Other meanings of Null hypothesis
STATISTICS
The null hypothesis is a statistical hypothesis that an observed effect is absent or attributable to chance. Researchers test it against an alternative hypothesis using data, a test statistic, and a prespecified decision rule; the result is evidence for or against the null, not proof that either hypothesis is true.
The null hypothesis provides a reference model in which the effect under investigation is zero, absent, or no different from a specified value. In a clinical trial, for example, it may state that two treatments have equal mean outcomes; in genetics, it may state that two variables are independent. The competing alternative hypothesis describes an effect or relationship that the study is designed to detect.1
Testing begins by translating a research question into formal hypotheses and identifying the population parameter at issue. The null is often written H₀, while the alternative is H₁ or Hₐ. It is not necessarily a claim that nothing in the world is happening: it is a deliberately defined probability model against which the observed data can be compared.
Null-hypothesis testing measures how incompatible the observed data are with the null model. A test statistic summarizes the data, and the p-value gives the probability, assuming H₀ is true, of obtaining a result at least as extreme as the one observed under the chosen test.2
Before examining the result, investigators commonly select a significance level, α, such as 0.05. If the p-value is at or below α, they reject H₀; otherwise, they fail to reject it. “Fail to reject” does not mean “accept”: a nonsignificant result can reflect a genuinely small effect, noisy measurements, a small sample, or inadequate statistical power. The threshold is a convention and does not measure the probability that H₀ is true.1
A statistically significant result indicates incompatibility with the null model at the chosen analytical threshold, not practical importance or a guaranteed discovery. Effect size, uncertainty, study design, measurement quality, and subject-matter context are needed to judge the result; a confidence interval can show a range of effects compatible with the data and make precision more visible than a decision alone.3
Two errors are central. A Type I error rejects a true null hypothesis, while a Type II error fails to reject a false one. Repeated testing, optional stopping, selective reporting, and unplanned subgroup analyses can increase false-positive findings, especially when many hypotheses are examined without adjustment.4 The American Statistical Association therefore recommends reporting estimates, uncertainty, design choices, and scientific context rather than treating a p-value as a binary verdict.1
The null hypothesis need not assert complete absence: it can specify a nonzero benchmark, such as a treatment difference of exactly two units, or a particular correlation. Tests can also be one-sided or two-sided, depending on whether departures in one direction or either direction are scientifically relevant. The choice must be made from the research question, not selected after seeing the data.
Modern practice distinguishes Fisher’s evidence-oriented use of significance tests from the Neyman–Pearson framework of long-run error rates and decision rules; many textbooks combine elements of both. A p-value is also sensitive to sample size: very large studies can make trivial effects statistically significant, while small studies may miss consequential effects. Large-scale research additionally requires attention to multiple comparisons, replication, preregistration, and the possibility that a published finding is a false positive within a broader research record.45
A null-hypothesis test is a model-based inferential procedure; its conclusions depend on the design, assumptions, analysis, and data-generating context.
Help improve the encyclopedia. Reports go straight to the site manager.