Improving Equivalence Tests in Fit Y by X

[awg]

Suggestion would be to allow an automatic solve for the difference value that gives equivalence for a given alpha.

In the example example-of-an-equivalence-test there would be an option to solve for the difference value that predicts equivalence for a given alpha value (it addition to the current option for supplying a specific difference).   Many times this difference value is specified inconsistently among a given department or functional group (often with many differing opinions).  So, instead of running multiple differences to show the p-values for equivalence it would be very convenient to solve for the difference that would predict equivalence given an alpha value.

[jszarka]

If you have an alpha=.05, a 90% confidence interval on the difference in means between 2 groups will align with the boundary difference value you specify. Within the Fit Y by X platform you can change the alpha level to say .10 and use the appropriate t-test for making that comparison in the Big Class data set.

In general though, it's not a good practice to get the confidence interval then determine whether or not that difference level is meaningful. I can't tell from your thread whether this debate occurs before or after the data occurs; naturally we want that agreed upon in advance so that when the data comes back the group can be held accountable to the preselected threshold.

[PatrickGiuliano]

I agree that a back-solver would be handy but it goes against the intent of an equivalence test which is to specify a practical difference threshold before running the experiment.  

This is kind of like cherry-picking the practical difference to "pass."