causalis.scenarios.unconfoundedness.refutation.score.residual_plots

Residual diagnostic plots for nuisance models g0/g1 and m.

Module Contents

Functions

plot_residual_diagnostics

Plot residual diagnostics for nuisance models.

Data

__all__

API

causalis.scenarios.unconfoundedness.refutation.score.residual_plots.plot_residual_diagnostics(estimate: causalis.data_contracts.causal_estimate.CausalEstimate, data: Optional[causalis.dgp.causaldata.CausalData] = None, *, clip_propensity: float = 1e-06, n_bins: int = 20, marker_size: float = 12.0, alpha: float = 0.35, max_scatter_points: Optional[int] = None, random_state: Optional[int] = None, figsize: Tuple[float, float] = (14.0, 4.8), dpi: int = 220, font_scale: float = 1.1, save: Optional[str] = None, save_dpi: Optional[int] = None, transparent: bool = False) → matplotlib.pyplot.Figure

Plot residual diagnostics for nuisance models.

Panels

  1. Treated-only: u1 = y - g1 vs g1.

  2. Control-only: u0 = y - g0 vs g0.

  3. Binned calibration error: E[d - m | m in bin] vs binned m.

Notes

These plots check the nuisance pieces directly:

  • outcome residuals :math:u_1 = Y - \hat g_1(X) and :math:u_0 = Y - \hat g_0(X) should look roughly centered around zero without strong patterns against fitted values,

  • treatment residuals :math:D - \hat m(X) should average near zero within propensity bins.

Clear trends usually mean the nuisance models still leave structure in the data, which can show up later as unstable score diagnostics.

Parameters

estimate : CausalEstimate Estimate with diagnostic data (m_hat, g0_hat; optionally g1_hat, y, d). data : CausalData, optional Optional fallback source for y and d when missing in diagnostic data. clip_propensity : float, default 1e-6 Clipping epsilon for propensity values in the treatment-residual panel. n_bins : int, default 20 Number of quantile bins for the binned-mean trend overlays. marker_size : float, default 12.0 Scatter marker size. alpha : float, default 0.35 Scatter opacity. max_scatter_points : int, optional Maximum number of points drawn in each residual scatter panel. When set, scatter points are sampled uniformly without replacement, while the binned-mean overlays and calibration panel still use all observations. random_state : int, optional Random seed used when max_scatter_points triggers scatter sampling. figsize : tuple, default (14.0, 4.8) Figure size. dpi : int, default 220 Dots per inch. font_scale : float, default 1.10 Font scaling factor. save : str, optional Path to save the figure. save_dpi : int, optional DPI for saving. transparent : bool, default False Whether to save with transparency.

Returns

matplotlib.figure.Figure The generated figure.

Examples

from sklearn.ensemble import RandomForestClassifier, RandomForestRegressor from causalis.dgp import obs_linear_26_dataset from causalis.scenarios.unconfoundedness.model import IRM data = obs_linear_26_dataset( … n=1000, … seed=3141, … include_oracle=False, … return_causal_data=True, … ) irm = IRM( … data=data, … ml_g=RandomForestRegressor( … n_estimators=200, … max_depth=6, … min_samples_leaf=5, … random_state=3141, … ), … ml_m=RandomForestClassifier( … n_estimators=200, … max_depth=6, … min_samples_leaf=5, … random_state=3141, … ), … n_folds=3, … random_state=3141, … ) estimate = irm.fit().estimate(score=”ATE”) fig = plot_residual_diagnostics(estimate, data=data) # doctest: +SKIP

causalis.scenarios.unconfoundedness.refutation.score.residual_plots.__all__

[‘plot_residual_diagnostics’]