Navigation:  FAQ >

 

What should I do when posterior mode estimation does not converge?

 

Print this Topic Previous pageReturn to chapter overviewNext page

There are a number of tools in YADA that affect the behavior of posterior mode estimation of the DSGE model parameters.

Change the DSGE model solver. AiM is fast but may have problems when it needs to use numeric shiftrights instead of exact shiftrights. The QZ decomposition based algorithms of Klein (2000) and Sims (2002) do not suffer from this specific issue; see the DSGE Model Data frame on the DSGE Data tab for details.
Change the Kalman algorithm. The standard Kalman filter is quite fast, but may be affected by numerical errors due to covariance matrices not being positive semidefinite. The square root filter is slower but guarantees that the covariance matrices are all numerically positive semidefinite; see the Kalman Filter Selections frame on the Settings tab for details.
Change the tolerance level or the method for initializing the inverse Hessian (csminwel). The options for these properties are located in the Optimization frame on the Options tab.
Change the initial values of the estimated parameters. If you have problems determining initial values from which, e.g., csminwel converges to an optimum, you may first want to try the gmhmaxlik "optimization" routine. The choice of optimizer and parameters to optimize with respect to are located in the Optimization frame on the Options tab. Once you have optained parameter values from this routine, you may try them as initial values for the csminwel optimizer.
Check for identification issues through the Information matrix functions.

You may also consider rewriting the DSGE model equations in the AiM or the dynare model file. The numerical properties of model solvers are likely to improve when you avoid redundant equations and simplify parametric functions by declaring a new parameter for such expressions. The new parameter can/should be defined in the file with parameters to update (or initialize).

It is also possible to affect the behavior of the model solvers by changing the numerical tolerance. This option is available on the Kalman Filter Selections frame on the Settings tab for details. However, this option should generally be avoided, but may be tested when all else has failed.

 

 


Page url: http://www.texlips.net/yada/help/index.html?what_should_i_do_when_posterio.htm