One-Tailed vs. Two-Tailed Tests: How to Choose, With a Worked Example

The hypothesis testing framework covers the six-step process broadly, but one decision inside that process deserves its own close look: whether a test is one-tailed or two-tailed. Get this choice wrong — or make it after peeking at the data — and the same test statistic can lead to a different, and sometimes wrong, conclusion about statistical significance.

What "Tails" Means

Every test statistic (z, t, χ², or F) has a sampling distribution, and the "tails" are the extreme regions at either end of that distribution — the values far enough from what the null hypothesis predicts to be considered surprising. Whether one or both tails count as evidence against the null hypothesis depends entirely on how the alternative hypothesis, H₁, is phrased.

How the Choice Changes the Math

For a fixed significance level α (conventionally 0.05), splitting the rejection region across two tails versus concentrating it in one tail changes exactly where the critical value falls:

The one-tailed critical value is smaller in magnitude because all 5% of the rejection region is concentrated in a single tail instead of split as 2.5% on each side — making a one-tailed test more powerful for detecting an effect in the predicted direction, at the cost of being completely unable to detect a genuine effect in the opposite direction (a one-tailed test testing for "greater than" will never reject H₀, no matter how convincingly the true effect turns out to be in the other direction).

Worked Example: Same Data, Different Conclusion

Problem: A fertilizer company claims its new product increases average corn yield above the current regional average of 50 bushels/acre (known population standard deviation σ = 8 bushels/acre, from long-running regional data). A trial on n = 40 plots produces a sample mean of 52.21 bushels/acre. Test at α = 0.05, both as a one-tailed and a two-tailed test, and compare the conclusions.

Standard error = σ/√n = 8/√40 ≈ 1.265.

z = (52.21 − 50) / 1.265 ≈ 1.75

As a two-tailed test (H₀: μ = 50, H₁: μ ≠ 50 — testing only whether the fertilizer changes yield, without presupposing direction): the two-tailed p-value for z = 1.75 is ≈ 0.080. Since 0.080 > 0.05, the result does not reach statistical significance — there isn't strong enough evidence to reject H₀ at the conventional 5% level.

As a one-tailed test (H₀: μ = 50, H₁: μ > 50 — testing specifically whether the fertilizer increases yield, which matches the company's actual claim): the one-tailed p-value for the same z = 1.75 is exactly half the two-tailed value, ≈ 0.040. Since 0.040 < 0.05, this version of the test does reach statistical significance.

The exact same trial, the exact same z-statistic, produces two different verdicts depending entirely on which test was chosen — which is precisely why the choice can't be made after seeing which one gives the "better" answer.

Choosing Before Seeing the Data

The one-tailed test is legitimate here specifically because the company's original claim was directional — "increases yield," not just "changes yield" — and that claim existed before the trial ran. Choosing the test type based on which one happens to produce a significant result after looking at the data is a serious methodological error sometimes called p-hacking: it inflates the true false-positive rate well above the stated α, because you're implicitly trying both directions and reporting only the one that worked. The correct practice is to decide the hypothesis's direction (or lack of one) based on the research question and any existing theory or prior evidence, before the data is collected, and state it in advance — in a pre-registered study, this decision is documented and locked in before the data is even seen.

When Each Type Is Actually Appropriate

Use a two-tailed test when a change in either direction would be meaningful and there's no strong prior reason to expect one direction over the other — the default choice in most exploratory or confirmatory research, and the safer default when in doubt.

Use a one-tailed test only when a result in the "wrong" direction is either impossible in principle or would be treated identically to "no effect" for the purposes of the decision being made — for example, testing whether a new manufacturing process reduces defect rates, where an increase in defects would simply mean rejecting the new process just as readily as no change would. Genuine one-tailed scenarios are less common in practice than the two-tailed default, which is why some fields and journals discourage one-tailed tests altogether unless the directional justification is made explicit and defended in advance.

Common Mistakes

Switching to a one-tailed test after seeing a two-tailed result fall just short of significance. This is the exact p-hacking scenario described above — it looks tempting precisely because it works arithmetically (halving the p-value), which is exactly why it's considered a serious violation of good statistical practice.

Using a one-tailed critical value while still describing the hypothesis as non-directional. The critical value and the phrasing of H₁ must match — a mismatch here silently changes the actual significance level being used, often making a test more liberal than the stated α suggests.

Assuming one-tailed is always more powerful and therefore always better. One-tailed tests are only more powerful for effects in the predicted direction — if the true effect turns out to run the opposite way, a one-tailed test looking in the wrong direction will fail to detect it no matter how large that opposite effect actually is.