In the test for equivalence (or non-inferiority / non-superiority) we define a "minimal effect size" $d$, and test that the measured effect is not larger that $d$

$$X_1 \sim PDF1$$ $$X_2 \sim PDF2$$ $$TOST: mean(X1 - X2) < |d|$$

The difference from a t-test being that we do assume an effect size $> 0$, but we set a boundary under which to neglect this effect size.

However, in the TOST procedure, the null hypothesis claims that the effect size is larger that $d$, and the alternative hypothesis is that it is smaller. This is a bit against the usual null hypothesis that assumes randomness, non-relatedness, and absent effect sizes.

If we assume that the treatments that generate $X_1$ and $X_2$ are meaningless, and both are random variables drawn from the same distribution $PDF1 = PDF2$, which is what we would assume doing a t-test, the null of the TOST should also be the absence of an effect (i.e the effect size being smaller than a pre-defined $d$).

Why does the TOST procedure have a contra-intuitive null?


The working hypothesis is: "the effect is scientifically negligible". This is what the researcher claims. The null hypothesis is simply the contrary, i.e. there is a scientifically non-negligible effect.

I find this very intuitive (in contrast to the fact that no multiple testing issue occurs with the two one-sided tests).

  • $\begingroup$ This makes it sound as if the choice of the null and alternative is an arbitrary choice of whatever the researcher 'claims', and are interchangeable by just re-arranging the wording of the research question. I assume there is a deeper reason why the null is designed that way, related to what can be tested with evidence and what can't. We choose the null in t-testing to be non-relatedness, as is the only claim we can either reject or fail to reject (we can't accept/prove absence of effect in a noisy world). Am I confused about this? $\endgroup$ – hirschme Apr 23 '20 at 18:22
  • $\begingroup$ Indeed, this is what hypothis testing is about. The researcher claims something, then does a nice study and verifies (or not) his claim by statistics. Another researcher might claim the contrary. His hypotheses would be flipped. Of course this second researcher will need to perform a new study to check his flipped working hypothesis. $\endgroup$ – Michael M Apr 23 '20 at 19:13
  • $\begingroup$ I think the point is that each test tackles a specific claim, so each test has its specific "type of" null hypothesis. A t-test H0 has to be under random and non-effect assumption, and this can't be flipped. The TOST is designed to test for non-equivalence, so its H0 is based on the assumption of a true effect. Your comment makes sense in this case: yes, hypothesis are interchangeable, but each one then suggests a different and appropriate test. Does this agree with you thoughts? $\endgroup$ – hirschme Apr 23 '20 at 19:49

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.