The null hypothesis says there is no effect, difference, or relationship of interest; the alternative hypothesis says there is an effect, difference, or relationship. In a statistical test, researchers assess how well the observed data fit the null, then use that evidence to decide whether to reject it in favor of the alternative—not to prove either hypothesis with certainty.
What the Null and Alternative Hypotheses Mean
A hypothesis is a claim that can be evaluated using data. In hypothesis testing, the null hypothesis, usually written as H₀, is the reference claim used to calculate what results would be expected if there were no effect of the kind being tested. It often states that a population parameter equals a particular value, or that two groups have no difference.
The alternative hypothesis, often written as H₁ or Hₐ, describes a departure from the null. It represents the effect, difference, or association the analysis is designed to detect. The alternative may say only that a value differs from the null value, or it may specify a direction, such as being greater or smaller.
These are competing statistical statements about a population or process, not descriptions of individual observations. A sample supplies evidence relevant to the claims, but the hypotheses themselves concern the broader quantity being studied.
Side-by-Side Comparison
| Feature | Null hypothesis (H₀) | Alternative hypothesis (H₁ or Hₐ) |
|---|---|---|
| Role | Reference claim evaluated by the test | Competing claim that the test may support |
| Typical wording | No effect, no difference, or a specified value | An effect, a difference, or a directional change |
| Example for a mean | The population mean equals a stated value | The population mean differs from, exceeds, or falls below that value |
| What a test concludes | Reject it, or fail to reject it | Evidence may support it if the null is rejected |
| What it does not establish | That the null is certainly true | That the alternative is certainly true or that an effect is important |
In many introductory tests, equality belongs in the null statement. For instance, a null might specify that a population mean equals a benchmark, while the alternative says it is not equal to that benchmark. The exact form depends on the research question and statistical method.
Why Hypothesis Testing Starts With the Null
A test needs a defined reference model. The null provides one: assuming it is true, the method describes how much sample variation would be expected from random sampling or other modeled randomness. The test statistic summarizes the data in a way that can be compared with this reference.
A p-value is the probability, calculated under the null model and the test’s assumptions, of getting a result at least as incompatible with the null as the result observed. A small p-value can count as evidence against the null under a decision rule chosen for the analysis. It is not the probability that the null is true, nor the probability that the result happened “by chance” in an unrestricted sense.
Researchers commonly set a significance level, denoted by alpha, before interpreting the test. If the p-value is at or below that threshold, the usual decision is to reject the null. Otherwise, the decision is to fail to reject it. The threshold is a rule for controlling a type of error in repeated use of the procedure; it does not turn uncertain evidence into proof.
How to Choose the Hypotheses
Begin with the research question and identify the population quantity being tested: a mean, proportion, difference between groups, or relationship between variables, for example. Then define the null value or no-effect claim and decide whether the question calls for a directional or nondirectional alternative.
- Use a two-sided alternative when a difference in either direction matters. Its form is “not equal to.”
- Use a one-sided alternative when the question specifically predicts a direction, such as “greater than” or “less than,” and a change in the opposite direction would not count as evidence for the stated claim.
- Specify the measure and population so the claim is clear. “The groups differ” is incomplete unless it is clear which outcome and which groups are meant.
- Choose the direction before examining the results when possible. Picking a favorable direction after seeing the data can distort the interpretation of the test.
The hypotheses should match the analysis. A test about a population mean does not, by itself, test a claim about a median. Likewise, a test of association does not automatically establish a causal effect.
Concrete Examples of Each Hypothesis
Null example: A course instructor wants to evaluate whether a new study format changes average quiz performance compared with the usual format. The null could state that the population mean score is the same for students using either format. This is a claim of no difference in the particular outcome being measured; it does not claim every student earns the same score.
Alternative example: For that same question, a two-sided alternative would state that the population mean scores differ between formats. If the research question had specified an expected increase in advance, a directional alternative could state that the mean under the new format is higher. The direction must reflect the question, not merely whichever direction the sample happens to show.
These examples also show why “the null is always no change” is an oversimplification. A null hypothesis can specify a nonzero benchmark, such as a population mean equal to a stated value. “No effect” is shorthand for no departure from the particular effect or value encoded in the model.
What Rejecting or Failing to Reject Means
When a test rejects the null, the data are considered sufficiently inconsistent with the null under the test’s assumptions and decision threshold. This is evidence favoring the alternative as framed by the test. It does not prove that the alternative is true, identify the cause of an observed pattern, or establish that the size of the effect matters in practice.
When a test fails to reject the null, it means the evidence did not meet the chosen standard for rejection. It does not show that the null is true or that there is no meaningful effect. The study may have too little information, measurements may be noisy, or the effect may be smaller than the analysis can reliably detect.
It is often useful to report an effect estimate and a confidence interval alongside the test result. These provide information about the estimated magnitude and uncertainty, which a reject-or-fail-to-reject decision alone cannot communicate. Statistical significance and practical importance are related questions, but they are not the same.
Common Mistakes and Misconceptions
- “The p-value is the chance the null is true.” It is calculated assuming the null model, and instead concerns how unusual the data would be under that model.
- “Failing to reject proves no effect.” It only says the test did not provide enough evidence to reject the null by its chosen rule.
- “Rejecting the null proves the alternative.” A test provides evidence with uncertainty; its conclusion depends on assumptions, study design, measurement, and the chosen analysis.
- “A significant result must be important.” A small effect can be statistically detectable, while an important effect may be estimated imprecisely. Interpret magnitude and context too.
- “The alternative is always directional.” It can be nondirectional, stating that a parameter differs from the null value in either direction.
- “A hypothesis test proves causation.” Statistical significance alone does not establish causality. Study design and potential confounding matter.
How to Remember the Difference
Think of the null as the test’s default reference claim and the alternative as the departure the researcher wants to evaluate. The null says what the data would be compared against; the alternative says what kind of difference or relationship would challenge that reference.
A quick check is to ask: “What exact claim would count as no departure?” That belongs in the null. Then ask: “What departure is the question asking me to detect?” That belongs in the alternative. Check that the wording matches the parameter and that the alternative’s direction reflects the research question.
Frequently Asked Questions
Can the null hypothesis be accepted?
In standard significance testing, the careful wording is “fail to reject the null,” rather than “accept it.” Not rejecting does not prove the null; it indicates that the test did not find sufficient evidence against it.
Is the alternative hypothesis the researcher’s prediction?
It can express the research prediction, especially when it specifies a direction. But it is a statistical claim, not a guarantee of what the data will show. A nondirectional alternative can represent a question about any difference, without predicting which way it goes.
Can the null hypothesis say there is a relationship?
Yes. The null is not always phrased as “nothing is related.” Depending on the test, it may specify a particular relationship or parameter value. The key is that it defines the reference claim against which the alternative is assessed.
What is the difference between a one-sided and two-sided alternative?
A one-sided alternative specifies a direction, such as greater than a benchmark. A two-sided alternative says the value differs in either direction. The choice should follow the question and be made before looking at results, rather than chosen to make a result appear significant.
Do null and alternative hypotheses apply only to experiments?
No. They are used in many statistical analyses, including studies that compare groups, estimate proportions, or examine relationships. The interpretation still depends on how the data were collected and whether the test’s assumptions are reasonable.