identification analysis incongruency

[francis]

Hello,

I have a question concerning the output of the identification command.

When I run the identification analysis on my model it tells me that all parameters are identified by the moments and by the model (H=J(q) and J=J_2 are full rank). However, in the identification strength plot no bar is shown for several parameters (their sensitivity component is positive and large). By inspecting the Dynare scripts behind the identification command I could find that the reason is that some columns of the analytic hessian are found to be linearly dependent and that the identification strength for the parameters for which no bar is shown are stored as NaNs.

After reading some papers of reference mentioned in the documentation by Marco Ratto, I learned that a full rank information matrix is a condition for local identification and that this condition can be checked by inspecting the rank of the matrix J(q), the Jacobian of the moments w.r.t. model parameters. In this case, however, checking the rank of J(q) does not return the same answer as checking the rank of the hessian (and, thus, of the information matrix).

Ultimately, I do not understand whether the parameters in my model for which no bar is shown are not strictly identified or whether they suffer from weak identification. Given the evidence presented above it seems to be the case that they are not strictly identifiable but I thought that the Jacobian matrices would tell me that.

Many thanks in advance,
Francis

[jpfeifer]

Hi,
could you please try the current unstable version and report back. If the problem persists, please provide me with the files.

[francis]

Thank you very much for your prompt reply Mr. Pfeifer.

Using the unstable version I do not run into incongruencies anymore.

May I ask you what has changed in the identification analysis? Was there a bug?

Best,
Francis

[francis]

Dear Mr. Pfeifer,

using a different parameterization of the model the problem is still there. There are more parameters for which bars are not shown than parameters for which identification is reported to fail according to the analysis of the Jacobians.

Thank you,
Francis

[jpfeifer]

Without the mod-file it is hard to tell. There have been recent changes related to the treatment of unit roots. Moreover, identification is a local phenomenon. Are you sure that you are comparing the right results (i.e. figures and tables at the prior mean as opposed to one at the prior mean and the other from MC sampling)?

[jpfeifer]

The issue should be solved in tomorrow's unstable version. The reason was https://github.com/DynareTeam/dynare/pull/1238