5D Library › 指—脑皮层放大、远端微针传入与卒中后认知网络重建的计算神经学 · English
第六章 漂移扩散、天花板效应与论文数据核查
试验的设计
证据
研究内容二的设计(据论文第 2 章)。 在天津市滨海新区中医医院针灸脑病科公开招募;共筛选 105 例,排除 37 例(其中不配合 12 例),68 例按 1:1 随机分组,随机序列由独立的生物统计人员用 SPSS 25.0 生成(第 2.2.1.1 节),两组各 34 例完成全部疗程(第 2.3.1 节)。治疗组:「醒脑开窍」针刺 + 基础治疗 + 双手十二井穴揿针;对照组:「醒脑开窍」针刺 + 基础治疗。疗程 4 周,另有治疗结束后 4 周的随访。主要疗效指标为 MoCA;次要指标为 MMSE 与「认知功能评估系统」0.6.1 版中的色词测试(斯特鲁普范式)、图形—数字测试、图形—计算测试、注意力测试(舒尔特方格)与数字广度测试(第 2.2.5 节)。纳入 MoCA 18–25 分、CDR 0–0.5 分、病程 2 周至 6 个月的 30–75 岁患者(第 2.1.3 节)(Lu 2024)。
由此可以确定试验的几个基本性质。它是开放标签的:对照组没有假针,患者与施针者都知道分组;论文正文中也没有报告结局评估者是否设盲。它是加载设计:两组都接受同样的针刺与基础治疗,比较的是「加上揿针与每天按揉」这一整体。它是单中心、每组 34 例的。此外,以本书所用的答辩稿文本为准,正文中未见伦理批件编号、试验注册号与不良事件结果的报告;招募时间在第 2.1.1 节写作「2023 年 5 月至 2024 年 4 月」,在第 2.3.1 节写作「2023 年 6 月份至 2024 年 5 月份」。
主要结果:组内前后比较不能代替组间比较
证据
主要指标的组间结果(论文第 2.3.2.3 节、第 2.3.3 节、表 2-5、表 2-8、摘要)。 论文对 MoCA 给出两种组间结果。其一,按「有效率」作四格表卡方检验,p = 0.287,论文据此写道「可以认为揿针刺激手十二井穴在改善患者卒中后轻度认知功能障碍方面较针刺疗效相当」。其二,治疗后 MoCA 评分的组间比较 P = 0.005(第 2.3.3 节;表 2-8 治疗组标一个星号),治疗结束 4 周随访时 P = 0.002。次要指标中,治疗后 MMSE 组间 P < 0.001,延迟回忆 P = 0.012,色词、图形—数字、图形—计算与注意力测试也有组间差异(第 2.3.3–2.3.4 节)。摘要另据两组各自的组内前后检验写道「疗效优于对照组」。论文结论第 2 条并列两句:「疗效与对照组相当;治疗组在改善认知评估系统的疗效优于对照组」(Lu 2024)。
「两组组内都显著」推不出「两组之间有差异」。Bland 与 Altman 2011 年指出,在随机分组内各自与基线比较、再据此推断组间差别,是常见而严重误导的做法,正确的做法是直接比较两组,通常用以基线为协变量的协方差分析 (Bland and Altman 2011);Gelman 与 Stern 2006 年讨论的是相近而不同的情形,即一组显著、一组不显著时,二者之差本身未必显著 (Gelman and Stern 2006)。论文的治疗后评分比较(P = 0.005)确实是组间比较,比组内检验更有分量,但仍不足以下结论:它没有调整基线;对照组那一格 21.76(24.00, 26.25)本身不自洽(见下文第 2 条);同一主要指标按有效率比较 p = 0.287;而试验是开放标签、单中心、没有假针对照,两组都处在卒中后的自然恢复期。
次要指标的组间差异要加三点限定:开放标签下治疗组患者预期自己在接受额外治疗,这会影响依赖努力的计时测验;计时测验有明显的练习效应,测试次数相同并不能消除动机的差别;五项计算机测验、两个量表与多个分项,论文没有说明多重比较的处理。
计时测验与漂移扩散模型
作者框架把舒尔特方格的提速读作「认知网络重建、漂移率回升」。论文表 2-26 的数字是:治疗组 80.06 秒降到 57.91 秒,对照组 80.84 秒降到 74.30 秒(论文标为中位数及四分位数)。
漂移扩散模型把一次二选一的判断描述为证据在两条边界之间随机游走的累积过程:证据积累的平均速度是漂移率 ,两条边界的距离 反映谨慎程度,非决策时间 包括知觉编码与动作执行 (Ratcliff 1978; Ratcliff and McKoon 2008)。Wagenmakers 等人 2007 年的 EZ 扩散模型给出正确率与平均反应时的闭式解 (Wagenmakers et al. 2007):
斯特鲁普色词干扰 (Stroop 1935) 与舒尔特方格都不是简单的二选一任务,下面的程序只用一个示意性的二选一任务说明一个一般的道理:同样幅度的提速可以来自三个完全不同的参数。
推导
# Chapter 6. Faster is not the same as "better cognition": the drift-diffusion model.
# Closed forms of the EZ-diffusion model (Wagenmakers et al. 2007), scale s = 0.1,
# unbiased start: accuracy Pc = 1/(1 + exp(-a v / s^2)) and mean decision time
# (a / 2v) (1 - exp(-a v / s^2)) / (1 + exp(-a v / s^2)); mean RT = Ter + that.
# A two-choice task with baseline v = 0.20, a = 0.12, Ter = 0.45 s is made 27.7 %
# faster, the ratio of the treatment-arm Schulte times in Lu (2024) Table 2-26
# (57.91 / 80.06), by changing ONE parameter at a time. An Euler simulation checks the
# formulas. The task and parameters are illustrative, not fitted to the thesis.
import math, random
S2 = 0.01
def ez(v, a, ter):
e = math.exp(-a * v / S2)
return 1 / (1 + e), ter + a / (2 * v) * (1 - e) / (1 + e)
def solve(f, lo, hi, target):
for _ in range(80):
mid = (lo + hi) / 2
if (f(mid) - target) * (f(lo) - target) <= 0:
hi = mid
else:
lo = mid
return (lo + hi) / 2
def simulate(v, a, ter, n, rng, dt=0.001):
rts, ok = [], 0
for _ in range(n):
x, t = a / 2, 0.0
while 0 < x < a:
x += v * dt + 0.1 * math.sqrt(dt) * rng.gauss(0, 1); t += dt
ok += x >= a; rts.append(t + ter)
return ok / n, math.fsum(rts) / n
if __name__ == "__main__":
v0, a0, t0 = 0.20, 0.12, 0.45
pc0, rt0 = ez(v0, a0, t0)
target = rt0 * 57.91 / 80.06
print(f"baseline: v {v0}, a {a0}, Ter {t0} s -> accuracy {pc0:.3f},"
f" mean RT {rt0:.3f} s")
print(f"target mean RT {target:.3f} s (ratio {57.91 / 80.06:.3f})")
v1 = solve(lambda v: ez(v, a0, t0)[1], v0, 5.0, target)
a1 = solve(lambda a: ez(v0, a, t0)[1], 0.001, a0, target)
t1 = target - (rt0 - t0)
for lab, (v, a, t) in (("drift v up", (v1, a0, t0)),
("boundary a down", (v0, a1, t0)),
("non-decision Ter down", (v0, a0, t1))):
pc, rt = ez(v, a, t)
print(f" {lab:22s} v {v:.3f} a {a:.3f} Ter {t:.3f} -> accuracy {pc:.3f},"
f" RT {rt:.3f}")
print(f"control-arm ratio {74.30 / 80.84:.3f} (Table 2-26):"
" the control arm got faster too")
pcs, rts = simulate(v0, a0, t0, 3000, random.Random(1978))
print(f"Euler check (3,000 trials, dt 1 ms): accuracy {pcs:.3f}, mean RT {rts:.3f} s")
# baseline: v 0.2, a 0.12, Ter 0.45 s -> accuracy 0.917, mean RT 0.700 s
# target mean RT 0.506 s (ratio 0.723)
# drift v up v 1.064 a 0.120 Ter 0.450 -> accuracy 1.000, RT 0.506
# boundary a down v 0.200 a 0.049 Ter 0.450 -> accuracy 0.729, RT 0.506
# non-decision Ter down v 0.200 a 0.120 Ter 0.256 -> accuracy 0.917, RT 0.506
# control-arm ratio 0.919 (Table 2-26): the control arm got faster too
# Euler check (3,000 trials, dt 1 ms): accuracy 0.923, mean RT 0.710 s基线平均反应时 0.700 秒、正确率 0.917。要使反应时缩短到原来的 0.723 倍(即治疗组舒尔特用时的比例),可以把漂移率从 0.20 提高到 1.064(正确率升到 1.000),也可以把边界从 0.12 降到 0.049(正确率降到 0.729),还可以只把非决策时间从 0.450 秒缩短到 0.256 秒(正确率不变)。三者给出完全相同的平均用时。只有同时记录正确率与反应时的分布,才能区分「信息处理更快」「更不谨慎」与「动作更熟练」;舒尔特方格只记总用时,没有这些信息。对照组的用时也缩短了 8%(74.30/80.84 = 0.919)。欧拉法模拟(3,000 次)的平均反应时比闭式解长约 1.4%,来自离散时间步长造成的越界偏差。
MoCA 的天花板与正确的分析方法
MoCA 满分 30 分,原作者建议以 26 分为正常界值,它对轻度认知障碍比 MMSE 更敏感 (Nasreddine et al. 2005)。论文纳入 MoCA 18–25 分的患者;治疗组治疗后的 MoCA 标为中位数 27.67(25.00, 28.00)(表 2-13),已经接近满分。Wu 等人 2019 年在 65 名接受康复治疗的卒中患者中研究了 MoCA 的反应度与最小临床重要差异,4–5 周治疗后的标准化反应均数为 0.67,按锚定法与分布法分别有 50.77% 与 30.77% 的患者超过最小临床重要差异 (Wu et al. 2019)(本轮核查未取得该文给出的具体分值,故第八章不直接引用其数值)。
有上限的量表会带来两个问题。其一,接近上限的患者没有改善的空间,组间差别被压缩。其二,把连续分数切成「有效/无效」会丢掉大量信息。正确的做法是以基线分数为协变量的协方差分析 (Vickers and Altman 2001);当相当一部分人达到上限时,可以用托宾模型把 30 分视为删失值 (Tobin 1958)。下面的程序是作者的模型:模拟一个与论文纳入范围相同的试验,比较五种分析方法的把握度。
推导
# Chapter 6. MoCA ceiling and the choice of analysis (the author's model). Baseline latent
# ~ N(21.5, 2.5^2), observed = round(latent + N(0,1)) kept if 18-25; after 4 weeks the
# latent rises 4.0 +- 2.0 (+ delta if treated); observed = min(30, round(...)).
# Tests (two-sided 0.05): responder rule (post - pre)/pre >= 25 % by chi-square; Mann-
# Whitney on post; t on change; ANCOVA; ANCOVA on the uncensored latent score.
# B: large sample, ANCOVA estimate vs a Tobit model (EM algorithm, censoring at 29.5).
import math, random, statistics
ND = statistics.NormalDist()
def patient(rng, delta):
while True:
lat = rng.gauss(21.5, 2.5); pre = round(lat + rng.gauss(0, 1))
if 18 <= pre <= 25:
break
post_lat = lat + rng.gauss(4.0, 2.0) + delta + rng.gauss(0, 1)
return pre, min(30, round(post_lat)), post_lat
def solve3(A, b):
m = [A[i][:] + [b[i]] for i in range(3)]
for c in range(3):
p = max(range(c, 3), key=lambda r: abs(m[r][c])); m[c], m[p] = m[p], m[c]
for r in range(3):
if r != c:
f = m[r][c] / m[c][c]
m[r] = [x - f * y for x, y in zip(m[r], m[c])]
return [m[i][3] / m[i][i] for i in range(3)]
def ancova(y, arm, base):
X = [(1.0, a, b) for a, b in zip(arm, base)]
XtX = [[math.fsum(x[i] * x[j] for x in X) for j in range(3)] for i in range(3)]
Xty = [math.fsum(x[i] * v for x, v in zip(X, y)) for i in range(3)]
beta = solve3(XtX, Xty)
res = [v - math.fsum(bb * xx for bb, xx in zip(beta, x)) for x, v in zip(X, y)]
s2 = math.fsum(r * r for r in res) / (len(y) - 3)
inv1 = solve3(XtX, [0.0, 1.0, 0.0])[1]
return beta[1], beta[1] / math.sqrt(s2 * inv1)
def mwu_z(a, b):
allv = sorted([(v, 0) for v in a] + [(v, 1) for v in b]); ranks = {}; i = 0; tie = 0
while i < len(allv):
j = i
while j + 1 < len(allv) and allv[j + 1][0] == allv[i][0]:
j += 1
ranks[allv[i][0]] = (i + j) / 2 + 1; t = j - i + 1; tie += t ** 3 - t; i = j + 1
n1, n2 = len(a), len(b); n = n1 + n2
U = math.fsum(ranks[v] for v in a) - n1 * (n1 + 1) / 2
var = n1 * n2 / 12 * ((n + 1) - tie / (n * (n - 1)))
return (U - n1 * n2 / 2) / math.sqrt(var)
def trial(rng, n, delta):
rows = [(1, *patient(rng, delta)) for _ in range(n)] + \
[(0, *patient(rng, 0.0)) for _ in range(n)]
arm = [r[0] for r in rows]; pre = [r[1] for r in rows]; post = [r[2] for r in rows]
resp = [[sum(1 for r in rows if r[0] == g and (r[2] - r[1]) / r[1] >= 0.25),
sum(1 for r in rows if r[0] == g and (r[2] - r[1]) / r[1] < 0.25)]
for g in (1, 0)]
R = [n, n]; C = [resp[0][0] + resp[1][0], resp[0][1] + resp[1][1]]
chi = math.fsum((resp[i][j] - R[i] * C[j] / (2 * n)) ** 2 / (R[i] * C[j] / (2 * n))
for i in range(2) for j in range(2) if C[j] > 0)
z = mwu_z([r[2] for r in rows if r[0]], [r[2] for r in rows if not r[0]])
ch = [r[2] - r[1] for r in rows]
c1 = [c for c, a in zip(ch, arm) if a]; c0 = [c for c, a in zip(ch, arm) if not a]
m1, m0 = math.fsum(c1) / n, math.fsum(c0) / n
sp = (math.fsum((c - m1) ** 2 for c in c1) + math.fsum((c - m0) ** 2 for c in c0))
tchg = (m1 - m0) / math.sqrt(sp / (2 * n - 2) * 2 / n)
t_anc = ancova(post, arm, pre)[1]; t_lat = ancova([r[3] for r in rows], arm, pre)[1]
crit = 1.9966 if n == 34 else 1.9790 # t quantiles, df 66 / 126
return (chi > 3.8415, abs(z) > 1.96, abs(tchg) > crit, abs(t_anc) > crit,
abs(t_lat) > crit, sum(1 for r in rows if r[0] and r[2] == 30) / n)
def tobit_em(y, arm, base, c=29.5, it=200):
beta, s = [10.0, 0.0, 0.8], 2.0
for _ in range(it):
ys, e2 = [], []
for v, a, b in zip(y, arm, base):
mu = beta[0] + beta[1] * a + beta[2] * b
if v >= 30:
al = (c - mu) / s; lam = ND.pdf(al) / max(1e-300, 1 - ND.cdf(al))
m1 = mu + s * lam; v2 = s * s * (1 + al * lam - lam * lam)
ys.append(m1); e2.append(v2)
else:
ys.append(v); e2.append(0.0)
X = [(1.0, a, b) for a, b in zip(arm, base)]
XtX = [[math.fsum(x[i] * x[j] for x in X) for j in range(3)] for i in range(3)]
beta = solve3(XtX, [math.fsum(x[i] * v for x, v in zip(X, ys)) for i in range(3)])
s = math.sqrt(math.fsum((v - beta[0] - beta[1] * a - beta[2] * b) ** 2 + w
for v, a, b, w in zip(ys, arm, base, e2)) / len(y))
return beta[1]
if __name__ == "__main__":
rng = random.Random(30)
labs = ("responder chi2", "Mann-Whitney post", "t on change", "ANCOVA",
"ANCOVA, no ceiling")
for n, delta in ((34, 1.5), (34, 0.0), (64, 1.5)):
res = [trial(rng, n, delta) for _ in range(800)]
pw = [sum(r[k] for r in res) / len(res) for k in range(5)]
top = math.fsum(r[5] for r in res) / len(res)
print(f"n = {n}/arm, delta = {delta}: share of treated at 30 = {top:.2f}")
print(" power: " + "; ".join(f"{l} {p:.2f}" for l, p in zip(labs[:3], pw)))
print(" " + "; ".join(f"{l} {p:.2f}" for l, p in zip(labs[3:], pw[3:])))
rows = [(1, *patient(rng, 1.5)) for _ in range(1500)] + \
[(0, *patient(rng, 0.0)) for _ in range(1500)]
arm = [r[0] for r in rows]; pre = [r[1] for r in rows]; post = [r[2] for r in rows]
lat = ancova([r[3] for r in rows], arm, pre)[0]
print("B. 1,500/arm, true latent effect 1.5: ANCOVA on observed scores"
f" {ancova(post, arm, pre)[0]:.2f};")
print(f" Tobit (EM) {tobit_em(post, arm, pre):.2f}; ANCOVA on the latent scores"
f" {lat:.2f}")
# n = 34/arm, delta = 1.5: share of treated at 30 = 0.21
# power: responder chi2 0.43; Mann-Whitney post 0.52; t on change 0.68
# ANCOVA 0.69; ANCOVA, no ceiling 0.72
# n = 34/arm, delta = 0.0: share of treated at 30 = 0.09
# power: responder chi2 0.04; Mann-Whitney post 0.05; t on change 0.05
# ANCOVA 0.05; ANCOVA, no ceiling 0.05
# n = 64/arm, delta = 1.5: share of treated at 30 = 0.20
# power: responder chi2 0.69; Mann-Whitney post 0.78; t on change 0.91
# ANCOVA 0.92; ANCOVA, no ceiling 0.94
# B. 1,500/arm, true latent effect 1.5: ANCOVA on observed scores 1.43;
# Tobit (EM) 1.60; ANCOVA on the latent scores 1.60在每组 34 例、真实效应 1.5 分的模拟中,治疗组约 21% 的人治疗后达到 30 分。按论文的有效率规则(符号更正后)作卡方检验,把握度只有 0.43;对治疗后分数作曼—惠特尼检验为 0.52;对变化值作 t 检验为 0.68;协方差分析为 0.69;如果没有天花板,协方差分析可达 0.72。没有效应时,五种方法的第一类错误率都在 0.04–0.05。每组 64 例时,协方差分析的把握度为 0.92,有效率卡方检验仍只有 0.69。B 部分用大样本说明天花板的另一个后果:真实的潜在效应为 1.5 分时,对观测分数作协方差分析估计为 1.43,被上限压低;用 EM 算法拟合的托宾模型估计为 1.60,与对潜在分数作协方差分析的 1.60 一致(两者都略高于 1.5,在每组 1,500 例的抽样误差范围内)。
逐表核查
论文完整地印出了大量表格,这使得不接触原始数据也能检查数字之间是否自洽。Brown 与 Heathers 2017 年提出的 GRIM 检验利用的正是这类约束:整数数据的均数只能取某些值 (Brown and Heathers 2017)。中位数有类似的约束:34 个整数的中位数是第 17 与第 18 个值的平均,只能是整数或 .5。从中位数与四分位数估计均数的方法另有讨论 (Wan et al. 2014)。下面的程序只使用论文印出的数字。
推导
# Chapter 6. Checks on summaries printed in Lu (2024) as "M (P25, P75)". DATA: all 60
# integer-scale cells of Tables 2-4, 2-8 to 2-12, 2-14, 2-22 (integer scoring of computer
# tests assumed): table, item, treatment M P25 P75, control M P25 P75. C1 granularity,
# C2 IQR and item maximum, C3 duplicates, C4 Table 2-5, C5 baseline tests, C6 response
# rule, C7 sample size, C8 Table 1-3, C9 p columns.
import math, statistics
DATA = """2-4 MoCA 20.24 19 21 19.94 18 21.25|2-4 MMSE 21.24 19 23 20.91 19.8 22|
2-4 CDR 0.247 0 0.5 0.247 0 0.5|2-4 SDS 32.03 28 35 33.50 30 36.25|
2-8 MoCA 27.67 25 28 21.76 24 26.25|2-8 MMSE 23.35 20 26 20.91 13 16|
2-9 visuospatial 3.32 3 3.25 3.32 3 4|2-9 naming 2.47 2 3 2.41 2 3|
2-9 attention 4.26 3 5 4.26 3.75 5|2-9 language 1.83 1 2.25 1.91 1 2.25|
2-9 abstraction 1.88 2 2 1.79 2 2|2-9 recall 1.64 1 2 1.61 1 2|
2-9 orientation 4.82 4 6 4.79 4 5|2-10 visuospatial 4.06 3 5 3.82 3 4|
2-10 naming 2.97 3 3 3.06 3 3|2-10 attention 5.23 5 6 5.09 5 6|
2-10 language 2.85 3 3 2.94 3 3|2-10 abstraction 2.00 2 2 2.09 2 2|
2-10 recall 3.55 3 4 3.05 3 4|2-10 orientation 5.03 4 6 5.09 5 5.25|
2-11 colour-word 17.61 13 23 16.85 14 19.25|
2-11 figure-number 11.74 8 15 10.09 7 14|
2-11 figure-calc 7.09 5 8.25 6.76 5 8.25|2-11 digit-span 3.50 3 4 3.44 3 4|
2-12 colour-word 23.35 20 26 20.91 13 16|
2-12 figure-number 17.08 16 19 14.67 13 16|
2-12 figure-calc 9.41 8 11 7.91 6 9|2-12 digit-span 5.35 5 6 5.24 4 6|
2-14 MoCA 25.50 23 27 24.24 22.5 25|2-22 MMSE 27.44 27 28 23.73 22 24.25"""
MAXI = {"visuospatial": 5, "naming": 3, "attention": 6, "language": 3, "abstraction": 2,
"recall": 5, "orientation": 6}
N = 34
CELLS = []
for rec in DATA.replace("\n", "").split("|"):
t, item, *v = rec.split()
for arm, k in (("treat", 0), ("ctrl", 3)):
CELLS.append((t, item, arm, *map(float, v[k:k + 3])))
def chi2_p(x, df):
if df == 1:
return math.erfc(math.sqrt(x / 2))
if df == 3:
return math.erfc(math.sqrt(x / 2)) + math.sqrt(2 * x / math.pi) * math.exp(-x / 2)
raise ValueError
def pearson(rows):
R = [sum(r) for r in rows]; C = [sum(c) for c in zip(*rows)]; T = sum(R)
return math.fsum((rows[i][j] - R[i] * C[j] / T) ** 2 / (R[i] * C[j] / T)
for i in range(len(R)) for j in range(len(C)))
def t_p(t, df): # two-sided p of Student t, exact series for even df
th = math.atan(abs(t) / math.sqrt(df)); c2 = math.cos(th) ** 2; term = acc = 1.0
for k in range(1, df // 2):
term *= c2 * (2 * k - 1) / (2 * k); acc += term
return 1 - math.sin(th) * acc
step = lambda item: 0.5 if item == "CDR" else 1.0
half = lambda c: c[3] / (step(c[1]) / 2)
med = [c for c in CELLS if abs(half(c) - round(half(c))) < 1e-9]
print(f"C1 {len(CELLS)} cells: {len(CELLS) - len(med)} cannot be the median of 34 scores;"
f" the {len(med)} that can:\n " + ", ".join(f"{c[3]:.2f}" for c in med))
def mean_ok(v, st):
return any(abs(math.floor(100 * k * st / N) / 100 - v) < 1e-9 or
abs(round(k * st / N, 2) - v) < 1e-9 for k in range(int(40 * N / st)))
odd = [c for c in CELLS if not mean_ok(c[3], step(c[1]))]
print(" as a mean of 34 scores (2 decimals, cut or rounded), incompatible:\n " +
", ".join(f"{c[3]:g} ({c[0]})" for c in odd))
print("C2 M outside its own IQR (Q1, Q3), or above the item maximum:")
for t, item, arm, m, q1, q3 in CELLS:
over = item in MAXI and t in ("2-9", "2-10") and m > MAXI[item]
if m < q1 - 1e-9 or m > q3 + 1e-9 or over:
flag = f" ABOVE MAXIMUM {MAXI[item]}" if over else ""
print(f" Table {t:5s} {item + ', ' + arm:20s} {m:6.2f} ({q1:g}, {q3:g}){flag}")
lo, mid, hi = 8, 18, 8 # observations below P25 / inside / above P75
need_low = (N * 21.76 - mid * 24 - hi * 26.25) / lo
need_high = (N * 20.91 - lo * 13 - mid * 16) / hi
print(f" if 21.76 (24, 26.25) were a mean, the {lo} lowest would average"
f" <= {need_low:.1f}"
"\n (all were 18-25 at entry);")
print(f" if 20.91 (13, 16) were a mean, the {hi} highest colour-word scores would"
" average"
f"\n >= {need_high:.1f} (treatment-arm P75 after treatment: 26)")
row = lambda t, it: [c[3:] for c in CELLS if c[0] == t and c[1] == it]
print("C3 Table 2-8 'MMSE' = Table 2-12 'colour-word', both arms:",
row("2-8", "MMSE") == row("2-12", "colour-word"))
cw = row("2-12", "colour-word")[1]
print(" its control 20.91 = control baseline MMSE, Table 2-4:",
cw[0] == row("2-4", "MMSE")[1][0])
print(" its control (13, 16) = control figure-number IQR, Table 2-12:",
cw[1:] == row("2-12", "figure-number")[1][1:])
tx = (0, 2, 26, 8); ct = (0, 0, 22, 12)
print(f"C4 Table 2-5, treatment counts {tx}: sum {sum(tx)} (n = 34);"
f"\n (2 + 26)/34 = {28 / 34:.1%} vs 'total effective 76.4 %'")
for lab, a in (("26 vs 22 effective", 26), ("28 vs 22 effective", 28)):
x = pearson([[a, 34 - a], [22, 12]])
print(f" {lab}: Pearson chi2 = {x:.3f}, p = {chi2_p(x, 1):.3f}")
rows = [[26, 8], [22, 12]]; R = [34, 34]; C = [48, 20]
yates = math.fsum((abs(rows[i][j] - R[i] * C[j] / 68) - 0.5) ** 2 / (R[i] * C[j] / 68)
for i in range(2) for j in range(2))
print(f" 26 vs 22 with Yates correction: chi2 = {yates:.3f},"
f" p = {chi2_p(yates, 1):.3f}")
print(" -> p = 0.287 fits 26/34 effective (uncorrected); counts summing to 34 with 26"
"\n effective: (0, 2, 24, 8) or (0, 0, 26, 8)")
print("C5 baseline tests re-computed from the printed summaries")
for lab, m1, s1, m2, s2, rep in (
("age (Table 2-3)", 61.70, 9.98, 65.52, 7.94, "-1.747, 0.085"),
("days since stroke", 52.67, 51.04, 37.88, 36.60, "1.373, 0.174")):
t = (m1 - m2) / math.sqrt((s1 ** 2 + s2 ** 2) / N)
print(f" {lab:17s} t = {t:+.3f}, p = {t_p(t, 2 * N - 2):.3f} (reported {rep})")
x = pearson([[24, 10], [23, 11]]); x2 = pearson([[7, 20, 5, 2], [10, 13, 7, 4]])
print(f" sex: chi2 = {x:.3f}, p = {chi2_p(x, 1):.3f} (Table 2-6: 0.069, 0.793;"
" text calls 0.069 the p)")
print(f" education: chi2 = {x2:.3f}, p = {chi2_p(x2, 3):.3f} (Table 2-7: 3.014, 0.389)")
print("C6 response rule: the printed (pre - post)/pre is negative for any improvement.")
print(" Sign reversed, (post - pre)/pre, the largest rate reachable with MoCA <= 30:")
print(" baseline: " + " ".join(f"{b}:{(30 - b) / b:.0%}" for b in range(18, 26)))
print(" 'cured' (>= 75 %) is unreachable for every eligible patient; 'effective'"
" (>= 25 %)\n needs a perfect 30 at baseline 24 and is impossible at 25")
z = statistics.NormalDist().inv_cdf
d = (25.88 - 22.96) / math.sqrt((4.83 ** 2 + 2.88 ** 2) / 2)
for lab, za in (("two-sided 0.05", z(0.975)), ("one-sided 0.05", z(0.95))):
nn = 2 * (za + z(0.80)) ** 2 / d ** 2 + za ** 2 / 4
print(f"C7 sample size, d = {d:.3f}, 80 % power, {lab}: {math.ceil(nn)} per arm"
if lab.startswith("two") else f" {lab}: {math.ceil(nn)} per arm"
" (thesis: 25)")
print(f" 20 % dropout: 25 / 0.8 = {25 / 0.8:.2f} -> 32 per arm (thesis: 25 x 1.2 = 30)")
T13 = (("Baihui", 127, 73.83), ("Sishencong", 68, 39.53), ("Shenting", 63, 36.72),
("Zusanli", 51, 29.65), ("Fengchi", 47, 27.33), ("Neiguan", 43, 40.03),
("Sanyinjiao", 40, 23.26), ("Shenmen", 38, 22.09), ("Yintang", 23, 13.37))
off = [(p, f, r, 100 * f / 172) for p, f, r in T13 if abs(100 * f / 172 - r) > 0.06]
print(f"C8 Table 1-3, f/172 against the printed %:"
f" {len(T13) - len(off)} of {len(T13)} agree;")
for p, f, r, c in off:
print(f" {p}: {f}/172 = {c:.2f} %, printed {r} %")
print("C9 p columns of Tables 2-8, 2-10, 2-12 are unlabelled; they match the treatment")
print(" arm's within-group p (2-10 orientation 0.407 = Table 2-21; 2-12 all < 0.001 =")
print(" Tables 2-23 to 2-27). Conflicting baseline p: MMSE 0.511 (Table 2-4) vs 0.661")
print(" (2.3.4(1)); digit span 0.648 (Table 2-11) vs 0.759 (2.3.4(6))")
# C1 60 cells: 56 cannot be the median of 34 scores; the 4 that can:
# 33.50, 2.00, 3.50, 25.50
# as a mean of 34 scores (2 decimals, cut or rounded), incompatible:
# 0.247 (2-4), 0.247 (2-4), 1.83 (2-9)
# C2 M outside its own IQR (Q1, Q3), or above the item maximum:
# Table 2-8 MoCA, ctrl 21.76 (24, 26.25)
# Table 2-8 MMSE, ctrl 20.91 (13, 16)
# Table 2-9 visuospatial, treat 3.32 (3, 3.25)
# Table 2-9 abstraction, treat 1.88 (2, 2)
# Table 2-9 abstraction, ctrl 1.79 (2, 2)
# Table 2-10 naming, treat 2.97 (3, 3)
# Table 2-10 naming, ctrl 3.06 (3, 3) ABOVE MAXIMUM 3
# Table 2-10 language, treat 2.85 (3, 3)
# Table 2-10 language, ctrl 2.94 (3, 3)
# Table 2-10 abstraction, ctrl 2.09 (2, 2) ABOVE MAXIMUM 2
# Table 2-12 colour-word, ctrl 20.91 (13, 16)
# if 21.76 (24, 26.25) were a mean, the 8 lowest would average <= 12.2
# (all were 18-25 at entry);
# if 20.91 (13, 16) were a mean, the 8 highest colour-word scores would average
# >= 39.9 (treatment-arm P75 after treatment: 26)
# C3 Table 2-8 'MMSE' = Table 2-12 'colour-word', both arms: True
# its control 20.91 = control baseline MMSE, Table 2-4: True
# its control (13, 16) = control figure-number IQR, Table 2-12: True
# C4 Table 2-5, treatment counts (0, 2, 26, 8): sum 36 (n = 34);
# (2 + 26)/34 = 82.4% vs 'total effective 76.4 %'
# 26 vs 22 effective: Pearson chi2 = 1.133, p = 0.287
# 28 vs 22 effective: Pearson chi2 = 2.720, p = 0.099
# 26 vs 22 with Yates correction: chi2 = 0.637, p = 0.425
# -> p = 0.287 fits 26/34 effective (uncorrected); counts summing to 34 with 26
# effective: (0, 2, 24, 8) or (0, 0, 26, 8)
# C5 baseline tests re-computed from the printed summaries
# age (Table 2-3) t = -1.747, p = 0.085 (reported -1.747, 0.085)
# days since stroke t = +1.373, p = 0.174 (reported 1.373, 0.174)
# sex: chi2 = 0.069, p = 0.793 (Table 2-6: 0.069, 0.793; text calls 0.069 the p)
# education: chi2 = 3.014, p = 0.389 (Table 2-7: 3.014, 0.389)
# C6 response rule: the printed (pre - post)/pre is negative for any improvement.
# Sign reversed, (post - pre)/pre, the largest rate reachable with MoCA <= 30:
# baseline: 18:67% 19:58% 20:50% 21:43% 22:36% 23:30% 24:25% 25:20%
# 'cured' (>= 75 %) is unreachable for every eligible patient; 'effective' (>= 25 %)
# needs a perfect 30 at baseline 24 and is impossible at 25
# C7 sample size, d = 0.734, 80 % power, two-sided 0.05: 31 per arm
# one-sided 0.05: 24 per arm (thesis: 25)
# 20 % dropout: 25 / 0.8 = 31.25 -> 32 per arm (thesis: 25 x 1.2 = 30)
# C8 Table 1-3, f/172 against the printed %: 7 of 9 agree;
# Shenting: 63/172 = 36.63 %, printed 36.72 %
# Neiguan: 43/172 = 25.00 %, printed 40.03 %
# C9 p columns of Tables 2-8, 2-10, 2-12 are unlabelled; they match the treatment
# arm's within-group p (2-10 orientation 0.407 = Table 2-21; 2-12 all < 0.001 =
# Tables 2-23 to 2-27). Conflicting baseline p: MMSE 0.511 (Table 2-4) vs 0.661
# (2.3.4(1)); digit span 0.648 (Table 2-11) vs 0.759 (2.3.4(6))以下逐条说明,按「表号、原文、问题、可能的更正」的顺序写。我们无法接触原始数据,所以这些都是需要论文作者核对的问题,而不是对数据真实性的判断;其中多数很可能是排版、抄录或统计软件输出整理中的笔误。
- 「中位数」的取值(C1)。 论文以「M(P25, P75)」报告量表与测验结果。程序检查了表 2-4、2-8 至 2-12、2-14、2-22 的全部 60 个整数计分格(不含舒尔特计时;计算机测验按整数计分是本书的假设),其中 56 个不可能是 34 个整数的中位数,例如 MoCA 20.24、27.67,CDR 0.247;可以是中位数的只有 33.50、2.00、3.50、25.50 四格。SPSS 按组中值插值的中位数也会给出非 .5 的值,但它与整数的四分位数和 M 落在四分位间距之外两点不相容。「M」很可能是均数:摘要原文写的是「评分均值较入组前提高」,图 2-6 是带误差线的柱图,60 格中除 CDR 0.247 与 1.83(同一格在表 2-19 印作 1.82)外,都与「34 个整数之均数、取两位小数」相容。建议: 注明这些数字是均数还是中位数,并报告标准差。
- 落在四分位间距之外或超出满分的数值(C2)。 有 11 格落在自己的四分位间距之外。多数是 0–6 分的 MoCA 分项(如抽象 1.88(2.00, 2.00)),若 M 是均数,偏态分布中可以出现。但有四格不成立或很难成立。表 2-10 对照组治疗后命名 3.06(3.00, 3.00)与抽象 2.09(2.00, 2.00)超过分项满分(北京版命名 3 分、抽象 2 分),无论作为何种汇总量都不可能。表 2-8 与表 2-13 对照组治疗后 MoCA 21.76(24.00, 26.25)若为均数,最低 8 人平均须不高于约 12.2 分,而入组时都在 18–25 分,对照组组内改善又标为 P < 0.001;图 2-6 中这一柱目测约为 25。表 2-12 与表 2-23 对照组治疗后色词 20.91(13.00, 16.00)若为均数,最高 8 人平均须不低于约 39.9,而治疗组治疗后的 P75 只有 26。建议: 核对这四格的原始输出。
- 重复与串行(C3)。 表 2-8 中 MMSE 一行(治疗组 23.35(20.00, 26.00),对照组 20.91(13.00, 16.00))与表 2-12 中色词一行完全相同;而这一格对照组的 20.91 等于表 2-4 对照组 MMSE 基线,(13.00, 16.00) 等于表 2-12 对照组图形—数字的四分位数,很可能是从其他行串入的。表 2-22 给出的 MMSE 治疗后数值 27.44(27.00, 28.00)与 23.73(22.00, 24.25)与图 2-6 一致。建议: 表 2-8 的 MMSE 行换为表 2-22 的数字,并核对色词对照组一格。
- 有效率表(C4,表 2-5)。 治疗组「痊愈 0、显效 2、有效 26、无效 8」之和为 36,而每组 34 例;(2 + 26)/34 = 82.4%,与「总有效率 76.4%」不符;另有「64.7.7%」「35.3.3%」「n=0.0」等排版笔误。程序显示,报告的 p = 0.287 恰好是「有效 26/34 对 22/34」的未校正皮尔逊卡方检验(χ² = 1.133),而「28/34 对 22/34」会给出 p = 0.099,加耶茨校正则为 p = 0.425。一种可能的更正: 治疗组「有效」应为 24 例(70.6%),总有效 26 例,76.5%(论文印作 76.4%,符合截断);另一种可能是「显效」应为 0 例。两者都使计数之和为 34、总有效为 26。
- 疗效公式(C6)。 第 2.2.5.3 节的「减分率 =(治疗前 MoCA − 治疗后 MoCA)/ 治疗前分数 × 100%」是下降型量表的写法,用于 MoCA 时任何改善都得负值;同节「同时满足 MoCA 量表评分下降至 12.68 ± 8.75 以下」也与 MoCA 的方向不符。即使把公式改为(治疗后 − 治疗前)/ 治疗前,在 30 分的上限下,基线 18 分的患者最多只能达到 67%,任何患者都不可能达到「痊愈」(≥ 75%);基线 24 分的患者须达到满分才算「有效」,基线 25 分的患者不可能「有效」。建议: 主要分析改用以基线为协变量的协方差分析;如需报告应答率,应按事先确定的、与基线无关的界值(例如提高若干分)定义。
- 样本量(C7)。 第 2.2.1.2 节按前期研究的均数与标准差(22.96 ± 4.83 对 25.88 ± 2.88)、用 PASS 计算得到每组 25 例。按这组数字,标准化效应为 0.734,双侧 α = 0.05、80% 把握度需要每组约 31 例,单侧时约 24 例;论文印出的公式是图片,未能核对其所用的单双侧与把握度。另外,20% 脱落的正确换算是 25/0.8 = 31.25,即每组 32 例,而不是 25 × 1.2 = 30 例;样本量按均数的 t 检验计算,主要分析却用有效率卡方检验与秩和检验,两者不匹配。还值得注意的是,在所引的前期研究中,治疗后「治疗组」的均数(22.96)低于「对照组」(25.88)。
- 网络部分的频率(C8,表 1-3)。 9 个可核对的频率中 7 个与「频次/172」一致;神庭 63/172 应为 36.63%(印作 36.72%),内关 43/172 应为 25.00%(印作 40.03%)。神门在表中重复出现。第 1.1.7.7 节又写神庭「出现频数为 68」、风池「49 次」,与表 1-3 的 63 与 47 不一致。
- p 值列的含义与两处基线矛盾(C9)。 论文各表中星号表示组间比较、井号表示与干预前比较,但表 2-8、2-10、2-12 的 p 值列没有注明含义。其数值与治疗组组内 p 高度吻合:表 2-10 定向力的 0.407 即表 2-21 治疗组组内 p,表 2-12 五项的 < 0.001 与表 2-23 至 2-27 的治疗组组内 p 一致,表 2-8 的 < 0.001 与表 2-13、2-22 的组内 p 一致(只有表 2-10 抽象一项例外)。按这种读法,表 2-8 的 MoCA「< 0.001」与正文 P = 0.005、单星号并不矛盾,数字广度与图形—计算也不矛盾。真正的矛盾只有两处基线 p:MMSE 表 2-4 为 0.511,第 2.3.4(1) 节为 0.661;数字广度表 2-11 为 0.648,第 2.3.4(6) 节为 0.759。建议: 在表注中写明 p 值列是何种比较,并统一两处基线 p。
- 其他文字上的矛盾。 第 2.3.2.1 节把性别比较的卡方值 0.069 写成 p 值(表 2-6 为 χ² = 0.069、p = 0.793);第 2.3.1 节把治疗组的年龄 61.70 ± 9.98 岁与病程 52.67 ± 51.04 天写成全部 68 例的数值;第 2.3.3 节「无显著性降低(P = 0.002 < 0.05)」、第 2.3.4(2) 节「无著性升高(P < 0.001)」自相矛盾;第 2.3.3(3) 节「P = 0.012 > 0.05」不等号方向有误;表 2-17 对照组抽象基线 1.61(1.00, 2.00)与表 2-9 的 1.79(2.00, 2.00)不一致。
致论文作者
论文完整印出这么多表格,本身就是对读者负责的做法,上面的核对也因此才成为可能;所列问题大多很可能来自表格整理与排版。但它们恰好落在决定结论的位置:主要疗效的定义与计数、主要指标的组间比较、MoCA 与 MMSE 的治疗后数值。任何正式结论,包括本书的结论,都需要原始数据。我们邀请论文作者公开去标识化的逐例数据(分组,基线与治疗后的 MoCA、MMSE 与各项计算机测验),以便任何人用协方差分析重新检验。
本章小结
研究内容二是开放标签、加载设计、单中心、每组 34 例的试验。论文对 MoCA 给出两种组间结果:有效率 p = 0.287(「疗效相当」),治疗后评分 P = 0.005;后者未调整基线,对照组数值不自洽,又无假针对照,仍不足以下结论;摘要「疗效优于对照组」部分依据组内前后检验,而组内与基线比较不能代替组间比较。计时测验的提速可以来自漂移率、边界或非决策时间中的任何一个。MoCA 有 30 分的上限,有效率卡方检验浪费了大部分信息,协方差分析的把握度明显更高,托宾模型可以校正上限造成的低估。逐表核查发现:60 个整数计分格中 56 个不可能是中位数;四格即使作为均数也不成立或很难成立(两格超出分项满分);一行 MMSE 与色词数据重复;有效率表计数不一致,而 p = 0.287 对应有效 26/34 对 22/34;疗效公式方向相反;样本量计算与分析方法不匹配;p 值列未注明含义,两处基线 p 矛盾。这些问题需要论文作者用原始数据核对。
参考文献
Fan, Y. (2026). Finger–Brain Cortical Magnification, Distal Microneedle Afference and Computational Neurology of Post-Stroke Cognitive Network Reconstruction: From Complex-Network Pinning Control and Thalamocortical 10 Hz Alpha Restoration to Drift-Diffusion Models, Read Beside the Twelve Jing-Well Points of the Lingshu [指—脑皮层放大、远端微针传入与卒中后认知网络重建的计算神经学] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23839915 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).