Model unable to output any successful draws in Gibbs sampling

Question

Alex 2024-8-4

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/2142786-model-unable-to-output-any-successful-draws-in-gibbs-sampling

回答： UDAYA PEDDIRAJU 2024-8-23

I am trying to estimate a structural VAR model using traditional sign and zero restrictions for the identification of shocks and Minnesota priors. When I adapt the (working) code to run on my data, I am unable to generate any successful draws in the Gibbs sampling phase. I have tried playing around with the settings of the Minnesota prior, as well as number of observations, number of variables, number of restrictions, running the model in First-Differences, and still can't get any successful draws. Does anyone have any ideas as to why this might be happening or anything I can try next?

Thanks in advance,

Alex

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

dpb 2024-8-4

Without a clue as to the data nor even which specific functions, no, not really...

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

UDAYA PEDDIRAJU 2024-8-23

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/2142786-model-unable-to-output-any-successful-draws-in-gibbs-sampling#answer_1503909

Hi Alex,

I understand you're facing issues with Gibbs sampling in your structural VAR model, here are some common reasons for this problem and potential remedies:

1. Prior Misspecification:

Overly tight priors: Relax the priors to allow for more variability in the posterior.
Incorrect prior form: Ensure the prior is conjugate to the likelihood for efficient sampling.

2. Data Issues:

Non-stationarity: Transform the data (e.g., first differencing) to achieve stationarity.
Outliers or missing values: Clean the data to remove outliers or handle missing values appropriately.

3. Model Misspecification:

Overly complex model: Simplify the model if it's too large or includes unnecessary variables.
Identification issues: Double-check your identification strategy and ensure it's appropriate for the model.

4. Gibbs Sampler Settings:

Insufficient iterations: Increase the number of iterations to allow the sampler to converge.
Burn-in period: Ensure a sufficient burn-in period to discard initial draws that may not represent the posterior distribution.

5. Other Considerations:

Initialization: Try different starting values for the Gibbs sampler.
Alternative samplers: Consider using other sampling methods like Metropolis-Hastings or Hamiltonian Monte Carlo.

Documentation reference: https://www.mathworks.com/help/econ/introduction-to-vector-autoregressive-var-models.html