So, I have a population which is believed to have $H_0:\mu = 4620$ and I should test for $H_1: \mu < 4620$. Given a sample of $n=81, \bar{X}=4450, S=900$, the p-value is: $pv=\Pr\left(\frac{\bar{X}-4620}{900/\sqrt{81}}\leq\frac{4450-4620}{900/\sqrt{81}}\right)\\=\Pr\left(Z\leq-\frac{170}{100}\right)\\\approx0.0445<\alpha=0.05$

So I reject the null hypothesis.

Then, I'm being asked to find a confidence interval for $\alpha=0.1$, which is: $CI_{0.9}\left(\mu\right)=\left[\bar{X}\pm\frac{S}{\sqrt{n}}\cdot t_{80,0.95}\right]=\left[4283.588,4616.412\right]$ and to infer from it, again, that the null should be rejected with $\alpha=0.05$.

Now, clearly $\mu_{H_{0}}>\sup CI_{0.9}$ so we can reject with $\alpha=0.1$. Furthermore, if I take the one-sided confidence interval with $\alpha =0.05$ it has the same upper bound as the two-sided interval. It seems like all we have done is to expand the non-rejection region to the left tail of the two-sided interval.

But what I'm having trouble with is how to formulate the argument - why does it make sense? Is it enough to state that the one-sided interval overlaps the two-sided one? That the values in the left tail, for the two-sided case, "don't matter"?

Why is this transformation of the two-sided interval into a suitable one-sided interval valid, in this case? Is there a general method/explaination of why these intervals are equivelant, if at all?

Edit: Found this answer Matching Confidence limits with One-Sided Hypothesis tests which explains it the other way around - why the two-sided interval should have been constructed with $\alpha=0.9$. But even after drawing the bell curve and coloring the expanded acceptance region, I can't understand the logic behind "merging" the left rejection area to the interval. It fits the desired $\alpha$, that's fine, but why would anyone test a one-sided hypothesis using a two-sided interval in the first place?


The thing which is confusing you is where to put the lower bound. In the two-sided 90% interval you give the lower bound is finite (4283.588). If you want a 95% one-sided interval it has the same upper bound (as you say) but the upper bound is infinite. So it is from $-\infty$ to 4616.412.

  • $\begingroup$ @mdewwy but the interval you described includes the null $\mu$, so I can't reject the null hypothesis, which contradicts the p-value test. How is that possible? $\endgroup$ – gbi1977 Jan 6 at 13:52
  • $\begingroup$ @mdewwy Thanks, as I said this is the same result I got - I included the left rejection area of the two-sided interval. My question is why the 2 intervals are equivelant? just because they have the same $\alpha$ now? I'm not satisfied with this kind of answer... $\endgroup$ – gbi1977 Jan 6 at 13:59
  • $\begingroup$ What goes against my intuition, I guess, is why the proper one-sided inteval, which is also the non-rejection area, includes smaller possible values for $\mu$? Isn't the whole purpose of the non-rejection area to be "closer" to the null $\mu$? $\endgroup$ – gbi1977 Jan 6 at 14:35
  • $\begingroup$ The two intervals are not equivalent. One is a two-sided 90% the other one-sided 95% so both have 5% in one tail and one of them has a further 5% in the other tail $\endgroup$ – mdewey Jan 6 at 14:38
  • $\begingroup$ I think this is the understanding I lacked. It's counter-intuitive. $\endgroup$ – gbi1977 Jan 6 at 14:53

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.