What Is a Null Hypothesis?
The null hypothesis (H₀) is a statement about a population parameter that assumes no effect, no difference, or a specific baseline value. It is the claim your statistical test evaluates. The alternative hypothesis (Hₐ or H₁) represents the departure from H₀ that your research is designed to detect.
The null hypothesis H₀ is the default assumption about a population parameter. It typically states that no difference, no effect, or no relationship exists, and it contains the equality condition used to calculate the test statistic.
The verbal template "there is no statistically significant difference between X and Y" is a useful starting point, but the precise mathematical form depends on the test and parameter. H₀ always refers to a population parameter, not a sample statistic. You test μ (population mean), not x̄ (sample mean).
A researcher tests whether a new drug lowers cholesterol. The null hypothesis is H₀: μtreatment = μcontrol (no difference in population means). The alternative is Hₐ: μtreatment < μcontrol (treatment group has lower mean). If p < 0.05, you reject H₀.
Null Hypothesis Examples at a Glance
The table below shows 10 research questions, their null hypotheses, and the corresponding alternative hypotheses. Scan this first to get the pattern, then read the field-specific sections for deeper context.
| # | Research Question | H₀ | Hₐ | Direction |
|---|---|---|---|---|
| 1 | Does study time improve exam scores? | H₀: ρ = 0 | Hₐ: ρ > 0 | Right-tailed |
| 2 | Does a drug lower blood pressure? | H₀: μT = μC | Hₐ: μT < μC | Left-tailed |
| 3 | Does teaching method affect test scores? | H₀: μA = μB | Hₐ: μA ≠ μB | Two-tailed |
| 4 | Does exercise reduce resting heart rate? | H₀: μd = 0 | Hₐ: μd > 0 | Right-tailed |
| 5 | Does an ad campaign increase sales? | H₀: μ ≤ 500 | Hₐ: μ > 500 | Right-tailed |
| 6 | Does sleep duration affect productivity? | H₀: β₁ = 0 | Hₐ: β₁ ≠ 0 | Two-tailed |
| 7 | Does fertilizer increase plant height? | H₀: μF = μC | Hₐ: μF > μC | Right-tailed |
| 8 | Does CBT reduce anxiety scores? | H₀: μd = 0 | Hₐ: μd > 0 | Right-tailed |
| 9 | Does training improve employee output? | H₀: μ₁ = μ₂ | Hₐ: μ₁ ≠ μ₂ | Two-tailed |
| 10 | Does product design affect satisfaction? | H₀: p₁ = p₂ | Hₐ: p₁ ≠ p₂ | Two-tailed |
Null vs Alternative Hypothesis
H₀ and Hₐ work as a complementary pair. Together they must cover every possible value of the parameter. If H₀ says μ = 70, then Hₐ says μ ≠ 70. If H₀ says μ ≤ 100, then Hₐ says μ > 100. They cannot overlap, and there is no third option.
| Feature | Null Hypothesis (H₀) | Alternative Hypothesis (Hₐ) |
|---|---|---|
| Symbol | H₀ | Hₐ or H₁ |
| Typical claim | No difference, no effect, or a specific value | Difference or effect exists |
| Equality | Contains = (or ≤ / ≥ for one-tailed) | Uses ≠, >, or < |
| What is tested | Directly tested with p-value | Supported when H₀ is rejected |
| Parameter | Population parameter (μ, p, ρ, β) | Same parameter, different claim |
| Decision language | Reject or fail to reject H₀ | Not accepted or rejected directly |
The correct phrase is "fail to reject H₀," not "accept H₀." Failing to reject means the data do not provide sufficient evidence against H₀ at the chosen significance level. It does not establish that H₀ is true.
How to Write a Null Hypothesis
Follow these seven steps to convert any research question into correctly stated hypotheses.
Identify the population
Define whose outcomes you are measuring. "All patients receiving Drug X," "all students using the new curriculum," "all social media users aged 18-34." The hypothesis is about this group, not your sample.
Identify the parameter
Choose the parameter being measured: population mean (μ), proportion (p), correlation (ρ), or regression slope (β). Write what the parameter symbol represents explicitly before your hypotheses.
Determine the comparison
Are you comparing a mean to a known value (one-sample), two group means (two-sample), proportions, or testing association (correlation/regression)? This determines your test and notation.
Write the alternative hypothesis first
The alternative captures what you are trying to find evidence for. Is it directional (one-tailed: μ > 50, μ < 50) or non-directional (two-tailed: μ ≠ 50)? Match the direction to your research question.
Construct H₀ as the complement
H₀ gets the equality. If Hₐ: μ > 50, then H₀: μ ≤ 50. If Hₐ: μ ≠ 50, then H₀: μ = 50. For the purpose of computing the test statistic, the boundary value (μ = 50) is used in either case.
Check your notation refers to the population
H₀: μ = 70 is correct. H₀: x̄ = 70 is incorrect because x̄ is a sample statistic. Population parameters in hypotheses: μ (mean), p (proportion), ρ (correlation), β (regression coefficient).
Write a plain-English interpretation
Good academic writing pairs the formula with a sentence. Example: H₀: μnew = μstandard means "The population mean exam score using the new method is the same as the standard method's population mean."
One-Tailed vs Two-Tailed Null Hypothesis Examples
The direction of Hₐ determines whether your test is one-tailed or two-tailed. The null hypothesis form adjusts accordingly.
Two-Tailed
H₀: μ = μ₀ Hₐ: μ ≠ μ₀Reject in either direction. Critical regions in both tails.
Right-Tailed
H₀: μ ≤ μ₀ Hₐ: μ > μ₀Reject only if the statistic falls in the upper tail.
Left-Tailed
H₀: μ ≥ μ₀ Hₐ: μ < μ₀Reject only if the statistic falls in the lower tail.
Null Hypothesis Examples by Field
Education
Does class size affect student performance?
H₀: μsmall = μlarge Hₐ: μsmall > μlargeRight-tailed t-test comparing mean scores of students in small (≤20) vs large (≥30) classes.
Does tutoring increase reading scores above 75?
H₀: μ ≤ 75 Hₐ: μ > 75One-sample t-test. Let μ = population mean reading score after tutoring.
Does online vs in-person instruction produce different outcomes?
H₀: μonline = μin-person Hₐ: μonline ≠ μin-personTwo-tailed independent t-test. No direction specified; either could score higher.
Is daily attendance correlated with final grade?
H₀: ρ = 0 Hₐ: ρ > 0Right-tailed Pearson correlation test. ρ = population correlation between attendance rate and grade.
Psychology
Does CBT reduce anxiety scores?
H₀: μd = 0 Hₐ: μd > 0Paired t-test. μd = mean difference (pre minus post score). Positive difference = reduction.
Does sleep deprivation worsen memory recall?
H₀: μdeprived = μrested Hₐ: μdeprived < μrestedLeft-tailed two-sample t-test. Sleep-deprived group expected to score lower.
Is social media use associated with lower wellbeing scores?
H₀: β₁ = 0 Hₐ: β₁ < 0Left-tailed regression test. β₁ = slope for social media hours predicting wellbeing.
Does exercise type affect depression scores across three groups?
H₀: μ₁ = μ₂ = μ₃ Hₐ: at least one μ differsOne-way ANOVA. Groups: aerobic, resistance, and control. See the ANOVA guide.
Healthcare
Does Drug A lower systolic blood pressure more than Drug B?
H₀: μA = μB Hₐ: μA < μBLeft-tailed two-sample t-test. Lower systolic value = more reduction. These are statistical examples; consult a physician for clinical guidance.
Does a screening test detection rate equal 90%?
H₀: p = 0.90 Hₐ: p ≠ 0.90Two-tailed one-proportion z-test. p = population proportion detected correctly.
Does physical therapy reduce pain scores after 6 weeks?
H₀: μd = 0 Hₐ: μd > 0Paired t-test. Pain measured before and after for the same patients. μd = mean pre-minus-post difference.
Is complication rate the same across four surgical techniques?
H₀: p₁ = p₂ = p₃ = p₄ Hₐ: at least one pᵢ differsChi-square test of homogeneity. Each pᵢ is the population complication proportion for technique i.
Biology
Does nitrogen fertilizer increase plant height?
H₀: μF = μC Hₐ: μF > μCRight-tailed two-sample t-test. μF = mean height with fertilizer, μC = mean height without fertilizer.
Does temperature affect bacterial growth rate?
H₀: ρ = 0 Hₐ: ρ ≠ 0Two-tailed correlation test. ρ = population correlation between temperature (°C) and colony growth rate.
Is genotype distribution consistent with Hardy-Weinberg proportions?
H₀: observed = expected frequencies Hₐ: they differChi-square goodness-of-fit test. The null specifies the expected frequencies from Hardy-Weinberg theory.
Business
Does a new ad campaign lift average weekly sales above $500?
H₀: μ ≤ 500 Hₐ: μ > 500Right-tailed one-sample t-test. μ = population mean weekly sales after the campaign.
Does a pricing change affect customer retention rate?
H₀: p₁ = p₂ Hₐ: p₁ ≠ p₂Two-tailed two-proportion z-test. p₁ = retention at old price, p₂ = retention at new price.
Does a website redesign increase conversion rate?
H₀: pnew ≤ pold Hₐ: pnew > poldRight-tailed two-proportion z-test. This is an A/B test structure. See the A/B test calculator.
Does training program type affect employee output across departments?
H₀: μ₁ = μ₂ = μ₃ Hₐ: at least one μ differsOne-way ANOVA comparing mean output across three training programs.
Null Hypothesis Examples by Statistical Test
Each statistical test has a standard null hypothesis form. Matching the correct form to your test prevents the most common mistake: writing a population claim that the test cannot evaluate.
One-Sample t-Test
Scenario: A factory claims its batteries last 100 hours on average. Quality control samples 30 batteries (x̄ = 97 h, s = 8 h). Is the claim accurate?
H₀: μ = 100 (population mean lifetime equals 100 hours)
Hₐ: μ ≠ 100 (two-tailed; mean could be higher or lower)
See the full one-sample t-test guide for the complete calculation procedure.
Independent-Samples t-Test
Scenario: Do men and women spend different average amounts on groceries weekly?
H₀: μ₁ = μ₂ (population mean spending is equal between groups)
Hₐ: μ₁ ≠ μ₂ (two-tailed; no prior expectation of direction)
See the two-sample t-test guide and the equal vs unequal variance guide.
Paired Samples t-Test
Scenario: Does a mindfulness intervention reduce stress scores in the same employees? Measured before and after an 8-week program.
H₀: μd = 0 (mean of paired differences equals zero; no change)
Hₐ: μd > 0 (pre-score minus post-score is positive; stress decreased)
See the paired samples t-test guide.
One-Proportion Test
Scenario: A company claims 60% of customers are satisfied. A survey of 200 customers finds 112 satisfied (56%). Test the claim.
H₀: p = 0.60 (population satisfaction proportion equals 60%)
Hₐ: p ≠ 0.60 (two-tailed)
Two-Proportion Test
Scenario: Does email subject line A produce a higher open rate than B?
H₀: p₁ = p₂ (no difference in population open rates)
Hₐ: p₁ > p₂ (right-tailed; testing if A beats B specifically)
One-Way ANOVA
Scenario: Do four different diets produce different weight-loss outcomes after 12 weeks?
H₀: μ₁ = μ₂ = μ₃ = μ₄ (all four population group means are equal)
Hₐ: at least one μᵢ differs from the others
ANOVA only tells you that a difference exists, not which groups differ. Use post-hoc tests (Tukey HSD, Bonferroni) afterward. See the ANOVA guide.
Pearson Correlation
Scenario: Is there a linear relationship between years of experience and salary?
H₀: ρ = 0 (no linear correlation in the population)
Hₐ: ρ > 0 (right-tailed; expecting positive correlation)
See the Pearson correlation guide.
Chi-Square Test of Independence
Scenario: Is there an association between smoking status and lung disease diagnosis in a sample of 500 patients?
H₀: Smoking status and lung disease are independent (no association in the population)
Hₐ: Smoking status and lung disease are not independent (association exists)
See the chi-square test guide and the chi-square table.
Linear Regression
Scenario: Does hours of study predict exam score? Testing whether the regression slope is meaningful.
H₀: β₁ = 0 (population slope = 0; study hours have no linear relationship with exam score)
Hₐ: β₁ > 0 (positive slope; more study time predicts higher scores)
See the simple linear regression guide.
Logistic Regression
Scenario: Does age predict the probability of a customer purchasing a premium plan?
H₀: β₁ = 0 (equivalently, odds ratio OR = 1; age has no effect on purchase odds)
Hₐ: β₁ ≠ 0 (equivalently, OR ≠ 1; age affects purchase odds)
A zero log-odds coefficient corresponds to an odds ratio of exactly 1. See the logistic regression guide.
Worked Example: From Research Question to Decision
This example walks through all eight steps, from identifying the parameter to writing the conclusion in academically correct language.
Research question: Does a new online tutoring program increase average math scores above 70?
Population and parameter: Let μ = the population mean math score for all students who complete the tutoring program. The comparison value is μ₀ = 70.
State the hypotheses:
H₀: μ ≤ 70 (the program does not increase the mean above 70)
Hₐ: μ > 70 (the program does increase the mean above 70)
This is a right-tailed test because the research question asks whether scores are above 70.
Significance level: α = 0.05
Sample data: n = 25 students, x̄ = 74.2, s = 9.3. The population standard deviation is unknown, so the one-sample t-test applies. df = 24.
Test statistic:
SE = s/√n = 9.3/√25 = 9.3/5 = 1.86
t = (x̄ − μ₀) / SE = (74.2 − 70) / 1.86 = 4.2 / 1.86 = 2.258
P-value: For t = 2.258, df = 24, one-tailed: p ≈ 0.017. Critical value from the t-distribution table: t* = 1.711.
Decision: p = 0.017 < α = 0.05 → Reject H₀. Also: t = 2.258 > t* = 1.711 confirms rejection.
Conclusion (academically correct): "There is sufficient evidence at the 5% significance level to reject H₀ (t(24) = 2.258, p = 0.017, one-tailed). The sample mean of 74.2 is statistically significantly greater than 70. This does not prove the program works in all contexts, and effect size (Cohen's d ≈ 0.45) suggests a moderate practical effect."
Phrases to avoid: "We proved Hₐ." / "H₀ is false." / "The p-value is the probability H₀ is true."
✅ Reject H₀. The data provide sufficient statistical evidence that the program raises mean scores above 70, though effect size and replication should be considered before drawing policy conclusions.
Null Hypothesis Builder
Select a test type, enter your variables, and the builder generates the correct H₀ and Hₐ with plain-English wording.
Null Hypothesis Builder
Hypothesis Testing Visualizer
Drag the slider to move the observed test statistic and see how the p-value and decision change. The shaded red region is the rejection zone.
Rejection Region and p-value
A small p-value means the observed result would be unusual if H₀ were true. It is not the probability that H₀ is true.
Reject vs Fail to Reject H₀
These two outcomes have specific meanings in statistics. Using the wrong language signals a misunderstanding of the procedure.
The observed result is statistically unlikely under H₀. There is sufficient evidence against H₀ at the chosen significance level. This supports Hₐ but does not prove it.
The data do not provide sufficient evidence against H₀. H₀ is not proven true. The result may reflect a real null effect, low statistical power, or a small sample.
| Outcome | Correct interpretation | Incorrect interpretation |
|---|---|---|
| Reject H₀ | Evidence against H₀ is sufficient at α | H₀ is proven mathematically false |
| Fail to reject H₀ | Evidence against H₀ is insufficient | H₀ has been proven true |
| p = 0.03, α = 0.05 | Reject H₀. The result is statistically significant. | There is a 3% chance H₀ is true |
| p = 0.08, α = 0.05 | Fail to reject H₀. Insufficient evidence at this α. | Accept H₀. No effect exists. |
Can You Prove a Null Hypothesis?
No. In conventional null hypothesis significance testing (NHST), failing to reject H₀ only means the data did not provide enough evidence against it at the chosen significance level. A sample that produces p = 0.30 does not tell you that H₀ is true; it tells you the sample was not unusual enough under H₀ to reject it.
This matters practically. A study with n = 10 might fail to reject H₀ because the sample was too small to detect a real effect, not because no effect exists. Statistical power — the probability of detecting an effect when one exists — is the relevant consideration.
When researchers need to demonstrate equivalence or similarity (for example, showing a generic drug is as effective as a brand-name drug), they use equivalence testing or non-inferiority testing, not conventional hypothesis testing. Bayesian methods can also provide evidence in support of H₀ through Bayes factors.
A result can be statistically significant (p < 0.05) with a tiny, practically irrelevant effect if the sample size is very large. Always report effect size (Cohen's d, r, η²) alongside p-values. See the effect size guide.
Null Hypothesis Formula Cheat Sheet
| Parameter | Test / Context | H₀ | Hₐ (two-tailed) |
|---|---|---|---|
| Population mean | One-sample t/z | H₀: μ = μ₀ | Hₐ: μ ≠ μ₀ |
| Difference in means | Two-sample t | H₀: μ₁ − μ₂ = 0 | Hₐ: μ₁ − μ₂ ≠ 0 |
| Paired mean difference | Paired t | H₀: μd = 0 | Hₐ: μd ≠ 0 |
| Population proportion | One-proportion z | H₀: p = p₀ | Hₐ: p ≠ p₀ |
| Difference in proportions | Two-proportion z | H₀: p₁ − p₂ = 0 | Hₐ: p₁ − p₂ ≠ 0 |
| Correlation | Pearson/Spearman | H₀: ρ = 0 | Hₐ: ρ ≠ 0 |
| Regression slope | Linear regression | H₀: β₁ = 0 | Hₐ: β₁ ≠ 0 |
| Odds ratio | Logistic regression | H₀: OR = 1 | Hₐ: OR ≠ 1 |
| Multiple means | One-way ANOVA | H₀: μ₁ = ... = μk | Hₐ: at least one μ differs |
| Categorical association | Chi-square | H₀: variables are independent | |
Common Mistakes When Writing H₀
| # | Mistake | Correct Approach |
|---|---|---|
| 1 | Writing H₀: x̄ = 70 (sample statistic) | Write H₀: μ = 70 (population parameter) |
| 2 | Setting H₀ = research hypothesis (what you expect to find) | H₀ is the skeptical default (no effect); Hₐ is what you are testing for |
| 3 | Missing the equality condition in H₀ | H₀ always contains = (or ≤/≥ for one-tailed) |
| 4 | Saying "accept H₀" after failing to reject | Say "fail to reject H₀"; it is not proven true |
| 5 | "p-value is the probability H₀ is true" | p-value is P(data this extreme or more | H₀ is true) |
| 6 | Non-significant result proves no effect | It may reflect low power or small sample; report power and effect size |
| 7 | Choosing one-tailed after seeing results | Specify tail direction before data collection |
| 8 | Confusing statistical and practical significance | Always report effect size alongside p-value |
Practice Questions
Write H₀ and Hₐ for each scenario, then reveal the answer.
A coffee shop claims its average cup size is 12 oz. You suspect cups are smaller. What are H₀ and Hₐ?
Left-tailed one-sample t-test. You suspect cups are smaller (below 12 oz), so the alternative points left. μ = population mean cup size in oz.
A researcher wants to know if a coin is fair. They plan to test if the probability of heads differs from 0.5. What are H₀ and Hₐ?
Two-tailed one-proportion z-test. p = population probability of heads. "Differs from" means no direction is specified, so two-tailed.
A psychologist tests whether relaxation training reduces cortisol levels in 20 participants, measuring cortisol before and after. What are H₀ and Hₐ?
Paired t-test. μd = mean of (pre minus post) differences. A positive μd means cortisol decreased after training. Right-tailed because direction is specified.
A university compares dropout rates for online and in-person programs. They want to know if the rates differ. What are H₀ and Hₐ?
Two-tailed two-proportion z-test. p₁ = population dropout rate (online), p₂ = population dropout rate (in-person). "Differ" implies no direction; two-tailed.
An ecologist investigates whether water temperature (°C) is linearly related to fish abundance. What are H₀ and Hₐ for a Pearson correlation test?
Two-tailed Pearson correlation test. ρ = population linear correlation between temperature and fish abundance. No direction is specified; either positive or negative correlation is of interest.
A food manufacturer tests whether salt content differs across four production lines. What are H₀ and Hₐ?
One-way ANOVA. μᵢ = population mean salt content for production line i. ANOVA tests the omnibus null that all group means are equal.
A market analyst uses simple linear regression to test whether advertising spend (in thousands) predicts sales revenue. What is the hypothesis for the slope?
Right-tailed regression slope test. β₁ = population slope of sales on advertising spend. The analyst expects positive returns on ad spend, so Hₐ is directional.
A pharmaceutical company tests a generic drug for non-inferiority against a branded version. What structure does H₀ take, and why is this different from a standard equivalence test?
In a non-inferiority test, Δ is the pre-specified non-inferiority margin. H₀ says the generic is inferior by more than Δ. Rejection of H₀ supports non-inferiority. This is the reverse of standard hypothesis testing logic: rejecting H₀ supports the claim you want to make. See equivalence testing literature (Schuirmann, 1987).
In logistic regression predicting loan default (yes/no), you want to test whether income (continuous predictor) is associated with default probability. State H₀ in two equivalent forms.
A zero log-odds coefficient implies an odds ratio of e&sup0; = 1, meaning income has no multiplicative effect on the odds of default. Both forms are tested using the Wald test or likelihood ratio test in logistic regression output.
A chi-square goodness-of-fit test is used to assess whether observed blood type frequencies match theoretical Hardy-Weinberg proportions. What does H₀ state?
Chi-square goodness-of-fit. H₀ specifies the exact expected proportions from theory. df = number of categories minus 1 minus the number of parameters estimated from the data. See the chi-square guide.
Frequently Asked Questions
The null hypothesis (H₀) is a default statement about a population parameter, typically claiming no effect, no difference, or a specific value. It is what your statistical test evaluates. If the data are sufficiently unlikely under H₀, you reject it.
H₀: μ = 70. This says the population mean equals 70. For instance: "The population mean exam score for students using the new curriculum is 70." The alternative might be Hₐ: μ ≠ 70 (two-tailed) or Hₐ: μ > 70 (right-tailed).
Not always. H₀ can specify any claimed parameter value, not just zero. For example, H₀: μ = 100 (testing whether a mean equals a known benchmark) or H₀: p = 0.60 (testing a specific proportion). The "no effect" form (e.g., H₀: μ₁ = μ₂) is a common special case where the specified value is zero difference.
Yes, for one-tailed tests. H₀: μ ≤ 50 paired with Hₐ: μ > 50 is a valid right-tailed formulation. In practice, the test statistic is calculated using the boundary equality (μ = 50), and the direction of Hₐ determines the rejection region.
H₀: ρ = 0. This states that the population correlation coefficient equals zero; there is no linear relationship between the two variables. ρ (rho) is the population parameter. The sample correlation r is the estimate of ρ.
H₀: μ₁ = μ₂ = ... = μk. The omnibus ANOVA null states that all k population group means are equal. Hₐ states that at least one mean differs. ANOVA does not specify which groups differ; post-hoc tests (e.g., Tukey HSD) do that.
It means the sample data do not provide sufficient evidence against H₀ at the chosen significance level. This is not the same as proving H₀ is true. Low power, small samples, or high variability can all cause a failure to reject a false H₀ (Type II error).
A two-tailed test has H₀: μ = μ₀ and Hₐ: μ ≠ μ₀. Rejection can occur in either direction. The significance level α is split between both tails (α/2 in each). Use this when you are interested in any departure from H₀, not just a specific direction.
For an individual predictor: H₀: β₁ = 0. This states that the population regression slope is zero, meaning the predictor has no linear relationship with the outcome after accounting for other predictors in the model. The corresponding Hₐ is β₁ ≠ 0 (two-tailed) or > 0 / < 0 (one-tailed).
For a chi-square test of independence: H₀: the two categorical variables are independent (no association). For a goodness-of-fit test: H₀: the observed frequencies match the theoretical/expected distribution. Chi-square tests are always right-tailed (larger χ² values indicate more deviation from H₀).