Roadmap

Goals for 2022

Examples

Add examples for the following cases:

• Offsets #418.
• Potentials #384. Don't forget to explain what is a potential and where the name comes from.
• Beta family #369.
• Categorical family #436.
• Posterior predictive check #252. I think we already use it in some places, but it is not advertised enough.
• Piecewise regression #437.
• Add example of posterior predictive sampling (and or check) #252

Book ports

• Bayes Rules book is a good candidate for Bambi. It uses rstanarm quite a lot and we have (almost?) all the features needed to reproduce the book.

• Assess status of current ports

  • Regression and other stories
  • Statistical Rethinking

Add formula object

This formula object will allow us to pass multiple formulas for different components of the model. We take the distributional model term from brms (this vignette), which says:

[...] refer to a model, in which we can specify predictor terms for all parameters of the assumed response distribution.

In other words, we not only have predictors for the mean, but also predictors for other parameters such as the dispersion. For example, this enables heteroskedastic linear regression models.

Distributional models

The type of models just mentioned above.

Prior class

It would be really nice if we could move forward with what it is in #365. At some point, we're going to face it. For example, if we allow distributional models and we use the same predictor for the mean and the standard deviation. How do we allow users to use different priors for each case?

A more efficient approach for categorical predictors

At the moment, both numerical and categorical predictors go into the same design matrix X and they contribute to the model via pm.math.dot(X, beta). It would be more efficient to do a regular dot product for numerical variables, but we could do it more efficiently if we use sparse multiplication for categorical predictors. There's still some research to do, though.

The same idea applies to group-specific effects. It's more likely to have high sparse settings here.

Add BART

Bayesian additive regression trees are often praised in the literature as model that "generally works" with almost no tuning needed from the user (often the comparison is made against GPs). So it may be good to have an option in Bambi to run them. This should be done after we move Bambi to use PyMC 4.x. As BART offers an heuristic for variable importance it could be nice to see how it behaves compared to projpred

Add projection predictive variable selection

Projection predictive variable selection for generalized linear models

Other topics

• INLA support
• Gaussian processes
• Allow "R-side" covariance structures and covariance priors in general (for varying effects too) #110