With the standard Binomial probability distribution we consider n trials each with a probability of success p. This can be somewhat generalized to the Beta-Binomial Distribution which is effectively the same, but this time the probability of success is a random variable with a Beta Distribution. My question is:

Could I in theory use a p which samples from some arbitrary probability distribution (continuous or discrete) so long as the support is on [0,1], or is there something special about the Beta distribution which allows for such a distribution?


The Beta is just conjugate to the Binomial which makes life easier in a number of ways. If we are in a Bayesian setting with a Binomial likelihood, there is nothing limiting you, theoretically, from applying any prior distribution on $p$ that you like. The problem is that you may end up with an improper posterior, that is, a posterior distribution that does not integrate to one. So if you choose to use such a prior you should check that your posterior is a proper density. But perhaps most importantly, you should be prepared to defend your use of such a prior in that it should make sense for your specific application.

| cite | improve this answer | |
  • $\begingroup$ My main concern is that my modified Binomial 'distribution' would not integrate to one with my choice of distribution for p. I guess what I was wondering is if there existed any work to prove whether or not any distribution could be used as the seed for the Binomial probability so long as this p had its support on [0,1] and was in fact a probability distribution. $\endgroup$ – mjnichol Jul 10 '14 at 19:39
  • $\begingroup$ mjnichol - perhaps you should clarify your question with this information; it sounds like the heart of your question is really just a fairly straightforward question about mixture distributions. $\endgroup$ – Glen_b Jul 10 '14 at 23:23
  • $\begingroup$ It would help to clarify what you mean by "could be used". And perhaps clarify any specific case that you may be interested in. $\endgroup$ – Zoë Clark Jul 11 '14 at 1:36
  • $\begingroup$ Okay. I have a situation where I run n experiments and so long as a single one of those experiments succeeds I am in the clear. That led me to the Binomial Distribution as a natural way to explore the problem; however, the probability of a success is not constant and is picked from a distribution I crafted from real data at my place of work. Is it okay to simply sample from this distribution in place of a constant p? Will this preserve the integrity of the modified binomial? If anything is unclear please let me know. Also thank you for all the help! $\endgroup$ – mjnichol Jul 11 '14 at 16:27

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.