statistics
One-tailed vs two-tailed tests: how to choose correctly
Choose a one-tailed test only when your hypothesis specifies a direction before you collect data and that direction is the only outcome you would act on; otherwise use a two-tailed test, which splits significance across both ends of the distribution and is the safer default for most coursework and research.
By UtilBloom · Published 2026-10-05 · Updated 2026-10-05
What "direction" means in a hypothesis test
Every hypothesis test starts with a null hypothesis (H0, usually "no effect" or "no difference") and an alternative hypothesis (H1) that you are trying to find evidence for. The alternative can be directional, "the new process increases output", or non-directional, "the new process changes output, up or down, we don't know which". That choice of wording is not cosmetic; it decides how you place the rejection region on the sampling distribution and therefore how large your test statistic needs to be to count as significant. A two-tailed test splits your significance level (commonly α = 0.05) into two equal rejection regions, one in each tail of the distribution, 0.025 on each side for a standard 5% test. A one-tailed test puts the entire 0.05 in a single tail, left or right depending on the direction you specified. That difference matters enormously at the boundary: a result can be significant at α = 0.05 one-tailed while failing to reach significance two-tailed, because the one-tailed critical value is less extreme.
How to choose before you look at the data
The direction of your test must come from your hypothesis, theory, or prior research, decided before you run the analysis, not from eyeballing the sample mean and picking whichever tail makes your p-value smaller. That second approach is a classic form of p-hacking: it inflates your real false-positive rate well above the nominal 5%, because you are implicitly giving yourself two chances (either direction) while only paying the statistical price of one. A genuinely one-tailed situation looks like: a drug is expected to lower blood pressure, and a result showing it raises blood pressure would be treated identically to "no effect" for your decision, you would not approve the drug either way. If an unexpected result in the "wrong" direction would still be interesting or actionable, you need a two-tailed test, because you have not actually ruled out that direction in advance. When in doubt, most journals, textbooks, and advisors default to two-tailed tests precisely because the directional commitment is easy to fake after the fact.
Critical values and p-values differ by test type
For a z-test at α = 0.05, the two-tailed critical values are ±1.96, while the one-tailed critical value is 1.645 (or −1.645 for a left-tailed test). For a t-test, the exact critical value depends on degrees of freedom, but the same logic holds: look up the one-tailed or two-tailed column depending on your hypothesis, not whichever gives you the answer you want. Statistics Hub's critical-value and t-table calculators let you pull both columns side by side so you can see exactly how much the threshold shifts. The p-value calculation also changes. For a two-tailed test, you double the one-sided tail probability to account for both directions. A test statistic that gives a one-tailed p-value of 0.03 becomes p = 0.06 two-tailed, which crosses the conventional 0.05 line from significant to not significant. This is the single most common source of confusion when comparing results across papers or software that default to different conventions, always check which convention a reported p-value uses before comparing it to your own threshold.
Common mistakes that inflate false positives
The most common mistake is switching to a one-tailed test after seeing a two-tailed result land just above 0.05, then reporting the easier one-tailed p-value as if it were the planned analysis. Reviewers and statistics instructors specifically watch for this, because it silently doubles your effective false-positive rate. A second mistake is using a one-tailed test "to be conservative" when the researcher actually means the opposite, a one-tailed test is less conservative in the hypothesized direction, not more; it is only more powerful if your directional assumption is correct and never need defending against the other direction. A third mistake is applying a one-tailed test to a two-sided real-world question, such as "is this coin fair", where both "biased toward heads" and "biased toward tails" are meaningful alternative outcomes. If you cannot honestly say you would ignore a result in the opposite direction, use two-tailed.
Worked example: one-tailed vs two-tailed on the same data
Suppose a sample of 36 students scores a mean of 78 on a new teaching method, against a known population mean of 75 and population standard deviation of 9. The z-statistic is (78 − 75) / (9 / √36) = 3 / 1.5 = 2.0. Two-tailed at α = 0.05, the critical value is ±1.96, so z = 2.0 exceeds 1.96 and the result is significant. One-tailed (testing only "the new method increases scores") at α = 0.05, the critical value is 1.645, and z = 2.0 is further past that threshold too, so this particular result is significant either way. Now drop the sample mean to 77.5: z = (77.5 − 75) / 1.5 = 1.67. Two-tailed, 1.67 does not exceed 1.96, so you fail to reject H0. One-tailed in the hypothesized direction, 1.67 does exceed 1.645, so you would reject H0. This is exactly the borderline zone where the choice of test type, made honestly before data collection, determines the conclusion you report.
FAQ
Is a one-tailed test ever the "correct" default?
Only when the opposite direction is truly irrelevant to your decision. Most academic and business contexts default to two-tailed because a surprising reversal is still informative, even if it changes what you do next.
Does a one-tailed test need a smaller sample size?
A one-tailed test is more powerful for detecting an effect in the hypothesized direction at the same sample size and alpha, which is sometimes why it is chosen in power analysis, but that choice must still be justified by the research question, not by convenience.
Can I convert a reported two-tailed p-value to one-tailed?
If the effect is in the direction you would have hypothesized, divide the two-tailed p-value by two. If it is in the opposite direction, the one-tailed p-value for your original hypothesis is 1 minus half the two-tailed value, effectively non-significant.
Do confidence intervals change between one-tailed and two-tailed?
A standard two-sided 95% confidence interval corresponds to a two-tailed test at α = 0.05. One-tailed tests are more naturally paired with one-sided confidence bounds (an upper or lower limit only), which are reported less often but are not wrong, just a different question.