A null hypothesis (denoted H₀ or H0) is the default claim in a statistical test that no effect, no difference, or no relationship exists between the variables being studied. It is the position a researcher assumes true until the data provide sufficient evidence to reject it. Every standard significance test — a t-test, chi-square test, ANOVA, correlation test, or regression coefficient test — is built around evaluating evidence against a stated null hypothesis, not around directly proving the researcher’s actual hypothesis of interest.
This guide covers what a null hypothesis actually claims, why hypothesis testing is built around it rather than around proving an effect exists, how to write one correctly, worked examples across disciplines, and the misconceptions that trip up even experienced researchers when interpreting test results.
What the null hypothesis actually claims
An observation counts as an instance of “the null hypothesis” if it makes a precise, testable statement that a specified parameter equals a specific value — almost always zero effect, zero difference, or independence. Common forms include:
- No difference between groups: “The mean test score of Group A equals the mean test score of Group B” (μ₁ = μ₂).
- No relationship between variables: “There is no correlation between hours of sleep and reaction time” (ρ = 0).
- No effect of an intervention: “The new drug produces no change in blood pressure relative to placebo.”
- Independence: “Voting preference is independent of age group” (used in chi-square tests of independence).
The null hypothesis is always stated as an equality or “no effect” claim, never as an inequality — that role belongs to the alternative hypothesis, discussed below.
Why hypothesis testing is built around the null, not the alternative
The logic traces back to philosopher of science Karl Popper’s falsifiability criterion: a scientific claim is only meaningful if it is capable, in principle, of being contradicted by observation. Popper argued theories can be falsified by evidence but never conclusively proven — no finite amount of confirming data rules out every possible future counter-example. Modern null-hypothesis significance testing, developed from Ronald Fisher’s and later Jerzy Neyman and Egon Pearson’s frameworks, applies this logic directly: it is statistically far more tractable to show that data are inconsistent with a single, precise “no effect” claim than to directly demonstrate that a specific effect exists.
In practice this means a significance test never sets out to prove the researcher’s hypothesis of interest. Instead it asks: if the null hypothesis were actually true, how likely is it that we would see data this extreme, purely by chance? That conditional probability is the p-value. When the p-value falls below a pre-specified threshold (the significance level, conventionally α = 0.05), the result is treated as strong enough evidence to reject the null hypothesis. When it does not, the correct conclusion is that the test fails to reject the null hypothesis — not that the null hypothesis has been proven true.
Null hypothesis vs. alternative hypothesis
Every hypothesis test pairs a null hypothesis with an alternative hypothesis (H₁ or Ha), the claim that an effect, difference, or relationship does exist. The two must be mutually exclusive and collectively exhaustive for the parameter in question. A test can result in one of two outcomes: reject H0 in favor of Ha, or fail to reject H0. It cannot “accept” either hypothesis outright, because absence of evidence against H0 is not evidence that H0 is exactly true.
For a full side-by-side breakdown of how the two are defined, symbolized, and interpreted — including when to use a directional versus non-directional alternative — see CASRAI’s dedicated comparison: Null Hypothesis vs. Alternative Hypothesis.
How to write a null hypothesis
A well-formed null hypothesis follows a consistent structure regardless of field:
- Identify the parameter of interest. A population mean, proportion, correlation coefficient, regression coefficient, or odds ratio — whatever quantity the study is actually testing.
- State the “no effect” value for that parameter. Usually zero difference, zero correlation, or a stated ratio of 1 (for odds/risk ratios).
- Write it as a precise equality, in testable, falsifiable language. “There is no difference in recovery time between Treatment A and Treatment B” is testable; “Treatment A doesn’t really help” is not, because it isn’t tied to a measurable parameter.
- Match it to a directional or non-directional alternative. A non-directional (two-tailed) alternative simply claims the parameter differs from the null value in either direction; a directional (one-tailed) alternative predicts the direction of the difference in advance (e.g., “the new method produces a higher mean score”). The choice should be made before data collection, based on genuine prior expectation, not after seeing which direction the data happen to point — choosing the tail after seeing the result inflates the false-positive rate.
Worked examples
The following are illustrative, generic examples used to demonstrate the format — they are not drawn from any specific published study.
| Research question | Null hypothesis (H0) | Alternative hypothesis (Ha) |
|---|---|---|
| Does a new teaching method affect exam scores? | Mean exam score is the same under both teaching methods (μ₁ = μ₂) | Mean exam scores differ between methods (μ₁ ≠ μ₂) |
| Is caffeine intake related to typing speed? | There is no correlation between caffeine intake and typing speed (ρ = 0) | Caffeine intake is correlated with typing speed (ρ ≠ 0) |
| Does a drug lower blood pressure more than placebo? | Mean blood-pressure reduction is equal for drug and placebo (μdrug = μplacebo) | Mean blood-pressure reduction is greater for drug than placebo (μdrug > μplacebo, one-tailed) |
| Is job satisfaction independent of department? | Job satisfaction category is independent of department | Job satisfaction category is associated with department |
Common misconceptions
“Fail to reject” does not mean “accept” or “prove”
A non-significant result means the data did not provide strong enough evidence against H0 — it does not establish that H0 is true. The study may simply lack statistical power to detect a real, smaller effect. Journals and reporting guidelines (see CASRAI’s guide to reporting p-values) consistently caution against phrases like “proves no effect” for this reason.
A significant result does not measure the size or importance of an effect
Rejecting the null hypothesis tells you an effect is unlikely to be exactly zero; it says nothing about how large or practically meaningful that effect is. The American Statistical Association’s 2016 statement on p-values states directly that “smaller p-values do not necessarily imply the presence of larger or more important effects, and larger p-values do not imply a lack of importance or even lack of effect.” Reporting an effect size alongside the test result, ideally with a confidence interval, is the standard corrective.
The null hypothesis is not automatically “no difference exists in reality”
It is a specific statistical statement chosen because it is testable, not a claim about what the researcher expects or hopes to find. Researchers typically hope to reject it.
P-hacking and hypothesizing after results are known (HARKing)
Deciding on, or switching, a directional alternative hypothesis after inspecting the data — or running many tests and reporting only the significant ones — inflates the true false-positive rate above the nominal significance level. Pre-registration of the null/alternative hypothesis pair before data collection is one of the standard safeguards against this in open-science practice.
The null hypothesis and statistical significance
The decision to reject or fail to reject H0 depends on three interacting quantities:
- Significance level (α): the pre-specified threshold, conventionally 0.05, below which a p-value is treated as sufficient evidence to reject H0.
- Type I error: incorrectly rejecting a null hypothesis that is actually true (a false positive); its probability is bounded by α.
- Type II error: incorrectly failing to reject a null hypothesis that is actually false (a false negative); its probability depends on the true effect size, sample size, and statistical power.
Because rejecting H0 is a probabilistic decision, not a certainty, every hypothesis test carries some residual chance of both error types. Adequately powering a study before data collection — choosing a sample size large enough to detect an effect of a plausible, meaningful size — is the primary way researchers control Type II error risk.
Frequently asked questions
What is the symbol for the null hypothesis?
The null hypothesis is conventionally written H₀ (H-naught) or H0. The alternative hypothesis is written H₁ or Ha.
Can you ever prove the null hypothesis true?
No. A statistical test can only provide evidence against H0 (leading to rejection) or insufficient evidence against it (failing to reject). Failing to reject H0 is not proof that no effect exists — it may simply reflect low statistical power.
What happens if you reject a true null hypothesis?
This is a Type I error, a false positive. Its probability is controlled by the chosen significance level (α), conventionally set at 0.05, meaning a 5% chance of this error when H0 is actually true.
Is the null hypothesis always “no effect”?
In practice, yes — it is almost always framed as no difference, no relationship, or no change, because that is the precise, testable claim that makes significance testing mathematically tractable. Some specialized tests (e.g., equivalence testing) invert this logic and test a null of “a meaningful difference exists,” but this is the exception, not the default.
How is the null hypothesis different from a research hypothesis?
A research hypothesis is the substantive prediction a study is designed to investigate — usually the claim that ultimately corresponds to the alternative hypothesis in a formal statistical test. See CASRAI’s guide on how to write a strong research hypothesis for the distinction and worked examples.
