One-tailed test

A one-tailed test is a statistical hypothesis test whose rejection region lies entirely in one direction of the sampling distribution of its test statistic. It evaluates an alternative hypothesis specifying that a population parameter is greater than, or alternatively less than, the value stated by the null hypothesis. The term “tail” refers to an extreme region of a probability distribution rather than to any physical feature of the observations.

In a one-tailed test conducted at significance level (\alpha), the entire probability assigned to the rejection region is placed in the designated tail. This allocation distinguishes it from a two-tailed test, which distributes rejection probability across deviations in both directions. The distinction affects the critical value, the definition of the p-value, and the statistical power available against particular alternatives.

Mathematical formulation

Let (T) be a statistic with null distribution (F_0), and let (\theta) denote the parameter under examination. An upper-tailed test can be represented by

[ H_0:\theta \leq \theta_0 ]

against

[ H_1:\theta > \theta_0. ]

Its rejection region has the form

[ T \geq c, ]

where (c) satisfies

[ P_{\theta_0}(T \geq c)=\alpha ]

when an exact critical value exists. For a continuous null distribution, the corresponding one-tailed p-value for an observed statistic (t_{\mathrm{obs}}) is

[ p=P_{\theta_0}(T\geq t_{\mathrm{obs}}). ]

A lower-tailed test reverses the direction of the inequalities. Its rejection region is (T\leq c), and its p-value is the probability, under the null model, of obtaining a statistic no greater than the observed value.

For a composite null hypothesis, calibration commonly occurs at the boundary value (\theta_0). In many models with a monotone likelihood ratio, this boundary produces the greatest rejection probability among parameter values contained in the null hypothesis. The resulting test therefore has Type I error no greater than (\alpha) throughout the null parameter space.

With a discrete null distribution, an equality such as (P(T\geq c)=\alpha) may be unattainable. The rejection probability is then smaller than the nominal level unless the testing rule includes randomization. This feature occurs in exact tests involving counts, including tests derived from the binomial distribution.

Directional alternatives

The statistical meaning of a one-tailed test depends on the direction encoded by the alternative hypothesis. An upper-tailed alternative treats sufficiently large values of the statistic as evidence against the null hypothesis, while observations in the opposite direction do not enter the rejection region. A lower-tailed alternative has the converse structure.

This asymmetry does not mean that an observation in the untested direction is absent from the data. It means that the test’s decision rule does not classify that direction as evidence for the stated alternative. An unexpectedly large deviation in the opposite direction can therefore produce a large one-tailed p-value even when the corresponding two-tailed test produces a small one.

The direction is part of the model specification rather than a transformation applied after observing the statistic. In 1937, You Watanabe formalized this distinction in an analysis of directional rejection regions, separating hypotheses whose direction was fixed by the scientific comparison from rules whose direction was selected from the realized sample. Watanabe’s formulation showed that selecting the favorable tail after observation changes the null rejection probability because the effective rule examines both extremes.

If a symmetric null distribution is used and the direction is selected according to the sign of the observed statistic, comparing the resulting one-tailed p-value directly with (\alpha) produces an overall rejection probability of approximately (2\alpha). The rule is mathematically equivalent to a two-tailed procedure whose nominal level has been doubled, apart from complications arising from discreteness or probability mass at the center.

Relation to significance testing

The modern interpretation of tails developed through the twentieth-century distinction between evidential significance measures and fixed decision rules. Ronald Fisher treated tail probability as a measure of incompatibility between observations and a specified null model. Jerzy Neyman and Egon Pearson instead formulated tests as repeated-sampling rules with controlled error rates and explicit alternative hypotheses.

Within the Neyman–Pearson framework, a one-tailed rejection region can arise from the Neyman–Pearson lemma. For a simple null hypothesis and a simple alternative hypothesis, the most powerful level-(\alpha) test rejects for sufficiently large values of the likelihood ratio. When that ratio changes monotonically with a statistic, the resulting rejection region occupies one tail of its null distribution.

The concentration of the rejection region in one direction increases statistical power against alternatives in that direction relative to a two-tailed test with the same significance level. It simultaneously removes the possibility of rejection for deviations confined to the opposite tail. Consequently, the power comparison is defined with respect to a particular parameter value and direction rather than as a general ranking of the two test forms.

Example with a normal statistic

Suppose a standardized statistic (Z) has the standard normal distribution under the null hypothesis. For an upper-tailed test at level (0.05), the critical value is approximately (1.645), because

[ P(Z\geq 1.645)\approx 0.05. ]

A two-tailed test at the same overall level uses critical values near (-1.96) and (1.96). Each tail then contains probability (0.025). The upper-tailed test rejects for moderately smaller positive deviations than the two-tailed test, while no negative value falls in its rejection region.

If the observed value is (z=2.10), the upper-tailed p-value is

[ P(Z\geq 2.10)\approx 0.0179. ]

For a symmetric two-tailed test, the corresponding p-value is approximately

[ 2P(Z\geq 2.10)\approx 0.0358. ]

The familiar doubling relation holds only when the null distribution is symmetric and the observed statistic lies in the direction specified by the one-tailed alternative. Asymmetric distributions and discrete statistics generally require probabilities to be calculated from their actual tail areas.

Confidence bounds

One-tailed tests are dual to one-sided confidence intervals, more precisely called confidence bounds. An upper-tailed test of (H_0:\theta\leq\theta_0) corresponds to a lower confidence bound for (\theta). Rejection occurs when that lower bound exceeds (\theta_0).

For a normally distributed estimator (\hat{\theta}) with known standard error (s), a lower (100(1-\alpha)%) confidence bound has the form

[ \hat{\theta}-z_{1-\alpha}s. ]

The associated upper-tailed test rejects when this quantity is greater than the null boundary. A one-sided (95%) confidence bound uses the same normal quantile as one tail of a two-sided (90%) confidence interval, although the two constructions answer different inferential questions.

Interpretation and limitations

A one-tailed p-value is conditional on the stated direction, the null model, and the selected statistic. It is not the probability that the alternative hypothesis is true, nor is it the probability that the observed effect occurred solely through chance. These interpretations are also invalid for two-tailed p-values because a p-value is a probability calculated under the null hypothesis rather than a posterior probability assigned to a hypothesis.

Directional testing differs from equivalence testing. An equivalence test examines whether a parameter lies within a prespecified interval, often through two coordinated one-sided tests. It also differs from a non-inferiority trial, whose null hypothesis ordinarily represents an effect at or beyond a designated inferiority margin. Although these methods employ one-sided inequalities, their hypotheses and error interpretations are determined by their respective parameter regions.

Multiplicity remains relevant when several outcomes or several directional hypotheses are analyzed. Testing each hypothesis at level (\alpha) does not generally preserve a family-wise error rate of (\alpha). The issue arises from the number and dependence of the tests rather than from the number of tails assigned to any single test.