I know that in the frequentist approach, the confidence interval contains the true parameter $\theta$ with some minimum probability (e.g. 95%); while in the bayesian approach, the credible interval includes 95% of the posterior distribution.

My question is, how would we relate the credible interval (with left and right bounds L and R) to the true parameter? In other words, how would we find $P(L(X) \leq \theta \leq R(X)$?

To put my question in another way: given a set of data, we can derive some posterior distribution (and thus credible interval), but that doesn't guarantee the true parameter is captured by this credible interval. How would I find the probability that the true parameter lies within the credible interval?


1 Answer 1


Bayesian credible interval. In Bayesian estimation of binomial success probability $\theta,$ suppose we begin with the prior distribution $\theta \sim \mathsf{Beta}(2. 2).$

Upon observing $x = 45$ successes in $n = 100$ trials, we have likelihood proportional to $\theta^{45}(1-\theta)^{55}.$

Then according to Bayes' Theorem, the posterior distribution is $\theta \sim \mathsf{Beta}(47. 57),$ and 95% Bayesian posterior probability interval (credible interval) $(0.358, 0.548).$

qbeta(c(.025, .975), 47, 57) 
[1] 0.3578055 0.5477928

If we believe the prior and the integrity of the data, then be must believe the posterior probability distribution is correct. According to the posterior distribution, $P(0.358 < \theta < 0.548) = 0.95.$

I'm puzzled by your sentence, "[T]hat doesn't guarantee the true parameter is captured by this credible interval." There is no absolute "guarantee" attached to any interval estimate [other than $(0,1)$]. However, under the posterior probability distribution, there is probability 95% that the credible interval covers $\theta.$

Agresti-Cooll frequntist confidence interval. For the same data as above we have a 95% A-C confidence interval %(0.356, 0.548).$

p = 47/104; p+qnorm(c(.025,.975))*sqrt(p*(1-p)/104)
[1] 0.3562732 0.5475730

One frequentist interpretation is that the process by which this CI was obtained would contain the true unknown value of $\theta$ over the long run in future repetitions of the experiment that produced the data.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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