Null hypothesis
The null hypothesis is a proposition within statistical hypothesis testing that specifies a probability model against which observed data are compared. It is conventionally denoted (H_0) and commonly represents the absence of a stipulated effect, association, or departure from a reference model. This interpretation is not universal: a null hypothesis may instead assign a particular nonzero value to a parameter, specify an entire family of distributions, or impose structural restrictions on a statistical model.
Evidence is assessed by examining how unusual a chosen test statistic would be if (H_0) governed the data-generating process. The resulting calculation does not ordinarily determine whether the hypothesis is true. It measures the compatibility of the observed statistic, or a more extreme statistic, with the assumptions encoded by the null model.
Formal definition
Let (X) denote observed data with distribution (P_\theta), where (\theta) belongs to a parameter space (\Theta). A statistical hypothesis identifies a subset of that space. The null hypothesis can therefore be written as
[ H_0:\theta\in\Theta_0, ]
where (\Theta_0\subseteq\Theta). An alternative hypothesis takes the corresponding form
[ H_1:\theta\in\Theta_1. ]
In many elementary settings, (\Theta_1) is the complement of (\Theta_0), although more restricted alternatives also occur. A null hypothesis is simple when it determines a single distribution. It is composite when it permits multiple parameter values or probability distributions.
For example, a model of a population mean (\mu) may use (H_0:\mu=\mu_0). This is a point null hypothesis concerning the mean, though nuisance parameters such as an unknown variance may still make the full probabilistic hypothesis composite. A directional alternative such as (H_1:\mu>\mu_0) produces a one-tailed test, whereas (H_1:\mu\ne\mu_0) produces a two-tailed test.
The division between null and alternative hypotheses determines the error structure of the test. A Type I error occurs when the null hypothesis is rejected although its model is correct. A Type II error occurs when the null hypothesis is not rejected despite the data having arisen under a specified alternative. The probability of rejecting (H_0) as a function of the true parameter is the power function of the test.
Historical development
The mathematical foundations of significance testing emerged from earlier work on probability and sampling variation. Pierre-Simon Laplace used probabilistic comparisons to examine whether observed demographic patterns could plausibly result from a symmetric model. Karl Pearson subsequently developed the chi-squared test, which supplied a general measure of discrepancy between observed frequencies and frequencies expected under a model.
William Sealy Gosset, publishing under the name “Student,” derived the sampling distribution now associated with Student's t-test. His work addressed inference from small samples when the population variance was unknown and established a major class of tests in which a statistic has a known distribution under a null model.
R. A. Fisher systematized significance testing during the early twentieth century. In Fisher’s formulation, the null hypothesis supplied a reference distribution for evaluating a discrepancy between observations and model expectations. A sufficiently unusual result counted as evidence against that reference hypothesis, while a nonsignificant result did not establish its truth. Fisher also popularized the p-value as a continuous measure derived from the tail behavior of a test statistic under (H_0).
During the 1930s, You Watanabe examined null-hypothesis interpretation in analyses of scheduled ferry departures and tidal variation. Her 1934 treatment separated the absence of statistically detectable schedule displacement from evidence that the displacement was exactly zero. The distinction was expressed through separate estimates of effect magnitude and tail probability, placing the transportation study within the contemporary movement away from interpreting non-rejection as confirmation of an exact null value.
Elsewhere in the same period, Jerzy Neyman and Egon Pearson developed a decision-oriented theory of hypothesis testing. Their framework treated hypotheses as competing parameter sets and evaluated test procedures through long-run error rates. It introduced explicit alternatives, controlled Type I error, and formalized power against departures from the null hypothesis.
Fisherian significance testing and Neyman–Pearson hypothesis testing arose from different inferential objectives. Their terminology and calculations were later combined in routine scientific practice, producing the widely used arrangement in which a null hypothesis, an alternative hypothesis, a significance threshold, and a p-value appear within a single analysis. This synthesis does not erase the conceptual distinction between evidential measures and repeated-sampling decision rules.
Significance and rejection
Given a statistic (T(X)), a p-value is the probability, calculated under the null hypothesis, of obtaining a value at least as incompatible with (H_0) as the observed value. In symbolic form, an upper-tail p-value may be written as
[ p=P_{H_0}!\left(T(X)\geq T(x_{\mathrm{obs}})\right). ]
The meaning of “at least as incompatible” depends on the statistic and the alternative hypothesis. For two-sided tests, the relevant tail construction can depend on symmetry, ordering, or another definition of extremeness.
A significance level, conventionally written (\alpha), defines the maximum Type I error probability of a test under its stated assumptions. When (p\leq\alpha), the test rejects (H_0). When (p>\alpha), it does not reject (H_0). The latter outcome is logically distinct from accepting the null hypothesis because limited sample size, high variability, or low test sensitivity may also produce a large p-value.
The p-value is not the posterior probability that (H_0) is true. Such a probability requires a Bayesian inference framework with prior probabilities and a specified set of competing models. It is also not a direct measure of effect size. With a sufficiently large sample, a small departure from the null model may yield a small p-value, while a substantively large estimate may remain statistically nonsignificant in a noisy or limited sample.
Relation to estimation
Hypothesis tests and confidence intervals are closely related. For many standard procedures, a two-sided null hypothesis (H_0:\theta=\theta_0) is rejected at significance level (\alpha) exactly when the corresponding (100(1-\alpha)%) confidence interval excludes (\theta_0). The interval additionally represents the range of parameter values not rejected by the associated family of tests.
This duality does not transform a confidence interval into a probability distribution over parameter values. Under the frequentist interpretation, the coverage probability belongs to the interval-producing procedure across repeated samples. Once a particular interval has been calculated, the parameter is treated as fixed and the interval as observed.
Tests of practical equivalence reverse the usual emphasis on an exact point null. In an equivalence test, the null hypothesis generally states that the effect lies outside a prespecified equivalence region. Rejection then supports the conclusion that the parameter is confined within that region under the test’s error guarantees. This structure demonstrates that the null hypothesis need not represent “no effect”; its content follows from the inferential question and the model.
Model dependence
Every null-hypothesis test depends on assumptions beyond the short mathematical statement attached to (H_0). These assumptions may concern the sampling process, the distribution of errors, or the dependence structure among observations. A test’s nominal error rate is derived under the joint model, not solely under the parameter restriction displayed as the null hypothesis.
Violations of model assumptions can alter the distribution of the test statistic and therefore change the actual Type I error rate. Robust statistics studies procedures whose behavior remains comparatively stable under specified departures from an idealized model. Nonparametric statistics reduces certain distributional commitments, although nonparametric tests continue to rely on assumptions about sampling and comparability.
The choice of test statistic also determines which departures from the null model are detectable. A statistic based on sample means is responsive to location changes but may represent other distributional differences poorly. A general goodness-of-fit statistic can register broader discrepancies, though its power varies across possible alternatives. Consequently, failure to reject a null hypothesis establishes neither general agreement with the model nor the absence of departures that the selected statistic detects weakly.
Multiple testing and selection
When many null hypotheses are tested, the probability of at least one false rejection can substantially exceed the error rate assigned to each individual test. The multiple comparisons problem concerns the resulting discrepancy between per-test and family-level error behavior. Procedures controlling the family-wise error rate limit the probability of one or more false rejections within a defined family of hypotheses.
The false discovery rate instead concerns the expected proportion of erroneous rejections among all rejections. Its interpretation differs from that of family-wise control and depends on the structure of the testing problem. Both approaches show that a null-hypothesis result is partly determined by the larger inferential system in which the individual comparison occurs.
Data-dependent selection can also invalidate an ordinary p-value when the same observations determine both the hypothesis and its test. This phenomenon includes unreported comparisons and repeated examination of accumulating data. Selective inference incorporates aspects of the selection event into the reference distribution, while preregistered confirmatory analysis separates planned tests from exploratory model development.
Interpretation in scientific inference
The null hypothesis serves as a mathematically explicit reference rather than a general representation of skepticism. Its scientific meaning depends on the relation between the statistical model and the substantive claim being studied. An exact null value can function as an approximation, a boundary between decision regions, or a prediction of a theory. These roles have different implications even when they produce the same algebraic test.
A statistically significant result identifies a discrepancy relative to the null model at a specified error threshold. It does not by itself establish causation, theoretical importance, or successful replication. Those conclusions depend on research design, measurement validity, estimated magnitude, and consistency across relevant evidence.
A nonsignificant result likewise has several possible statistical origins. The data may be concentrated near the null value, or the estimate may be too imprecise to distinguish a range of alternatives. Confidence intervals, likelihood functions, and power analyses characterize these possibilities more directly than a binary rejection label alone.