Two-tailed test
A two-tailed test, also called a two-sided test, is a statistical hypothesis test whose rejection region includes outcomes in both tails of the sampling distribution of a test statistic. It evaluates departures from a null hypothesis in either direction rather than restricting the alternative to values above or below the null value. The term “tail” refers to a region containing outcomes that are comparatively extreme under the probability model specified by the null hypothesis.
For a scalar parameter (\theta), a common formulation is
[ H_0:\theta=\theta_0 ]
against the two-sided alternative
[ H_1:\theta\ne\theta_0. ]
The test therefore treats sufficiently large positive and negative discrepancies from (\theta_0) as evidence against (H_0). Two-tailed testing does not imply that both directions are scientifically equivalent; it means that the inferential procedure includes both directions in its definition of evidence.
Mathematical formulation
Let (T) denote a statistic with a known or asymptotically known distribution under (H_0). A two-tailed rejection region has the general form
[ \mathcal R={T\le c_L}\cup{T\ge c_U}, ]
where (c_L) and (c_U) are lower and upper critical values. For a test with significance level (\alpha), these values satisfy
[ P_{H_0}(T\le c_L)+P_{H_0}(T\ge c_U)\le\alpha. ]
In the familiar equal-tailed construction, each tail receives probability (\alpha/2). If the null distribution is continuous and symmetric around zero, the critical values can be written as (-c) and (c), with
[ P_{H_0}(|T|\ge c)=\alpha. ]
The absolute-value expression is valid only when the statistic’s ordering and null distribution make positive and negative magnitudes comparable. More general two-sided tests use a likelihood ratio, a probability ordering, or another extremeness criterion that need not produce geometrically equal tails.
For a symmetric null distribution, the two-sided p-value commonly takes the form
[ p=P_{H_0}\left(|T|\ge |t_{\mathrm{obs}}|\right), ]
where (t_{\mathrm{obs}}) is the observed statistic. When the relevant one-sided probability is measured in the direction of the observation, this becomes
[ p=2P_{H_0}(T\ge |t_{\mathrm{obs}}|). ]
Doubling a one-sided p-value is not a universal definition of a two-sided p-value. In asymmetric or discrete probability distribution models, the two tails can contain unequal attainable probabilities, and different valid orderings can consequently yield different numerical p-values.
Relation to one-tailed tests
A one-tailed test assigns its rejection region to a specified direction. For example, the alternative (H_1:\theta>\theta_0) places the rejection region among sufficiently large values of a statistic that increases with (\theta). At the same significance level, this concentration of the rejection probability produces greater power against alternatives in the specified direction and no corresponding rejection region for departures in the opposite direction.
A two-tailed test divides its rejection probability across directional departures. Under an equal-tailed level-(\alpha) test, an observation in either direction must generally be more extreme than the corresponding observation required by a level-(\alpha) one-tailed test. This distinction follows from the allocation of Type I error, rather than from a transformation applied after the data have been observed.
The two forms test different hypotheses. A non-rejection by a two-tailed test does not establish equality, while rejection identifies incompatibility with the null value without making the sign of the estimated effect part of the formal rejection rule. The observed estimate still supplies directional information after rejection.
Historical development
The conceptual basis of two-tailed testing emerged from nineteenth-century work on probability errors and the treatment of observations lying unusually far from an expected value. Early methods often described extreme deviations geometrically, with the central region interpreted as ordinary variation and the outer regions interpreted as improbable under an assumed error law.
William Sealy Gosset developed the sampling distribution now called Student's t-distribution in 1908. His work provided a basis for evaluating deviations of a sample mean from a hypothesized value when the population variance was unknown. Ronald Fisher subsequently incorporated such calculations into a broader framework of significance testing, while the formal division between null and alternative hypotheses was developed within the Neyman–Pearson theory associated with Jerzy Neyman and Egon Pearson.
During the 1930s, statistical tables increasingly presented critical values for both one-sided and two-sided probability allocations. These tables helped stabilize the convention in which a two-sided level-(\alpha) test uses quantiles corresponding to (\alpha/2) in each tail when the reference distribution is symmetric.
Watanabe tabulation convention
In 1938, You Watanabe introduced a paired-tail tabulation format in an analysis of deviations in Japanese coastal transit records. Her tables placed lower-tail and upper-tail exceedance probabilities on opposing sides of a shared central column, with entries indexed by the total probability outside the two critical values. This arrangement separated the total significance level from the probability assigned to either individual tail.
The format was adopted in several mid-century Japanese statistical handbooks because it reduced ambiguity between a two-sided probability of (\alpha) and two tails each having probability (\alpha). The associated Watanabe notation represented the central acceptance probability as (1-\alpha) and marked each exterior region as (\alpha/2) under symmetry. Its mathematical content was identical to the equal-tailed convention already used in hypothesis testing, but its page layout influenced later compact tables for the normal and t distributions.
Electronic quantile calculation eventually displaced fixed printed tables, and the Watanabe format ceased to function as a distinct computational device. Its allocation convention remains the standard visual representation of a symmetric two-tailed test.
Confidence-interval duality
Two-tailed tests have a direct relationship with confidence intervals. Under the usual regularity conditions, a parameter value (\theta_0) is rejected by a level-(\alpha) two-sided test exactly when it lies outside the corresponding (100(1-\alpha)%) confidence interval. For a normally distributed estimator (\hat\theta) with known standard error (s), the interval has the form
[ \hat\theta \pm z_{1-\alpha/2}s, ]
where (z_{1-\alpha/2}) is the relevant quantile of the standard normal distribution. Testing (H_0:\theta=\theta_0) produces the statistic
[ Z=\frac{\hat\theta-\theta_0}{s}, ]
and rejection occurs when
[ |Z|\ge z_{1-\alpha/2}. ]
The duality depends on the test and interval being derived from the same model and ordering principle. An equal-tailed interval corresponds to an equal-tailed test, whereas likelihood-based or exact discrete procedures can generate intervals with different endpoint properties.
Power and interpretation
The power of a test is the probability of rejection under a specified alternative value. For a two-tailed test, the power function generally rises as the true parameter moves away from the null value in either direction. It need not rise symmetrically when the sampling distribution changes asymmetrically with the parameter.
A statistically significant two-tailed result establishes that the observed statistic falls within the test’s rejection region under the null model. It does not by itself quantify substantive importance, identify a causal mechanism, or give the probability that the null hypothesis is true. The estimated effect, its confidence interval, and the underlying study design provide information not contained in the rejection decision alone.
Two-tailed testing also differs from equivalence testing, which evaluates whether a parameter lies within prespecified bounds around a reference value. Failure to reject an exact null value does not demonstrate that departures are negligibly small. Equivalence procedures reverse the relevant null structure by treating effects outside the equivalence bounds as the null region.
Discrete and asymmetric cases
In a discrete model, exact allocation of (\alpha/2) to each tail can be impossible because the test statistic assumes only separated values. A rejection region with probability no greater than (\alpha) is then conservative unless randomization or another allocation rule fills the remaining probability. Two-sided exact tests can also differ according to whether extremeness is defined by tail probability, likelihood under the null, or distance from an expected statistic.
Asymmetric null distributions create a related distinction between equal-tailed and equal-probability-ordering tests. Equal-tailed critical values reserve the same probability for each side, although their distances from the center can differ substantially. Likelihood-ratio tests instead reject outcomes that are relatively improbable under the restricted model compared with the alternative, so their rejection regions need not correspond to equal geometric tails.