1
$\begingroup$

I have two probabilities $p$ and $q$. $p>q$, and they aren't correlated. I'm going to calculate $i$ such that $p^i=q$, which is easily done as $\log_p(q)$.

Now, I'd like to also calculate a confidence interval for $i$, which is necessarily going to be a function of both $p$ and $q$'s confidence intervals. My first approach was to do

p.min <- qbeta(0.025, 152, 29)
p.max <- qbeta(0.975, 152, 29)

q.min <- qbeta(0.025, 37, 19)
q.max <- qbeta(0.975, 37, 19)

## q.max and p.min have the smallest difference
i.min <- log(q.max, base = p.min)

## q.min and p.max have the largest difference
i.max <- log(q.min, base = p.max)

But it occurs to me that 95% confidence intervals for $p$ and $q$ independently probably produces too large a confidence interval for $i$, because the joint confidence interval for $p$ and $q$ will be narrower.

So, how do I go about figuring out the joint confidence interval of $p$ and $q$. They're uncorrelated, which should make things easier. It is as simple as narrowing the quantiles in qbeta()? By how much?

$\endgroup$
  • 1
    $\begingroup$ What does it mean when you say that probabilities aren't correlated? $\endgroup$ – Dilip Sarwate May 16 '12 at 15:18
  • $\begingroup$ Upon observing that $\log_p{q}$ is defined to be $\log{q}/\log{p}$ you could apply the approximate method (for finding a CI for a ratio of random variables) described at stats.stackexchange.com/questions/21298. Additional guidance is available by searching. $\endgroup$ – whuber May 16 '12 at 15:45
  • $\begingroup$ @DilipSarwate, I mean that the value of q has no effect on the probability of a value for p, and vice versa. So, it's not like a small value for p is more likely of q is also small. $\endgroup$ – JoFrhwld May 16 '12 at 20:11
2
$\begingroup$

I think there is a confusion between confidence interval and probability interval here.

In the R code, you are indicating that $p\sim Beta(152,29)$ and $q\sim Beta(37,19)$, then you can calculate the distribution of $i=log(q)/log(p)$ using a change of variable and then obtain the corresponding probability interval for $i$ using this distribution.

Another possibility is to approximate this probability interval by Monte Carlo simulation. In this case this interval is approximately $(1.30, 4.23)$

i=log(rbeta(100000, 37, 19))/log(rbeta(100000, 152, 29))
c(quantile(i,0.025),quantile(i,0.975))

In order to construct a confidence interval for $i$ you would require that $p$ and $q$ are parameters of a sampling model.

$\endgroup$
  • $\begingroup$ I agree with Procrastinator. I think there is a problem with terminology. p and q are parameters not random variables. It is the sample estimates that are the random variables. p and q are functionally related. Statistical independence is a property of teh estimates. So independence depends on how you select the estimates. It seems that you would want to exploit the relationship and choose estimates that are correlated. Also a joint confidence region for p and q does not need to be a pair of intervals.It could be an ellipse or some other 2-dimensional region. $\endgroup$ – Michael Chernick May 17 '12 at 11:54
  • 1
    $\begingroup$ If you do want confidence intervals you have to define the estimators for p, q and i. $\endgroup$ – Michael Chernick May 17 '12 at 11:54

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.