关于「最小二乘与生存曲线:小村庄的噪声、尾部与一百二十二岁」,《死亡率倍增与蓝区审计》给出的定量结论是什么?
结论:冈珀茨参数可以直接换成生存曲线:众数与中位数稳定,尾部却对斜率极端敏感。小样本中,泊松似然优于丢弃零死亡年龄的最小二乘;300 人的村庄单凭偶然就能让倍增时间在 7 年到 9 年之间摆动。卡尔芒的 122 岁在纯冈珀茨外推下几乎不可能出现,在平台模型下则只是罕见;统计上的突出不能代替档案核验。下一章正式检验误差注入:多少年龄夸大,能把一条冈珀茨曲线压成平台。
数据口径:书中参数为示意;全部计算由书中程序生成;本书不是医疗建议,不宣称任何产品或方法的效应,自我试验须先咨询执业医师。 书中以标注框区分已发表证据、可运行计算(22 段程序)、原典与作者假说。
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: 300
| people | method | mean MRDT | 2.5% | 97.5% |
| 300 | OLS | 7.92 | 6.90 | 9.05 |
| 300 | WLS | 8.12 | 7.20 | 9.18 |
| 300 | Poisson ML | 8.01 | 7.17 | 9.00 |
| 2000 | OLS | 7.98 | 7.58 | 8.43 |
| 2000 | WLS | 8.02 | 7.67 | 8.38 |
| 2000 | Poisson ML | 8.00 | 7.65 | 8.36 |
| 20000 | OLS | 8.01 | 7.83 | 8.14 |
| 20000 | WLS | 8.01 | 7.84 | 8.11 |
| 20000 | Poisson ML | 8.00 | 7.84 | 8.11 |
查看完整验算程序
# Chapter 3. p31: Least squares versus Poisson likelihood in a village.
# Model: pure Gompertz from 60, mu(65) = 0.012, B = ln2/8 (示意). A
# "village" of n people aged 60 is followed to extinction; deaths and
# exact person-years are tabulated by single year of age 60-104.
# Three estimators of B: (1) OLS of log m(x) on age, dropping ages
# with zero deaths; (2) the same, weighted by deaths; (3) Poisson
# maximum likelihood with exposure offsets, by Newton-Raphson.
# 300 villages per size, fixed seed; we report the mean and the 2.5%
# and 97.5% points of the implied doubling time (true value 8.00).
import math
import random
M65, B = 0.012, math.log(2) / 8.0
A60 = M65 * math.exp(-5 * B)
rng = random.Random(20261112)
def village(n):
d, ex = [0] * 45, [0.0] * 45
for _ in range(n):
e = -math.log(1.0 - rng.random())
t = math.log(1.0 + B * e / A60) / B
for i in range(min(int(t), 45)):
ex[i] += 1.0
if t < 45:
ex[int(t)] += t - int(t)
d[int(t)] += 1
return d, ex
def wls(d, ex, weighted):
pts = [(i + 0.5, math.log(d[i] / ex[i]), d[i] if weighted else 1)
for i in range(45) if d[i] > 0]
w = math.fsum(p[2] for p in pts)
mx = math.fsum(p[0] * p[2] for p in pts) / w
my = math.fsum(p[1] * p[2] for p in pts) / w
sxy = math.fsum(p[2] * (p[0] - mx) * (p[1] - my) for p in pts)
sxx = math.fsum(p[2] * (p[0] - mx) ** 2 for p in pts)
return sxy / sxx
def mle(d, ex):
a, b = math.log(A60), 0.05
for _ in range(30):
lam = [ex[i] * math.exp(a + b * (i + 0.5)) for i in range(45)]
g0 = math.fsum(d[i] - lam[i] for i in range(45))
g1 = math.fsum((i + 0.5) * (d[i] - lam[i]) for i in range(45))
h00 = math.fsum(lam)
h01 = math.fsum((i + 0.5) * lam[i] for i in range(45))
h11 = math.fsum((i + 0.5) ** 2 * lam[i] for i in range(45))
det = h00 * h11 - h01 * h01
a += (h11 * g0 - h01 * g1) / det
b += (h00 * g1 - h01 * g0) / det
return b
def summary(vals):
vals = sorted(math.log(2) / v for v in vals)
k = len(vals)
return (math.fsum(vals) / k, vals[int(0.025 * k)],
vals[int(0.975 * k) - 1])
print("people method mean MRDT 2.5% 97.5%")
for n in (300, 2000, 20000):
res = {"OLS": [], "WLS": [], "Poisson ML": []}
reps = 300 if n < 20000 else 60
for _ in range(reps):
d, ex = village(n)
res["OLS"].append(wls(d, ex, False))
res["WLS"].append(wls(d, ex, True))
res["Poisson ML"].append(mle(d, ex))
for name in ("OLS", "WLS", "Poisson ML"):
m, lo, hi = summary(res[name])
print("{:6d} {:12s} {:9.2f} {:8.2f} {:8.2f}".format(
n, name, m, lo, hi))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《冈珀茨死亡率倍增定律、超级百岁老人年龄误报与长寿蓝区数据审计母典》第 3 章「最小二乘与生存曲线:小村庄的噪声、尾部与一百二十二岁」(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-193 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026193,
author = {Fan, Yafu},
title = {Gompertz Mortality Doubling, Age Misreporting among Supercentenarians and Data Audits of Longevity Blue Zones},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}