artest-stats 0.2.0__tar.gz

artest_stats-0.2.0/LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

artest_stats-0.2.0/PKG-INFO

@@ -0,0 +1,20 @@
+Metadata-Version: 2.4
+Name: artest-stats
+Version: 0.2.0
+Summary: Experiment statistics: proportions, segment comparison, power analysis and the SRM check.
+License-Expression: Apache-2.0
+License-File: LICENSE
+Requires-Dist: pydantic>=2.14
+Requires-Dist: scipy>=1.16
+Requires-Python: >=3.12
+Project-URL: Repository, https://gitlab.com/Arta_Shoja/artest
+Description-Content-Type: text/markdown
+
+# artest-stats
+
+Experiment statistics: proportions, segment comparison, power analysis and the SRM check.
+
+Pure numerical functions with no I/O and no language models. Every number Artest works out for itself
+to drive a decision comes from here; an experiment's verdict is GrowthBook's.
+
+Import as `artest.stats`. See the [decision record](https://gitlab.com/Arta_Shoja/artest/-/blob/main/docs/adr/0018-statistics.md).

artest_stats-0.2.0/README.md

@@ -0,0 +1,8 @@
+# artest-stats
+
+Experiment statistics: proportions, segment comparison, power analysis and the SRM check.
+
+Pure numerical functions with no I/O and no language models. Every number Artest works out for itself
+to drive a decision comes from here; an experiment's verdict is GrowthBook's.
+
+Import as `artest.stats`. See the [decision record](https://gitlab.com/Arta_Shoja/artest/-/blob/main/docs/adr/0018-statistics.md).

artest_stats-0.2.0/pyproject.toml

@@ -0,0 +1,22 @@
+[project]
+name = "artest-stats"
+version = "0.2.0"
+description = "Experiment statistics: proportions, segment comparison, power analysis and the SRM check."
+readme = "README.md"
+license = "Apache-2.0"
+license-files = ["LICENSE"]
+requires-python = ">=3.12"
+dependencies = [
+    "pydantic>=2.14",
+    "scipy>=1.16",
+]
+
+[project.urls]
+Repository = "https://gitlab.com/Arta_Shoja/artest"
+
+[build-system]
+requires = ["uv_build>=0.13,<0.14"]
+build-backend = "uv_build"
+
+[tool.uv.build-backend]
+module-name = "artest.stats"

artest_stats-0.2.0/pyproject.toml.orig

@@ -0,0 +1,22 @@
+[project]
+name = "artest-stats"
+version = "0.2.0"
+description = "Experiment statistics: proportions, segment comparison, power analysis and the SRM check."
+readme = "README.md"
+license = "Apache-2.0"
+license-files = ["LICENSE"]
+requires-python = ">=3.12"
+dependencies = [
+    "pydantic>=2.14",
+    "scipy>=1.16",
+]
+
+[project.urls]
+Repository = "https://gitlab.com/Arta_Shoja/artest"
+
+[build-system]
+requires = ["uv_build>=0.13,<0.14"]
+build-backend = "uv_build"
+
+[tool.uv.build-backend]
+module-name = "artest.stats"

artest_stats-0.2.0/src/artest/stats/__init__.py

@@ -0,0 +1,46 @@
+"""Experiment statistics: proportions, segment comparison, power analysis and the SRM check.
+
+Pure numerical functions with no I/O and no language models. Every number that drives a decision
+in Artest comes from here.
+"""
+
+from artest.stats.power import (
+    SampleSize,
+    SampleSizeRange,
+    cohens_h,
+    duration_days,
+    sample_size,
+    sample_size_for_lift,
+)
+from artest.stats.proportions import Proportion, step_rates, wilson_interval
+from artest.stats.segments import (
+    Counts,
+    SegmentComparison,
+    StepCounts,
+    TwoProportionTest,
+    benjamini_hochberg,
+    compare_segments,
+    two_proportion_test,
+)
+from artest.stats.srm import SrmCheck, check_srm
+
+__all__ = [
+    "Counts",
+    "Proportion",
+    "SampleSize",
+    "SampleSizeRange",
+    "SegmentComparison",
+    "SrmCheck",
+    "StepCounts",
+    "TwoProportionTest",
+    "benjamini_hochberg",
+    "check_srm",
+    "cohens_h",
+    "compare_segments",
+    "duration_days",
+    "sample_size",
+    "sample_size_for_lift",
+    "step_rates",
+    "two_proportion_test",
+    "wilson_interval",
+]

artest_stats-0.2.0/src/artest/stats/base.py

@@ -0,0 +1,16 @@
+from pydantic import BaseModel, ConfigDict
+
+
+class StatsModel(BaseModel):
+    """Base for every result.
+
+    Immutable, and refuses NaN and infinity, so a division by zero fails where it happens instead of
+    reaching a decision as a number.
+    """
+
+    model_config = ConfigDict(
+        frozen=True,
+        extra="forbid",
+        allow_inf_nan=False,
+        json_schema_serialization_defaults_required=True,
+    )

artest_stats-0.2.0/src/artest/stats/power.py

@@ -0,0 +1,125 @@
+"""How many users a test needs, and how long they take to arrive."""
+
+import math
+from fractions import Fraction
+
+from pydantic import Field
+from scipy.stats import norm
+
+from artest.stats.base import StatsModel
+from artest.stats.proportions import Proportion
+
+# Below this expected count the normal approximation breaks down (Cochran 1954), as for the segment test.
+_MIN_EXPECTED_COUNT = 5
+
+
+class SampleSize(StatsModel):
+    """The users each arm needs to detect a move from `baseline` to `target`, if the move is really that big.
+
+    Both rates are taken as given. `target` is what the test is sized to catch, not what a change will do: a
+    smaller real effect is missed more often than `power` says. A baseline that's itself an estimate moves
+    this number with it, about as one over the rate when the rate is small: `sample_size_for_lift` says how
+    far.
+
+    `small_sample` means an arm expects too few successes or failures for the normal approximation behind
+    `per_arm`. The exact test that runs at such counts needs more users: in the cases worked, a size flagged
+    so reached 52% to 79% power where 80% was asked. Read it as a rough lower bound.
+    """
+
+    baseline: float = Field(gt=0, lt=1)
+    target: float = Field(gt=0, lt=1)
+    effect_size: float = Field(
+        gt=0, description="Cohen's h, the distance between the rates on the arcsine scale"
+    )
+    alpha: float = Field(gt=0, lt=1)
+    power: float = Field(gt=0, lt=1)
+    per_arm: int = Field(ge=1)
+    small_sample: bool
+
+
+def cohens_h(rate_a: float, rate_b: float) -> float:
+    """Cohen's effect size for two proportions: how far apart they are, with the variance made even."""
+    return abs(2 * math.asin(math.sqrt(rate_a)) - 2 * math.asin(math.sqrt(rate_b)))
+
+
+def sample_size(baseline: float, target: float, *, alpha: float = 0.05, power: float = 0.8) -> SampleSize:
+    """Users per arm for a two-sided two-proportion test with equal arms (Cohen 1988, chapter 6).
+
+    A rate of 0% or 100% is refused: no variation means nothing to size a test from, and the approximation
+    says nothing there.
+    """
+    for name, rate in (("baseline", baseline), ("target", target)):
+        if not 0 < rate < 1:
+            raise ValueError(f"the {name} rate must lie strictly between 0 and 1, got {rate}")
+    if baseline == target:
+        raise ValueError(f"the baseline and the target are both {baseline}, so there's no effect to detect")
+    if not 0 < alpha < 1:
+        raise ValueError(f"alpha must lie between 0 and 1, got {alpha}")
+    if not alpha < power < 1:
+        raise ValueError(f"power must lie between alpha ({alpha}) and 1, got {power}")
+
+    effect = cohens_h(baseline, target)
+    z = float(norm.ppf(1 - alpha / 2)) + float(norm.ppf(power))
+    per_arm = math.ceil(2 * (z / effect) ** 2)
+    rarest = min(baseline, 1 - baseline, target, 1 - target)
+    return SampleSize(
+        baseline=baseline,
+        target=target,
+        effect_size=effect,
+        alpha=alpha,
+        power=power,
+        per_arm=per_arm,
+        small_sample=per_arm * rarest < _MIN_EXPECTED_COUNT,
+    )
+
+
+class SampleSizeRange(StatsModel):
+    """The size at a measured rate, and how far it moves across that rate's interval at the same lift.
+
+    The size depends on the rate, and a measured rate is uncertain: 30 of 1,000 is 2.1% to 4.3%, and at a 20%
+    lift that's anywhere from 9,663 to 19,950 users per arm around the 13,887 at 3%. `fewest` and `most` cover
+    the interval's ends. An end whose lift would pass 100% isn't counted, as no test could reach it.
+    """
+
+    likely: SampleSize
+    fewest: int = Field(ge=1)
+    most: int = Field(ge=1)
+
+
+def sample_size_for_lift(
+    rate: Proportion, relative_lift: float, *, alpha: float = 0.05, power: float = 0.8
+) -> SampleSizeRange:
+    """Users per arm to detect `relative_lift` on a measured `rate`, with the range its interval allows.
+
+    A rate of 0% or 100%, or none, sizes no test: it's refused, and so is a lift past 100%.
+    """
+    if rate.rate is None:
+        raise ValueError("there's no measured rate to size a test from")
+    if relative_lift <= 0:
+        raise ValueError(f"the lift must be above zero, got {relative_lift}")
+    likely = sample_size(rate.rate, rate.rate * (1 + relative_lift), alpha=alpha, power=power)
+    sizes = [
+        likely.per_arm,
+        *(
+            sample_size(end, end * (1 + relative_lift), alpha=alpha, power=power).per_arm
+            for end in (rate.lower, rate.upper)
+            if end * (1 + relative_lift) < 1
+        ),
+    ]
+    return SampleSizeRange(likely=likely, fewest=min(sizes), most=max(sizes))
+
+
+def duration_days(per_arm: int, users: int, days: int = 1, *, arms: int = 2) -> int | None:
+    """Whole days until every arm has its users, when the `users` seen over `days` all enter the test.
+
+    None when no users came, as such a test never ends. It assumes traffic stays as it is, and says nothing
+    of weekly cycles, so run a test for whole weeks whatever this says. In whole numbers, as a rate per day
+    rounds: 300 users over 7 days fill 600 places in exactly 14 days, where floats say 15.
+    """
+    if per_arm < 1 or arms < 1:
+        raise ValueError(f"a test needs users and arms, got {per_arm} per arm and {arms} arms")
+    if users < 0 or days < 1:
+        raise ValueError(f"traffic is users over at least a day, got {users} over {days}")
+    if users == 0:
+        return None
+    return math.ceil(Fraction(arms * per_arm * days, users))

artest_stats-0.2.0/src/artest/stats/proportions.py

@@ -0,0 +1,63 @@
+"""Rates and how far to trust them."""
+
+import math
+from collections.abc import Sequence
+from itertools import pairwise
+
+from pydantic import Field
+from scipy.stats import norm
+
+from artest.stats.base import StatsModel
+
+
+class Proportion(StatsModel):
+    """How many of `trials` succeeded, with a Wilson score interval.
+
+    With no trials there's no rate to state: `rate` is None and the interval spans everything, which is
+    also the limit Wilson's formula reaches as trials fall to zero.
+
+    The interval covers sampling noise and nothing else. It doesn't say why the rate is what it is, or
+    what a change would do to it.
+    """
+
+    successes: int = Field(ge=0)
+    trials: int = Field(ge=0)
+    rate: float | None = Field(ge=0, le=1)
+    lower: float = Field(ge=0, le=1)
+    upper: float = Field(ge=0, le=1)
+    confidence: float = Field(gt=0, lt=1)
+
+
+def wilson_interval(successes: int, trials: int, *, confidence: float = 0.95) -> Proportion:
+    """The rate `successes / trials` with its Wilson score interval (Wilson 1927).
+
+    Unlike the textbook Wald interval, it stays inside [0, 1] and keeps a real width at 0% and 100%. Across
+    rates it covers the true rate about as often as it says, even at small samples (Newcombe 1998), though a
+    95% interval dips to about 90% close to 0% and 100% (Brown, Cai and DasGupta 2001).
+    """
+    if not 0 <= successes <= trials:
+        raise ValueError(f"{successes} successes out of {trials} trials isn't a proportion")
+    if not 0 < confidence < 1:
+        raise ValueError(f"confidence must lie between 0 and 1, got {confidence}")
+    if trials == 0:
+        return Proportion(successes=0, trials=0, rate=None, lower=0.0, upper=1.0, confidence=confidence)
+
+    z = float(norm.ppf(0.5 + confidence / 2))
+    rate = successes / trials
+    denominator = trials + z**2
+    centre = (successes + z**2 / 2) / denominator
+    half_width = z / denominator * math.sqrt(successes * (trials - successes) / trials + z**2 / 4)
+    # At 0 and 1 the limit is exact by the formula; the arithmetic alone can land a rounding error beside it.
+    lower = 0.0 if successes == 0 else centre - half_width
+    upper = 1.0 if successes == trials else centre + half_width
+    return Proportion(
+        successes=successes, trials=trials, rate=rate, lower=lower, upper=upper, confidence=confidence
+    )
+
+
+def step_rates(counts: Sequence[int], *, confidence: float = 0.95) -> list[Proportion]:
+    """The rate from each count to the next, for counts of users in the order they pass through.
+
+    Users can only be lost along the way, so a count above the one before it isn't a rate and raises.
+    """
+    return [wilson_interval(after, before, confidence=confidence) for before, after in pairwise(counts)]

artest_stats-0.2.0/src/artest/stats/segments.py

@@ -0,0 +1,153 @@
+"""Does a segment convert differently from everyone else at a step?"""
+
+import math
+from collections.abc import Mapping, Sequence
+from typing import Literal, Self
+
+from pydantic import Field, model_validator
+from scipy.stats import false_discovery_control, fisher_exact
+
+from artest.stats.base import StatsModel
+from artest.stats.proportions import Proportion, wilson_interval
+
+# Below this expected count the normal approximation behind the z-test can't be trusted (Cochran 1954).
+_MIN_EXPECTED_COUNT = 5
+
+
+class TwoProportionTest(StatsModel):
+    """A two-sided test of whether two rates differ.
+
+    `z` is above zero when the first rate is the higher one, and None for the exact test, which has no z.
+    `p_value` reads exactly 0 once it falls below the smallest float, from a z of about 38.
+    """
+
+    method: Literal["z_test", "fisher_exact"]
+    z: float | None
+    p_value: float = Field(ge=0, le=1)
+
+
+def two_proportion_test(
+    successes_a: int, trials_a: int, successes_b: int, trials_b: int
+) -> TwoProportionTest:
+    """A two-proportion z-test with the pooled variance, or Fisher's exact test when counts are small.
+
+    Exact whenever an expected count falls below 5. That includes two rates that are both 0% or both 100%,
+    where the z-test divides by zero and the exact test gives the honest answer: p = 1, no evidence of any
+    difference.
+    """
+    for arm_successes, arm_trials in ((successes_a, trials_a), (successes_b, trials_b)):
+        if not 0 <= arm_successes <= arm_trials or arm_trials == 0:
+            raise ValueError(f"can't test {arm_successes} successes out of {arm_trials} trials")
+
+    trials = trials_a + trials_b
+    successes = successes_a + successes_b
+    # The smallest expected count, in whole numbers, as rounding would land a count of 5 just below it.
+    if min(trials_a, trials_b) * min(successes, trials - successes) < _MIN_EXPECTED_COUNT * trials:
+        table = [[successes_a, trials_a - successes_a], [successes_b, trials_b - successes_b]]
+        return TwoProportionTest(method="fisher_exact", z=None, p_value=float(fisher_exact(table).pvalue))
+
+    pooled = successes / trials
+    standard_error = math.sqrt(pooled * (1 - pooled) * (1 / trials_a + 1 / trials_b))
+    z = (successes_a / trials_a - successes_b / trials_b) / standard_error
+    return TwoProportionTest(method="z_test", z=z, p_value=math.erfc(abs(z) / math.sqrt(2)))
+
+
+def benjamini_hochberg(p_values: Sequence[float]) -> list[float]:
+    """Benjamini-Hochberg adjusted p-values: those at or below q are discoveries at false discovery rate q.
+
+    The guarantee holds for independent or positively dependent tests (Benjamini and Yekutieli 2001).
+    Comparisons in one funnel share users, so read it as close control and not as a proof.
+    """
+    return [float(adjusted) for adjusted in false_discovery_control(p_values, method="bh")]
+
+
+class Counts(StatsModel):
+    """Of `trials` users, `successes` went on."""
+
+    successes: int = Field(ge=0)
+    trials: int = Field(ge=0)
+
+    @model_validator(mode="after")
+    def _possible(self) -> Self:
+        if self.successes > self.trials:
+            raise ValueError(f"{self.successes} successes out of {self.trials} trials isn't a proportion")
+        return self
+
+
+class StepCounts(StatsModel):
+    """One step's counts: everyone's and each segment's. Segments may overlap; each faces the rest."""
+
+    step: str
+    total: Counts
+    segments: Mapping[str, Counts]
+
+
+class SegmentComparison(StatsModel):
+    """One segment against everyone else at one step.
+
+    `test` is None when one side has no users, so there's nothing to compare, and such a comparison doesn't
+    count towards the correction. `significant` means the gap is unlikely to be sampling noise, across all
+    the comparisons made. It's an observation and not an effect: these users may differ from the rest in
+    many ways, so it doesn't say a change would close the gap, or how far.
+    """
+
+    step: str
+    segment: str
+    segment_rate: Proportion
+    rest_rate: Proportion
+    test: TwoProportionTest | None
+    adjusted_p_value: float | None = Field(ge=0, le=1)
+    significant: bool
+
+
+def compare_segments(
+    steps: Sequence[StepCounts], *, fdr: float = 0.05, confidence: float = 0.95
+) -> list[SegmentComparison]:
+    """Every segment against the rest at every step, with Benjamini-Hochberg across all those comparisons.
+
+    `confidence` sets the rates' intervals only; they aren't corrected for the number of comparisons.
+    """
+    if not 0 < fdr < 1:
+        raise ValueError(f"the false discovery rate must lie between 0 and 1, got {fdr}")
+    if not 0 < confidence < 1:
+        raise ValueError(f"confidence must lie between 0 and 1, got {confidence}")
+
+    pairs = [
+        (step.step, name, segment, _rest(step, name, segment))
+        for step in steps
+        for name, segment in step.segments.items()
+    ]
+    tests = [
+        two_proportion_test(segment.successes, segment.trials, rest.successes, rest.trials)
+        if segment.trials and rest.trials
+        else None
+        for _, _, segment, rest in pairs
+    ]
+    adjusted = iter(benjamini_hochberg([test.p_value for test in tests if test is not None]))
+
+    comparisons: list[SegmentComparison] = []
+    for (step, name, segment, rest), test in zip(pairs, tests, strict=True):
+        adjusted_p = None if test is None else next(adjusted)
+        comparisons.append(
+            SegmentComparison(
+                step=step,
+                segment=name,
+                segment_rate=wilson_interval(segment.successes, segment.trials, confidence=confidence),
+                rest_rate=wilson_interval(rest.successes, rest.trials, confidence=confidence),
+                test=test,
+                adjusted_p_value=adjusted_p,
+                significant=adjusted_p is not None and adjusted_p <= fdr,
+            )
+        )
+    return comparisons
+
+
+def _rest(step: StepCounts, name: str, segment: Counts) -> Counts:
+    trials = step.total.trials - segment.trials
+    successes = step.total.successes - segment.successes
+    if trials < 0 or successes < 0 or successes > trials:
+        raise ValueError(
+            f"at {step.step!r}, segment {name!r} ({segment.successes} of {segment.trials}) doesn't fit "
+            f"inside the total ({step.total.successes} of {step.total.trials})"
+        )
+    return Counts(successes=successes, trials=trials)

artest_stats-0.2.0/src/artest/stats/srm.py

@@ -0,0 +1,80 @@
+"""Did the users split the way the experiment planned?"""
+
+from collections.abc import Sequence
+
+from pydantic import Field
+from scipy.stats import binomtest, chi2
+
+from artest.stats.base import StatsModel
+
+# The threshold GrowthBook flags at: a real mismatch shows very clearly, and a false alarm halts a good test.
+_THRESHOLD = 0.001
+# Below this expected count the chi-squared approximation reads far too sure (Cochran 1954).
+_MIN_EXPECTED_COUNT = 5
+
+
+class SrmCheck(StatsModel):
+    """A sample ratio mismatch check: the users each arm got, against what the planned split expected.
+
+    A `mismatch` means assignment or tracking broke, so the experiment's results can't be trusted whatever
+    they say (Fabijan et al. 2019). No mismatch isn't proof that all is well: the check finds only a problem
+    in the split, and a small one only with many users.
+
+    `p_value` is the chi-squared test's, as GrowthBook reports it. When an arm expects too few users for that
+    approximation, `exact_p_value` decides instead: an exact binomial test of each arm, Bonferroni corrected.
+    It's None otherwise. An arm planned for no traffic takes no part in either, though users in it are a
+    mismatch whatever the numbers say.
+    """
+
+    observed: tuple[int, ...]
+    expected: tuple[float, ...]
+    chi_square: float = Field(ge=0)
+    p_value: float = Field(ge=0, le=1)
+    exact_p_value: float | None = Field(ge=0, le=1)
+    mismatch: bool
+
+
+def check_srm(
+    observed: Sequence[int], planned: Sequence[float], *, threshold: float = _THRESHOLD
+) -> SrmCheck:
+    """Chi-squared goodness of fit of `observed` users against the `planned` split, in any scale."""
+    if len(observed) != len(planned) or len(observed) < 2:
+        raise ValueError(
+            f"need the same two or more arms, got {len(observed)} counts and {len(planned)} shares"
+        )
+    if any(users < 0 for users in observed):
+        raise ValueError(f"an arm can't have fewer than no users, got {list(observed)}")
+    if any(share < 0 for share in planned) or not any(share > 0 for share in planned):
+        raise ValueError(f"the shares can't be negative, and one must be above zero, got {list(planned)}")
+    if not 0 < threshold < 1:
+        raise ValueError(f"the threshold must lie between 0 and 1, got {threshold}")
+
+    # Only the arms that were planned take part. Users in any other decide the check by themselves.
+    shares = tuple(share / sum(planned) for share in planned)
+    live = [(o, share) for o, share in zip(observed, shares, strict=True) if share > 0]
+    users = sum(o for o, _ in live)
+    expected = tuple(share * users for share in shares)
+    statistic = sum((o - share * users) ** 2 / (share * users) for o, share in live) if users else 0.0
+    p_value = float(chi2.sf(statistic, len(live) - 1)) if users and len(live) > 1 else 1.0
+    exact = (
+        _exact_p_value(live, users) if min(share for _, share in live) * users < _MIN_EXPECTED_COUNT else None
+    )
+    return SrmCheck(
+        observed=tuple(observed),
+        expected=expected,
+        chi_square=statistic,
+        p_value=p_value,
+        exact_p_value=exact,
+        mismatch=users < sum(observed) or (p_value if exact is None else exact) < threshold,
+    )
+
+
+def _exact_p_value(arms: Sequence[tuple[int, float]], users: int) -> float:
+    """The smallest of the arms' exact binomial p-values, corrected for how many different tests they are.
+
+    Two arms are one test: each is the other's complement. No users, no evidence.
+    """
+    if not users:
+        return 1.0
+    tests = 1 if len(arms) == 2 else len(arms)
+    return min(1.0, tests * min(float(binomtest(o, users, share).pvalue) for o, share in arms))