cb4b6869fc
Encodes the four lead-paper confirmatory tests exactly per the frozen prereg §6, calibrated on synthetic ground truth only (never re-fit to cell data): - RQ1-P1 leak: correlation AUC + BCa CI; gate = CI excludes 0.5 (D2); 0.60 CI lower bound = separate materiality label (material / weak-but-real), not the gate. - RQ1-P2 padding: paired ΔAUC = AUC(no-pad) − AUC(pad) over circuits; effective iff CI > 0. - RQ2-P1 anonymity set: ΔH = H(federated) − H(single, matched N) with Miller-Madow per-circuit entropy; two-sided grow / honest-shrink / inconclusive by CI sign. - RQ2-P3 mechanism: Spearman ρ(top-3 bridge concentration, per-circuit H) + CI. apply_holm corrects the reported RQ1/RQ2 subset against the full frozen family of 7 (family_size default 7) — never re-optimised to the 4 reported. Bootstrap p-values order the Holm step-down only; every decision is a CI gate, never a bare p. stats.py: bootstrap CIs can now return their resample distribution so the Holm ordering p-value comes from the same resamples as the CI. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
348 lines
14 KiB
Python
348 lines
14 KiB
Python
"""Pre-registered inferential statistics for the RQ1/RQ2 confirmatory analysis.
|
||
|
||
This module is the *turnkey* implementation of the frozen prereg §6 analysis plan
|
||
(`sor-consent-prereg.md`), written **before** any confirmatory data exists so the
|
||
analysis precedes the data (`rigor-standards §Statistics`). It computes nothing
|
||
study-specific by itself — it is a small, stdlib-only inference toolkit:
|
||
|
||
* :func:`bootstrap_ci` / :func:`two_sample_diff_ci` — BCa (bias-corrected and
|
||
accelerated) bootstrap 95% CIs, 10,000 resamples by default, with a percentile
|
||
fallback when the acceleration/bias terms are degenerate (§6: "Bootstrap:
|
||
10,000 resamples, BCa intervals; seed spot-check");
|
||
* :func:`miller_madow_entropy_bits` — plug-in Shannon entropy with the
|
||
Miller–Madow finite-sample bias correction (§3 estimator [APPROVAL]);
|
||
* :func:`spearman` — Spearman rank correlation (RQ2-P3 mechanism test);
|
||
* :func:`holm_bonferroni` — Holm step-down multiplicity correction over the
|
||
frozen confirmatory family, with an explicit ``family_size`` so a lead paper
|
||
that reports a subset of the 7 pre-registered tests still corrects against the
|
||
full family (§6; never re-optimised to the reported subset).
|
||
|
||
It performs no I/O, moves no traffic, and spawns no engine. Resampling is seeded
|
||
(``random.Random(seed)``) and the seed is returned in every :class:`CIResult`, so
|
||
a CI is reproducible and the §6 three-seed spot-check is mechanical.
|
||
"""
|
||
|
||
from __future__ import annotations
|
||
|
||
import math
|
||
import random
|
||
from dataclasses import dataclass
|
||
from statistics import NormalDist
|
||
from typing import Callable, List, Sequence, TypeVar
|
||
|
||
T = TypeVar("T")
|
||
|
||
_NORM = NormalDist(0.0, 1.0)
|
||
DEFAULT_RESAMPLES = 10_000
|
||
DEFAULT_ALPHA = 0.05
|
||
|
||
|
||
@dataclass(frozen=True)
|
||
class CIResult:
|
||
"""A point estimate with a bootstrap confidence interval and full provenance
|
||
(method actually used, resample count, seed, alpha) so it is reproducible and
|
||
auditable. ``excludes(v)`` is the pre-registered gate primitive: True iff the
|
||
whole interval lies on one side of ``v`` (e.g. RQ1-P1's "CI excludes 0.5")."""
|
||
|
||
point: float
|
||
lo: float
|
||
hi: float
|
||
alpha: float
|
||
n_resamples: int
|
||
method: str # "bca" | "percentile"
|
||
seed: int
|
||
|
||
def excludes(self, v: float) -> bool:
|
||
return (self.lo > v) or (self.hi < v)
|
||
|
||
def strictly_greater(self, v: float) -> bool:
|
||
return self.lo > v
|
||
|
||
def strictly_less(self, v: float) -> bool:
|
||
return self.hi < v
|
||
|
||
def as_dict(self) -> dict:
|
||
return {
|
||
"point": self.point,
|
||
"ci_lo": self.lo,
|
||
"ci_hi": self.hi,
|
||
"alpha": self.alpha,
|
||
"n_resamples": self.n_resamples,
|
||
"method": self.method,
|
||
"seed": self.seed,
|
||
}
|
||
|
||
|
||
def _percentile(sorted_vals: Sequence[float], q: float) -> float:
|
||
"""Linear-interpolation percentile of an already-sorted sequence, ``q`` in
|
||
[0, 1]. Empty -> NaN; clamps out-of-range q to the endpoints."""
|
||
n = len(sorted_vals)
|
||
if n == 0:
|
||
return float("nan")
|
||
if q <= 0:
|
||
return float(sorted_vals[0])
|
||
if q >= 1:
|
||
return float(sorted_vals[-1])
|
||
pos = q * (n - 1)
|
||
lo = int(math.floor(pos))
|
||
hi = int(math.ceil(pos))
|
||
if lo == hi:
|
||
return float(sorted_vals[lo])
|
||
frac = pos - lo
|
||
return float(sorted_vals[lo]) * (1 - frac) + float(sorted_vals[hi]) * frac
|
||
|
||
|
||
def _bca_endpoints(
|
||
thetas: Sequence[float], theta_hat: float, jack: Sequence[float], alpha: float
|
||
) -> tuple[float, float, str]:
|
||
"""Return (q_lo, q_hi, method) adjusted percentiles for a BCa interval, or the
|
||
plain (alpha/2, 1-alpha/2, "percentile") pair when the bias/acceleration terms
|
||
are degenerate (all resamples equal, or zero jackknife spread)."""
|
||
b = len(thetas)
|
||
n_less = sum(1 for t in thetas if t < theta_hat)
|
||
# Bias correction z0. If every resample is on one side, z0 is undefined ->
|
||
# fall back to the percentile interval rather than emit a garbage bound.
|
||
if n_less == 0 or n_less == b:
|
||
return alpha / 2.0, 1.0 - alpha / 2.0, "percentile"
|
||
z0 = _NORM.inv_cdf(n_less / b)
|
||
|
||
# Acceleration from the jackknife leave-one-out estimates.
|
||
jbar = sum(jack) / len(jack)
|
||
diffs = [jbar - j for j in jack]
|
||
denom = 6.0 * (sum(d * d for d in diffs) ** 1.5)
|
||
if denom == 0.0:
|
||
return alpha / 2.0, 1.0 - alpha / 2.0, "percentile"
|
||
a = sum(d ** 3 for d in diffs) / denom
|
||
|
||
z_lo = _NORM.inv_cdf(alpha / 2.0)
|
||
z_hi = _NORM.inv_cdf(1.0 - alpha / 2.0)
|
||
|
||
def adjust(z: float) -> float:
|
||
num = z0 + z
|
||
return _NORM.cdf(z0 + num / (1.0 - a * num))
|
||
|
||
q_lo = adjust(z_lo)
|
||
q_hi = adjust(z_hi)
|
||
if not (0.0 < q_lo < q_hi < 1.0):
|
||
return alpha / 2.0, 1.0 - alpha / 2.0, "percentile"
|
||
return q_lo, q_hi, "bca"
|
||
|
||
|
||
def two_sided_bootstrap_p(thetas: Sequence[float], null: float) -> float:
|
||
"""A bootstrap two-sided p-value for H0: θ = ``null`` from a bootstrap
|
||
distribution ``thetas`` — 2·min(P(θ* ≤ null), P(θ* ≥ null)), capped at 1.0.
|
||
Used only to *order* the Holm family; the pre-registered decision gates are
|
||
the CIs, never a p-value alone (§6)."""
|
||
b = len(thetas)
|
||
if b == 0:
|
||
return 1.0
|
||
le = sum(1 for t in thetas if t <= null) / b
|
||
ge = sum(1 for t in thetas if t >= null) / b
|
||
return min(1.0, 2.0 * min(le, ge))
|
||
|
||
|
||
def bootstrap_ci(
|
||
units: Sequence[T],
|
||
statistic: Callable[[Sequence[T]], float],
|
||
*,
|
||
n_resamples: int = DEFAULT_RESAMPLES,
|
||
alpha: float = DEFAULT_ALPHA,
|
||
seed: int = 0,
|
||
method: str = "bca",
|
||
return_dist: bool = False,
|
||
):
|
||
"""One-sample bootstrap CI of ``statistic`` over ``units`` (the unit of
|
||
analysis — a circuit-pair for RQ1, a circuit for RQ2). Resamples ``units`` with
|
||
replacement ``n_resamples`` times. ``method="bca"`` applies bias-correction +
|
||
acceleration (falling back to percentile if degenerate); ``"percentile"``
|
||
forces the plain interval. With ``return_dist=True`` also returns the sorted
|
||
bootstrap distribution (so a p-value can be derived from the same resamples)."""
|
||
n = len(units)
|
||
if n == 0:
|
||
res = CIResult(float("nan"), float("nan"), float("nan"),
|
||
alpha, n_resamples, "empty", seed)
|
||
return (res, []) if return_dist else res
|
||
theta_hat = float(statistic(units))
|
||
rng = random.Random(seed)
|
||
thetas: List[float] = []
|
||
for _ in range(n_resamples):
|
||
sample = [units[rng.randrange(n)] for _ in range(n)]
|
||
thetas.append(float(statistic(sample)))
|
||
thetas.sort()
|
||
|
||
if method == "bca" and n > 1:
|
||
jack = [float(statistic([units[j] for j in range(n) if j != i])) for i in range(n)]
|
||
q_lo, q_hi, used = _bca_endpoints(thetas, theta_hat, jack, alpha)
|
||
else:
|
||
q_lo, q_hi, used = alpha / 2.0, 1.0 - alpha / 2.0, "percentile"
|
||
|
||
res = CIResult(theta_hat, _percentile(thetas, q_lo), _percentile(thetas, q_hi),
|
||
alpha, n_resamples, used, seed)
|
||
return (res, thetas) if return_dist else res
|
||
|
||
|
||
def two_sample_diff_ci(
|
||
units_a: Sequence[T],
|
||
units_b: Sequence[T],
|
||
statistic: Callable[[Sequence[T]], float],
|
||
*,
|
||
n_resamples: int = DEFAULT_RESAMPLES,
|
||
alpha: float = DEFAULT_ALPHA,
|
||
seed: int = 0,
|
||
method: str = "bca",
|
||
return_dist: bool = False,
|
||
):
|
||
"""Bootstrap CI for the difference ``statistic(A) - statistic(B)`` of two
|
||
independent arms (RQ2-P1: ΔH = H(federated) − H(single-house, matched N)).
|
||
Each arm is resampled independently. BCa uses a combined leave-one-out
|
||
jackknife across both arms (each point dropped from its own arm). With
|
||
``return_dist=True`` also returns the sorted bootstrap distribution."""
|
||
na, nb = len(units_a), len(units_b)
|
||
if na == 0 or nb == 0:
|
||
res = CIResult(float("nan"), float("nan"), float("nan"),
|
||
alpha, n_resamples, "empty", seed)
|
||
return (res, []) if return_dist else res
|
||
theta_hat = float(statistic(units_a)) - float(statistic(units_b))
|
||
rng = random.Random(seed)
|
||
thetas: List[float] = []
|
||
for _ in range(n_resamples):
|
||
sa = [units_a[rng.randrange(na)] for _ in range(na)]
|
||
sb = [units_b[rng.randrange(nb)] for _ in range(nb)]
|
||
thetas.append(float(statistic(sa)) - float(statistic(sb)))
|
||
thetas.sort()
|
||
|
||
if method == "bca" and na > 1 and nb > 1:
|
||
stat_b_full = float(statistic(units_b))
|
||
stat_a_full = float(statistic(units_a))
|
||
jack: List[float] = []
|
||
for i in range(na):
|
||
jack.append(float(statistic([units_a[j] for j in range(na) if j != i])) - stat_b_full)
|
||
for i in range(nb):
|
||
jack.append(stat_a_full - float(statistic([units_b[j] for j in range(nb) if j != i])))
|
||
q_lo, q_hi, used = _bca_endpoints(thetas, theta_hat, jack, alpha)
|
||
else:
|
||
q_lo, q_hi, used = alpha / 2.0, 1.0 - alpha / 2.0, "percentile"
|
||
|
||
res = CIResult(theta_hat, _percentile(thetas, q_lo), _percentile(thetas, q_hi),
|
||
alpha, n_resamples, used, seed)
|
||
return (res, thetas) if return_dist else res
|
||
|
||
|
||
def mean(xs: Sequence[float]) -> float:
|
||
"""Arithmetic mean; 0.0 for an empty sequence (a bootstrap resample is never
|
||
empty, but degenerate jackknife folds can be)."""
|
||
xs = list(xs)
|
||
return sum(xs) / len(xs) if xs else 0.0
|
||
|
||
|
||
# --------------------------------------------------------------------------- #
|
||
# RQ2 — Miller–Madow bias-corrected entropy (§3 estimator [APPROVAL]).
|
||
# --------------------------------------------------------------------------- #
|
||
def miller_madow_entropy_bits(counts: Sequence[float]) -> float:
|
||
"""Shannon entropy in bits with the Miller–Madow bias correction.
|
||
|
||
H_MM = H_plugin + (K − 1) / (2 N ln 2), where K is the number of observed
|
||
(non-zero) categories and N the total count. This corrects the systematic
|
||
downward bias of the plug-in (MLE) estimator at finite N. Empty/all-zero -> 0.
|
||
"""
|
||
vals = [c for c in counts if c > 0]
|
||
total = float(sum(vals))
|
||
if total <= 0:
|
||
return 0.0
|
||
h = 0.0
|
||
for c in vals:
|
||
p = c / total
|
||
h -= p * math.log2(p)
|
||
k = len(vals)
|
||
return h + (k - 1) / (2.0 * total * math.log(2.0))
|
||
|
||
|
||
# --------------------------------------------------------------------------- #
|
||
# RQ2-P3 — Spearman rank correlation.
|
||
# --------------------------------------------------------------------------- #
|
||
def _rankdata(values: Sequence[float]) -> List[float]:
|
||
"""Fractional ranks (ties get the average of the ranks they span)."""
|
||
order = sorted(range(len(values)), key=lambda i: values[i])
|
||
ranks = [0.0] * len(values)
|
||
i = 0
|
||
n = len(values)
|
||
while i < n:
|
||
j = i
|
||
while j + 1 < n and values[order[j + 1]] == values[order[i]]:
|
||
j += 1
|
||
avg = (i + j) / 2.0 + 1.0 # 1-based average rank over the tie block
|
||
for k in range(i, j + 1):
|
||
ranks[order[k]] = avg
|
||
i = j + 1
|
||
return ranks
|
||
|
||
|
||
def _pearson(xs: Sequence[float], ys: Sequence[float]) -> float:
|
||
n = len(xs)
|
||
if n == 0 or n != len(ys):
|
||
return 0.0
|
||
mx = sum(xs) / n
|
||
my = sum(ys) / n
|
||
num = sum((x - mx) * (y - my) for x, y in zip(xs, ys))
|
||
dx = math.sqrt(sum((x - mx) ** 2 for x in xs))
|
||
dy = math.sqrt(sum((y - my) ** 2 for y in ys))
|
||
if dx == 0.0 or dy == 0.0:
|
||
return 0.0
|
||
return num / (dx * dy)
|
||
|
||
|
||
def spearman(xs: Sequence[float], ys: Sequence[float]) -> float:
|
||
"""Spearman ρ = Pearson correlation of the fractional ranks. Returns 0.0 for
|
||
empty, length-mismatched, or zero-variance inputs."""
|
||
if len(xs) != len(ys) or not xs:
|
||
return 0.0
|
||
return _pearson(_rankdata(xs), _rankdata(ys))
|
||
|
||
|
||
# --------------------------------------------------------------------------- #
|
||
# Multiplicity — Holm–Bonferroni over the frozen confirmatory family.
|
||
# --------------------------------------------------------------------------- #
|
||
@dataclass(frozen=True)
|
||
class HolmResult:
|
||
name: str
|
||
p: float
|
||
p_adjusted: float
|
||
reject: bool
|
||
rank: int # 1-based ascending rank among the reported tests
|
||
multiplier: int # Holm denominator actually used (family_size − rank + 1)
|
||
|
||
|
||
def holm_bonferroni(
|
||
pvalues: dict, alpha: float = DEFAULT_ALPHA, family_size: int | None = None
|
||
) -> List[HolmResult]:
|
||
"""Holm step-down correction.
|
||
|
||
``pvalues`` maps test-name -> raw p. ``family_size`` is the size of the frozen
|
||
confirmatory family (default = number of tests supplied). When a lead paper
|
||
reports a **subset** of the pre-registered family (e.g. the 4 RQ1/RQ2 tests of
|
||
a 7-test frozen family), pass ``family_size=7``: the k-th smallest reported p
|
||
is then tested against ``alpha / (family_size − k + 1)`` — i.e. the reported
|
||
tests are treated as occupying the *smallest* slots of the full family, giving
|
||
the largest (most conservative) Holm multipliers. This is strictly no less
|
||
stringent than the true embedded correction and can never re-optimise the
|
||
family down to the reported subset (which would inflate the false-rejection
|
||
rate and constitute p-hacking).
|
||
|
||
Adjusted p-values are made monotone non-decreasing in rank (standard Holm) and
|
||
capped at 1.0. ``reject`` is ``p_adjusted <= alpha``.
|
||
"""
|
||
items = sorted(pvalues.items(), key=lambda kv: kv[1])
|
||
m = family_size if family_size is not None else len(items)
|
||
if m < len(items):
|
||
raise ValueError(f"family_size {m} < number of reported tests {len(items)}")
|
||
|
||
out: List[HolmResult] = []
|
||
running = 0.0
|
||
for k, (name, p) in enumerate(items, start=1):
|
||
mult = m - k + 1
|
||
adj = min(1.0, mult * p)
|
||
running = max(running, adj) # enforce step-down monotonicity
|
||
out.append(HolmResult(name=name, p=p, p_adjusted=running,
|
||
reject=running <= alpha, rank=k, multiplier=mult))
|
||
return out
|