I am looking to model a poisson process $X \sim Poisson(\lambda)$ and infer $\lambda$ using bayesian inference from $X$ e.g. $[0,0,0,2,1,0,0,2,1]$ could be my dataset. The data comes from the count of something occurring in a day.

I am considering $Exponential(\alpha)$ as my prior as it can represent positive real numbers, and I am wondering how to pick $\alpha$. If it is not possible to do so in an uninformative way, I would also like to know if there are methods to sensibly pick $\alpha$ or possibly whether there is a more suitable prior distribution.

The output would be a distribution around $\lambda$

  • 1
    $\begingroup$ It is not possible to pick $\alpha$ with no prior information about the scale of the problem. $\endgroup$
    – Xi'an
    May 16 '17 at 10:46
  • $\begingroup$ Thank you Xi'an, that's good to know. Given this, is there a logical way to go about it's selection? $\endgroup$ May 16 '17 at 10:57
  • $\begingroup$ $\alpha$ is the inverse of the prior scale parameter for $\lambda$. Hence you need a prior opinion on the scale of $\lambda$. To know that this is a daily count does not help, as it could be a rare or a frequent event... $\endgroup$
    – Xi'an
    May 16 '17 at 16:47

The conjugate prior for a Poisson likelihood is the Gamma distribution, conveniently described with a shape parameter and a rate parameter (the inverse of a scale parameter).

Your choice of an exponential distribution would correspond to a prior shape parameter of $1$, at which point you would need to choose a prior rate parameter, inserting information into your prior.

An alternative approach might be to choose a prior shape parameter of $0$ and a prior rate parameter of $0$, which is improper and largely meaningless in itself but would have the result that after observing $x_1,x_2,\ldots,x_n$ your posterior distribution would have a shape parameter of $\sum\limits_{i=1}^n x_i$ and a rate parameter of $n$. Providing that at least one of your observations was non-zero, this could be a fairly sensible posterior distribution with mean the intuitive $\frac{\sum x_i}{n}$, and easy to work with and update further; you can see how little prior information is involved.

In your example, your Gamma posterior distribution for $\lambda$ would then have shape parameter $6$ and rate parameter $9$ under this alternative approach; compare this with starting with an exponential prior where the posterior would have shape parameter $7$ and rate parameter $9$ plus whatever rate you had chosen initially for your exponential.


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.