One-Tailed vs Two-Tailed Test
By CalcMulti Editorial Team··6 min read
When running a hypothesis test, you must decide whether to test for an effect in one specific direction (one-tailed) or any direction (two-tailed). This choice affects your critical values, your p-value, and the conclusions you can draw.
The decision must be made before collecting data based on your research question — not after seeing results. Switching to a one-tailed test because your two-tailed result came out at p = 0.07 is p-hacking.
Side-by-Side Comparison
| Property | One-Tailed Test | Two-Tailed Test |
|---|---|---|
| Alternative hypothesis | H₁: μ > μ₀ (right) or H₁: μ < μ₀ (left) | H₁: μ ≠ μ₀ (either direction) |
| Critical region | One tail only (left or right) | Split between both tails |
| Critical z for α=0.05 | z* = 1.645 (right-tailed) | z* = ±1.960 |
| Critical z for α=0.01 | z* = 2.326 (right-tailed) | z* = ±2.576 |
| p-value relationship | p_one-tailed = p_two-tailed / 2 (if correct direction) | p_two-tailed = 2 × p_one-tailed |
| Statistical power | More power in predicted direction | Power split across both directions |
| Misses effects in | Opposite direction | Nothing — detects both directions |
| When to use | Strong directional prediction before data collection | When any difference is relevant |
| Risk of misuse | p-hacking if chosen after seeing results | Lower (safer default) |
| Default choice | — | Yes — most situations |
When to Use a One-Tailed Test
A one-tailed test is justified when you have a strong directional prediction established before collecting data AND when an effect in the opposite direction would be impossible, irrelevant, or would be treated identically to no effect.
Valid one-tailed scenarios: (1) You are testing whether a new drug lowers blood pressure — an increase would be a failure and you would not act on it. (2) Quality control: testing whether a batch exceeds a minimum strength threshold — only "too weak" matters. (3) Non-inferiority trials: testing whether a generic drug is "not worse" than the brand drug.
The key justification: what would you conclude if the effect were in the opposite direction? If "the new treatment is significantly worse" would be just as important to you as "no effect," then a two-tailed test is correct. If you would genuinely act the same way for no-effect and opposite-direction, a one-tailed test may be justified.
Why Two-Tailed Is the Default
Two-tailed tests are appropriate when you want to detect any difference, regardless of direction. This is the case in the vast majority of research.
Examples where two-tailed is correct: comparing means of two groups (either group could score higher); testing whether two proportions differ; measuring whether a correlation exists (could be positive or negative); checking whether a treatment has any effect (good or bad).
The two-tailed test is conservative — it requires a larger test statistic to reject H₀. This is appropriate because it protects against claiming effects in both directions when none exists.
When in doubt: always use a two-tailed test. You can always switch to a one-tailed test with strong justification, but you cannot justify it post-hoc.
P-Value Relationship Between One- and Two-Tailed Tests
For the same data, if the effect is in the predicted direction: p_one-tailed = p_two-tailed / 2.
Example: You run a t-test and get t = 1.85, df = 30. Two-tailed p = 0.074. If you had predicted the correct direction beforehand, one-tailed p = 0.037 (significant at α=0.05).
This halving of the p-value is why one-tailed tests are tempting — and why choosing one-tailed AFTER seeing the p-value is problematic. If you saw p = 0.074 and then switched to one-tailed to get p = 0.037, you have inflated your false positive rate to ~10% while reporting results as if α = 5%.
If the effect is in the opposite direction: one-tailed p = 1 − p_two-tailed/2. A strong effect in the wrong direction will have a very high one-tailed p-value.
Worked Example
Scenario: A researcher tests whether caffeine improves reaction time. Sample of n=25, mean reaction time 342ms vs control 360ms, SD=40ms pooled.
t = (342−360)/(40/√25) = −18/8 = −2.25, df=24.
Two-tailed: H₁: μ_caffeine ≠ μ_control. p = 0.034. Reject H₀ at α=0.05. Conclusion: caffeine significantly changes reaction time.
One-tailed (correct direction): H₁: μ_caffeine < μ_control (caffeine improves = lower time). p = 0.017. Reject H₀ at α=0.05 with more power.
One-tailed (wrong direction): H₁: μ_caffeine > μ_control. p = 0.983. Fail to reject H₀ — the one-tailed test in the wrong direction completely misses the real effect.
Lesson: if you had incorrectly predicted caffeine would slow reaction time, the one-tailed test would fail to detect the actual improvement.
The P-Hacking Warning
Choosing one-tailed vs two-tailed AFTER seeing your results is a form of p-hacking that inflates the false positive rate. If your two-tailed p = 0.07 and you switch to one-tailed to get p = 0.035, you are misrepresenting your actual Type I error rate.
The justification for a one-tailed test must be pre-specified: written in your research protocol, pre-registration, or analysis plan before data collection begins.
Many reviewers and journals require explicit justification for one-tailed tests. If you cannot justify it before seeing the data, use two-tailed.
Summary
Use two-tailed tests by default. Use one-tailed only with a pre-specified, strongly justified directional hypothesis where an effect in the opposite direction would be irrelevant or impossible.
- Two-tailed tests detect effects in either direction — safer and more general
- One-tailed tests have more power in the predicted direction, but completely miss opposite effects
- p_one-tailed = p_two-tailed / 2 (if correct direction) — do not exploit this post-hoc
- Critical value advantage: z* = 1.645 (one-tailed) vs 1.960 (two-tailed) at α=0.05
- When in doubt: always two-tailed — the cost is slightly lower power, the benefit is scientific integrity
Related Calculators
Frequently Asked Questions
Educational use only. Content is based on publicly documented mathematical formulas and reviewed for accuracy by the CalcMulti Editorial Team. Last updated: March 2026.