I have a random events generator. I know in advance the set of event that can be generated (in my case I have only three possible events). The probabilities of the events are not known. I need to estimate these probabilities. For that I run an experiment. For example I generate 20 events. So, I have a sequence of events. For example: a,a,b,a,c,c,c,a,b,....c. Having the sequence I can count for the number of every event (so, in my case I get three integer numbers $n_1$, $n_2$ and $n_3$).

I can calculate the probabilities of every event in the following way:

$\nu_1 = \frac{n_1}{n_1+n_2+n_3}$ $\nu_2 = \frac{n_2}{n_1+n_2+n_3}$ $\nu_3 = \frac{n_3}{n_1+n_2+n_3}$

But these values are approximate. For example, I can have probability of 0.25 for the event a, so if I generate 20 events, I have to get approximately 5 a-events. But just by chance I can get 10 or 0 a-events.

So, I want to have a density distribution of probabilities. Since, $\nu_1$, $\nu_2$ and $\nu_3$ are dependent (the sum of them is equal to one) I going to use $\nu_1$ and $\nu_2$. So, I want to have an explicit form for the density distribution of $\nu_1$ and $\nu_2$.

$\rho (\nu_1,\nu_2) = F (n_1,n_2,n_3)$.

Does anybody know where I can get it?

  • $\begingroup$ The terminology you are using in your question is a little bit nonstandard: for example, your title asks for a "likelihood" but your problem description asks for a probability distribution on $(\nu_{1},\nu_{2})$. In short, a Bayesian solution would start with a (conjugate) Dirichlet prior on $(\nu_{1},\nu_{2})$, then you would observe a multinomial likelihood with $(n_{1},n_{2})$, then the posterior would be Dirichlet again, equivalent to your $F(n_{1},n_{2},n_{3})$, above. This gives you a density, but it isn't entirely clear from your description that this is what you want. $\endgroup$
    – user1108
    Dec 17 '10 at 1:59
  • $\begingroup$ By the way, the solution outlined above is written out in more detail at the bottom of this other problem: stats.stackexchange.com/questions/3194/… $\endgroup$
    – user1108
    Dec 17 '10 at 2:04

You should look at the multinomial distribution.


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.