I am running a fully Bayesian MCMC procedure to estimate some time series models, and my model has a lot of parameter estimates. In particular, one of these parameters, $\phi$, is $\in [-1,1]$. The conditional density is nonstandard, so I have implemented an independance-chain Metropolis-Hastings with a truncated normal $N(\hat{\phi},\sigma^{2}_{\phi})1(|\phi|<1)$ where the mean and variance are based on some other parameters. I was a little unsure how reliable my model was, so I have run some simulated data through it where I have fixed $\phi=0.50$.

I have a simulated sample size of $500$ a loop size of $20,000$, and a burn-in period of $5,000$ and a computation time of around 60 minutes (and yes my matrices are sparse).

Surprisingly (or maybe not), the posterior is skewed to the left and the posterior mean/median are well below $0.50$ (typically around $0.4$). Obviously, when the sampler gets close to one, it gets pushed to the left and can wander off to the left near negative values. I do not believe this is "bad", but the mean and median won't do a good job here in summarising $\hat{\phi}$. The mode looks like a much more sensible estimate, but the creditability intervals are wide (too wide).

Wikipedia thinks using the posterior mode is a bad idea: http://en.wikipedia.org/wiki/Maximum_a_posteriori_estimation

Is there a usual treatment for summarising the posterior when dealing with a constrained parameter space?

  • 1
    $\begingroup$ I don't understand the question. Are you asking what is the best way to summarize a posterior distribution using a single value? Why do you want to do that in the first place? $\endgroup$ – Tom Minka Nov 28 '14 at 15:51
  • $\begingroup$ Please check the edit. I need to summarise it as I am conducting a simulation study to validate the authenticity of my model $\endgroup$ – akkp Nov 28 '14 at 21:36
  • $\begingroup$ How do you know that the mean won't do a good job? $\endgroup$ – Tom Minka Nov 28 '14 at 22:48
  • $\begingroup$ I've done that (as I wrote above). The posterior distributions are skewed to the left - apologies if this was not clear from the question $\endgroup$ – akkp Nov 29 '14 at 13:16
  • $\begingroup$ The posterior median has the property that it is equally likely to be above or below the true $\phi$, when $\phi$ is sampled from the prior. It sounds like you want this property to be true for a given fixed value of $\phi$. However, that is a frequentist property, not a Bayesian one. $\endgroup$ – Tom Minka Nov 30 '14 at 22:28

To increase accuracy near the boundaries of the parameter space, a common approach is to make a variable transformation $\theta = f(\phi)$, where $f$ maps $[-1,1]$ into $[-\infty,\infty]$, compute the posterior mean of $\theta$, then transform this estimate back to $\phi$. This is equivalent to using Bayesian decision theory where the loss function emphasizes errors near the boundary of the original parameter space.

  • $\begingroup$ Thanks. This makes sense. I suppose that a suitable transform would be $\frac{e^{x}}{1-e^{x}}$. How would I go about actually doing this? How does the Jacobian come into play here? Suppose my posterior was normal. How does this change the implementation? Thanks Tom! $\endgroup$ – akkp Dec 3 '14 at 11:37

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.