Type I and type II errors

Every test can be wrong in two ways. Drag the controls to see why you can't shrink one error without paying for it with the other, unless you collect more data.

H₀H₁
α (false positive) 5% β (miss) 51% power 49%
Type I error (α)
5%
Type II error (β)
50.9%
Power (1 − β)
49.1%

Try it: slide α down to 0.01 and watch β grow. Now raise n: both curves stay put in shape, but H₁ moves away from H₀ and β shrinks. Set d to 0 and “power” collapses to α, because there is nothing to find. (Normal approximation to a two-sample test.)

The four possible outcomes

H₀ is true (no effect)H₀ is false (real effect)
Reject H₀Type I error (false positive), probability αCorrect detection, probability 1 − β (power)
Fail to reject H₀Correct, probability 1 − αType II error (false negative), probability β

Examples

  • Clinical trial: concluding a new drug lowers blood pressure when it doesn't (type I) vs dismissing a drug that works (type II).
  • Screening test: a healthy person flagged as ill (type I) vs a disease missed (type II).
  • Education study: reporting that a tutoring programme raised scores when the difference was chance (type I) vs a small pilot of 15 students per group that misses a genuine d = 0.4 improvement (type II; its power is only about 18%).

Why thresholds mislead intuition

The critical value is a cut-off, and cut-offs are where intuition fails: p = .049 and p = .051 are practically identical evidence, yet land on opposite sides of the line. People misread cut-offs everywhere, and a classic case is income tax, where many believe earning a dollar past a bracket raises the tax on all their pay; ahaboo's narrated walkthrough of US tax brackets untangles that one.

For research, the practical lesson is to report the exact p-value, the effect size and its confidence interval, and to plan power in advance so a “non-significant” result is informative rather than a type II error waiting to happen.

Multiple testing inflates type I errors

Run 20 independent tests at α = .05 on pure noise and the chance of at least one false positive is 1 − 0.95²⁰ ≈ 64%. Corrections such as Bonferroni (α/m) or Holm keep the family-wise rate at 5%; the false discovery rate approach of Benjamini and Hochberg (1995) is less conservative for large numbers of tests.

Frequently asked questions

What is a type I error?

Rejecting a null hypothesis that is actually true: a false positive. Its probability is α, the significance level you choose, usually 5%.

What is a type II error?

Failing to reject a null hypothesis that is actually false: a false negative, missing a real effect. Its probability is β; power is 1 − β.

Which error is worse?

It depends on the consequences. Approving an ineffective drug (type I) and abandoning an effective one (type II) carry different costs. Fields set α and target power to reflect that balance; the conventional 5% vs 20% treats a false positive as four times as costly.

How do I reduce both errors at once?

Increase the sample size, reduce measurement noise, or study a larger effect. With a fixed sample, lowering α always raises β.