1 tailed vs. 2 tailed p-values

[trinker]

I have been encountering some discussion lately on 1 vs. 2 tailed p-values. The 1 tailed test it seems is not preferred even with an a priori directional hypothesis. The 1 tailed hypothesis test and p-value does not make sense for confidence interval. I'm struggling with this myself and have not formed an opinion one way or the other. What are people's thoughts about 1 vs. 2 tailed with an a priori directional hypothesis?

[Dason]

I personally don't mind using a 1-tailed test if you a priori have a direction in mind. There is a caveat to that though. I have to imagine that a situation came up where for some reason we observed the opposite of what we expected - then I ask myself/{whoever I'm working with} - would we want to test to see if the reversed trend would be significant. So let's say I hypothesize that people perform a certain task faster with their right arm than they do with their left arm. If for some reason the data shows the trend the other way (left arm being faster than the right arm) am I going to say "forget it - looks like I was wrong" or am I going to say "that's interesting... I'll test if that's significant". If it's the first situation then I have no problem with the 1-tailed test. If it's the second (and you really do need to be honest with yourself) then the two-tailed test is the way to go.

But honesty is a hard thing to judge (especially a priori...)

[Jake]

I was going to say pretty much exactly what Dason said. More often than not it probably just ends up being a sneaky way of doing a two-tailed test with alpha=.10.

But directional hypotheses and 1-tailed tests would be kind of silly even in a field full of saints. We should always allow the data to surprise us. We don't allow for that possibility when we ignore effects going in one direction.