I have constructed a gamma glm modelling hormone concentration with three predictors: treatment (treatment or control), sex (male or female) and site (thyroid, parathyroid). The work is mostly exploratory, and I had no a priori hypotheses regarding the effects of sex or site, but I have seen previously that the experimental treatment group produced higher hormone concentrations than the control, so I hypothesised a main effect of increased hormone concentrationin the experimental treatment group, over the control.

In my results I have reported the t statistic and p value for each predictor, but for treatment, my supervisor said this should be one-sided as I have an a priori hypothesis (that hormone concentration will be larger in treatment compared to control) and therefore my p value should be halved to turn it into a one-tailed value.

I know GLMs report a t statistic, but I believe it’s a Wald’s test it’s reporting which is slightly different from a t-test? Does the one-tailed vs two-tailed argument apply in GLMs in the same way it would in a Student’s t-test? Is it reasonable for me to simply halve my p-value in this case and report it as a one-tailed value?


1 Answer 1


(With a GLiM, strictly speaking, it's a $z$ statistic, not a $t$ statistic.)

The arguments about one-tailed vs. two-tailed work exactly the same in this context as the conventional STATS 101 $t$-test context. For what it's worth, I typically default away from one-tailed tests. In your case, I think you should have had a strong a-priori hypothesis, and have built that decision into the design of the study—it's sketchier to decide you want a one-tailed test after you've seen the results.


Your Answer

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

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