两次单尾检验替代双尾检验能否提升显著性？兼论双侧单尾检验可行性

Hey there, these are really thoughtful questions about one-tailed vs two-tailed hypothesis testing—super common points of confusion, so let’s unpack them clearly:

1. Does using two one-tailed tests instead of one two-tailed test boost significance? #

Short answer: It looks like it, but only because you’re cutting corners on Type I error control.

Here’s the breakdown: When you run two separate one-tailed tests (e.g., one checking for a positive effect, another for a negative effect), you’re effectively doubling your allowed false positive rate. Let’s say you set α=0.05 for each test. A standard two-tailed test splits that 0.05 evenly across both tails (0.025 each) to keep the total false positive risk at 5%. But with two uncorrected one-tailed tests, your overall chance of a false positive jumps to 0.05 + 0.05 = 10%.

So if a two-tailed test gives a p-value of 0.06 (just above the 0.05 threshold), one of the one-tailed tests might spit out a p-value of 0.03—making it look "significant." But this isn’t a real increase in meaningful significance; it’s just you letting yourself get away with more false positives. This is called α inflation, and it’s a big red flag in rigorous statistical practice.

2. Why not run one-tailed tests for each side (and add a reverse test if needed) to get the smaller sample size benefit? #

This approach falls apart for two critical reasons:

• Uncontrolled α inflation: As noted above, running two uncorrected one-tailed tests blows up your false positive rate. If you try to fix this by correcting α (e.g., using Bonferroni correction to set α=0.025 per test), you’ve essentially recreated the two-tailed test. The corrected significance threshold will match what you’d get from a standard two-tailed test, so you lose the "smaller sample size" advantage entirely.
• Violates one-tailed test assumptions: One-tailed tests are only valid if you have a priori (before seeing data) justification to ignore one direction of effect. If you run a positive one-tailed test, don’t find significance, then run a negative one, that’s post-hoc testing—you’re essentially fishing for a significant result. The smaller sample size isn’t a free win—it’s a trade-off: you give up the ability to detect effects in the opposite direction in exchange for more power in your pre-specified direction. If you’re not willing to make that trade-off, a two-tailed test is the appropriate choice.

You can’t "fix" the one-tailed test’s directional assumption by adding a reverse test later because that defeats the purpose of the one-tailed test in the first place. The power/sample size benefit comes from narrowing your focus to one direction, but if you’re willing to check both, you don’t get to keep that benefit without sacrificing statistical rigor.

内容的提问来源于stack exchange，提问作者Xerxes