关于「带宽选择:偏差与方差的交换、蒙特卡洛最优与边界交叉验证」,《会员门槛的断点回归》给出的定量结论是什么?

结论:带宽为 2 时只用了 1801 人,估计 2.723,标准误 0.479,区间从 1.784 到 3.662,很宽。带宽从 4 到 16,估计在 2.676 到 2.954 之间摆动,区间全部包住真值,标准误从 0.345 降到 0.180。带宽放到 24,估计掉到 2.217,区间上限 2.519 已经够不到真值;放到 40,几乎用上了全部 19581 人,估计只剩 0.873,区间从 0.604 到 1.142,离真值很远。

数据口径:落笔前本书核对原文并作蓝图修正:1960 年的论文研究的是奖学金荣誉证书而非会员体系,2008 年的指南是方法论文,二者都没有给出任何会员权益的效应量,书中的高估倍数全部来自固定种子的示意数据;命题也必须带边界陈述:只识别门槛处的局部效应,客人能精确凑单时设计失效,估计对带宽敏感,不宜用全局高阶多项式。 书中参数为示意;全部计算由书中程序生成。

English version

对照数据表

下表数字由本章 Python 程序原样打印,可重新运行复现。 程序只用 Python 标准库与结论吻合的打印值: 0.180, 0.345, 0.479, 0.604, 0.873, 1.142

hestimatesen95% interval
22.7230.4791801[ 1.784,3.662]
42.9540.3453533[ 2.279,3.630]
62.8300.2815260[ 2.279,3.382]
82.7640.2456888[ 2.284,3.244]
122.8050.2039885[ 2.407,3.203]
162.6760.18012549[ 2.324,3.028]
242.2170.15416505[ 1.914,2.519]
400.8730.13719581[ 0.604,1.142]
true local effect 2.5769
查看完整验算程序
# Chapter 3. p31: one data set, many bandwidths.
# Data as in p11 (seed 20261113, 20000 customers, line c = 60, true
# local effect 2.5769). Local linear RD with a triangular kernel for
# bandwidths from 2 to 40 units; for each we print the estimate, the
# sandwich standard error, the customers used and the 95% interval.
# Narrow windows are noisy; wide windows lean on the curvature of
# E[y | x] (the cubic term) far from the line and drift.
import math
import random

C = 60.0
LOCAL = 2.0 + 12.0 / 208.0 * (C - 50.0)


def draw(rng, n, sel=1.0, c=C):
    rows = []
    for _ in range(n):
        u = rng.gauss(0.0, 1.0)
        x = 50.0 + 12.0 * u + 8.0 * rng.gauss(0.0, 1.0)
        d = 1 if x >= c else 0
        g = 0.15 * (x - c) + 0.0003 * (x - c) ** 3
        y = 20.0 + 6.0 * sel * u + g + (2.0 + u) * d
        rows.append((x, y + rng.gauss(0.0, 3.0), u))
    return rows


def wls_line(pts, c, h):
    xs, ys, ws = [], [], []
    for x, y in pts:
        t = abs(x - c) / h
        if t < 1.0:
            xs.append(x - c)
            ys.append(y)
            ws.append(1.0 - t)
    s0 = math.fsum(ws)
    s1 = math.fsum(w * x for w, x in zip(ws, xs))
    s2 = math.fsum(w * x * x for w, x in zip(ws, xs))
    t0 = math.fsum(w * y for w, y in zip(ws, ys))
    t1 = math.fsum(w * x * y for w, x, y in zip(ws, xs, ys))
    det = s0 * s2 - s1 * s1
    a = (s2 * t0 - s1 * t1) / det
    b = (s0 * t1 - s1 * t0) / det
    va = math.fsum((w * (s2 - s1 * x) / det * (y - a - b * x)) ** 2
                   for w, x, y in zip(ws, xs, ys))
    return a, va, len(xs)


if __name__ == "__main__":
    rows = draw(random.Random(20261113), 20000)
    left = [(r[0], r[1]) for r in rows if r[0] < C]
    right = [(r[0], r[1]) for r in rows if r[0] >= C]
    print("   h   estimate     se      n     95% interval")
    for h in (2.0, 4.0, 6.0, 8.0, 12.0, 16.0, 24.0, 40.0):
        al, vl, nl = wls_line(left, C, h)
        ar, vr, nr = wls_line(right, C, h)
        tau, se = ar - al, math.sqrt(vl + vr)
        print("{:4.0f}  {:8.3f}  {:6.3f}  {:6d}  [{:6.3f}, {:6.3f}]".format(
            h, tau, se, nl + nr, tau - 1.96 * se, tau + 1.96 * se))
    print("true local effect {:.4f}".format(LOCAL))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《断点回归设计与会员等级门槛、消费满额升级线两侧的局部因果识别母典》第 3 章「带宽选择:偏差与方差的交换、蒙特卡洛最优与边界交叉验证」(版本 1.0.0). 5D Library. DOI 待发布

DOI 待发布 · 专著页正式发布后上线(本站暂无专著页) · 5DL-B-303 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026303,
  author    = {Fan, Yafu},
  title     = {Regression-Discontinuity Designs at Membership-Tier Thresholds: Local Causal Identification on Either Side of a Spend-Based Upgrade Line},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论