0
$\begingroup$

Hi everyone

The question is the following:

A certain event may or may not take place. So we say that if we focus on it one time, it has a probability p of being satisfied (0 <= p < 1)

If we observe it multiple times, and we find out that it occurred zero times among our n observation, what can we say about p? What is the most likely value for p? Or better, find a function k(x, n) that returns the probability that p = x

How does it all change if among our n observation the event occurred m times? I know that if n tens to infinity then p tends to m/n but that is only part of the question

Thank you

$\endgroup$
4
  • 2
    $\begingroup$ Probably this is off-topic for this site. In any case, no answer can be given without more information, such as the joint distribution of the observations. $\endgroup$ Sep 29, 2012 at 13:24
  • $\begingroup$ Thanks for the comment, but I am not a mathematician yet, so I do not understand what's missing.. $\endgroup$
    – Ant
    Sep 29, 2012 at 13:45
  • $\begingroup$ If the events are not independent, what happens? For example, if all observations will surely be the same. Then: observing 0 multiple times is no better than observing it once. $\endgroup$ Sep 29, 2012 at 17:30
  • $\begingroup$ well I should have mentioned that but of course I meant that the events are independent from each other.. $\endgroup$
    – Ant
    Sep 29, 2012 at 18:00

1 Answer 1

0
$\begingroup$

I'll interpret "most likely" as Maximum Likelihood Estimation, that is, given some observations, what is the value of $p$ that makes the probability of those observations the largest. For example, if we observe something occurring $0$ out of $n$ times, the maximum likelihood estimate for $p$ is $0$, because that gives our observations a probability of $1$. This is only unreasonable if you have some prior information about $p$ (such as "it's probably around $0.5$, but maybe a little higher or lower" or "it's very close to either $1$ or $0$, but I don't know which").

Suppose we run $n$ trials, and find that the event occurred $n$ times, and we assume that each trial is independent with the same probability $p$ of success. Then the probability of our observation is $$P(p) = {n \choose m} p^m (1-p)^{n-m},$$ since the probability of each of the $m$ successes is $p$, the probability of the $n-m$ failures is $1-p$, and there are ${n\choose m}$ ways for $m$ of the $n$ trials to be the successful trials. Assuming $m$ and $n-m$ are both nonzero, this probability vanishes when $p=0$ or $1$, so the maximizing value of $p$ will be somewhere in between. Then we can find this maximizing value of $p$ by taking the derivative of $P$ and setting the result equal to $0$. We can take the derivative: $$P'(p) = {n\choose m} \left[mp^{m-1}(1-p)^{n-m} - (n-m)p^m(1-p)^{n-m-1}\right]$$ $$ = {n\choose m} \left[m(1-p) - (n-m)p\right]p^{m-1}(1-p)^{n-m-1}$$ This vanishes when $m(1-p)-(n-m)p=0$, i.e. when $m = np$ or $p = m/n$.

So if you have no prior information about what $p$ should be, but you observe $m$ successes in $n$ independent trials, the value of $p$ that best matches your observation is $p=m/n$.

$\endgroup$
2
  • $\begingroup$ That was easier than i thought, thank you! Though I do not understand why they closed the question as off-topic since it's not; one on mathoverflow can only ask something about Abelian rings to be answered? $\endgroup$
    – Ant
    Sep 30, 2012 at 9:28
  • $\begingroup$ @Ant, I'd say that it was closed because it's quite far from research level... $\endgroup$ Sep 30, 2012 at 13:05

Not the answer you're looking for? Browse other questions tagged or ask your own question.