Potentially revise `repeat` parameter or introduce `trials`

[ReneSkukies]

As far as I understand, right now the repeat parameter works by simply repeating the full factorial of the indicated design. Meaning two things:

1. If I want a certain number of trials, am forced to work with multiples of the full factorial of my condition dictionary
2. I cannot have fewer trials than the full factorial of my design.

The second point can quickly become an issue, given that if I only have 4 covariates in my model, like so:

cond_dict = Dict(
        :condition => ["face", "bike"],
        :sac_amplitude => range(0, 5, length = 10),
        :evidence => range(1, 5, length = 8),
        :duration => range(2, 8, length = 12)
        )

I already end up with 1920 trials, more covariates would explode this even more and quickly leads to a nonsensical amount of trials. The above is actually realistic in a 10-min free viewing task if you use saccades as events of interest, but if I add one more covariate such as :luminence => range(1, 31, length = 30) I will already have about 57.000 trials. This also quickly becomes a computational problem.

I am not sure how to better handle it (currently I draw random events without replacement to achieve a specified amount of trials), but I thought it would be good to have a discussion around implementing a way to get a specified number of trials back from UnfoldSim. It would probably be better to introduce a new way to do this than to completely revise the repeat parameter.

[jschepers]

@ReneSkukies Thanks a lot for your issue. I had not yet thought of this in this way. I will think about it once I have time. Maybe we can have another "Wrapper"-design PartialDesign in which one can specify which (number of) combinations to simulate.
We had a related discussion at some point when we thought about balanced vs imbalanced designs.

[ReneSkukies]

Maybe it would also be a good idea to throw a warning when a user creates a design with a high number of covariates (i.e. when the full factorial exceeds a certain number)? I.e. before simulating the design

[behinger]

The key issue seems to me is the use of continuous variables as if they are categorical.

I think a wrapper-design that simulates continuous covariates would help a lot. I already had plans for a long time how do do it, but haven't come around to implement it.

E.g. one specifies a name and a Distribution.jl, and automatically covariates are pulled.

Some conceptual issue i had:

• It is a second function to be discovered
• best syntax for multiple covariates?
• how to do correlated covariates

Nothing serious, but something to think about

To issue one, e.g.
ContinuousCovariateDesign(RepeatDesign(LinearModelDesign)) looks verbose...

PS: just subsampling a design will still generate the large one first - not a huge problem IMHO because data frames is quite performant, but not ideal