One-Tailed vs Two-Tailed Tests: Decide Before You See the Result

Start with the research question, not the observed statistic

Suppose a researcher is comparing a new programme with an existing programme. Let

Δ = μ_new − μ_existing.

The statistical test begins by specifying a null hypothesis and an alternative hypothesis. Hypotheses concern population parameters or models rather than the particular sample outcome that happens to be observed (Moore et al., 2021; Devore & Berk, 2012).

Newton and Rudestam (1999) describe a one-tailed test as appropriate when the research hypothesis specifies the direction of the hypothesized difference; a two-tailed formulation states that a difference exists without specifying its direction. Meier et al. (2014) likewise distinguish a one-tailed test for a hypothesis that specifies a direction from a two-tailed test for a problem in which departure in either direction matters.

The practical distinction

The wording of the substantive question should determine which row applies—not the sign of the estimate after the data have been analyzed (Moore et al., 2021).

What changes when you choose one tail or two?

The choice changes which sample outcomes count as sufficiently strong evidence against H₀.

A rejection region is the set of values of the test statistic for which the null hypothesis will be rejected. Devore and Berk (2012) describe a hypothesis-testing procedure in terms of a test statistic and a rejection region: H₀ is rejected when the observed test statistic falls within that region.

These distinctions follow from the alternative hypothesis (Devore & Berk, 2012; Moore et al., 2021).

Alpha is allocated differently

The significance level α controls the probability of rejecting the null hypothesis when it is true under the specified testing procedure. Changing from a one-sided to a two-sided alternative changes where that probability is placed in the null distribution (Moore et al., 2021; Devore & Berk, 2012).

Consider a standard Normal test with α = .05.

Devore and Berk (2012) formulate two-tailed rejection regions using upper and lower critical values based on α/2, whereas one-tailed tests place the rejection region in the direction specified by the alternative.

The practical consequence is important: with the same overall α, an observation does not have to travel as far into the specified tail to enter the rejection region of a one-sided test.

One-tailed and two-tailed p-values answer different questions

A p-value is calculated assuming the null hypothesis and measures how extreme the observed test statistic is in the direction or directions specified by the hypotheses. The alternative hypothesis determines which outcomes count as “as extreme or more extreme” than the observed result (Moore et al., 2021).

For a positive observed z-statistic and an upper-tailed alternative,

H_A: θ > θ₀,

the one tailed p value is the probability in the upper tail at or beyond the observed statistic.

For a symmetric null distribution and a two-sided alternative,

H_A: θ ≠ θ₀,

extreme results in both directions count. Moore et al. (2021), for example, calculate a two-sided Normal p-value by adding the probabilities in the two symmetric tails; equivalently in that setting, it is twice the corresponding one-tail probability.

Numerical example: one-sided significant, two-sided not significant

Directional hypothesis

Suppose first that the prespecified research question is directional:

H₀: θ ≤ θ₀

H_A: θ > θ₀.

The upper-tail probability beyond z = 1.75 is approximately

p_one-sided ≈ .040.

Because .040 < .05, the result falls in the rejection region for the one-sided test. The researcher rejects H₀ in favor of the prespecified positive-direction alternative.

Nondirectional hypothesis

Now suppose instead that the research question is nondirectional:

H₀: θ = θ₀

H_A: θ ≠ θ₀.

For the symmetric standard Normal distribution, the corresponding two tailed p value is approximately

p_two-sided ≈ 2(.040) = .080.

Because .080 > .05, the result does not meet the .05 criterion for the two-sided test. The researcher therefore fails to reject H₀.

Why these conclusions are not contradictory

The conclusions differ because the hypotheses are different.

Those are not the same inferential question. They assign the significance level differently, define different rejection regions, and calculate extremeness according to different alternatives (Moore et al., 2021; Devore & Berk, 2012).

It would therefore be misleading to describe the example as one statistical test giving two contradictory answers. Two different hypotheses were tested.

What happens to power?

Power is the probability that a test rejects the null hypothesis when a specified alternative is true; equivalently, it is 1 − β, where β is the probability of a Type II error for that alternative (Newton & Rudestam, 1999; Devore & Berk, 2012).

The one-sided design concentrates its rejection region in the direction specified by the research hypothesis. At the same overall α, this produces a less extreme critical threshold in that direction than the corresponding symmetric two-sided test. Consequently, a properly specified one-sided test can have greater ability to detect alternatives in its designated direction. The tradeoff is that the testing rule is specifically constructed around that direction rather than departures on both sides (Devore & Berk, 2012).

This is the source of the apparent attraction of a one-sided test: if the true effect lies in the prespecified direction, concentrating the rejection region there can increase power.

Power itself also depends on the particular alternative being considered. Effects farther from the null are easier to distinguish, and sample size, variability, significance level, and the alternative effect all influence a study's ability to reject the null (Moore et al., 2021; Devore & Berk, 2012; Newton & Rudestam, 1999).

When to use a one tailed test

The defensible question is not “Would a one-tailed test give me significance?” It is:

Before seeing the data, was the research hypothesis genuinely directional?

A one-sided test is most defensible when the research question specifies a direction in advance and the inferential claim being tested is genuinely about that direction. A two-sided test is appropriate when departures in either direction answer the research question or when no specific direction is firmly justified before examining the data (Meier et al., 2014; Moore et al., 2021; Newton & Rudestam, 1999).

Moore et al. (2021) make the timing requirement especially explicit: the alternative should express the hopes or suspicions brought to the data, and framing the alternative after looking at the observations to match what they show is improper. If a specific direction is not firmly established beforehand, they instruct the researcher to use a two-sided alternative.

Newton and Rudestam (1999) similarly emphasize objectivity in significance testing and discuss selecting the cutoff before “peeking at the results.”

Why “I predicted the direction” is not enough after the fact

Suppose a researcher originally asks whether an intervention changes an outcome. The data then show a positive estimate with a two-sided p-value of .08. Recasting the analysis afterward as a positive one-sided test because the resulting p-value is approximately .04 changes the question after seeing which direction makes the evidence look strongest.

The problem is not that .04 was calculated incorrectly. The problem is that the directional alternative was selected using information from the result it is supposed to test.

That undermines the prespecified interpretation of the significance level. Moore et al. (2021) explicitly warn against looking at the data first and then framing the alternative to fit the observations. More generally, their discussion of the use and abuse of significance tests cautions against searching for significance (Moore et al., 2021).

A practical decision framework

Before running the test, document these decisions:

1. What population parameter is being tested? State the parameter or contrast clearly.
2. What is the null hypothesis? Specify the null value or boundary.
3. What substantive claim would answer the research question? Is the question about any difference, or specifically an increase or decrease?
4. Would a result in the opposite direction still be scientifically or practically relevant to the question? If so, a two-sided alternative is usually aligned with that question.
5. Was the direction specified before examining the result? Do not choose the tail after seeing the estimate, test statistic, or p-value.
6. What significance level will be used? State α before testing and recognize that one- and two-sided procedures allocate it differently.
7. What p-value corresponds to the actual alternative? Interpret the p-value for the hypothesis that was prespecified rather than selecting whichever version is smaller.
8. What is the correct decision language? Reject H₀ when the evidence satisfies the prespecified criterion; otherwise fail to reject H₀. Do not translate p > α into “accept the null” (Moore et al., 2021; Devore & Berk, 2012; Newton & Rudestam, 1999).

Common mistakes

The decision rule to remember

The choice between a one tailed vs two tailed test belongs at the hypothesis-design stage.

Use a one-sided alternative when the research question itself is genuinely directional and that direction is specified before examining the result. Use a two-sided alternative when the question concerns differences in either direction or when a directional alternative cannot be justified in advance (Meier et al., 2014; Moore et al., 2021; Newton & Rudestam, 1999).

A one-sided test can produce a smaller p-value and greater power in its specified direction because the rejection region is concentrated in that tail. That is not a statistical shortcut. It is the consequence of asking a narrower question (Devore & Berk, 2012).

References

Devore, J. L., & Berk, K. N. (2012). Modern mathematical statistics with applications (2nd ed.). Springer.

Meier, K. J., Brudney, J. L., & Bohte, J. (2014). Applied statistics for public and nonprofit administration (9th ed.). Cengage Learning.

Moore, D. S., McCabe, G. P., & Craig, B. A. (2021). Introduction to the practice of statistics (10th ed.). Macmillan Learning.

Newton, R. R., & Rudestam, K. E. (1999). Your statistical consultant: Answers to your data analysis questions. SAGE Publications.