Null & Alternative Hypotheses — Null and Alternative Hypotheses - College Statistics

SECTION 1

Historical Context & Motivation

The practice of formulating competing claims about population parameters and then using data to adjudicate between them is so central to modern science that it is easy to forget how recently the framework was codified. Before the twentieth century, researchers relied on informal judgment and qualitative reasoning to decide whether observed patterns were genuine or merely artifacts of chance variation. The development of hypothesis testing provided a rigorous, repeatable procedure for quantifying the strength of evidence against a default assumption—transforming statistical inference from an art into a discipline with formal rules.

The conceptual seeds were planted in the early 1700s when John Arbuthnot published what many regard as the first formal significance test, examining whether the observed excess of male births in London parish records could be explained by chance alone. Over the next two centuries, mathematicians and scientists refined these ideas, but the modern framework crystallized primarily through the work of three intellectual giants: Ronald A. Fisher, Jerzy Neyman, and Egon Pearson. Their sometimes conflicting visions of how inference should proceed were eventually merged—sometimes awkwardly—into the hybrid procedure taught in most statistics courses today.

1710 • Arbuthnot's Divine Providence Test

John Arbuthnot examined 82 years of London christening records and calculated the probability of observing more male than female births every single year under a fair-chance assumption, finding it astronomically small. This is widely considered the first published significance test.

1925 • Fisher's Statistical Methods for Research Workers

Ronald A. Fisher formalized the concept of a null hypothesis (H₀) as the default claim to be tested, introduced p-values as a continuous measure of evidence, and proposed the 0.05 significance threshold that remains ubiquitous.

1933 • Neyman–Pearson Lemma

Jerzy Neyman and Egon Pearson introduced the alternative hypothesis (H₁ or Hₐ) as an explicit competing claim, formalized Type I and Type II errors, and developed the framework of statistical power—ideas that gave hypothesis testing its decision-theoretic backbone.

1950s • The Hybrid Framework Emerges

Textbooks began merging Fisher's significance-testing philosophy with the Neyman–Pearson decision-theoretic approach, producing the hybrid null-versus-alternative procedure that dominates applied research. Debates about the coherence of this merger continue in the philosophy-of-statistics literature.

2016 • ASA Statement on P-Values

The American Statistical Association issued its first formal guidance on the interpretation of p-values, emphasizing that no single number can substitute for careful scientific reasoning. This statement reignited interest in the proper formulation and understanding of null and alternative hypotheses.

The central question that motivated these developments remains the same one you will face every time you encounter data: Is the pattern I observe in my sample strong enough to conclude that something real is happening in the population, or could chance alone have produced it? The null and alternative hypotheses give that question a precise mathematical structure.

SECTION 2

Core Principles & Definitions

Hypothesis testing begins by translating a research question into two mutually exclusive and collectively exhaustive statements about a population parameter. These statements—the null hypothesis and the alternative hypothesis—serve as the logical scaffolding upon which the entire inferential procedure rests. Understanding their roles, their asymmetric relationship, and the conventions that govern their formulation is essential before computing any test statistic or p-value.

Null Hypothesis (H₀)

The default position asserting that no effect, no difference, or no relationship exists in the population. It typically states that a parameter equals a specific value (e.g., μ = μ₀). The null serves as the claim that the researcher attempts to find evidence against.

Alternative Hypothesis (Hₐ or H₁)

The competing claim asserting that an effect, difference, or relationship does exist. It can be one-sided (μ > μ₀ or μ < μ₀) or two-sided (μ ≠ μ₀). Accepting this hypothesis implies the data provide sufficient evidence to reject the null.

Significance Level (α)

The pre-specified probability threshold for Type I error—the risk of incorrectly rejecting a true null hypothesis. Common choices are α = 0.05, 0.01, and 0.10. The level must be set before data collection to maintain the integrity of the test.

Burden of Evidence

The null hypothesis enjoys a privileged status analogous to 'innocent until proven guilty.' It is retained unless the data provide sufficiently strong evidence against it. We never 'prove' the null true; we either reject it or fail to reject it.

Directionality

A one-tailed (directional) test specifies the direction of the expected effect in Hₐ (e.g., μ > 100). A two-tailed (non-directional) test allows for deviations in either direction (μ ≠ 100). The choice must be justified by the research context and stated prior to analysis.

✦

KEY TAKEAWAY

Think of hypothesis testing like a criminal trial. The null hypothesis is the presumption of innocence—the defendant (status quo) is assumed not guilty until the prosecution (the data) presents evidence strong enough to convince the jury beyond a reasonable doubt. The alternative hypothesis is the prosecution's claim that guilt exists. A 'not guilty' verdict does not prove innocence; it simply means the evidence was insufficient to convict. Similarly, failing to reject H₀ does not prove the null is true—it means the sample data were not convincing enough to warrant rejecting it.

SECTION 3

Visual Explanation — The Hypothesis Testing Decision Framework

The relationship between the null hypothesis, alternative hypothesis, and the decision outcomes can be visualized as a flowchart that traces the logic of the testing procedure from formulation through conclusion. The following diagram illustrates how a research question is translated into formal hypotheses, how data are collected and summarized into a test statistic, and how the resulting p-value is compared against the significance level to reach a decision.

The flowchart traces the hypothesis testing procedure from research question to decision. The null hypothesis (H₀) and alternative hypothesis (Hₐ) are formulated first, then data are collected and summarized into a test statistic. The resulting p-value is compared to the pre-set significance level α at the diamond decision node.

Notice the asymmetry in the final decision boxes. When the p-value falls at or below the significance level, we reject H₀ and conclude that the data provide sufficient evidence to support the alternative hypothesis. When the p-value exceeds α, we fail to reject H₀—a carefully chosen phrase that acknowledges we have not proven H₀ true, but rather that the evidence was not strong enough to discard it. This linguistic precision matters: 'fail to reject' and 'accept' are not synonyms, and conflating them is one of the most common errors in applied statistics.

SECTION 4

Mathematical Framework — Formulating Hypotheses

The null and alternative hypotheses are formal mathematical statements about one or more population parameters—most commonly the population mean (μ), population proportion (p), or population variance (σ²). The null hypothesis always contains an equality condition, while the alternative hypothesis is defined by the complement of the null with respect to the parameter space.

One-Sample Tests for a Population Mean

TWO-TAILED TEST

H₀: μ = μ₀ versus Hₐ: μ ≠ μ₀

where μ is the population mean, μ₀ is the hypothesized value under the null. This test is appropriate when the researcher suspects a difference from μ₀ but has no prior reason to expect the direction of that difference.

LEFT-TAILED TEST

H₀: μ ≥ μ₀ versus Hₐ: μ < μ₀

Used when the research question asks whether the population mean is less than a specific value. Some textbooks equivalently write H₀: μ = μ₀ since the critical region only involves values below μ₀.

RIGHT-TAILED TEST

H₀: μ ≤ μ₀ versus Hₐ: μ > μ₀

Used when the research question asks whether the population mean is greater than a specific value. The rejection region lies entirely in the upper tail of the sampling distribution.

One-Sample Tests for a Population Proportion

PROPORTION HYPOTHESES

H₀: p = p₀ versus Hₐ: p ≠ p₀ (or p < p₀, or p > p₀)

where p is the population proportion and p₀ is the hypothesized proportion. The test statistic is z = (p̂ − p₀) / √(p₀(1 − p₀)/n), which follows approximately a standard normal distribution for large samples.

⚠

Convention Alert

The null hypothesis always contains the equality sign. Whether it is written as H₀: μ = μ₀, H₀: μ ≤ μ₀, or H₀: μ ≥ μ₀, the key point is that the boundary value μ₀ is always included in H₀. The alternative hypothesis never contains an equality. This convention ensures that the null specifies a precise probability model from which the sampling distribution of the test statistic can be derived.

SECTION 5

Detailed Breakdown — One-Tailed vs. Two-Tailed Tests

One of the most consequential decisions in hypothesis testing is whether to use a one-tailed or two-tailed alternative hypothesis. This choice directly affects the location of the rejection region, the magnitude of the p-value, and ultimately whether H₀ is rejected. The decision must be grounded in the research question and specified before the data are examined—post hoc switching from two-tailed to one-tailed (or vice versa) to achieve significance is a form of p-hacking.

Three normal distribution curves showing rejection regions for different test types. The two-tailed test splits α evenly between both tails, while one-tailed tests concentrate the entire α in a single tail. The shaded areas represent the rejection regions where extreme sample results would lead to rejecting H₀.

Feature	Two-Tailed Test	One-Tailed Test
Alternative Hypothesis	Hₐ: μ ≠ μ₀	Hₐ: μ > μ₀ or Hₐ: μ < μ₀
Rejection Region	Both tails (α/2 each)	One tail only (entire α)
Critical Value (α = 0.05)	z = ±1.96	z = 1.645 or z = −1.645
Statistical Power	Lower (α split across two regions)	Higher (all α concentrated in one direction)
When to Use	When no prior expectation of direction exists	When theory or prior evidence predicts a specific direction

Comparison of two-tailed and one-tailed hypothesis tests

SECTION 6

Worked Example — Formulating and Testing Hypotheses

A pharmaceutical company claims that its new sleep medication increases average nightly sleep duration. A researcher wants to test this claim against the known population average of 6.8 hours of sleep per night. She collects a random sample of n = 36 patients who use the medication and finds a sample mean of x̄ = 7.2 hours with a sample standard deviation of s = 1.2 hours. Test the company's claim at the α = 0.05 significance level.

Does the sleep medication increase average sleep duration?

1
Step 1 — Identify the Parameter and Hypothesized Value

The parameter of interest is the population mean sleep duration μ for patients taking the medication. The hypothesized value under the null is μ₀ = 6.8 hours, the known average for the general population.
2
Step 2 — State the Null and Alternative Hypotheses

Because the company claims the medication increases sleep duration, the research hypothesis is directional. We formulate a right-tailed test:

H₀: μ ≤ 6.8 Hₐ: μ > 6.8
3
Step 3 — Select the Significance Level

The problem specifies α = 0.05. Since this is a right-tailed test, the entire 5% rejection region lies in the upper tail of the sampling distribution.

α = 0.05, critical z-value = 1.645
4
Step 4 — Compute the Test Statistic

With n = 36 (large sample), we use the z-test. The test statistic is z = (x̄ − μ₀) / (s / √n) = (7.2 − 6.8) / (1.2 / √36) = 0.4 / (1.2 / 6) = 0.4 / 0.2 = 2.0.

z = 2.00
5
Step 5 — Determine the p-value

For a right-tailed test, the p-value is P(Z > 2.00). Using the standard normal table, P(Z < 2.00) = 0.9772, so the p-value = 1 − 0.9772 = 0.0228.

p-value = 0.0228
6
Step 6 — Make the Decision and State the Conclusion

Since the p-value (0.0228) is less than α (0.05), we reject the null hypothesis. There is statistically significant evidence at the 5% level to support the pharmaceutical company's claim that the medication increases average nightly sleep duration beyond 6.8 hours.

Reject H₀. The data support the claim that μ > 6.8 hours.

SECTION 7

Type I and Type II Errors — Strengths and Limitations

Because hypothesis testing is based on sample data, every decision carries a risk of error. The formal framework classifies these risks into two categories, each linked directly to the null and alternative hypotheses. Understanding these errors is essential for interpreting results correctly and for appreciating why the choice of significance level involves a genuine trade-off.

	H₀ Is Actually True	H₀ Is Actually False
Reject H₀	Type I Error (α) — False positive. We conclude an effect exists when it doesn't.	Correct Decision — Power (1 − β). We correctly detect a real effect.
Fail to Reject H₀	Correct Decision — We correctly retain a true null.	Type II Error (β) — False negative. We fail to detect a real effect.

Decision outcomes in hypothesis testing: a 2 × 2 classification

A Type I error occurs when we reject a null hypothesis that is in fact true—a false alarm. The probability of committing this error is exactly equal to the significance level α, which is why choosing a smaller α (say 0.01 instead of 0.05) reduces the false-positive rate. However, this protection comes at a cost: decreasing α while keeping sample size fixed necessarily increases the probability of a Type II error (β)—the failure to reject a false null hypothesis, thereby missing a real effect. The only way to reduce both error rates simultaneously is to increase the sample size, which narrows the sampling distribution and makes the test more sensitive.

✦

KEY TAKEAWAY

Think of the α–β trade-off like the sensitivity settings on a smoke detector. If you set the sensitivity very high (small α, hard to trigger a false alarm), the detector is less likely to go off when you burn toast—but it might also fail to detect a real fire starting slowly in the basement (high β). If you lower the threshold so it triggers easily (large α), you catch every genuine fire but also get woken up every time someone cooks bacon. Increasing sample size is like upgrading to a smarter detector that can better distinguish real smoke from cooking steam—reducing both error rates without compromising either.

SECTION 8

Connection to Confidence Intervals, Power Analysis, and Bayesian Inference

The null and alternative hypothesis framework does not exist in isolation; it connects to several more advanced inferential techniques. Understanding these connections enriches your interpretation of test results and prepares you for the methodological debates you will encounter in upper-division courses and research practice.

Concept	Relationship to H₀ / Hₐ	Key Insight
Confidence Intervals	A (1 − α)×100% confidence interval contains all values of the parameter that would not be rejected by a two-tailed test at level α.	If the hypothesized value μ₀ falls outside the confidence interval, reject H₀. The CI approach provides the same conclusion as the hypothesis test plus an estimate of effect magnitude.
Statistical Power (1 − β)	Power is the probability of correctly rejecting H₀ when Hₐ is true. It depends on α, sample size n, and the true effect size.	A priori power analysis determines the sample size needed to detect a specified effect with a desired probability—typically 0.80 or higher.
Effect Size	Effect size (e.g., Cohen's d = (μ − μ₀)/σ) quantifies the practical magnitude of the departure from H₀, independent of sample size.	A statistically significant result (reject H₀) may reflect a trivially small effect if n is very large. Always report effect size alongside the p-value.
Bayesian Hypothesis Testing	Instead of fixing α and computing a p-value, Bayesian approaches assign prior probabilities to H₀ and Hₐ and update them with data to obtain posterior probabilities.	The Bayes factor quantifies how much the data shift the odds between hypotheses, allowing direct probability statements about H₀ being true—something the frequentist framework explicitly avoids.

Connections between null/alternative hypotheses and advanced inferential methods

As you advance through your statistics curriculum, you will encounter methods such as ANOVA, chi-squared tests, and regression significance tests. Each of these involves specifying a null and alternative hypothesis, computing a test statistic, and making a decision—the same logical skeleton you have learned in this lesson. The parameters, distributions, and formulas change, but the inferential logic remains constant.

SECTION 9

Practice Problems

Question 1: CONCEPTUAL

A researcher writes: 'H₀: x̄ = 50.' Explain why this formulation is incorrect, and provide the corrected version. What is the fundamental distinction between the symbol used in H₀ and the one the researcher wrote?

Show answer

The null hypothesis must be a statement about a population parameter, not a sample statistic. The symbol x̄ denotes the sample mean, which is a known quantity once data are collected—it makes no sense to hypothesize about it. The correct formulation is H₀: μ = 50, where μ is the unknown population mean. Hypothesis testing is fundamentally about making inferences from sample statistics (like x̄) to population parameters (like μ), and confusing the two undermines the entire inferential framework. This is one of the most common conceptual errors students make, and it reflects a misunderstanding of the distinction between descriptive and inferential statistics.

Question 2: BASIC CALCULATION

A manufacturer claims that the mean lifetime of its light bulbs is at least 1,000 hours. A consumer protection agency wants to test whether the bulbs last less than 1,000 hours. State the null and alternative hypotheses, identify the type of test (one-tailed or two-tailed), and determine the critical z-value at α = 0.05.

Show answer

Since the consumer agency suspects the bulbs last less than claimed, the research hypothesis is directional (specifically, that μ is below 1,000). The hypotheses are H₀: μ ≥ 1,000 and Hₐ: μ < 1,000. This is a left-tailed (one-tailed) test because the rejection region lies entirely in the lower tail of the distribution. At α = 0.05 for a left-tailed z-test, the critical value is z = −1.645. We would reject H₀ if the computed test statistic falls below −1.645, providing evidence that the true mean lifetime is less than 1,000 hours.

Question 3: INTERMEDIATE

A university reports that the mean GPA of its graduates is 3.20. An alumni association surveys 49 recent graduates and finds x̄ = 3.35 with s = 0.56. At α = 0.01, is there sufficient evidence that the mean GPA has changed? Formulate the hypotheses, compute the test statistic, find the p-value, and state your conclusion.

Show answer

Because the question asks whether the GPA has 'changed' (not specifically increased or decreased), this requires a two-tailed test. The hypotheses are H₀: μ = 3.20 and Hₐ: μ ≠ 3.20. The test statistic is z = (x̄ − μ₀) / (s / √n) = (3.35 − 3.20) / (0.56 / √49) = 0.15 / 0.08 = 1.875. For a two-tailed test, the p-value = 2 × P(Z > 1.875) = 2 × (1 − 0.9696) = 2 × 0.0304 = 0.0608. Since the p-value (0.0608) exceeds α = 0.01, we fail to reject H₀. At the 1% significance level, there is insufficient evidence to conclude that the mean GPA has changed from 3.20. Note that the result would also not be significant at α = 0.05, but it would be marginally significant at α = 0.10—highlighting how the choice of α can change the conclusion.

Question 4: APPLIED

A clinical trial investigates whether a new vaccine reduces the infection rate below the current rate of 12% (p₀ = 0.12). In a sample of 200 vaccinated individuals, 16 become infected. Formulate the hypotheses, compute the test statistic, and determine whether the vaccine significantly reduces the infection rate at α = 0.05.

Show answer

The sample proportion is p̂ = 16/200 = 0.08. Since we are testing whether the vaccine reduces the infection rate, the hypotheses are H₀: p ≥ 0.12 and Hₐ: p < 0.12 (left-tailed). The test statistic is z = (p̂ − p₀) / √(p₀(1 − p₀)/n) = (0.08 − 0.12) / √(0.12 × 0.88 / 200) = −0.04 / √(0.000528) = −0.04 / 0.02298 ≈ −1.74. For a left-tailed test at α = 0.05, the critical value is z = −1.645. Since our test statistic (−1.74) is less than −1.645 (further into the rejection region), we reject H₀. The p-value = P(Z < −1.74) ≈ 0.0409, confirming rejection since 0.0409 < 0.05. There is statistically significant evidence at the 5% level that the vaccine reduces the infection rate below 12%.

Question 5: CRITICAL THINKING

A researcher conducts a two-tailed test of H₀: μ = 500 at α = 0.05 with a sample of n = 10,000 and obtains a test statistic of z = 2.05, yielding a p-value of 0.040. She reports: 'We found a highly significant result proving that the population mean differs from 500.' Identify at least three problems with this interpretation and explain what a more careful conclusion would state.

Show answer

First, hypothesis testing never 'proves' anything—it provides evidence for or against hypotheses. The correct language is that the data provide sufficient evidence to reject H₀, not that μ ≠ 500 has been proven. Second, a p-value of 0.040 at α = 0.05 is barely below the threshold; calling it 'highly significant' is misleading and conflates statistical significance with strength of evidence. Third, with n = 10,000, even a trivially small deviation from 500 can achieve statistical significance. The researcher should compute an effect size (e.g., Cohen's d) or construct a confidence interval to assess practical significance. If the sample mean were, say, 500.3 with σ = 50, then d = 0.3/50 = 0.006—a negligible effect. A more careful conclusion would state: 'At α = 0.05, we reject H₀ (z = 2.05, p = 0.040); however, the extremely large sample size means this result may reflect a difference too small to be practically meaningful. The effect size should be examined before drawing substantive conclusions.'

SUMMARY

Lesson Summary

Summary

Hypothesis testing provides a structured, evidence-based procedure for making decisions about population parameters using sample data. The null hypothesis (H₀) serves as the default claim—typically asserting no effect, no difference, or no relationship—while the alternative hypothesis (Hₐ) represents the competing claim that the researcher seeks evidence to support. The null always contains the equality condition, ensuring that a specific probability model can be used to derive the sampling distribution of the test statistic. The choice between a one-tailed and two-tailed test depends on whether the research question specifies a direction and must be made before examining the data.

Every test carries risks of error: a Type I error (α) occurs when a true H₀ is incorrectly rejected, and a Type II error (β) occurs when a false H₀ is not rejected. The significance level (α) is set in advance to control the false-positive rate, and increasing the sample size is the primary means of reducing both error types simultaneously. Finally, statistical significance alone does not imply practical importance—always complement your hypothesis test with effect sizes and confidence intervals to provide a complete picture of the evidence.