Features ideas

Plots

Below you will find a list of features that marginaleffects supports for comparisons and slopes whereas Bambi does not. The first bulleted list indicates functionality that would be shared across both functions. The second and third bulleted list are functionalities specific to that function.

comparisons and slopes:

• type argument indicating the scale of the predictions used to compute differences. Not only the response scale, but also the link scale for example
• Compute slopes or comparisons using posterior predictive samples: pps=True
• Compute slopes or comparisons for other model parameters (in distributional models)
• comparisons or slopes between marginal means (each regressor is specified at its mean or mode)

comparisons:

• cross comparisons / contrasts - we would like to know what happens when two (or more) predictors change at the same time
• hypothesis tests - the result can be interpreted as a “difference-in-differences”: moving from 1st to 3rd class in the Titanic model has a much larger negative effect on the probability of survival for a 50 year old woman than for a 50 year old man. This difference is statistically significant.
• comparison types - custom functions (user passed) for the type of comparisons to be computed such as adjusted risk ratio / difference, etc. custom differences - forward, backward, centered, and or user passed. By default, comparisons uses centered differences

slopes:

• counterfactual slopes - grid of values to generate a hypothetical dataset (I am not so sure how this is different from conditional marginal effects since a grid of values is being used to compute this also. Anytime a grid of values is used to compute a pairwise grid, it is a hypothetical dataset).

These lists were compiled by cross referencing the main vignette's for slopes and comparisons and identifying "main" areas of functionality differences. (edited)

Bayesian additive regression trees (BART)

Adding support for PyMC-BART.

General

• Shared hyperpriors #687
• R2D2 implementation #647
• Better priors for a log-link function with low frequency #643
• Use a sparse matrix for group specific effects #545
  • This is more a formulae thing than a Bambi thing, but smaller updates will be needed on Bambi too.
• Non-linear effects
  • There's some work done at https://github.com/tomicapretto/bambi/tree/nlexprs but it's not finished
• Examples
  • Potentials
  • Offset
  • Variational inference
  • JAX-based samplers
  • Piecewise regression