COLLEGE STATISTICS • HYPOTHESIS TESTING

Type I & Type II Errors — Type I and Type II Errors

Understanding the two fundamental ways a statistical hypothesis test can lead to an incorrect conclusion.

Historical Context & Motivation

The formal framework of hypothesis testing—and the classification of the errors that can arise within it—owes its existence to decades of debate among some of the twentieth century's most influential statisticians. In the early 1900s, researchers lacked a principled method for deciding whether observed data justified rejecting a scientific assumption. The concepts of Type I error (falsely rejecting a true null hypothesis) and Type II error (failing to reject a false null hypothesis) emerged from the intellectual rivalry between Ronald Fisher and the Neyman–Pearson school, and they remain foundational to every discipline that relies on statistical inference.

1925
Fisher's Significance Testing
Ronald Fisher publishes Statistical Methods for Research Workers, introducing the p-value and the idea of rejecting a null hypothesis when the data are sufficiently unlikely under it. Although Fisher did not formally name error types, his framework implicitly defined what would later be called a Type I error.
1928
Neyman & Pearson's Error Framework
Jerzy Neyman and Egon Pearson publish their first joint paper arguing that hypothesis testing should consider two competing hypotheses—null and alternative—and explicitly control the probabilities of two distinct kinds of error. This marks the formal birth of the Type I / Type II error taxonomy.
1933
The Neyman–Pearson Lemma
Neyman and Pearson prove the fundamental lemma showing that the likelihood-ratio test is the most powerful test for a given significance level, directly linking the control of Type I error rate (α) to maximum achievable power (1 − β). This result remains a cornerstone of optimal test theory.
1948
Wald's Decision Theory
Abraham Wald extends the Neyman–Pearson framework into general statistical decision theory, introducing loss functions that assign explicit costs to Type I and Type II errors. This generalization enables practitioners to balance error rates according to the real-world consequences of each mistake.
1960s–Present
Power Analysis & Modern Practice
Jacob Cohen and others popularize statistical power analysis, urging researchers to design studies that control both α and β. Today, regulatory agencies (FDA, EMA) and scientific journals mandate pre-specified significance levels and power calculations, making Type I and Type II error management a routine part of study design.

The central question these pioneers confronted remains just as relevant for any college statistics student today: when we make a decision based on sample data, how do we quantify and control the risk of making the wrong decision? Answering that question requires a precise understanding of the two mutually exclusive ways a hypothesis test can err—Type I and Type II errors—and the mathematical relationship between them.

Core Principles & Definitions

Before dissecting the two error types, it is essential to recall that every hypothesis test begins with a null hypothesis (H₀), which represents the default or status-quo claim, and an alternative hypothesis (H₁ or Hₐ), which represents the claim we seek evidence for. The test produces one of two decisions: reject H₀ or fail to reject H₀. Because the decision is based on a finite sample rather than the full population, there is always a nonzero probability that the decision is wrong. That probability manifests as either a Type I or a Type II error, depending on which hypothesis is actually true.

1

Type I Error (False Positive)

Occurs when H₀ is true but the test leads us to reject it. The probability of a Type I error is denoted α (alpha), also called the significance level of the test.
2

Type II Error (False Negative)

Occurs when H₀ is false but the test fails to reject it. The probability of a Type II error is denoted β (beta). The complement, 1 − β, is the statistical power of the test.
3

Significance Level (α)

A pre-chosen threshold (commonly 0.05, 0.01, or 0.10) that sets the maximum tolerable probability of a Type I error. If the p-value falls below α, we reject H₀. A smaller α reduces Type I error risk but, for a fixed sample size, increases Type II error risk.
4

Statistical Power (1 − β)

The probability of correctly rejecting H₀ when it is indeed false. Power depends on the true effect size, sample size, variability, and α. Researchers typically aim for power ≥ 0.80, meaning β ≤ 0.20.
5

The α–β Trade-off

For a given sample size and effect size, decreasing α (being more conservative about false positives) necessarily increases β (raising the chance of missing a real effect). The only way to reduce both simultaneously is to increase the sample size or study a larger effect.
KEY TAKEAWAY
Think of a hypothesis test like a smoke detector. A Type I error is a false alarm—the detector goes off even though there is no fire. A Type II error is a missed detection—there really is a fire, but the detector stays silent. Making the detector more sensitive (lowering α) reduces false alarms but may cause it to trigger on harmless toast (increasing β). The only way to simultaneously reduce both false alarms and missed fires is to install a better (higher-sample-size) detector.

Visual Explanation — Decision Matrix & Distributions

The Four Outcomes of a Hypothesis Test

Every hypothesis test results in one of four outcomes, determined by the intersection of two binary states: the true state of nature (H₀ is true or H₀ is false) and the decision made (reject H₀ or fail to reject H₀). The following diagram presents these outcomes as a 2 × 2 decision matrix, with the two correct decisions on the main diagonal and the two errors on the off-diagonal.

The decision matrix shows that a Type I error (α) occupies the top-left cell (rejecting a true H₀), while a Type II error (β) occupies the bottom-right cell (failing to reject a false H₀). The two green cells represent correct decisions.

Notice that the two error types are mutually exclusive in any given test: if H₀ is true, only a Type I error is possible; if H₀ is false, only a Type II error is possible. You can never commit both errors in the same test. The challenge of good study design lies in controlling both α and β simultaneously, which, as we will see, requires careful attention to sample size, effect size, and the choice of significance level.

Mathematical Framework

The mathematical definitions of Type I and Type II error probabilities follow directly from the conditional probability structure of hypothesis testing. Let us formalize each error rate and then examine the key equation that connects α, β, sample size, and effect size.

TYPE I ERROR PROBABILITY
α = P(Reject H₀ | H₀ is true)
α is the probability of rejecting the null hypothesis given that it is actually true. For a z-test at significance level α = 0.05 (two-tailed), this corresponds to the combined area in both tails of the standard normal distribution beyond z = ±1.96.
TYPE II ERROR PROBABILITY
β = P(Fail to reject H₀ | H₁ is true)
β is the probability of failing to reject the null hypothesis when the alternative hypothesis is true. β depends on the true parameter value under H₁, the sample size n, the population standard deviation σ, and the chosen α.
STATISTICAL POWER
Power = 1 − β = P(Reject H₀ | H₁ is true)
Power is the probability of correctly detecting a real effect. A study with power = 0.80 has a 20% chance of committing a Type II error (β = 0.20).
SAMPLE SIZE FOR A ONE-SAMPLE Z-TEST
n = ((z_α/₂ + z_β) × σ / δ)²
where zα/2 is the critical value for the desired significance level, zβ is the critical value for the desired power, σ is the population standard deviation, and δ = |μ₁ − μ₀| is the minimum detectable difference (effect size in raw units). This equation makes explicit how increasing n reduces both error probabilities.
⚠️ Why Can't We Set Both α and β to Zero?
For any finite sample, the sampling distributions of the test statistic under H₀ and H₁ overlap. The rejection region is a fixed boundary: moving it to shrink the α-tail under H₀ necessarily enlarges the β-region under H₁, and vice versa. Only by increasing the sample size (which narrows both distributions) or studying a larger effect (which pushes the distributions apart) can we simultaneously reduce α and β.

The α–β Trade-off & Power Curves

The relationship between Type I and Type II errors is most vividly understood by visualizing the sampling distributions under H₀ and H₁ side by side. The following diagram shows two normal distributions: the left distribution is the sampling distribution of the test statistic assuming H₀ is true, and the right distribution assumes H₁ is true at a particular effect size. The critical value (vertical dashed line) partitions each distribution into regions corresponding to α, β, and power.

The purple curve (H₀) and the cyan curve (H₁) overlap in the middle. The critical value c (yellow dashed line) separates the rejection region from the non-rejection region. The red-shaded tail of H₀ beyond c represents α; the pink-shaded portion of H₁ to the left of c represents β; and the green-shaded portion of H₁ to the right of c represents power.

This diagram reveals the fundamental trade-off. If we move the critical value c to the right (i.e., require stronger evidence to reject H₀), the α-shaded region under the H₀ curve shrinks—fewer false positives. However, the β-shaded region under the H₁ curve grows—more false negatives. Conversely, moving c to the left increases α while decreasing β. The only way to shrink both shaded regions simultaneously is to increase the separation between the two curves (larger effect size δ) or narrow both curves (larger sample size n, since the standard error σ/√n decreases with n).

How various design choices affect error rates and power
ActionEffect on αEffect on βEffect on Power
Decrease α (e.g., 0.05 → 0.01)↓ Decreases↑ Increases↓ Decreases
Increase sample size n— Unchanged (fixed by researcher)↓ Decreases↑ Increases
Larger true effect size δ— Unchanged↓ Decreases↑ Increases
Decrease population variability σ— Unchanged↓ Decreases↑ Increases
Switch from two-tailed to one-tailed test— Unchanged (same α, redistributed)↓ Decreases (for the specified direction)↑ Increases (in that direction)

Worked Example — Drug Efficacy Trial

A pharmaceutical company claims that a new drug lowers systolic blood pressure by at least 10 mmHg compared to placebo. From prior studies, the population standard deviation of blood-pressure changes is σ = 15 mmHg. The researchers set α = 0.05 (two-tailed) and desire power = 0.80 (so β = 0.20). They recruit n = 36 patients and observe a sample mean reduction of x̄ = 7.5 mmHg. We will determine the test outcome, identify which error (if any) could have been committed, and calculate β for a specific alternative.

Hypothesis Test for Mean Blood-Pressure Reduction
1
Step 1 — State the HypothesesH₀: μ = 10 mmHg (the drug reduces blood pressure by exactly 10 mmHg, as claimed). H₁: μ ≠ 10 mmHg (the drug's effect differs from the claim). This is a two-tailed test because we want to detect both under-performance and over-performance.
2
Step 2 — Compute the Standard ErrorThe standard error of the sample mean is SE = σ / √n = 15 / √36 = 15 / 6 = 2.5 mmHg.
SE = 2.5 mmHg
3
Step 3 — Calculate the Test Statisticz = (x̄ − μ₀) / SE = (7.5 − 10) / 2.5 = −2.5 / 2.5 = −1.00.
z = −1.00
4
Step 4 — Determine the Critical Values and DecisionFor α = 0.05 (two-tailed), the critical values are z* = ±1.96. Since |z| = 1.00 < 1.96, we fail to reject H₀. There is insufficient evidence to conclude that the drug's effect differs from 10 mmHg at the 5% significance level.
Decision: Fail to reject H₀
5
Step 5 — Identify the Possible ErrorBecause we failed to reject H₀, the only error we could have committed is a Type II error—failing to detect a real departure from μ₀ = 10. A Type I error is impossible here because we did not reject H₀.
6
Step 6 — Calculate β for a Specific AlternativeSuppose the true mean reduction is μ₁ = 7 mmHg (so δ = |10 − 7| = 3 mmHg). We need P(fail to reject H₀ | μ = 7). Under H₁, the test statistic follows z ~ N((7 − 10)/2.5, 1) = N(−1.2, 1). We fail to reject when −1.96 ≤ z ≤ 1.96. Transforming to the H₁ distribution: P(−1.96 ≤ z ≤ 1.96) when z ~ N(−1.2, 1) is P(−1.96 − (−1.2) ≤ z' ≤ 1.96 − (−1.2)) = P(−0.76 ≤ z' ≤ 3.16) where z' ~ N(0,1). From standard normal tables: Φ(3.16) − Φ(−0.76) ≈ 0.9992 − 0.2236 = 0.7756.
β ≈ 0.776, Power ≈ 0.224
7
Step 7 — Interpret the ResultWith n = 36, if the true effect is only 7 mmHg (3 mmHg less than claimed), the test has only about 22.4% power to detect this discrepancy. This means there is a 77.6% chance of committing a Type II error—failing to realize the drug under-performs. To achieve 80% power for this alternative, we would need a substantially larger sample, which we could compute using the sample-size formula from Section 4.

Real-World Consequences of Each Error

The severity of a Type I versus a Type II error depends entirely on the application context. In some domains—criminal justice, pharmaceutical approval—the consequences of one error type far outweigh the other, driving the choice of α and the required power. Understanding these asymmetries is essential for calibrating any hypothesis test to its stakes.

Error consequences vary dramatically across domains
DomainType I Error ConsequenceType II Error ConsequenceTypical α Choice
Criminal TrialConvicting an innocent person (very costly)Acquitting a guilty person (costly but more tolerable)Very low ("beyond reasonable doubt" ≈ α ≪ 0.05)
Drug Approval (FDA)Approving an ineffective or harmful drugFailing to approve a beneficial drugα = 0.05 (often 0.025 one-sided)
Quality ControlHalting production unnecessarily (financial loss)Shipping defective products (safety risk, recalls)Depends on cost ratio; often α = 0.01
Medical ScreeningFalse positive diagnosis → unnecessary procedures, anxietyMissed disease → delayed treatment, worse prognosisHigher α tolerated (prioritize sensitivity / low β)
Particle PhysicsClaiming a new particle exists when it doesn'tMissing a genuine discoveryExtremely low: 5σ → α ≈ 3 × 10⁻⁷
KEY TAKEAWAY
Choosing α is not a purely statistical exercise—it is a decision-theoretic judgment that reflects the relative costs of the two error types. In medical screening, we accept more false positives (higher α) to avoid missing true disease (lower β). In particle physics, the cost of a false discovery is so reputationally devastating that α is set near zero, tolerating substantial β. Always ask: which error is more harmful in this specific context?

Connection to Advanced Theory

The Type I / Type II error framework extends naturally into several advanced topics in statistical theory. Understanding these connections prepares you for more sophisticated methods encountered in upper-division courses and graduate study.

From introductory error concepts to advanced statistical theory
Introductory ConceptAdvanced ExtensionKey Idea
Fixed α (e.g., 0.05)Multiple testing correctionsWhen performing m tests simultaneously, the family-wise error rate (FWER) inflates to 1 − (1 − α)ᵐ. Bonferroni, Holm, and Benjamini–Hochberg procedures control different error rate metrics.
Type I error (α)False Discovery Rate (FDR)In high-dimensional settings (genomics, fMRI), FDR = E[V/R] controls the expected proportion of false discoveries among all rejections, offering a less conservative alternative to FWER control.
Power = 1 − β at a pointPower functions & UMP testsThe power function π(θ) = P(Reject H₀ | θ) maps every parameter value to its rejection probability. The Neyman–Pearson lemma identifies the uniformly most powerful (UMP) test for simple vs. simple hypotheses.
Type II error (β)Bayesian decision theoryIn a Bayesian framework, the prior probability of H₀ modulates the posterior probability of error. The frequentist α is replaced by the Bayes factor or posterior odds, which automatically incorporate prior information about both hypotheses.
Sample-size formula for z-testSequential & adaptive designsGroup-sequential designs (O'Brien–Fleming, Lan–DeMets) allow interim analyses that preserve overall α while potentially stopping early for efficacy or futility, reducing expected sample size.

Each of these extensions rests on the same foundational insight: any statistical decision procedure must reckon with the possibility of error, and the practitioner's task is to manage error probabilities in a way that reflects the scientific and practical context. Whether you are adjusting for thousands of simultaneous tests in a genome-wide association study or designing a sequential clinical trial with interim analyses, the logic of Type I and Type II errors remains the conceptual bedrock.

Practice Problems

PROBLEM 1CONCEPTUAL
A researcher tests H₀: μ = 50 vs. H₁: μ ≠ 50 at α = 0.05 and decides to reject H₀. If the true population mean is actually μ = 50, what type of error has the researcher committed? Explain why the other error type is impossible in this scenario.
PROBLEM 2BASIC CALCULATION
A one-sample z-test is conducted at α = 0.01 (two-tailed) with H₀: μ = 100, σ = 20, and n = 64. The sample mean is x̄ = 94.5. Compute the test statistic, state the decision, and identify which error type could have been made.
PROBLEM 3INTERMEDIATE
Using the same setup as Problem 2 (H₀: μ = 100, σ = 20, n = 64, α = 0.01 two-tailed), calculate the probability of a Type II error (β) if the true population mean is μ₁ = 95. Then compute the power of the test against this alternative.
PROBLEM 4APPLIED
An environmental agency tests whether the mean lead concentration in a water supply exceeds the safety threshold of 15 ppb: H₀: μ ≤ 15 vs. H₁: μ > 15 (one-tailed, α = 0.05). They collect n = 49 samples with σ = 7 ppb. (a) What is the minimum sample mean that would trigger a rejection of H₀? (b) If the true mean is μ₁ = 17 ppb, what is the probability of a Type II error? (c) Discuss the practical consequences of each error type in this context.
PROBLEM 5CRITICAL THINKING
A researcher conducts 20 independent hypothesis tests, each at α = 0.05, and all 20 null hypotheses are actually true. (a) What is the probability that at least one test produces a Type I error? (b) If the researcher applies a Bonferroni correction, what adjusted α should be used per test? (c) Critically evaluate: does controlling the family-wise Type I error rate via Bonferroni come at a cost? If so, what is that cost, and how might it be mitigated?

Summary — Type I & Type II Errors

Every hypothesis test involves a binary decision—reject or fail to reject H₀—and the possibility of error whenever the sample does not perfectly represent the population. A Type I error (α) occurs when a true null hypothesis is incorrectly rejected (a false positive), while a Type II error (β) occurs when a false null hypothesis is not rejected (a false negative). The significance level α is set by the researcher before data collection and caps the maximum Type I error rate. The complement of β, called statistical power (1 − β), measures the test's ability to detect a real effect and depends on the sample size, the true effect size, the population variability, and the chosen α.

The central trade-off of hypothesis testing is that, for a fixed sample size, decreasing α increases β and vice versa. The only way to reduce both error probabilities simultaneously is to increase the sample size or study a larger effect size. In practice, the appropriate balance between α and β is determined by the relative costs of each error in the given application domain—a principle that extends from basic z-tests to advanced topics such as multiple testing corrections, false discovery rate control, and sequential trial designs.

Varsity Tutors • College Statistics • Type I & Type II Errors