Skip to main content
29 events
when toggle format what by license comment
May 13, 2020 at 8:15 comment added Sextus Empiricus What I mean is that conjugate prior relates to this relationship between the pdf <-> likelihood (and related to that you refer to Bayes, which is not very clear and a bit handwaving). ---- The issue that is in this question is the relationship between cumulative distribution functions. You mention a relationship between the beta distribution and the binomial distribution, but that seems to me to be a different relationship. It is not clear what this has to do with the relationship between the CDF's.
May 13, 2020 at 8:08 comment added Amit Portnoy I don't know the actual historical progress, but I've seen it presented as starting from the likelihood and finding a constant that makes it into a PDF. The fact that it happens to be the conjugate prior may have been just a nice coincidence (or intuition on Bayes part).
May 13, 2020 at 8:00 comment added Sextus Empiricus I believe that the beta distribution is foremost the conjugate prior distribution of the binomial distribution. The connection here is the relation $$cdfBeta(I, N-I+1, F) + cdfBinomial(N, F, I-1) = 1$$ which is a different thing then the relationship as conjugate prior.
May 13, 2020 at 7:49 comment added Amit Portnoy Regarding "beta distribution is actually similar to the likelihood for p". I think this is how beta was viewed by Bayes. Start with the binomial likelihood for p, and find the constant that makes the integration works.
May 13, 2020 at 7:42 comment added Amit Portnoy I don't know, but I think it may have something to do with the support size of the Binomial. In any case, I'm interested to know if there is any known symmetry principle. $1/(n+1)$ comes up analytically from integrating the likelihood of p.
May 13, 2020 at 7:34 comment added Sextus Empiricus I am saying that it doesn't work for other cases, but I would actually have to check that out. Maybe there is some symmetry principle.
May 13, 2020 at 7:26 comment added Sextus Empiricus I see now. The first version made me confused. What you seem to be saying now is that the beta distribution is actually similar as the likelihood for $p$ with the binomial distribution. But why is that intuitive? It is not clear why it is like that. We can not do the same trick with other distributions so why does it work here while it doesn't for other cases? Is there some special property that makes it work, or is it just a coincidence? And why did you choose $n+1$? It is a parameter that normalizes the likelihood function, but is that the reason, or is there some more/other logic behind.
May 13, 2020 at 6:45 history edited Amit Portnoy CC BY-SA 4.0
added 68 characters in body
May 12, 2020 at 23:40 comment added Amit Portnoy Let us continue this discussion in chat.
May 12, 2020 at 23:40 history edited Amit Portnoy CC BY-SA 4.0
added 121 characters in body
May 12, 2020 at 23:10 comment added Amit Portnoy You're right. I edited my answer.
May 12, 2020 at 23:05 history edited Amit Portnoy CC BY-SA 4.0
deleted 9 characters in body
May 12, 2020 at 22:45 history edited Amit Portnoy CC BY-SA 4.0
deleted 9 characters in body
May 12, 2020 at 22:37 history edited Amit Portnoy CC BY-SA 4.0
deleted 225 characters in body
May 12, 2020 at 20:27 comment added Sextus Empiricus So it is a likelihood and can be above 1. But still where does the multiplication with a factor n+1 come from. Why is it not something like $$1-\left(1- {n\choose k} \pmb{p}^k (1-\pmb{p})^{n-k}\right)^{n+1}$$ according to $$p(\text{at least one event in k trials}) = 1 -(1 - p(\text{one event in one trial}))^k$$
May 12, 2020 at 19:49 history edited Amit Portnoy CC BY-SA 4.0
added 13 characters in body
May 12, 2020 at 19:43 comment added Amit Portnoy Yes, it's a PDF not a PMF. I'll rephrase.
May 12, 2020 at 19:21 comment added Sextus Empiricus "The chance of at least one of the experiments having exactly k successes is" I have difficulties to imagine how that equals $n+1$ time the probability of having exactly k success in a single run of the experiment. For instance when p=0.5, n=2 and k=1 then the probability equals 1.5 which makes no sense.
May 12, 2020 at 18:31 history edited Amit Portnoy CC BY-SA 4.0
added 99 characters in body
May 12, 2020 at 18:13 history edited Amit Portnoy CC BY-SA 4.0
added 44 characters in body
May 12, 2020 at 18:11 comment added Amit Portnoy This is supposed to be an intuitive explanation. Nothing more. (I did replace 1 / BetaFunc with the (N+1) * binomial coefficient, is this the issue?)
May 12, 2020 at 16:32 comment added Amit Portnoy @SextusEmpiricus I think Thomas Bayes derived it :). See youtu.be/UZjlBQbV1KU?t=1074 (though his story is different)
May 12, 2020 at 16:17 comment added Sextus Empiricus $$(n+1){n\choose k} \pmb{p}^k (1-\pmb{p})^{n-k}$$how did you derive this?
May 12, 2020 at 15:57 history edited Amit Portnoy CC BY-SA 4.0
added 312 characters in body
May 12, 2020 at 15:23 history edited Amit Portnoy CC BY-SA 4.0
edited body
May 11, 2020 at 9:46 history edited Amit Portnoy CC BY-SA 4.0
added 52 characters in body
May 11, 2020 at 9:25 history edited Amit Portnoy CC BY-SA 4.0
added 5 characters in body
May 11, 2020 at 9:14 history edited Amit Portnoy CC BY-SA 4.0
added 32 characters in body
May 11, 2020 at 9:08 history answered Amit Portnoy CC BY-SA 4.0