lisuo's picture
Add PFN-DiD Arena inference release
b7ab973 verified
Raw History Blame Contribute Delete
2.02 kB
"""Shared normalized panel DR score; nuisance-rate assumptions are required for IF inference."""
from statistics import NormalDist
import numpy as np
def validate_table(x, d, y):
x, d, y = (np.asarray(v, dtype=float) for v in (x, d, y))
if (
d.ndim != 1
or y.shape != d.shape
or x.ndim != 2
or len(x) != len(d)
or len(d) < 4
or not all(np.isfinite(v).all() for v in (x, d, y))
or not np.isin(d, [0, 1]).all()
or len(np.unique(d)) != 2
):
raise ValueError(
"Expected finite numeric X, binary D with both arms, and finite change"
)
return x, d.astype(int), y
def drdid_panel(y, d, propensity, outcome, clip=0.05, confidence=0.95):
_, d, y = validate_table(np.empty((len(d), 0)), d, y)
propensity, outcome = map(
lambda v: np.asarray(v, dtype=float), (propensity, outcome)
)
if (
propensity.shape != y.shape
or outcome.shape != y.shape
or not np.isfinite(propensity).all()
or not np.isfinite(outcome).all()
or ((propensity < 0) | (propensity > 1)).any()
or not 0 < clip < 0.5
or not 0 < confidence < 1
):
raise ValueError("Invalid nuisance predictions, clipping, or confidence")
e = np.clip(propensity, clip, 1 - clip)
weights = (1 - d) * e / (1 - e)
residual = y - outcome
treated_mean = np.average(residual, weights=d)
control_mean = np.average(residual, weights=weights)
att = float(treated_mean - control_mean)
influence = d / d.mean() * (residual - treated_mean) - weights / weights.mean() * (
residual - control_mean
)
se = float(influence.std(ddof=1) / np.sqrt(len(d)))
margin = NormalDist().inv_cdf((1 + confidence) / 2) * se
return dict(
att=att,
se=se,
lower=att - margin,
upper=att + margin,
clipped_fraction=float(np.mean(e != propensity)),
control_ess=float(weights.sum() ** 2 / (weights @ weights)),
)