1
$\begingroup$

I have a series of probabilities that event $E$ occurs at iteration $i$ with $i=1,2,....$. These are denoted $p_1,p_2,....$. These probabilities are that $E$ occurs on iteration $i$, but do not consider whether $E$ occurred before $i$.

I want to transform these $p$s into a new series such that $\hat{p}^i$ = probability of first occurrence of $E$ on $i$. My thought is that this is:

$$\hat{p}^i = (1-p_1)(1-p_2)....(1-p_{i-1})p_i$$

Assuming $p_i$ is independent of $p_j$ for $i>j$.

But how do I transform if they are not independent?

In my problem, the probability $E$ occurs over time approaches a stationary distribution and is increasing up to that point. E.g, after some iteration $T$ $p_{t} = X \forall t>T$. "An event waiting to happen"

$\endgroup$
  • $\begingroup$ +1 agree with your answer, and yes, it certainly assumes independence $\endgroup$ – MikeP May 19 '17 at 14:21
  • $\begingroup$ @MikeP I edited my question with new information asking about that. $\endgroup$ – Tommy May 19 '17 at 14:24
  • $\begingroup$ You would normally speak of independent events, rather than probabilities. What do you mean by "agnostic"? If the probability of an event at $j$ is the same whether or not an event occured at $i$, then the events are independent. $\endgroup$ – GeoMatt22 May 19 '17 at 15:00
1
$\begingroup$

They could still be independent. I can come up with an example, although very contrived.

Suppose that p-primes are 0.1, 0.2, 0.3, 0.4 (have to sum to 1)

We can work backwards and come up with ps assuming independence to match using:

$$p'(i)=\prod_{j=1}^{i-1}(1-p(j))\cdot p(i)$$

$$p(i)=\frac{p(i)}{\prod_{j=1}^{i-1}(1-p(j))}$$

Using Matlab ps are 0.1000, 0.2222, 0.4286, 1.0000

$\endgroup$
  • $\begingroup$ from summing my $p$s I have learned that I must not have independence. Thus I've edited my question to ask how to transform it not independent. $\endgroup$ – Tommy May 19 '17 at 18:51
1
$\begingroup$

You would normally talk about independent events rather than probabilities.

To be more precise, let $E_i$ be a binary indicator such that $$E_i=\begin{cases} 1 & \text{an event occurs at iteration }i \\ 0 & \text{otherwise} \end{cases}$$

Then the marginal probability of $E_i=1$, with respect to an event at some other iteration $j$, will be \begin{align}\Pr\big[E_i=1\big] &= \Pr\big[E_i=1\mid E_j=1\big]\times\Pr\big[E_j=1\big] \\ &+\Pr\big[E_i=1\mid E_j=0\big]\times\Pr\big[E_j=0\big]\end{align} If the event $E_i$ is independent of the event $E_j$, then you will have $$\Pr\big[E_i=1\mid E_j=1\big] = \Pr\big[E_i=1\mid E_j=0\big] = p_i$$ This intuitively would seem to correspond to your statement about "agnostic".

Mathematically, since you are specifying the probabilities $p_i$ without reference to the sequence of values $E_{j<i}$, your model is assuming independent events. (Otherwise you would need to marginalize over possible "prefix sequences" to determine the $p_i$.)

$\endgroup$
  • $\begingroup$ this answer helps. because my $p$s do not seem to add to 1, so I guess I am violating independence somewhere. $\endgroup$ – Tommy May 19 '17 at 18:34
  • $\begingroup$ can I re-ask, how to transorm if they are not independent? I edited my question $\endgroup$ – Tommy May 19 '17 at 18:39
  • $\begingroup$ Where do your $p_i$ values come from? My understanding was that multiple events can occur, so the sum of the $p$'s would not sum to $1$. (As distinguished from the $\hat{p}$'s; e.g. you have a variable-$p$ version of this.) $\endgroup$ – GeoMatt22 May 19 '17 at 18:53
  • $\begingroup$ more information: 1) only one event can occur. 2) i've updated the question with information about $p$s approaching a stationary distribution. $\endgroup$ – Tommy May 19 '17 at 18:58
  • $\begingroup$ yes it looks like a variable p version of that; is there a straitforward answer to that? $\endgroup$ – Tommy May 19 '17 at 19:06

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.