3  Estimator Contract

The package is organized around a small public contract: fit on observed data, then call predict(...) at requested treatment levels or treatment grids.

3.1 One Interface, Different Estimands

flowchart TD
  A["Binary or categorical treatment"] --> B["Average potential outcomes"]
  C["Continuous treatment"] --> D["Average dose-response function"]
  E["Density estimator"] --> F["Treatment assignment score"]

The method names stay compact, but the returned quantity depends on the estimator family.

Family Typical output Interpretation
Categorical response estimators One value per requested treatment level Average potential outcome
Continuous response estimators One value per requested grid point Average dose-response function
Density estimators One value per observed row Treatment density or stabilized density ratio

3.2 Prediction Requires A Requested Treatment Table

The package does not treat prediction as “just give me X”. It treats prediction as “give me the intervention I want to evaluate”, and average-response estimators then average over the covariate sample stored during fit.

estimator.fit(X, t, y)
estimated_effect = estimator.predict(requested_treatment)

That means the shape of requested_treatment matters:

  1. For binary or categorical problems, it usually contains a small set of treatment levels.
  2. For continuous problems, it usually contains a dense treatment grid.

3.3 Density Outputs Need Interpretation Discipline

Not every density-like quantity is a normalized density.

Important

In skcausal, some density estimators return a conditional density p(t | x) while others return a stabilized ratio p(t | x) / p(t). Always check estimator.get_tag("density_kind") before interpreting the output.

That distinction is why the density docs separate SkproDensityEstimator from PermutationWeighting.

3.4 Reading The Examples Correctly

The examples compare estimators on synthetic datasets, but they are not all trying to estimate the exact same intermediate object. They are trying to estimate the same causal target, sometimes through different nuisance models.

Continue to Continuous Workflows for the cleanest grid-based example.