Class Hierarchy

Models

Base Models:

• Model: These are standard models that just train on data and produce predictions
  • Examples: BoostedTrees, DebiasedLasso, LassoCV, NeuralNet
• LinearModel - (Inherits Model) Models that also provide coefficients on the features. Not a perfect name.
• CausalModel - (Inherits LinearModel) Models that also provide standard errors and a covariance matrix of the coefficients.
  • Examples: OLS, RidgeCV

SampleSplit/Fold-aware model. These are used directly by the estimation models which requir requires cross-fitting

• SampleSplitModel: Analogous to Model, but takes folds over the data for fit and predict.
  • Examples: CrossFitContainer (can wrap a basic Model), and ensemble models such as StackedSS ("Stacked Generalization", takes other SampleSplitModels) and BucketSS ("Bucket of Models", takes other SampleSplitModels)

Estimation Models:

• DoubleMLLikeModel: Models that have a baseline stage where treatment and outcome are predicted (using a SampleSplitModel) and then a causal stage where residual outcome is regressed (using a CausalModel). They take VarBuilder objects that dynamically create variables rather being passed explicit features. They also take a data splitter which is used to split the data in the baseline.
  • Examples: DoubleML, DynamicDML

Variable Generation

The user needs to be able to specify how features are generated in the baseline stages and how to construct treatments in the causal model. At it's core this is done by providing lists of VarBuilders. Interally these lists are managed by FeatureGenerators and TreatmentGenerators.

VarBuilder: Takes data (including residualized variables in the treatment stage) and produce output (either features or treatments). For the DynamicDML case they also can make lead-specific variables. At the baseline stage this lead-specific building is necessary to distinguish between variables that are known when they occur (most variables) and those that are pre-determined (e.g. holiday schedules). At the causal stage they can construct variables from multiple different lead-specific residualization models to generate time-related effects like "pull-forward" effects.

• Examples: OwnVar (for a simple coefficient), PToPVar (e.g. cross-product price effects)

Featurizer methods: These take a schema and produce lists of VarBuilders.

• Examples: default_dynamic_featurizer, default_panel_featurizer

Other

Schema: Stores the DataType of each column (e.g. is an integer really a category variable to be 1-hot encoded) as well as ColType (Outcome, Treatment, (normal control), or Predetermined (a control that is known in advance)).

EstimationDataset: Stores data, builds a multiindex, and stores the fold information when it's been fit using one of the Estimation Models

Predictions: Computers prediciton statistics from the Estimation Models

DDMLMarginalEffects: Helpful utility for complicated treatments