Model Comparison: How to constrain a model to positive mu-differences?

[Olaf]

@junpenglao

IIUC, you have condition, block, direction which potentially be different, and user could also display different strategy - in that case, I recommend you to try parsing the linear formula σproj∼(condition∗block∗direction*user) in patsy, and use indexing for trial (the random effect.)

That sounds about right. I don’t really care about trials, e.g. if all trials with number 1 differ or not. I did try to use patsy and got an outcome and a DesignMatrix. Unfortunately, I don’t have the slightest clue about what to do with the DesignMatrix object. I tried multiplying it with a distribution and only got an AssertionError. I also don’t know what to index where?

@junpenglao

In general, frequentist testing like ANOVA are GLMs

Yes, and you only have to drag some variables into the right slots in a GUI software most of the time.

@junpenglao

I would love to give you a bit more concrete suggestion, maybe it is easier you share a jupyter notebook with some simulation data?

That is very kind of you and much appreciated! I prepared a gist with some mocked up data. There’s not much of a model, because I don’t know how to use the DesignMatrix and indexing.

https://gist.github.com/OlafHaag/2e36125a3d34571284705d624e81b64f