2
$\begingroup$

Let's take for simplicity an iid sample from a Gaussian random variable $Y \sim \mathcal{N}(\mu, \sigma^2)$. Let $y_0 \in \mathbb{R}$. The $(1-\alpha, p)$-bound of the lower one-sided (i.e. right-sided) tolerance interval is a $(1-\alpha)$-lower confidence bound of the quantile $q_p$ of the normal distribution $\mathcal{N}(\mu, \sigma^2)$. Therefore, if this tolerance bound is higher than $y_0$, one can say that the exceedance probability $\Pr(Y > y_0)$ is $\alpha$-significantly higher than $p$.

Now take two real numbers $y_1 < y_2$. If the two-sided $(1-\alpha,p)$-tolerance interval falls inside the interval $[y_1, y_2]$, does it make sense to say that the probability $\Pr\bigl(Y \in [y_1, y_2]\bigr)$ is significantly higher than $p$ at the $\alpha$ level?

$\endgroup$
13
  • $\begingroup$ What happens as $y_1$ approaches arbitrarily close to $y_2$? The probability in question ought to grow vanishingly small. Any procedure that concludes this probability is significantly greater than any given positive $p$ would be awfully suspect! $\endgroup$
    – whuber
    Aug 29, 2017 at 14:59
  • $\begingroup$ @whuber I'm not comfortable with this question (that's why I ask it ^^) but I don't follow your reasoning. If the interval $[y_1,y_2]$ is small, and if the tolerance interval falls inside it, that means that the (estimate of the) standard deviation is small, and then the probability is not small. $\endgroup$ Aug 30, 2017 at 8:35
  • $\begingroup$ I don't follow your conclusion, and perhaps that amounts to not understanding what you mean by "the two-sided tolerance interval." Usually that is understood to be a confidence interval for the quantile $q_p$. Thus, you cannot draw any conclusion about the SD of $Y$: like most confidence intervals, that one shrinks down to zero as the amount of data grows, independently of whatever value the SD of $Y$ might be (even if the SD is infinite!). $\endgroup$
    – whuber
    Aug 30, 2017 at 12:21
  • $\begingroup$ A lower (upper) tolerance limit is a lower (upper) confidence bound about the quantile. But this is not true for the two-sided tolerance interval (usually called the "tolerance interval", shortly). $\endgroup$ Aug 30, 2017 at 12:32
  • $\begingroup$ Why don't you follow my conclusion ? The length of the tolerance interval is $2k\hat\sigma$. So I simply say that $\hat\sigma$ is small if this interval is contained in a small interval. $\endgroup$ Aug 30, 2017 at 12:34

1 Answer 1

2
$\begingroup$

No this is not true.

One-sided case

Therefore, if this tolerance bound is higher than $y_0$, one can say that the exceedance probability $\Pr(Y > y_0)$ is $\alpha$-significantly higher than $p$.

Let's check this claim with the help of simulations. We simulate $Y_i \sim_{\text{iid}} \mathcal{N}(0,1)$, we take $y_0=1$, and we take the lower tolerance bound with coverage $p=\Pr(Y_i>y_0)$. Then the probability that the lower tolerance bound is higher than $y_0$ should be $\alpha$.

y0 <- 1
p <- 1-pnorm(y0)
nsims <- 200000
n <- 10
k <- tolerance::K.factor(n, alpha=0.05, P=p, side=1)
test <- logical(nsims)
for(i in 1:nsims){ 
  Y <- rnorm(n)
  bound <- mean(Y) - k*sd(Y)
  test[i] <- bound > y0
}

This is confirmed by the simulations:

> mean(test)
[1] 0.049805

Two-sided case

Now take two real numbers $y_1 < y_2$. If the two-sided $(1-\alpha,p)$-tolerance interval falls inside the interval $[y_1,y_2]$, does it make sense to say that the probability $\Pr\bigl(Y \in [y_1, y_2]\bigr)$ is significantly higher than $p$ at the $\alpha$ level?

Let's check with simulations that this is not true. As for the one-sided case, if we take $p = \Pr\bigl(Y_i \in [y_1, y_2]\bigr)$, then the probability that the two-sided $(1-\alpha,p)$-tolerance interval falls inside the interval $[y_1,y_2]$ should be $\alpha$ if the claim were true.

y1 <- -1; y2 <- 2
p <- pnorm(y2)-pnorm(y1)
nsims <- 200000
n <- 10
k <- tolerance::K.factor(n, alpha=0.05, P=p, side=2, method="EXACT")
test <- logical(nsims)
for(i in 1:nsims){ 
  Y <- rnorm(n)
  bounds <- mean(Y) + k*c(-1,1)*sd(Y)
  test[i] <- bounds[1] > y1 && bounds[2] < y2
}

This is not confirmed:

> mean(test)
[1] 0.01193
$\endgroup$

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.