Extend functionality to cover fake/synthetic data subworflow

[aloctavodia]

Should we extend the functionality of this repo and turn it into a package to assist users to perform fake/synthetic/constructed data simulations?

Many times this synthetic data is data generated from a model, but sometimes we also want synthetic data without having or prior to contructing a model, like linear regression with specific R2 or a dataset were some of the covariates are noise. Or maybe some blobs (like scikit-learn functionality)...

[currocam]

Hi, I’m not really sure this is what you mean, but it would be great if the package supports some functionality to pass custom data-generation functions.

I guess often you are interested in knowing if your model is well calibrated at two different levels / in two different worlds:

• If all your assumptions hold (e.g. quantity X is truly normally distributed)
• Under some more complex model (which you might not know how to trace but you can simulate from) or mechanistic data generation process.

It would be nice if the API allows you to test both. I’m happy to make some PR or talk about the interface :)

I would go with an optional extra argument in the SBC function that is a user-defined function. The question is which interface the user-defined function should have.

[aloctavodia]

Hi @currocam, thanks for contacting us.

What you describe is definitely something we want to have! And your help is more than welcome!

Having said that, I just want to clarify that this issue is about something else. It's not about SBC; instead it's about convenience functions to generate data with specific features. For instance, we may need data with highly correlated predictors, or data with a lot of covariates, but most of which don't contribute to the response. We usually want this data to try new algorithms/models or for teaching/understanding methods.