One-tailed test following anova

[Kristen Valentine]

I'm running the anova command in STATA on a model with three main effects and all their interaction terms. My understanding of the significance levels in the output is that those are for two-tailed hypothesis tests. Is there a way to get STATA to provide the one-tailed significance levels or do I need to look up the F-stat in a table for every variable? I know with a simple t-test I could just divide the p-value by two, but the F-distribution isn't symmetric. I've seen examples on this forum/online for how to do one-tailed significance levels following the regress command, but not anova.

[Carlo Lazzaro]

Kristen:
without reading what you typed and what Stata gave you back (as recommended by the FAQ), the first advice that springs to my mind is that you can apply the solution proposed for OLS, being it a linear model just like ANOVA. As a more general aside, I would say that is (almost) nothing that OLS can't do better than ANOVA. This fact that OLS often outperforms ANOVA seems to be supported by the sharp decrease in Stata threads concerning the latter linear model during the last years.

[Marcos Almeida]

I fully agree with Carlo with regards to the ubiquitous use of linear regression instead of ANOVA.

That said, I wish to underline that one-tailed tests, at least in health sciences, generally speaking, tend to be less used.

To be sincere, except in non-inferiority studies (by the way, nota bene, though one-tailed, the alpha is usually set to 0.025), I fail to envisage a situation where it is clearly recommended.

Indeed, the New England Journal of Medicine (which has the highest impact factor in health sciences) states this warning in the guidelines to the authors:

Except when one-sided tests are required by study design, such as in noninferiority trials, all reported P values should be two-sided.

Once, I was challenged to present a true-to-type case where a one-tailed test must be the best alternative, and I failed to provide a good example in health sciences.

There are other concerns as well, such as a (dubious) strategy to "reap" power for a small sample size. A risky step, as well, for it can lead to an increase in type I error. That can skyrocket "positive results", shall we decide to make post hoc tests.