Skip to content

Causal

uncertainty_flow.causal

CausalUncertaintyEstimator

Bases: BaseUncertaintyModel

Estimate treatment effects with quantified uncertainty.

Supports doubly robust, S-learner, and T-learner methods. Wraps an outcome model and (optionally) a propensity score model to produce CATE estimates and prediction intervals on treatment effects. Outcome-dependent summary metrics (ATE / confidence interval) are computed via :meth:evaluate, not :meth:predict.

Examples:

>>> from sklearn.ensemble import GradientBoostingRegressor
>>> from uncertainty_flow.wrappers import ConformalRegressor
>>> from uncertainty_flow.causal import CausalUncertaintyEstimator
>>> import polars as pl
>>> import numpy as np
>>>
>>> rng = np.random.default_rng(0)
>>> n = 200
>>> df = pl.DataFrame({
...     "x1": rng.standard_normal(n),
...     "x2": rng.standard_normal(n),
...     "treatment": rng.binomial(1, 0.5, n),
...     "outcome": np.zeros(n),
... })
>>> # Y = 2*T + noise
>>> outcome_vals = (
...     df["treatment"].to_numpy() * 2.0
...     + df["x1"].to_numpy()
...     + rng.standard_normal(n) * 0.5
... )
>>> df = df.with_columns(outcome=pl.Series(outcome_vals))
>>> base = GradientBoostingRegressor(n_estimators=10, random_state=42)
>>> outcome = ConformalRegressor(base_model=base, random_state=42)
>>> model = CausalUncertaintyEstimator(
...     outcome_model=outcome,
...     method="doubly_robust",
...     random_state=42,
... )
>>> model.fit(df, target="outcome")
>>> pred = model.predict(df.drop("outcome"))
>>> metrics = model.evaluate(df)
>>> ate = metrics["ate"]
Source code in uncertainty_flow/causal/estimator.py 
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
class CausalUncertaintyEstimator(BaseUncertaintyModel):
    """
    Estimate treatment effects with quantified uncertainty.

    Supports doubly robust, S-learner, and T-learner methods.
    Wraps an outcome model and (optionally) a propensity score model to
    produce CATE estimates and prediction intervals on treatment effects.
    Outcome-dependent summary metrics (ATE / confidence interval) are
    computed via :meth:`evaluate`, not :meth:`predict`.

    Examples:
        >>> from sklearn.ensemble import GradientBoostingRegressor
        >>> from uncertainty_flow.wrappers import ConformalRegressor
        >>> from uncertainty_flow.causal import CausalUncertaintyEstimator
        >>> import polars as pl
        >>> import numpy as np
        >>>
        >>> rng = np.random.default_rng(0)
        >>> n = 200
        >>> df = pl.DataFrame({
        ...     "x1": rng.standard_normal(n),
        ...     "x2": rng.standard_normal(n),
        ...     "treatment": rng.binomial(1, 0.5, n),
        ...     "outcome": np.zeros(n),
        ... })
        >>> # Y = 2*T + noise
        >>> outcome_vals = (
        ...     df["treatment"].to_numpy() * 2.0
        ...     + df["x1"].to_numpy()
        ...     + rng.standard_normal(n) * 0.5
        ... )
        >>> df = df.with_columns(outcome=pl.Series(outcome_vals))
        >>> base = GradientBoostingRegressor(n_estimators=10, random_state=42)
        >>> outcome = ConformalRegressor(base_model=base, random_state=42)
        >>> model = CausalUncertaintyEstimator(
        ...     outcome_model=outcome,
        ...     method="doubly_robust",
        ...     random_state=42,
        ... )
        >>> model.fit(df, target="outcome")
        >>> pred = model.predict(df.drop("outcome"))
        >>> metrics = model.evaluate(df)
        >>> ate = metrics["ate"]
    """

    def __init__(
        self,
        outcome_model,
        propensity_model=None,
        treatment_col: str = "treatment",
        method: str = "doubly_robust",
        random_state: int | None = None,
    ):
        """
        Initialize CausalUncertaintyEstimator.

        Args:
            outcome_model: Any sklearn-like regressor with fit/predict.
                Doubly-robust mode currently does not support conformal
                wrappers as the direct outcome model.
            propensity_model: Optional sklearn classifier for propensity
                scores. Defaults to LogisticRegression.
            treatment_col: Name of the binary treatment column.
            method: One of "doubly_robust", "s_learner", "t_learner".
            random_state: Random seed for reproducibility.

        Raises:
            ValueError: If method is not one of VALID_METHODS.
        """
        if method not in VALID_METHODS:
            raise ConfigurationError(f"Invalid method '{method}'. Must be one of {VALID_METHODS}")
        self.outcome_model = outcome_model
        self.propensity_model = propensity_model
        self.treatment_col = treatment_col
        self.method = method
        self.random_state = random_state
        self._fitted = False
        self._effect_std_: float = 0.0

    # ------------------------------------------------------------------
    # fit
    # ------------------------------------------------------------------

    def fit(
        self,
        data: PolarsInput,
        target: TargetSpec | None = None,
        **kwargs,
    ) -> "CausalUncertaintyEstimator":
        """
        Fit the causal model to training data.

        Args:
            data: Polars DataFrame or LazyFrame with features, treatment
                indicator, and outcome.
            target: Outcome column name.
            **kwargs: Additional keyword arguments (unused).

        Returns:
            self (for method chaining).
        """
        data = materialize_lazyframe(data)

        if target is None:
            raise ConfigurationError("target is required for CausalUncertaintyEstimator")
        target_str = target if isinstance(target, str) else target[0]

        # Feature columns = everything except target and treatment
        self._target_col_ = target_str
        self._feature_cols_ = [
            col for col in data.columns if col != target_str and col != self.treatment_col
        ]

        feature_cols = self._feature_cols_
        t = to_numpy_series(data[self.treatment_col]).astype(float)
        y = to_numpy_series(data[target_str]).astype(float)
        x = to_numpy(data, feature_cols)

        if self.method == "doubly_robust":
            self._fit_doubly_robust(data, x, t, y, target_str, feature_cols)
        elif self.method == "s_learner":
            self._fit_s_learner(data, x, t, y, target_str, feature_cols)
        elif self.method == "t_learner":
            self._fit_t_learner(data, x, t, y, target_str, feature_cols)

        if self.method == "doubly_robust":
            self._effect_std_ = float(np.std(self._dr_scores_train, ddof=1))
        else:
            self._effect_std_ = float(np.std(self._cate_train, ddof=1))

        self._fitted = True
        return self

    # ------------------------------------------------------------------
    # predict
    # ------------------------------------------------------------------

    def predict(self, data: PolarsInput) -> DistributionPrediction:
        """
        Generate predictions with treatment effect estimates.

        Args:
            data: Polars DataFrame or LazyFrame with features and treatment.

        Returns:
            DistributionPrediction with CATE-based uncertainty bands.
        """
        if not self._fitted:
            raise ModelNotFittedError("CausalUncertaintyEstimator")

        data = materialize_lazyframe(data)

        feature_cols = self._feature_cols_
        x = to_numpy(data, feature_cols)

        if self.method == "doubly_robust":
            mu1, mu0 = self._predict_counterfactuals_dr(x, data, feature_cols)
        elif self.method == "s_learner":
            mu1, mu0 = self._predict_counterfactuals_sl(x, data, feature_cols)
        elif self.method == "t_learner":
            mu1, mu0 = self._predict_counterfactuals_tl(x)
        else:
            raise ConfigurationError(f"Unknown method: {self.method}")

        cate = mu1 - mu0

        # Build quantile matrix from CATE ± DR-score-based uncertainty
        cate_std = float(self._effect_std_)
        quantile_levels = _CAUSAL_QUANTILES
        z_scores = np.array([-1.2816, 0.0, 1.2816])  # normal quantiles for 0.1, 0.5, 0.9
        quantile_matrix = np.column_stack([cate + z * cate_std for z in z_scores])

        return DistributionPrediction(
            quantile_matrix=quantile_matrix,
            quantile_levels=quantile_levels,
            target_names=["treatment_effect"],
            treatment_info={
                "cate": cate,
                "treatment_col": self.treatment_col,
            },
        )

    def evaluate(self, data: PolarsInput) -> dict[str, float | tuple[float, float]]:
        """
        Compute outcome-dependent treatment-effect metrics on labeled data.

        Args:
            data: Data with features, treatment indicator, and observed outcome.

        Returns:
            Dictionary with ``ate`` and ``ate_ci``.
        """
        if not self._fitted:
            raise ModelNotFittedError("CausalUncertaintyEstimator")

        data = materialize_lazyframe(data)

        if self._target_col_ not in data.columns:
            raise ConfigurationError(
                f"evaluate() requires observed outcome column '{self._target_col_}'."
            )

        feature_cols = self._feature_cols_
        t = to_numpy_series(data[self.treatment_col]).astype(float)
        y = to_numpy_series(data[self._target_col_]).astype(float)
        x = to_numpy(data, feature_cols)

        if self.method == "doubly_robust":
            mu1, mu0 = self._predict_counterfactuals_dr(x, data, feature_cols)
            e = self._propensity_predict(x)
            e = np.clip(e, 0.01, 0.99)
            scores = mu1 - mu0 + t * (y - mu1) / e - (1 - t) * (y - mu0) / (1 - e)
        elif self.method == "s_learner":
            mu1, mu0 = self._predict_counterfactuals_sl(x, data, feature_cols)
            scores = mu1 - mu0
        elif self.method == "t_learner":
            mu1, mu0 = self._predict_counterfactuals_tl(x)
            scores = mu1 - mu0
        else:
            raise ConfigurationError(f"Unknown method: {self.method}")

        ate = float(np.mean(scores))
        se = float(np.std(scores, ddof=1) / np.sqrt(len(scores)))
        ate_ci = (ate - 1.96 * se, ate + 1.96 * se)
        return {"ate": ate, "ate_ci": ate_ci}

    # ------------------------------------------------------------------
    # Private helpers - fit
    # ------------------------------------------------------------------

    def _fit_doubly_robust(self, data, x, t, y, target_str, feature_cols):
        """Fit outcome model (features + treatment) and propensity model."""
        from sklearn.base import clone

        if hasattr(self.outcome_model, "base_model"):
            raise ConfigurationError(
                "doubly_robust currently does not support conformal wrapper outcome_model "
                "objects because conformalized DR intervals are not yet implemented. "
                "Pass a direct regressor outcome_model or use method='s_learner'/'t_learner'."
            )

        x_aug = np.column_stack([x, t])

        self._outcome_model_fitted = clone(self.outcome_model)
        self._outcome_model_fitted.fit(x_aug, y)

        if self.propensity_model is not None:
            self._propensity_fitted = clone(self.propensity_model)
        else:
            self._propensity_fitted = LogisticRegression(
                random_state=self.random_state, max_iter=1000
            )
        self._propensity_fitted.fit(x, t)

        e = self._propensity_fitted.predict_proba(x)[:, 1]
        e = np.clip(e, 0.01, 0.99)

        x_aug1 = np.column_stack([x, np.ones(len(t))])
        x_aug0 = np.column_stack([x, np.zeros(len(t))])
        self._mu1_train = self._outcome_model_fitted.predict(x_aug1)
        self._mu0_train = self._outcome_model_fitted.predict(x_aug0)

        self._dr_scores_train = (
            self._mu1_train
            - self._mu0_train
            + t * (y - self._mu1_train) / e
            - (1 - t) * (y - self._mu0_train) / (1 - e)
        )

    def _fit_s_learner(self, data, x, t, y, target_str, feature_cols):
        """Fit single outcome model with treatment as feature."""
        from sklearn.base import clone

        x_aug = np.column_stack([x, t])
        base = (
            self.outcome_model.base_model
            if hasattr(self.outcome_model, "base_model")
            else self.outcome_model
        )
        self._outcome_model_fitted = clone(base)
        self._outcome_model_fitted.fit(x_aug, y)
        x_aug1 = np.column_stack([x, np.ones(len(t))])
        x_aug0 = np.column_stack([x, np.zeros(len(t))])
        self._cate_train = self._outcome_model_fitted.predict(
            x_aug1
        ) - self._outcome_model_fitted.predict(x_aug0)

    def _fit_t_learner(self, data, x, t, y, target_str, feature_cols):
        """Fit separate outcome models for T=0 and T=1 groups."""
        from sklearn.base import clone

        mask1 = t == 1
        mask0 = t == 0

        if not np.any(mask1) or not np.any(mask0):
            raise ConfigurationError(
                "t_learner requires at least one treated and one control sample"
            )

        base = (
            self.outcome_model.base_model
            if hasattr(self.outcome_model, "base_model")
            else self.outcome_model
        )

        self._model_t1 = clone(base)
        self._model_t0 = clone(base)

        self._model_t1.fit(x[mask1], y[mask1])
        self._model_t0.fit(x[mask0], y[mask0])
        self._cate_train = self._model_t1.predict(x) - self._model_t0.predict(x)

    # ------------------------------------------------------------------
    # Private helpers - predict
    # ------------------------------------------------------------------

    def _propensity_predict(self, x):
        """Return propensity scores P(T=1|X)."""
        return self._propensity_fitted.predict_proba(x)[:, 1]

    def _predict_counterfactuals_dr(self, x, data, feature_cols):
        """Compute mu1 and mu0 for doubly robust method."""
        x_aug1 = np.column_stack([x, np.ones(x.shape[0])])
        x_aug0 = np.column_stack([x, np.zeros(x.shape[0])])
        mu1 = self._outcome_model_fitted.predict(x_aug1)
        mu0 = self._outcome_model_fitted.predict(x_aug0)
        return mu1, mu0

    def _predict_counterfactuals_sl(self, x, data, feature_cols):
        """Compute mu1 and mu0 for S-learner."""
        x_aug1 = np.column_stack([x, np.ones(x.shape[0])])
        x_aug0 = np.column_stack([x, np.zeros(x.shape[0])])
        mu1 = self._outcome_model_fitted.predict(x_aug1)
        mu0 = self._outcome_model_fitted.predict(x_aug0)
        return mu1, mu0

    def _predict_counterfactuals_tl(self, x):
        """Compute mu1 and mu0 for T-learner."""
        mu1 = self._model_t1.predict(x)
        mu0 = self._model_t0.predict(x)
        return mu1, mu0

__init__(outcome_model, propensity_model=None, treatment_col='treatment', method='doubly_robust', random_state=None)

Initialize CausalUncertaintyEstimator.

Parameters:

Name Type Description Default
outcome_model

Any sklearn-like regressor with fit/predict. Doubly-robust mode currently does not support conformal wrappers as the direct outcome model.

 required
propensity_model

Optional sklearn classifier for propensity scores. Defaults to LogisticRegression.

 None
treatment_col str

Name of the binary treatment column.

 'treatment'
method str

One of "doubly_robust", "s_learner", "t_learner".

 'doubly_robust'
random_state int | None

Random seed for reproducibility.

 None

Raises:

Type Description
ValueError

If method is not one of VALID_METHODS.
Source code in uncertainty_flow/causal/estimator.py 
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
def __init__(
    self,
    outcome_model,
    propensity_model=None,
    treatment_col: str = "treatment",
    method: str = "doubly_robust",
    random_state: int | None = None,
):
    """
    Initialize CausalUncertaintyEstimator.

    Args:
        outcome_model: Any sklearn-like regressor with fit/predict.
            Doubly-robust mode currently does not support conformal
            wrappers as the direct outcome model.
        propensity_model: Optional sklearn classifier for propensity
            scores. Defaults to LogisticRegression.
        treatment_col: Name of the binary treatment column.
        method: One of "doubly_robust", "s_learner", "t_learner".
        random_state: Random seed for reproducibility.

    Raises:
        ValueError: If method is not one of VALID_METHODS.
    """
    if method not in VALID_METHODS:
        raise ConfigurationError(f"Invalid method '{method}'. Must be one of {VALID_METHODS}")
    self.outcome_model = outcome_model
    self.propensity_model = propensity_model
    self.treatment_col = treatment_col
    self.method = method
    self.random_state = random_state
    self._fitted = False
    self._effect_std_: float = 0.0

fit(data, target=None, **kwargs)

Fit the causal model to training data.

Parameters:

Name Type Description Default
data PolarsInput

Polars DataFrame or LazyFrame with features, treatment indicator, and outcome.

 required
target TargetSpec | None

Outcome column name.

 None
**kwargs

Additional keyword arguments (unused).

 {}

Returns:

Type Description
'CausalUncertaintyEstimator'

self (for method chaining).
Source code in uncertainty_flow/causal/estimator.py 
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
def fit(
    self,
    data: PolarsInput,
    target: TargetSpec | None = None,
    **kwargs,
) -> "CausalUncertaintyEstimator":
    """
    Fit the causal model to training data.

    Args:
        data: Polars DataFrame or LazyFrame with features, treatment
            indicator, and outcome.
        target: Outcome column name.
        **kwargs: Additional keyword arguments (unused).

    Returns:
        self (for method chaining).
    """
    data = materialize_lazyframe(data)

    if target is None:
        raise ConfigurationError("target is required for CausalUncertaintyEstimator")
    target_str = target if isinstance(target, str) else target[0]

    # Feature columns = everything except target and treatment
    self._target_col_ = target_str
    self._feature_cols_ = [
        col for col in data.columns if col != target_str and col != self.treatment_col
    ]

    feature_cols = self._feature_cols_
    t = to_numpy_series(data[self.treatment_col]).astype(float)
    y = to_numpy_series(data[target_str]).astype(float)
    x = to_numpy(data, feature_cols)

    if self.method == "doubly_robust":
        self._fit_doubly_robust(data, x, t, y, target_str, feature_cols)
    elif self.method == "s_learner":
        self._fit_s_learner(data, x, t, y, target_str, feature_cols)
    elif self.method == "t_learner":
        self._fit_t_learner(data, x, t, y, target_str, feature_cols)

    if self.method == "doubly_robust":
        self._effect_std_ = float(np.std(self._dr_scores_train, ddof=1))
    else:
        self._effect_std_ = float(np.std(self._cate_train, ddof=1))

    self._fitted = True
    return self

predict(data)

Generate predictions with treatment effect estimates.

Parameters:

Name Type Description Default
data PolarsInput

Polars DataFrame or LazyFrame with features and treatment.

 required

Returns:

Type Description
DistributionPrediction

DistributionPrediction with CATE-based uncertainty bands.
Source code in uncertainty_flow/causal/estimator.py 
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
def predict(self, data: PolarsInput) -> DistributionPrediction:
    """
    Generate predictions with treatment effect estimates.

    Args:
        data: Polars DataFrame or LazyFrame with features and treatment.

    Returns:
        DistributionPrediction with CATE-based uncertainty bands.
    """
    if not self._fitted:
        raise ModelNotFittedError("CausalUncertaintyEstimator")

    data = materialize_lazyframe(data)

    feature_cols = self._feature_cols_
    x = to_numpy(data, feature_cols)

    if self.method == "doubly_robust":
        mu1, mu0 = self._predict_counterfactuals_dr(x, data, feature_cols)
    elif self.method == "s_learner":
        mu1, mu0 = self._predict_counterfactuals_sl(x, data, feature_cols)
    elif self.method == "t_learner":
        mu1, mu0 = self._predict_counterfactuals_tl(x)
    else:
        raise ConfigurationError(f"Unknown method: {self.method}")

    cate = mu1 - mu0

    # Build quantile matrix from CATE ± DR-score-based uncertainty
    cate_std = float(self._effect_std_)
    quantile_levels = _CAUSAL_QUANTILES
    z_scores = np.array([-1.2816, 0.0, 1.2816])  # normal quantiles for 0.1, 0.5, 0.9
    quantile_matrix = np.column_stack([cate + z * cate_std for z in z_scores])

    return DistributionPrediction(
        quantile_matrix=quantile_matrix,
        quantile_levels=quantile_levels,
        target_names=["treatment_effect"],
        treatment_info={
            "cate": cate,
            "treatment_col": self.treatment_col,
        },
    )

evaluate(data)

Compute outcome-dependent treatment-effect metrics on labeled data.

Parameters:

Name Type Description Default
data PolarsInput

Data with features, treatment indicator, and observed outcome.

 required

Returns:

Type Description
dict[str, float | tuple[float, float]]

Dictionary with ate and ate_ci.
Source code in uncertainty_flow/causal/estimator.py 
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
def evaluate(self, data: PolarsInput) -> dict[str, float | tuple[float, float]]:
    """
    Compute outcome-dependent treatment-effect metrics on labeled data.

    Args:
        data: Data with features, treatment indicator, and observed outcome.

    Returns:
        Dictionary with ``ate`` and ``ate_ci``.
    """
    if not self._fitted:
        raise ModelNotFittedError("CausalUncertaintyEstimator")

    data = materialize_lazyframe(data)

    if self._target_col_ not in data.columns:
        raise ConfigurationError(
            f"evaluate() requires observed outcome column '{self._target_col_}'."
        )

    feature_cols = self._feature_cols_
    t = to_numpy_series(data[self.treatment_col]).astype(float)
    y = to_numpy_series(data[self._target_col_]).astype(float)
    x = to_numpy(data, feature_cols)

    if self.method == "doubly_robust":
        mu1, mu0 = self._predict_counterfactuals_dr(x, data, feature_cols)
        e = self._propensity_predict(x)
        e = np.clip(e, 0.01, 0.99)
        scores = mu1 - mu0 + t * (y - mu1) / e - (1 - t) * (y - mu0) / (1 - e)
    elif self.method == "s_learner":
        mu1, mu0 = self._predict_counterfactuals_sl(x, data, feature_cols)
        scores = mu1 - mu0
    elif self.method == "t_learner":
        mu1, mu0 = self._predict_counterfactuals_tl(x)
        scores = mu1 - mu0
    else:
        raise ConfigurationError(f"Unknown method: {self.method}")

    ate = float(np.mean(scores))
    se = float(np.std(scores, ddof=1) / np.sqrt(len(scores)))
    ate_ci = (ate - 1.96 * se, ate + 1.96 * se)
    return {"ate": ate, "ate_ci": ate_ci}