Stochastic shocks in some periods- Iacoviello 2015

[jpfeifer]

Your mode_check-plots tend to look strange, suggesting you did not find the correct mode. There is no reason to use the diffuse filter in your case. There is no unit root. I ran the regular filter with mode_compute=9 and it returned

Code: Select all

Fval obtained by the minimization routine (minus the posterior/likelihood)): -339.518149

RESULTS FROM POSTERIOR ESTIMATION
parameters
           prior mean     mode    s.d. prior pstdev

psi_db         0.165   0.0447     NaN gamm  0.1250
psi_dh         0.120   0.0000     NaN gamm  0.1250
psi_k          1.166   1.1289     NaN gamm  0.5000
psi_m          0.173   0.8743     NaN gamm  0.1250
psi_a          0.396   0.2824     NaN gamm  0.1250
psi_h          0.892   0.3710     NaN gamm  0.5000
rho_b          0.219   0.1578     NaN beta  0.1000
rho_m          0.250   0.5014     NaN beta  0.1000
rho_TB         0.925   0.9890     NaN beta  0.1000

standard deviation of shocks
           prior mean     mode    s.d. prior pstdev

epsilon_TB_I   0.004   0.0017     NaN invg  0.0250


As you can see

Code: Select all

psi_dh

ran to 0, which may be due to the prior with an asymptote at 0.

[leonard007]

Thank you so much for running my codes.
Yes your right, I see the 0 mode for psi_dh that is problematic. That is quiet strange, Would you think there is a way to fix this issue?

[jpfeifer]

First of all, estimating such a model on only one observable series is a bit of a stretch. Maybe adding other observables will help. Regardless, you should aim at a prior with an interior peak instead of one with an asymptote. Play around with the variance of the prior.

[leonard007]

Dear Professor Jpfeifer,
Many Thanks for your great hint. To add some other observable series,(considering I have only one shock in my model), I used your very informative Guide line "A Guide to Specifying Observation Equations for the
Estimation of DSGE Models", and I used measurement errors to add 3 other observable series to my estimation. Then I tried to reduce the number of estimated parameters and add one by one (and run estimation after adding each parameter) to find which one causes the problem. I found the problem comes from "rho_b" (inertia for capital constraint) and "rho_m"(inertia for collateral constraint). Then as your comment, I changed their variances to see in which range, estimation works and Dynare can find the steady state and now the model works. I have 2 question:
1. In order to find the prior's mean and st error, At first, I used mode_compute=6 and I put the estimated mean and st error as the prior's mean and st error for my new estimation with mode_compute=4, Is this way ok or this causes problem for estimation?
2. after estimating model with mode_compute=4, I have 2 parameters that their posterior are very close to their priors, does it show something wrong? (Considering a line in the guideline which says if posterior is close to prior that shows a mistake)
I attache my .mod file here.
Sincerely,
Leo

[jpfeifer]

1. Your prior is your prior before seeing the data. What you describe is problematic, because it will result in wrong inference. You should simply choose a prior that looks sensible, i.e. has a single peak around e.g. 1 percent standard deviation and is somewhat diffuse.
2. That simply means your data is not very informative. If the identification command does not complain, this should be fine.

[leonard007]

Dear Professor Jpfeifer,
I am very grateful to you for your help. You are absolutely right. I changed the mean and st errors as your comment and I ran the estimation. Now ,I think, everything looks good.
Sincerely,
Leo