I had written a hypothesis that we expect x outcome variable to increase when y treatment is applied. A reviewer commented that based off of the hypothesis, we should be doing a one-sided test.

However, this doesn't make sense to me, since it is important to know from the data if x actually decreases when y is applied. Wouldn't doing a one-way hypothesis miss this?

If the written hypothesis must reflect what kind of testing was done, how should it be written to prevent it being boring/uninformative? For example, "We expected x would change when y is applied".

Thank you!

  • @EdV I think that this is sufficiently divorced from the contents of the OP's research to be on-topic. Its relevant for anyone whose research involves hypothesis testing.
    – nick012000
    Aug 24, 2021 at 0:31
  • 1
    @nick012000 Good to know: the boundary between on topic and off seems to be fractal. Anyway, Buffy’s answer is correct and the choice of one sided or two sided gets done before the testing.
    – Ed V
    Aug 24, 2021 at 0:36
  • 1
    This is probably on-topic on crossvalidated.stackexchange.com I don't think it's on-topic here, as it is fairly technical and substantive.
    – henning
    Aug 24, 2021 at 7:30
  • I was interested in the writing aspect - how to write the hypothesis at the end of the introduction. However, from Buffy's answer I see that I was missing the connection between the hypothesis I set out at the beginning and how it's written up.
    – user1762
    Aug 24, 2021 at 12:01
  • 1
    This is opinion based. Even if we knew exactly what the research was, it might still be opinion based. I do not see any reason not to do it the reviewer's way. Aug 24, 2021 at 18:48

1 Answer 1


If your hypothesis doesn't match your test, how can you say that carrying out the experiment gives you anything valid. Make them match. If you want a two sided hypothesis, use a two sided test. Similarly for one sided.

Of course, a close examination of the data might reveal whether the preponderance is on the high or low side, but that isn't really using statistics correctly.

Make them match. If you expect an increase, then test for that. If the test fails then reevaluate your assumptions and test again.

  • 2
    Right. To echo: claiming to be testing for several possibilities all at once is like p-value hacking and such: a mis-use of statistics. Aug 24, 2021 at 3:30
  • Thank you - this clarifies things for me. I think I missed the connection between the hypothesis I set out with and then what gets written in the manuscript. I appreciate the clarification!
    – user1762
    Aug 24, 2021 at 12:03
  • @paulgarrett People get very agitated about P-hacking, but thinking about it from a Bayesian perspective, I'd expect the effect of P-hacking on the inference process to be fairly weak - specifically, I'd expect the number of data points required to reach a given level of certainty to scale like the logarithm of the number of hypotheses being considered. Or am I missing something? Aug 24, 2021 at 18:17
  • @DanielHatton, I think you're right, but/and people operating in ignorance of such considerations have a big chance of messing up... Awareness of the pitfalls, I think, is important, and not reliably "publicized", it seems. Aug 24, 2021 at 18:20
  • 1
    @Andrew Yes, omitting part of the data set would be deeply problematic, as would allowing information from the data set to leak into the prior (either the prior between hypotheses or the prior over adjustable parameters within any one of the hypotheses) (although formulating hypotheses after seeing the data does not necessarily imply that one is guilty of such information leakage). Aug 25, 2021 at 9:36

Not the answer you're looking for? Browse other questions tagged .