Absurd results with Bayesian estimation
Posted: Wed Jul 23, 2014 12:37 pmDear Dynare users,
I am trying to estimate a model with procyclical leverage in the banking sector. The Bayesian estimation completes successfully, however, I get absurd estimates for the shock parameters (some standard errors have extremely high values) and the following warning:
Warning: Matrix is singular to working precision. Results may be inaccurate.
I believe the optimization routine is unable to find the right mode and this biases the results. I have tried all optimizers, especially mode_compute=6 and mode_compute=9, which I have also used successively (loading previous mode file). Please note that I only estimate the parameters that are well identified (using the identification command). I have checked my observation equations and shut off prior trunc. I have also tried the following:
- Considerably increase the informativeness of the priors (up to incredibly tight priors);
- With mode_compute=6, increase the number of MCMC run sequentially, the number of iterations in each MCMC routine, and the maximum number of iterations used for adjusting the scale parameter of the jumping distribution;
- Add and remove certain observable variables;
- Add and remove measurement errors for certain observable variables (always avoiding stochastic singularity);
- Estimate only the shock parameters and calibrate the rest;
None of these actions allow me to get rid of the warning message. Any help would be much appreciated.
I am trying to estimate a model with procyclical leverage in the banking sector. The Bayesian estimation completes successfully, however, I get absurd estimates for the shock parameters (some standard errors have extremely high values) and the following warning:
Warning: Matrix is singular to working precision. Results may be inaccurate.
I believe the optimization routine is unable to find the right mode and this biases the results. I have tried all optimizers, especially mode_compute=6 and mode_compute=9, which I have also used successively (loading previous mode file). Please note that I only estimate the parameters that are well identified (using the identification command). I have checked my observation equations and shut off prior trunc. I have also tried the following:
- Considerably increase the informativeness of the priors (up to incredibly tight priors);
- With mode_compute=6, increase the number of MCMC run sequentially, the number of iterations in each MCMC routine, and the maximum number of iterations used for adjusting the scale parameter of the jumping distribution;
- Add and remove certain observable variables;
- Add and remove measurement errors for certain observable variables (always avoiding stochastic singularity);
- Estimate only the shock parameters and calibrate the rest;
None of these actions allow me to get rid of the warning message. Any help would be much appreciated.