Skip to content

Evaluation

onlinecml.evaluation.progressive.progressive_causal_score(stream, model, metrics, step=100)

Evaluate a causal model progressively over a streaming dataset.

Implements the predict-before-learn protocol: for each observation, the model is scored before it sees the label, then updated. This gives an unbiased estimate of online generalisation performance.

At every step observations, each metric's current score is recorded.

Parameters:

Name Type Description Default
stream iterable of (x, treatment, outcome, true_cate)

Any OnlineCML dataset or iterable yielding 4-tuples.

 required
model BaseOnlineEstimator

An unfitted (or partially fitted) causal estimator. Must implement learn_one(x, treatment, outcome) and predict_one(x).

 required
metrics list

List of metric objects (e.g., [ATEError(), PEHE()]). Each must implement update(x, w, y, true_cate, cate_hat, model) and a score property.

 required
step int

Record metric scores every step observations. Default 100.

 100

Returns:

Name Type Description
results dict

Dictionary with key "steps" (list of checkpoint indices) and one key per metric class name (list of scores at each checkpoint).

Examples:

>>> from onlinecml.datasets import LinearCausalStream
>>> from onlinecml.reweighting import OnlineIPW
>>> from onlinecml.evaluation import progressive_causal_score
>>> from onlinecml.evaluation.metrics import ATEError, PEHE
>>>
>>> results = progressive_causal_score(
...     stream  = LinearCausalStream(n=500, seed=0),
...     model   = OnlineIPW(),
...     metrics = [ATEError(), PEHE()],
...     step    = 100,
... )
>>> len(results["steps"])
5
Source code in onlinecml/evaluation/progressive.py 
11
12
13
14
15
16
17
18
19
20
21
22
23
24
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
def progressive_causal_score(
    stream,  # noqa: ANN001
    model: "BaseOnlineEstimator",
    metrics: list,
    step: int = 100,
) -> dict[str, Any]:
    """Evaluate a causal model progressively over a streaming dataset.

    Implements the predict-before-learn protocol: for each observation, the
    model is scored **before** it sees the label, then updated. This gives an
    unbiased estimate of online generalisation performance.

    At every ``step`` observations, each metric's current ``score`` is recorded.

    Parameters
    ----------
    stream : iterable of (x, treatment, outcome, true_cate)
        Any OnlineCML dataset or iterable yielding 4-tuples.
    model : BaseOnlineEstimator
        An unfitted (or partially fitted) causal estimator. Must implement
        ``learn_one(x, treatment, outcome)`` and ``predict_one(x)``.
    metrics : list
        List of metric objects (e.g., ``[ATEError(), PEHE()]``). Each must
        implement ``update(x, w, y, true_cate, cate_hat, model)`` and a
        ``score`` property.
    step : int
        Record metric scores every ``step`` observations. Default 100.

    Returns
    -------
    results : dict
        Dictionary with key ``"steps"`` (list of checkpoint indices) and one
        key per metric class name (list of scores at each checkpoint).

    Examples
    --------
    >>> from onlinecml.datasets import LinearCausalStream
    >>> from onlinecml.reweighting import OnlineIPW
    >>> from onlinecml.evaluation import progressive_causal_score
    >>> from onlinecml.evaluation.metrics import ATEError, PEHE
    >>>
    >>> results = progressive_causal_score(
    ...     stream  = LinearCausalStream(n=500, seed=0),
    ...     model   = OnlineIPW(),
    ...     metrics = [ATEError(), PEHE()],
    ...     step    = 100,
    ... )
    >>> len(results["steps"])
    5
    """
    history: dict[str, list[float]] = {m.__class__.__name__: [] for m in metrics}
    steps_list: list[int] = []

    for i, (x, w, y, true_cate) in enumerate(stream):
        # Predict-before-learn
        cate_hat = model.predict_one(x)

        # Update each metric with the current prediction
        for m in metrics:
            m.update(x, w, y, true_cate, cate_hat, model)

        # Train the model
        model.learn_one(x, w, y)

        # Record checkpoint
        if (i + 1) % step == 0:
            steps_list.append(i + 1)
            for m in metrics:
                history[m.__class__.__name__].append(m.score)

    return {"steps": steps_list, **history}

Metrics

onlinecml.evaluation.metrics.ATEError

Running absolute error between the estimated and true ATE.

Accumulates the true CATE via a running mean (so it works for both constant-ATE and heterogeneous streams) and computes |model.predict_ate() - mean(true_cate)| at each checkpoint.

Examples:

>>> m = ATEError()
>>> m.score
0.0
Source code in onlinecml/evaluation/metrics.py 
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
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
class ATEError:
    """Running absolute error between the estimated and true ATE.

    Accumulates the true CATE via a running mean (so it works for both
    constant-ATE and heterogeneous streams) and computes
    ``|model.predict_ate() - mean(true_cate)|`` at each checkpoint.

    Examples
    --------
    >>> m = ATEError()
    >>> m.score
    0.0
    """

    def __init__(self) -> None:
        self._n: int = 0
        self._sum_true: float = 0.0
        self._last_ate_hat: float = 0.0

    def update(
        self,
        x: dict,
        w: int,
        y: float,
        true_cate: float,
        cate_hat: float,
        model,  # noqa: ANN001
    ) -> None:
        """Accumulate one observation.

        Parameters
        ----------
        x : dict
            Covariate dict (unused by this metric).
        w : int
            Treatment indicator (unused by this metric).
        y : float
            Observed outcome (unused by this metric).
        true_cate : float
            True CATE for this unit. Used to build a running mean of the
            population ATE.
        cate_hat : float
            Predicted CATE (unused by this metric; uses model.predict_ate()).
        model :
            The causal estimator. Must implement ``predict_ate() -> float``.
        """
        self._n += 1
        self._sum_true += true_cate
        self._last_ate_hat = model.predict_ate()

    @property
    def score(self) -> float:
        """Current |ATE_hat - ATE_true|.

        Returns ``0.0`` before any data has been seen.
        """
        if self._n == 0:
            return 0.0
        return abs(self._last_ate_hat - self._sum_true / self._n)

    def reset(self) -> None:
        """Reset all accumulated state."""
        self.__init__()  # type: ignore[misc]

score property

Current |ATE_hat - ATE_true|.

Returns 0.0 before any data has been seen.

reset()

Reset all accumulated state.

Source code in onlinecml/evaluation/metrics.py 
66
67
68
def reset(self) -> None:
    """Reset all accumulated state."""
    self.__init__()  # type: ignore[misc]

update(x, w, y, true_cate, cate_hat, model)

Accumulate one observation.

Parameters:

Name Type Description Default
x dict

Covariate dict (unused by this metric).

 required
w int

Treatment indicator (unused by this metric).

 required
y float

Observed outcome (unused by this metric).

 required
true_cate float

True CATE for this unit. Used to build a running mean of the population ATE.

 required
cate_hat float

Predicted CATE (unused by this metric; uses model.predict_ate()).

 required
model

The causal estimator. Must implement predict_ate() -> float.

 required
Source code in onlinecml/evaluation/metrics.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
def update(
    self,
    x: dict,
    w: int,
    y: float,
    true_cate: float,
    cate_hat: float,
    model,  # noqa: ANN001
) -> None:
    """Accumulate one observation.

    Parameters
    ----------
    x : dict
        Covariate dict (unused by this metric).
    w : int
        Treatment indicator (unused by this metric).
    y : float
        Observed outcome (unused by this metric).
    true_cate : float
        True CATE for this unit. Used to build a running mean of the
        population ATE.
    cate_hat : float
        Predicted CATE (unused by this metric; uses model.predict_ate()).
    model :
        The causal estimator. Must implement ``predict_ate() -> float``.
    """
    self._n += 1
    self._sum_true += true_cate
    self._last_ate_hat = model.predict_ate()

onlinecml.evaluation.metrics.PEHE

Running Precision in Estimation of Heterogeneous Effects.

Computes sqrt(mean((cate_hat - cate_true)^2)) incrementally using Welford's algorithm. The cate_hat passed by progressive_causal_score is the predict-before-learn CATE prediction from model.predict_one(x).

References

Hill, J. (2011). Bayesian nonparametric modeling for causal inference. Journal of Computational and Graphical Statistics, 20(1), 217-240.

Examples:

>>> m = PEHE()
>>> m.score
0.0
Source code in onlinecml/evaluation/metrics.py 
 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
class PEHE:
    """Running Precision in Estimation of Heterogeneous Effects.

    Computes ``sqrt(mean((cate_hat - cate_true)^2))`` incrementally using
    Welford's algorithm. The ``cate_hat`` passed by ``progressive_causal_score``
    is the predict-before-learn CATE prediction from ``model.predict_one(x)``.

    References
    ----------
    Hill, J. (2011). Bayesian nonparametric modeling for causal inference.
    Journal of Computational and Graphical Statistics, 20(1), 217-240.

    Examples
    --------
    >>> m = PEHE()
    >>> m.score
    0.0
    """

    def __init__(self) -> None:
        self._n: int = 0
        self._mean_sq: float = 0.0

    def update(
        self,
        x: dict,
        w: int,
        y: float,
        true_cate: float,
        cate_hat: float,
        model,  # noqa: ANN001
    ) -> None:
        """Accumulate one observation.

        Parameters
        ----------
        x : dict
            Covariate dict (unused by this metric).
        w : int
            Treatment indicator (unused by this metric).
        y : float
            Observed outcome (unused by this metric).
        true_cate : float
            True CATE for this unit.
        cate_hat : float
            Predicted CATE from ``model.predict_one(x)`` before learning.
        model :
            The causal estimator (unused by this metric).
        """
        self._n += 1
        err_sq = (cate_hat - true_cate) ** 2
        self._mean_sq += (err_sq - self._mean_sq) / self._n

    @property
    def score(self) -> float:
        """Current PEHE (sqrt of running mean squared CATE error).

        Returns ``0.0`` before any data has been seen.
        """
        if self._n == 0:
            return 0.0
        return math.sqrt(self._mean_sq)

    def reset(self) -> None:
        """Reset all accumulated state."""
        self.__init__()  # type: ignore[misc]

score property

Current PEHE (sqrt of running mean squared CATE error).

Returns 0.0 before any data has been seen.

reset()

Reset all accumulated state.

Source code in onlinecml/evaluation/metrics.py 
134
135
136
def reset(self) -> None:
    """Reset all accumulated state."""
    self.__init__()  # type: ignore[misc]

update(x, w, y, true_cate, cate_hat, model)

Accumulate one observation.

Parameters:

Name Type Description Default
x dict

Covariate dict (unused by this metric).

 required
w int

Treatment indicator (unused by this metric).

 required
y float

Observed outcome (unused by this metric).

 required
true_cate float

True CATE for this unit.

 required
cate_hat float

Predicted CATE from model.predict_one(x) before learning.

 required
model

The causal estimator (unused by this metric).

 required
Source code in onlinecml/evaluation/metrics.py 
 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
def update(
    self,
    x: dict,
    w: int,
    y: float,
    true_cate: float,
    cate_hat: float,
    model,  # noqa: ANN001
) -> None:
    """Accumulate one observation.

    Parameters
    ----------
    x : dict
        Covariate dict (unused by this metric).
    w : int
        Treatment indicator (unused by this metric).
    y : float
        Observed outcome (unused by this metric).
    true_cate : float
        True CATE for this unit.
    cate_hat : float
        Predicted CATE from ``model.predict_one(x)`` before learning.
    model :
        The causal estimator (unused by this metric).
    """
    self._n += 1
    err_sq = (cate_hat - true_cate) ** 2
    self._mean_sq += (err_sq - self._mean_sq) / self._n

onlinecml.evaluation.metrics.UpliftAUC

Area under the uplift curve (AUUC).

Accumulates (cate_hat, treatment, outcome) triples. At each call to score, it sorts the buffer by predicted CATE descending, computes the cumulative uplift curve, and returns the area via the trapezoidal rule, normalized to [0, 1].

The uplift at depth k is::

uplift(k) = mean_outcome_treated(top_k) - mean_outcome_control(top_k)

Parameters:

Name Type Description Default
max_buffer int

Maximum number of recent observations to retain. Older observations are dropped to keep memory bounded. Default 5000.

 5000

Examples:

>>> m = UpliftAUC()
>>> m.score
0.0
Source code in onlinecml/evaluation/metrics.py 
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
class UpliftAUC:
    """Area under the uplift curve (AUUC).

    Accumulates ``(cate_hat, treatment, outcome)`` triples. At each call to
    ``score``, it sorts the buffer by predicted CATE descending, computes the
    cumulative uplift curve, and returns the area via the trapezoidal rule,
    normalized to ``[0, 1]``.

    The uplift at depth ``k`` is::

        uplift(k) = mean_outcome_treated(top_k) - mean_outcome_control(top_k)

    Parameters
    ----------
    max_buffer : int
        Maximum number of recent observations to retain. Older observations
        are dropped to keep memory bounded. Default 5000.

    Examples
    --------
    >>> m = UpliftAUC()
    >>> m.score
    0.0
    """

    def __init__(self, max_buffer: int = 5000) -> None:
        self.max_buffer = max_buffer
        self._buffer: list[tuple[float, int, float]] = []

    def update(
        self,
        x: dict,
        w: int,
        y: float,
        true_cate: float,
        cate_hat: float,
        model,  # noqa: ANN001
    ) -> None:
        """Accumulate one observation.

        Parameters
        ----------
        x : dict
            Covariate dict (unused by this metric).
        w : int
            Treatment indicator (0 or 1).
        y : float
            Observed outcome.
        true_cate : float
            True CATE (unused by this metric).
        cate_hat : float
            Predicted CATE used to rank units.
        model :
            The causal estimator (unused by this metric).
        """
        self._buffer.append((cate_hat, w, y))
        if len(self._buffer) > self.max_buffer:
            self._buffer.pop(0)

    @property
    def score(self) -> float:
        """Current AUUC, normalized to ``[0, 1]``.

        Returns ``0.0`` when fewer than two observations have been seen or
        when all units are in one arm.
        """
        if len(self._buffer) < 2:
            return 0.0
        sorted_buf = sorted(self._buffer, key=lambda t: t[0], reverse=True)
        n = len(sorted_buf)

        cum_t_sum, cum_c_sum = 0.0, 0.0
        cum_t_n,   cum_c_n   = 0,   0
        uplift_vals = []

        for _, w, y in sorted_buf:
            if w == 1:
                cum_t_sum += y
                cum_t_n   += 1
            else:
                cum_c_sum += y
                cum_c_n   += 1
            mean_t = cum_t_sum / cum_t_n if cum_t_n > 0 else 0.0
            mean_c = cum_c_sum / cum_c_n if cum_c_n > 0 else 0.0
            uplift_vals.append(mean_t - mean_c)

        if not uplift_vals:
            return 0.0
        # Trapezoidal AUC over depth fractions [0, 1]
        auc = sum(uplift_vals) / n
        # Normalize by the range of outcomes so score is on a comparable scale
        all_y = [y for _, _, y in sorted_buf]
        y_range = max(all_y) - min(all_y)
        if y_range == 0.0:
            return 0.0
        return max(0.0, auc / y_range)

    def reset(self) -> None:
        """Reset all accumulated state."""
        self._buffer.clear()

score property

Current AUUC, normalized to [0, 1].

Returns 0.0 when fewer than two observations have been seen or when all units are in one arm.

reset()

Reset all accumulated state.

Source code in onlinecml/evaluation/metrics.py 
236
237
238
def reset(self) -> None:
    """Reset all accumulated state."""
    self._buffer.clear()

update(x, w, y, true_cate, cate_hat, model)

Accumulate one observation.

Parameters:

Name Type Description Default
x dict

Covariate dict (unused by this metric).

 required
w int

Treatment indicator (0 or 1).

 required
y float

Observed outcome.

 required
true_cate float

True CATE (unused by this metric).

 required
cate_hat float

Predicted CATE used to rank units.

 required
model

The causal estimator (unused by this metric).

 required
Source code in onlinecml/evaluation/metrics.py 
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
def update(
    self,
    x: dict,
    w: int,
    y: float,
    true_cate: float,
    cate_hat: float,
    model,  # noqa: ANN001
) -> None:
    """Accumulate one observation.

    Parameters
    ----------
    x : dict
        Covariate dict (unused by this metric).
    w : int
        Treatment indicator (0 or 1).
    y : float
        Observed outcome.
    true_cate : float
        True CATE (unused by this metric).
    cate_hat : float
        Predicted CATE used to rank units.
    model :
        The causal estimator (unused by this metric).
    """
    self._buffer.append((cate_hat, w, y))
    if len(self._buffer) > self.max_buffer:
        self._buffer.pop(0)

onlinecml.evaluation.metrics.QiniCoefficient

Qini coefficient (area under the Qini curve).

The Qini curve plots cumulative incremental gains vs. cumulative population fraction when units are ranked by predicted CATE descending. The Qini coefficient is the area under this curve minus the area under the random policy line, normalized by the maximum achievable Qini.

Parameters:

Name Type Description Default
max_buffer int

Maximum number of recent observations to retain. Default 5000.

 5000
References

Radcliffe, N.J. (2007). Using control groups to target on predicted lift. Direct Marketing Analytics Journal, 14-21.

Examples:

>>> m = QiniCoefficient()
>>> m.score
0.0
Source code in onlinecml/evaluation/metrics.py 
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
class QiniCoefficient:
    """Qini coefficient (area under the Qini curve).

    The Qini curve plots cumulative incremental gains vs. cumulative population
    fraction when units are ranked by predicted CATE descending. The Qini
    coefficient is the area under this curve minus the area under the random
    policy line, normalized by the maximum achievable Qini.

    Parameters
    ----------
    max_buffer : int
        Maximum number of recent observations to retain. Default 5000.

    References
    ----------
    Radcliffe, N.J. (2007). Using control groups to target on predicted lift.
    Direct Marketing Analytics Journal, 14-21.

    Examples
    --------
    >>> m = QiniCoefficient()
    >>> m.score
    0.0
    """

    def __init__(self, max_buffer: int = 5000) -> None:
        self.max_buffer = max_buffer
        self._buffer: list[tuple[float, int, float]] = []

    def update(
        self,
        x: dict,
        w: int,
        y: float,
        true_cate: float,
        cate_hat: float,
        model,  # noqa: ANN001
    ) -> None:
        """Accumulate one observation.

        Parameters
        ----------
        x : dict
            Covariate dict (unused by this metric).
        w : int
            Treatment indicator (0 or 1).
        y : float
            Observed outcome.
        true_cate : float
            True CATE (unused by this metric).
        cate_hat : float
            Predicted CATE used to rank units.
        model :
            The causal estimator (unused by this metric).
        """
        self._buffer.append((cate_hat, w, y))
        if len(self._buffer) > self.max_buffer:
            self._buffer.pop(0)

    @property
    def score(self) -> float:
        """Current normalized Qini coefficient.

        Returns ``0.0`` when fewer than two observations have been seen or
        when either arm is empty.
        """
        if len(self._buffer) < 2:
            return 0.0
        sorted_buf = sorted(self._buffer, key=lambda t: t[0], reverse=True)
        n = len(sorted_buf)

        n_t_total = sum(1 for _, w, _ in sorted_buf if w == 1)
        n_c_total = n - n_t_total
        if n_t_total == 0 or n_c_total == 0:
            return 0.0

        cum_t, cum_c = 0, 0
        qini_vals = [0.0]  # starts at 0

        for _, w, _ in sorted_buf:
            if w == 1:
                cum_t += 1
            else:
                cum_c += 1
            # Qini at this depth: treated_rate - control_rate * (n_t_total / n_c_total)
            qini_vals.append(cum_t / n_t_total - cum_c / n_c_total)

        # Area under Qini curve (trapezoidal)
        depths = [i / n for i in range(n + 1)]
        auc = sum(
            0.5 * (qini_vals[i] + qini_vals[i + 1]) * (depths[i + 1] - depths[i])
            for i in range(n)
        )
        # Normalize: maximum area is 0.5 (perfect model)
        return auc / 0.5

    def reset(self) -> None:
        """Reset all accumulated state."""
        self._buffer.clear()

score property

Current normalized Qini coefficient.

Returns 0.0 when fewer than two observations have been seen or when either arm is empty.

reset()

Reset all accumulated state.

Source code in onlinecml/evaluation/metrics.py 
337
338
339
def reset(self) -> None:
    """Reset all accumulated state."""
    self._buffer.clear()

update(x, w, y, true_cate, cate_hat, model)

Accumulate one observation.

Parameters:

Name Type Description Default
x dict

Covariate dict (unused by this metric).

 required
w int

Treatment indicator (0 or 1).

 required
y float

Observed outcome.

 required
true_cate float

True CATE (unused by this metric).

 required
cate_hat float

Predicted CATE used to rank units.

 required
model

The causal estimator (unused by this metric).

 required
Source code in onlinecml/evaluation/metrics.py 
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
def update(
    self,
    x: dict,
    w: int,
    y: float,
    true_cate: float,
    cate_hat: float,
    model,  # noqa: ANN001
) -> None:
    """Accumulate one observation.

    Parameters
    ----------
    x : dict
        Covariate dict (unused by this metric).
    w : int
        Treatment indicator (0 or 1).
    y : float
        Observed outcome.
    true_cate : float
        True CATE (unused by this metric).
    cate_hat : float
        Predicted CATE used to rank units.
    model :
        The causal estimator (unused by this metric).
    """
    self._buffer.append((cate_hat, w, y))
    if len(self._buffer) > self.max_buffer:
        self._buffer.pop(0)

onlinecml.evaluation.metrics.CIWidth

Running mean width of the confidence interval on the ATE.

Computes the mean of (upper - lower) for each CI returned by model.predict_ci(alpha) at each observation.

Parameters:

Name Type Description Default
alpha float

Significance level for the CI. Default 0.05 (95% CI).

 0.05

Examples:

>>> m = CIWidth()
>>> m.score
0.0
Source code in onlinecml/evaluation/metrics.py 
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
class CIWidth:
    """Running mean width of the confidence interval on the ATE.

    Computes the mean of ``(upper - lower)`` for each CI returned by
    ``model.predict_ci(alpha)`` at each observation.

    Parameters
    ----------
    alpha : float
        Significance level for the CI. Default 0.05 (95% CI).

    Examples
    --------
    >>> m = CIWidth()
    >>> m.score
    0.0
    """

    def __init__(self, alpha: float = 0.05) -> None:
        self.alpha = alpha
        self._n: int = 0
        self._mean_width: float = 0.0

    def update(
        self,
        x: dict,
        w: int,
        y: float,
        true_cate: float,
        cate_hat: float,
        model,  # noqa: ANN001
    ) -> None:
        """Accumulate one observation.

        Parameters
        ----------
        x : dict
            Covariate dict (unused by this metric).
        w : int
            Treatment indicator (unused by this metric).
        y : float
            Observed outcome (unused by this metric).
        true_cate : float
            True CATE (unused by this metric).
        cate_hat : float
            Predicted CATE (unused by this metric).
        model :
            The causal estimator. Must implement ``predict_ci(alpha) -> tuple``.
        """
        lo, hi = model.predict_ci(alpha=self.alpha)
        width = hi - lo
        self._n += 1
        self._mean_width += (width - self._mean_width) / self._n

    @property
    def score(self) -> float:
        """Current mean CI width.

        Returns ``0.0`` before any data has been seen.
        """
        return self._mean_width

    def reset(self) -> None:
        """Reset all accumulated state."""
        self.__init__(alpha=self.alpha)  # type: ignore[misc]

score property

Current mean CI width.

Returns 0.0 before any data has been seen.

reset()

Reset all accumulated state.

Source code in onlinecml/evaluation/metrics.py 
404
405
406
def reset(self) -> None:
    """Reset all accumulated state."""
    self.__init__(alpha=self.alpha)  # type: ignore[misc]

update(x, w, y, true_cate, cate_hat, model)

Accumulate one observation.

Parameters:

Name Type Description Default
x dict

Covariate dict (unused by this metric).

 required
w int

Treatment indicator (unused by this metric).

 required
y float

Observed outcome (unused by this metric).

 required
true_cate float

True CATE (unused by this metric).

 required
cate_hat float

Predicted CATE (unused by this metric).

 required
model

The causal estimator. Must implement predict_ci(alpha) -> tuple.

 required
Source code in onlinecml/evaluation/metrics.py 
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
def update(
    self,
    x: dict,
    w: int,
    y: float,
    true_cate: float,
    cate_hat: float,
    model,  # noqa: ANN001
) -> None:
    """Accumulate one observation.

    Parameters
    ----------
    x : dict
        Covariate dict (unused by this metric).
    w : int
        Treatment indicator (unused by this metric).
    y : float
        Observed outcome (unused by this metric).
    true_cate : float
        True CATE (unused by this metric).
    cate_hat : float
        Predicted CATE (unused by this metric).
    model :
        The causal estimator. Must implement ``predict_ci(alpha) -> tuple``.
    """
    lo, hi = model.predict_ci(alpha=self.alpha)
    width = hi - lo
    self._n += 1
    self._mean_width += (width - self._mean_width) / self._n

onlinecml.evaluation.metrics.CIcoverage

Empirical coverage of the ATE confidence interval.

At each checkpoint, checks whether the true population ATE (running mean of true_cate) falls within model.predict_ci(alpha). Returns the running fraction of checkpoints where the CI covered the true ATE.

Parameters:

Name Type Description Default
alpha float

Significance level. Default 0.05 (95% CI; target coverage = 0.95).

 0.05

Examples:

>>> m = CIcoverage()
>>> m.score
0.0
Source code in onlinecml/evaluation/metrics.py 
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
class CIcoverage:
    """Empirical coverage of the ATE confidence interval.

    At each checkpoint, checks whether the true population ATE (running mean
    of ``true_cate``) falls within ``model.predict_ci(alpha)``. Returns the
    running fraction of checkpoints where the CI covered the true ATE.

    Parameters
    ----------
    alpha : float
        Significance level. Default 0.05 (95% CI; target coverage = 0.95).

    Examples
    --------
    >>> m = CIcoverage()
    >>> m.score
    0.0
    """

    def __init__(self, alpha: float = 0.05) -> None:
        self.alpha = alpha
        self._n_obs: int = 0
        self._sum_true: float = 0.0
        self._n_checks: int = 0
        self._n_covered: int = 0

    def update(
        self,
        x: dict,
        w: int,
        y: float,
        true_cate: float,
        cate_hat: float,
        model,  # noqa: ANN001
    ) -> None:
        """Accumulate one observation and check CI coverage.

        Parameters
        ----------
        x : dict
            Covariate dict (unused by this metric).
        w : int
            Treatment indicator (unused by this metric).
        y : float
            Observed outcome (unused by this metric).
        true_cate : float
            True CATE. Used to estimate the population ATE via running mean.
        cate_hat : float
            Predicted CATE (unused by this metric).
        model :
            The causal estimator. Must implement ``predict_ci(alpha) -> tuple``.
        """
        self._n_obs += 1
        self._sum_true += true_cate
        true_ate = self._sum_true / self._n_obs
        lo, hi = model.predict_ci(alpha=self.alpha)
        self._n_checks += 1
        if lo <= true_ate <= hi:
            self._n_covered += 1

    @property
    def score(self) -> float:
        """Fraction of checkpoints where the CI covered the true ATE.

        Returns ``0.0`` before any data has been seen.
        """
        if self._n_checks == 0:
            return 0.0
        return self._n_covered / self._n_checks

    def reset(self) -> None:
        """Reset all accumulated state."""
        self.__init__(alpha=self.alpha)  # type: ignore[misc]

score property

Fraction of checkpoints where the CI covered the true ATE.

Returns 0.0 before any data has been seen.

reset()

Reset all accumulated state.

Source code in onlinecml/evaluation/metrics.py 
479
480
481
def reset(self) -> None:
    """Reset all accumulated state."""
    self.__init__(alpha=self.alpha)  # type: ignore[misc]

update(x, w, y, true_cate, cate_hat, model)

Accumulate one observation and check CI coverage.

Parameters:

Name Type Description Default
x dict

Covariate dict (unused by this metric).

 required
w int

Treatment indicator (unused by this metric).

 required
y float

Observed outcome (unused by this metric).

 required
true_cate float

True CATE. Used to estimate the population ATE via running mean.

 required
cate_hat float

Predicted CATE (unused by this metric).

 required
model

The causal estimator. Must implement predict_ci(alpha) -> tuple.

 required
Source code in onlinecml/evaluation/metrics.py 
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
def update(
    self,
    x: dict,
    w: int,
    y: float,
    true_cate: float,
    cate_hat: float,
    model,  # noqa: ANN001
) -> None:
    """Accumulate one observation and check CI coverage.

    Parameters
    ----------
    x : dict
        Covariate dict (unused by this metric).
    w : int
        Treatment indicator (unused by this metric).
    y : float
        Observed outcome (unused by this metric).
    true_cate : float
        True CATE. Used to estimate the population ATE via running mean.
    cate_hat : float
        Predicted CATE (unused by this metric).
    model :
        The causal estimator. Must implement ``predict_ci(alpha) -> tuple``.
    """
    self._n_obs += 1
    self._sum_true += true_cate
    true_ate = self._sum_true / self._n_obs
    lo, hi = model.predict_ci(alpha=self.alpha)
    self._n_checks += 1
    if lo <= true_ate <= hi:
        self._n_covered += 1