRFs. Those are different. See the manual.
9

@jpfeifer Hello professor, in my Bayesian estimation, since I only have one shock in my model so I can only have one data. I want to use the ex-post real interest rate as the input. For this case, lets denote nominal rate as R and inflation as Pi. In the model section, to build the variable, should I use R(-1)/Pi or R/Pi(+1) or they will make no difference?

10

The ex-post real interest rate would be R(-1)/Pi. R/Pi(+1) is the ex-ante one.

1 Like