[lisualsethelp]

I know some of you guys are good at math and stats so if you could help my question is pretty basic.

Say you're doing a hypothesis test of two sample groups to determine if there is a statistically significant difference between the means. You have no reason to believe they should be different so by default your hypothesis would be no difference between the means of group x and y. Say you run this test and you find there is a statistically significant difference. And say your mean for x was 70 and your y mean was 60 (I'm just making these numbers up). Would it then be OK to run an addtional hypothesis test of X mean > Y mean and see if there is statistical significance? Or can you only run the X mean no different than Y mean and that's it?

[gfkasjhfg]

Well, in general you set your hypothesis before you collect data.

In your example, you are technically wrong. You can only accept (fail to reject) or reject a hypothesis - rejecting a hypothesis does not make the contrary true. By rejecting the hypothesis, you are saying that your particular dataset does not support your hypothesis.

In this case, it is binary, so the opposite is always true, but don't get into the habit of thinking the choices are accept null hypothesis or accept alternative hypothesis. That could get you murdered by a pedantic.

To answer your question, there are different schools of thought on this. Purists (my professor was one these) believe that you can only run hypothesis tests once per set of data, others do not care.

Personally, I think that if you are changing your hypothesis around you are probably missing the point.

[Esmeralfaf]

Well, in general you set your hypothesis before you collect data. In your example, you are technically wrong. You can only accept (fail to reject) or reject a hypothesis - rejecting a hypothesis does not make the contrary true. By rejecting the hypothesis, you are saying that your particular dataset does not support your hypothesis. In this case, it is binary, so the opposite is always true, but don't get into the habit of thinking the choices are accept null hypothesis or accept alternative hypothesis. That could get you murdered by a pedantic. To answer your question, there are different schools of thought on this. Purists (my professor was one these) believe that you can only run hypothesis tests once per set of data, others do not care. Personally, I think that if you are changing your hypothesis around you are probably missing the point.

Borrrrrrrring. How you can manage to get off on this **** I'll never know.

[mdUzAMbG]

I know some of you guys are good at math and stats so if you could help my question is pretty basic.

Say you're doing a hypothesis test of two sample groups to determine if there is a statistically significant difference between the means. You have no reason to believe they should be different so by default your hypothesis would be no difference between the means of group x and y. Say you run this test and you find there is a statistically significant difference. And say your mean for x was 70 and your y mean was 60 (I'm just making these numbers up). Would it then be OK to run an addtional hypothesis test of X mean > Y mean and see if there is statistical significance? Or can you only run the X mean no different than Y mean and that's it?

There are certain circumstances where that's allowed. In medical research this approach is usually expressly forbidden, because it leads to hypothesis drift - the idea of flipping through multiple hypotheses until you find one that fits your data in such a way as to make it statistically-significant. Because this makes the hypothesis another variable and changes the truth value, it's generally at the very least frowned-upon.

[JetePlentuara]

There are certain circumstances where that's allowed. In medical research this approach is usually expressly forbidden, because it leads to hypothesis drift - the idea of flipping through multiple hypotheses until you find one that fits your data in such a way as to make it statistically-significant. Because this makes the hypothesis another variable and changes the truth value, it's generally at the very least frowned-upon.

It's my understanding that you should only do a one-tailed test when you have prior evidence to believe that a one-tailed test is warranted. But what if you were trying to figure out something that had no prior research? Say for example what proportion of FM members preferred coke and what proportion preferred pepsi? Obviously there is no prior evidence to go on, so you would only be able to do a two-tailed test in this situation and all you could ever say was if the proportion liking coke was equal or not equal to the proportion liking pepsi. But you could never say the proportion liking coke > or < the proportion liking pepsi?