General dynare questions - non linear new keynesian model

Estimation ML/Bayesian estimation
1

Hello,

I hope this message finds you well. For my master thesis, I am estimating 3 versions of a non-linear new keynesian model, a baseline, one with external habit formation, and one with liquidity constrained consumers. it’s not log linearized and i specified the steady state. I am quite new to this so I have some general questions:

  1. should the acceptance ratio for the MH algorithm be more or less the same across all models? or is it ok to just have the acceptance ratio in the range 0.2 - 0.4? or what should be the recommended range?
  2. what is the exact role of the parameter calibration at the beginning, since in the estim block i suggest a prior distribution? Are the calibrated parameter values just taken form the literature? Should the calibrated value at the beginning and the initial value in the estim_parameter block be aligned?
  3. How should i choose the starting value in the “estimated_params;” block?
    For example: theta, 0.75, 0.01, 0.999, BETA_PDF, 0.6, 0.2; why would i specifically choose 0.75 as starting value? could it be the posterior mean of an estimated paper i take as a reference?
  4. after i run the code, is it possible to change the suggested initial value for a parameter after looking at the data? I guess it’s a wrong move right?

Thank you very much

2
  1. The exact acceptance rate and number of draws is not important as long as there is convergence. Aiming for 0.25 is often a good idea.
  2. The initial value serves as the starting point of the Markov Chain during estimation. If the chain properly converged, it will not matter. But starting at a likely parameter value improves efficiency. Hence, you should pick a good value, which is often the calibrated one.
  3. Yes, you could take the estimated posterior means as the starting values as opposed to values you consider likely.
  4. See point 2: the starting value should not matter asymptotically. However, the chain may get stuck at a local mode. Trying different starting values is therefore often a good idea. What would not be OK is modifying the actual prior distribution after seeing the data.
1 Like
3

Thank you very much for your reply.

I wuould have an additional question, You see that the bar for parameter eta is going down instead of going up. Is this a big issue? What does it mean? Thank you

figure_33
figure_33875×656 23.3 KB

4

No, that is not problematic. It only means the logarithm of an identification strength smaller than 1 was taken.

1 Like