ML Wiki
Machine Learning Wiki - A collection of ML concepts, algorithms, and resources.

Confidence Intervals and Statistical Tests

Relationship between Confidence Intervals and Tests

How to connect Hypothesis Testing and Confidence Intervals?

Example

  • Remember the beer cap flipping experiment?
  • We test: $H_0: p = 0.5, H_A: p \neq 0.5$ (2-sided) Observations:
  • $n = 1000$
  • $\hat{p} = 0.576$
  • under $H_0$, is difference $ \hat{p} - p = \hat{p} - 0.5 = 0.076$ too large to reject H_0?

We calculate $p$-value $P(\mid \hat{p} - 0.5\mid \geqslant 0.076) \approx P(\mid N(0, 1)| \geqslant 4.81 ) \approx 1 / 663000 \leqslant 0.05$ And reject the $H_0$ because the $p$-value is small

Now let’s calculate 95% CI:

  • $\hat{p} \pm 1.96 \sqrt{p(1-p)/n} = [0.532, 0.620]$
  • The CI misses 0.5
  • and this is not a coincidence |

    General Test

    Suppose we have a test of the following form

  • $H_0: \mu = \mu_0, H_A: \mu \neq \mu_0$
  • Our observations are: $n$, $\bar{X}$, $s$,
  • We’re interested in the difference between observed mean and the true mean: $\Delta = | \bar{X} - \mu_0|$ So we reject $H_0$ if
  • $P(\mid \bar{X} - \mu_0\mid \geqslant \Delta) \leqslant \alpha$
  • $P\left(\left \cfrac{\bar{X} - \mu_0}{\sqrt{s^2 / n}} \right \geqslant \cfrac{\Delta}{\sqrt{s^2 / n}} \right) = P\left(\left t_{n - 1} \right \geqslant \cfrac{\Delta}{\sqrt{s^2 / n}} \right) \leqslant \alpha$

This will only happen if

  • $\cfrac{\Delta}{\sqrt{s^2 / n}} \geqslant T_{\alpha/2, n-1}$
  • where $T_{\alpha/2, n-1}$ is critical value s.t. $P(\mid t_{n-1}\mid \geqslant T_{\alpha/2, n-1} ) = \alpha$

And $(1 - \alpha)$ Confidence Intervals for $\mu$ is

  • $\bar{X} \pm T_{\alpha/2, n-1} \cdot \sqrt{s^2 / n}$
  • This misses $\mu_0$ when $| \bar{X} - \mu_0 | \geqslant T_{\alpha/2, n-1} \cdot \sqrt{s^2 / n}$

So these are equivalent: Reject $H_0$ when C.I. misses $\mu_0$

See also

Sources