Statistics Concepts
One-Tailed vs Two-Tailed Tests: What Is the Difference?
A two-tailed test asks whether a value differs from the null value in either direction and splits α between both tails. A one-tailed test asks about one direction only and puts all of α in one tail. When the result goes the predicted way, the one-tailed p-value is half the two-tailed one: a z of 1.80 has p = 0.0359 one-tailed but 0.0719 two-tailed. Choose the direction before you see the data.
The three ways to state the alternative
| Test | Null hypothesis | Alternative hypothesis | Rejection region |
|---|---|---|---|
| Two-tailed | μ = μ₀ | μ ≠ μ₀ | Both tails, α/2 in each |
| Right-tailed | μ ≤ μ₀ (or μ = μ₀) | μ > μ₀ | Upper tail, all of α |
| Left-tailed | μ ≥ μ₀ (or μ = μ₀) | μ < μ₀ | Lower tail, all of α |
The alternative hypothesis is what you set out to show; the null hypothesis is the claim of no difference. The null and alternative hypothesis guide covers how to write them, and hypothesis testing the full procedure.
What changes in the numbers
| Quantity | Two-tailed | One-tailed |
|---|---|---|
| z critical value, α = 0.05 | ±1.960 | 1.645 |
| z critical value, α = 0.01 | ±2.576 | 2.326 |
| t critical value, α = 0.05, 20 df | ±2.086 | 1.725 |
| p-value for z = 1.80 | 0.0719 | 0.0359 |
| p-value for t = 2.10, 15 df | 0.0531 | 0.0265 |
The one-tailed p-value is exactly half the two-tailed one for a symmetric distribution, when the observed effect lies in the predicted direction. Use the p-value calculator to see the tails for any statistic.
A case where the choice flips the conclusion
Suppose a t-test gives t = 2.10 with 15 degrees of freedom and the test is run at α = 0.05.
- Two-tailed: p = 0.0531, which is above 0.05, so the result is not statistically significant.
- One-tailed, in the predicted direction: p = 0.0265, which is below 0.05, so the result is significant.
This is the temptation: the same data are significant or not depending on the tail. That is why the choice must be justified and recorded before the data are collected. Switching to the one-tailed test after seeing a p-value of 0.0531 is the same as changing α to 0.10 for a two-tailed test, and it inflates the false-positive rate. See type 1 and type 2 errors.
When a one-tailed test is legitimate
- The direction is fixed in advance, for a stated scientific or practical reason, and written in the analysis plan.
- An effect in the other direction would be treated the same as no effect. For example, a cheaper supplier is only worth adopting if it is not worse, and a result that shows it is much better changes nothing about the decision.
- The test is naturally one-sided, such as a non-inferiority test or a chi-square or F test, where only large values count.
When in doubt, use the two-tailed test. It is the conventional default, and it protects you against effects you did not expect.
Running each version in software
| Tool | Two-tailed | One-tailed (greater) |
|---|---|---|
| Excel T.TEST | =T.TEST(A2:A7,B2:B7,2,3) | =T.TEST(A2:A7,B2:B7,1,3) — check the direction yourself |
| R | t.test(x, y) | t.test(x, y, alternative = "greater") |
| Python (SciPy) | stats.ttest_ind(x, y, equal_var=False) | stats.ttest_ind(x, y, equal_var=False, alternative='greater') |
Excel's one-tailed T.TEST is simply half the two-tailed value whichever way the difference points, whereas R and SciPy compute the tail you name, so they return a large p-value when the effect goes the other way. More in the t-test in Excel and the t-test in R and Python.
Try the P-Value Calculator
Get a left-tailed, right-tailed or two-tailed p-value from a z, t, chi-square or F statistic.
Frequently Asked Questions
Is a one-tailed test more powerful than a two-tailed test?
In the predicted direction, yes: at α = 0.05 the z critical value is 1.645 instead of 1.960, so a smaller effect reaches significance. The price is that the test has no power at all to detect an effect in the opposite direction. That trade is only sensible if a difference the other way would be treated exactly like no difference.
What happens if the result goes in the opposite direction?
The one-tailed p-value becomes large and the null hypothesis is not rejected, however extreme the result. For a right-tailed test with t = −2.10 and 15 degrees of freedom the p-value is 0.9735, not 0.0265. You cannot then switch to a left-tailed test, because the direction had to be chosen before seeing the data.
Can I use a one-tailed test in an A/B test?
Only if you would take the same action, do not ship the variant, whether it turns out equal to control or worse. If a decline would make you act, for example by rolling back a change, then both directions matter and the two-tailed test is the honest choice. Many experimentation teams default to two-tailed for exactly this reason.
Are chi-square and ANOVA tests one-tailed or two-tailed?
Their statistics, χ² and F, are always positive and only large values count as evidence against the null hypothesis, so the p-value is the area in the right tail. That is built into the test and is not the same choice as a one-sided t-test. The idea of a direction applies to comparisons of means or proportions, not to these overall tests.
How are one-tailed tests related to confidence intervals?
A one-tailed test at α = 0.05 matches a one-sided 95% confidence bound, and a two-tailed test at α = 0.05 matches a two-sided 95% interval. The lower limit of a one-sided 95% bound equals the lower limit of a two-sided 90% interval, which is why one-sided results look stronger.
Do I split α in half for a two-tailed test?
Yes: with α = 0.05 each tail holds 0.025, so the critical z values are −1.960 and +1.960. Equivalently you keep α at 0.05 and double the one-tail area to get the p-value. Both routes give the same decision.