In the sciences very often we are testing for the presence of a predicted relationship, for example, that two variables are positively correlated ($\beta > 0$). Under traditional frequentist testing, one might set the null hypothesis to be $H_0: \beta = 0$, with a compound alternative hypothesis of $H_{1,a}: \beta > 0$ and $H_{1,b}: \beta < 0$.

If one finds strong support for $H_{1,b}$ (that the slope is negative) is it correct to say that this is "evidence against $H_{1,a}$"? Typically we only state that we find evidence against the null hypothesis, but in the case of a directional test it seems intuitive that if you find strong evidence for a negative slope, then this should also count as strong evidence against a positive slope, not simply against the null of $H_0: \beta = 0$.

So is it correct to say use the term "evidence against" one of the other compound alternative hypotheses? Otherwise it seems to imply that no one can ever find evidence against a presupposed hypothesis, despite evidence to its contrary. Or is it that my null and alternative hypotheses are not formed correctly?

  • $\begingroup$ This issue is an artefact of using a point null in a one-sided situation. You have found evidence strongly consistent with nonpositive slope. $\endgroup$ Commented Mar 22, 2023 at 19:10
  • $\begingroup$ And no support for the hypothesis of positive slope. $\endgroup$ Commented Mar 22, 2023 at 19:13
  • $\begingroup$ But what is the correct null then? If we choose $H_0: \beta > 0$ (which is what we hypothesize to be true) then this somewhat contradicts the idea of the null being "no relationship". However, if we select $H_0: \beta < 0$ then we can never "reject" $\beta > 0$ since you typically can only reject the null. And so it seems paradoxical -- how do you frame the question so that you can reject the hypothesis of a positive slope? $\endgroup$ Commented Mar 23, 2023 at 9:38


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.