Through a MCMC Gibbs sampler I obtain $M$ chains of the parameters vector $\mathbf{\theta}$, meaning that each component of $\mathbf{\theta}$ is the value of one parameter at a given iteration.

Since I know the likelihood and the prior I found the maximum a posteriori estimate:


I'd like to have an error on each component of $\hat{\mathbf{\theta}}$.

One other estimate of the best parameters is the mode or the mean of MCMC chains for each parameter, for which I can consider the standard deviation as an error, but this is not what I'm interested in.


Since we are treating $\theta$ as vector of random variables i think the term error is thematically wrong, indeed one of the strength of Bayesian methods is the richness of information granted about $\theta$ through it's posterior distribution. For MCMC sampling methods (of which the Gibbs sampler is one kind) the posterior distribution of paramters $\theta$ can generally be obtained in this manner:

  1. Let the chains converge to stationary (if it's possible for the problem) - generally known as the burn-in period.
  2. Keep every x sample of $\theta$. Since the chain is correlated in time we only keep every xth value, this concept is known as thinning. The value of x is generally a trade-off between time and accuracy.
  3. The kept samples of $\theta$ describes it's posterior distribution for which any moment can be calculated if so desired.
  • $\begingroup$ Thank you for your answer, kept samples of θ they are the marginalized posterior for each parameter. With maximum a posteriori estimate, from those generated, I take the set of parameter more likely given the model and data. No moment of its marginalized posterior is reasonable as error of one of these parameters most probable $\endgroup$ – Johnpiton Sep 27 '19 at 18:44

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.