5D Library › 内感受解离、突显网络切换与内脏自主神经修复 · English

第六章 复合模型、不可识别性与个案证据

前五章逐个检验了链条的环节。本章把它们接成一个模型,然后做三件事:说明这个模型的结论是它自己假设的算术后果;证明它的关键参数无法从原说法所依赖的测量中识别;分析「脱敏个案」这一证据形式能支持什么。

把链条写成一个模型

三个状态,三个一阶方程。高阶精度 π\pi 向一个由练习设定的目标松弛;防御驱动 SS 被精度加权的威胁评估抬高、按速率 rSr_S 衰减;内脏传导度 GG 随 SS 下降、向上限松弛。

dπdt=πset−πτπ,dSdt=aSπA(t)−rSS,dGdt=Gmaxe−kGS−GτG.\frac{d\pi}{dt} = \frac{\pi_{\text{set}} – \pi}{\tau_\pi}, \qquad \frac{dS}{dt} = a_S\,\pi\, A(t) – r_S S, \qquad \frac{dG}{dt} = \frac{G_{\max} e^{-k_G S} – G}{\tau_G}.

A(t)A(t) 是内在事件的评估脉冲序列,在每一次运行中完全相同。这是作者的模型,每一个箭头都是一个假设。

推导

# Chapter 6. The author's composite model, written out so it can be
# criticised. Three coupled states, integrated with RK4 (dt = 0.01 s, 600 s):
#   higher-order precision   pi(t)  - driven toward a set point pi_set by
#   "practice" defensive drive          S(t)   - raised by precision-weighted
#   threat appraisal,
#                                     decays with rate rS
#   visceral conductance     G(t)   - falls with S (alpha-adrenergic), relaxes
#   to Gmax
#     dpi/dt = (pi_set - pi) / tau_pi
#     dS/dt  = aS * pi * A(t) - rS * S
#     dG/dt  = (Gmax * exp(-kG * S) - G) / tau_G
# A(t) is the stream of inner events (a fixed synthetic sequence of appraisal
# pulses, the same in every run). This is a MODEL: every arrow is an
# assumption, and the program's job is to show which conclusions follow from
# it and which do not.
import math

DT, T_END = 0.01, 600.0
GMAX, KG, TAU_G = 1.0, 0.8, 25.0
AS, RS = 1.0, 0.08
EVENTS = [
    (60, 20, 1.0),
    (150, 20, 1.4),
    (260, 20, 0.8),
    (380, 20, 1.6),
    (500, 20, 1.0),
]

def appraisal(t):
    for t0, dur, amp in EVENTS:
        if t0 <= t < t0 + dur:
            return amp
    return 0.0

def run(pi0, pi_set, tau_pi):
    pi, S, G = pi0, 0.0, GMAX
    out = []
    def f(state, t):
        p, s, g = state
        return ((pi_set - p) / tau_pi,
                AS * p * appraisal(t) - RS * s,
                (GMAX * math.exp(-KG * s) - g) / TAU_G)
    n = int(T_END / DT)
    for i in range(n):
        t = i * DT
        st = (pi, S, G)
        k1 = f(st, t)
        k2 = f(tuple(st[j] + DT / 2 * k1[j] for j in range(3)), t + DT / 2)
        k3 = f(tuple(st[j] + DT / 2 * k2[j] for j in range(3)), t + DT / 2)
        k4 = f(tuple(st[j] + DT * k3[j] for j in range(3)), t + DT)
        pi, S, G = (st[j] + DT / 6 * (k1[j] + 2 * k2[j] + 2 * k3[j] + k4[j])
                    for j in range(3))
        out.append((t, pi, S, G))
    return out

def summarise(tr):
    Ss = [r[2] for r in tr]
    Gs = [r[3] for r in tr]
    # conductance-seconds lost
    auc = math.fsum((GMAX - g) * DT for g in Gs)
    return max(Ss), math.fsum(Gs) / len(Gs), min(Gs), auc

if __name__ == "__main__":
    print(
        "Five appraisal pulses of 20 s each over a 600 s run; identical in"
        " every"
    )
    print("case.")
    print(
        "\ncase                                 peak S   mean G   min G  "
        " lost G.s"
    )
    cases = [("pi fixed at 1.0 (no practice)", 1.0, 1.0, 1e9),
             ("pi fixed at 0.5", 0.5, 0.5, 1e9),
             ("pi fixed at 0.2", 0.2, 0.2, 1e9),
             ("pi 1.0 -> 0.2, tau 60 s", 1.0, 0.2, 60.0),
             ("pi 1.0 -> 0.2, tau 300 s", 1.0, 0.2, 300.0),
             ("pi 1.0 -> 0.2, tau 1200 s", 1.0, 0.2, 1200.0)]
    for lab, p0, ps, tp in cases:
        pk, mg, mn, auc = summarise(run(p0, ps, tp))
        print(
            "{:36s} {:7.3f} {:8.3f} {:7.3f} {:9.2f}".format(
                lab, pk, mg, mn, auc
            )
        )

    print(
        "\nWithin the model, lowering pi lowers peak defensive drive"
        " roughly in"
    )
    print(
        "proportion and recovers most of the lost conductance. That is not a"
        " finding"
    )
    print(
        "about bodies; it is arithmetic, because pi multiplies the only term"
        " that"
    )
    print("raises S. The model cannot be evidence for its own first arrow.")

    print("\nWhat the model gets WRONG if read as physiology, three ways:")
    print(
        "1. G relaxes to Gmax with tau_G = {:.0f} s and no autoregulation;"
        " programme".format(TAU_G)
    )
    print("   p07 shows flow sensing gives back 34-60 % of such a change.")
    print(
        "2. S has a single decay rate; real sympathetic outflow to different"
        " beds is"
    )
    print("   not one variable (regional specificity).")
    print(
        "3. pi enters only multiplicatively, so the model cannot represent"
        " the case"
    )
    print(
        "   programme p01 found: an alarmed PRIOR, where lowering sensory"
        " precision"
    )
    print("   raises the posterior estimate instead of lowering it.")

    print(
        "\nThe third point is a genuine failure of this model, so here is the"
        " variant"
    )
    print(
        "that admits it: dS/dt = aS * (pi * A + (1 - pi) * B) - rS * S, where"
        " B is a"
    )
    print(
        "standing alarmed prior that the higher-order term no longer"
        " suppresses."
    )

    def run_prior(pi_set, B, tau_pi=300.0, pi0=1.0):
        pi, S, G = pi0, 0.0, GMAX
        n = int(T_END / DT)
        Gs, Ss = [], []
        for i in range(n):
            t = i * DT
            dpi = (pi_set - pi) / tau_pi
            dS = AS * (pi * appraisal(t) + (1 - pi) * B) - RS * S
            dG = (GMAX * math.exp(-KG * S) - G) / TAU_G
            pi += DT * dpi
            S += DT * dS
            G += DT * dG
            Gs.append(G)
            Ss.append(S)
        return max(Ss), math.fsum(Gs) / len(Gs)

    print("\n   B (standing      pi_set 1.0:            pi_set 0.2:")
    print("   alarmed prior)   peak S / mean G        peak S / mean G")
    for B in (0.0, 0.1, 0.3, 0.6):
        a = run_prior(1.0, B)
        b = run_prior(0.2, B)
        print("   {:14.2f}   {:9.3f} / {:7.3f} {:12.3f} / {:7.3f}"
              .format(B, a[0], a[1], b[0], b[1]))
    print(
        "\n   With a standing alarmed prior above about 0.3, lowering pi"
        " STOPS helping"
    )
    print(
        "   and then hurts: the practice removes the term that was"
        " suppressing the"
    )
    print(
        "   prior. This is the model's own boundary condition, and it is"
        " testable."
    )
# Five appraisal pulses of 20 s each over a 600 s run; identical in every
# case.
#
# case                                 peak S   mean G   min G   lost G.s
# pi fixed at 1.0 (no practice)         15.960    0.564   0.171    261.74
# pi fixed at 0.5                        7.980    0.646   0.257    212.24
# pi fixed at 0.2                        3.192    0.770   0.431    138.16
# pi 1.0 -> 0.2, tau 60 s                4.386    0.739   0.367    156.65
# pi 1.0 -> 0.2, tau 300 s               9.299    0.639   0.212    216.75
# pi 1.0 -> 0.2, tau 1200 s             12.557    0.584   0.181    249.65
#
# Within the model, lowering pi lowers peak defensive drive roughly in
# proportion and recovers most of the lost conductance. That is not a finding
# about bodies; it is arithmetic, because pi multiplies the only term that
# raises S. The model cannot be evidence for its own first arrow.
#
# What the model gets WRONG if read as physiology, three ways:
# 1. G relaxes to Gmax with tau_G = 25 s and no autoregulation; programme
#    p07 shows flow sensing gives back 34-60 % of such a change.
# 2. S has a single decay rate; real sympathetic outflow to different beds is
#    not one variable (regional specificity).
# 3. pi enters only multiplicatively, so the model cannot represent the case
#    programme p01 found: an alarmed PRIOR, where lowering sensory precision
#    raises the posterior estimate instead of lowering it.
#
# The third point is a genuine failure of this model, so here is the variant
# that admits it: dS/dt = aS * (pi * A + (1 - pi) * B) - rS * S, where B is a
# standing alarmed prior that the higher-order term no longer suppresses.
#
#    B (standing      pi_set 1.0:            pi_set 0.2:
#    alarmed prior)   peak S / mean G        peak S / mean G
#              0.00      15.965 /   0.564        9.301 /   0.639
#              0.10      15.965 /   0.564        9.709 /   0.439
#              0.30      15.965 /   0.564       10.525 /   0.247
#              0.60      15.965 /   0.564       11.749 /   0.152
#
#    With a standing alarmed prior above about 0.3, lowering pi STOPS helping
#    and then hurts: the practice removes the term that was suppressing the
#    prior. This is the model's own boundary condition, and it is testable.

第一张表的内容是:在模型内部,降低 π\pi 大致按比例降低防御驱动的峰值(15.960 → 3.192),并把损失的传导度大部分收回(261.74 → 138.16 传导度·秒)。但这不是关于身体的发现,而是算术:π\pi 乘在唯一抬高 SS 的项上,所以降低 π\pi 必然降低峰值。一个模型不能成为它自己第一个箭头的证据。这是第 33 本反复强调的一点,也是本书把「模型中的证明」限定在模型内部的理由。

程序随后列出这个模型若被读作生理学会错在哪里:GG 以 τG=25\tau_G = 25 秒向上限松弛而没有自动调节,而第五章的程序 p07 显示流量敏感的调节会吃掉 34% 到 60%;SS 只有一个衰减速率,而真实的交感输出不是一个变量 (Morrison 2001);最关键的是,π\pi 只以乘法进入,所以这个模型无法表示第一章的情形——先验本身警觉时,降低精度使估计上升。

最后一段给出承认这一缺陷的变体:dS/dt=aS(πA+(1−π)B)−rSSdS/dt = a_S(\pi A + (1-\pi)B) – r_S S,其中 BB 是一个常驻的警觉先验,高阶项不再压制它。结果是一个边界条件:B=0B = 0 时降低 π\pi 使平均传导度从 0.564 升到 0.639(有益);B=0.3B = 0.3 时降到 0.247,B=0.6B = 0.6 时降到 0.152(有害)。也就是说,当常驻的警觉先验超过约 0.3 时,降低 π\pi 由有益转为有害,因为练习移除的正是原本压制该先验的那一项。这是模型自己的边界条件,而且它是可检验的——第八章把基线的身体警觉作为预先指定的效应修饰变量,正是为了检验它。

测量识别不了要紧的那个参数

现在是本章最重要的结果。原说法用心率变异度支持关于内脏/盆腔的结论。第四章说明这两者是不同的解剖系统。本节把这件事写成一个结构可识别性的结果:即使接受整个模型,心脏的观测在原则上也不能确定内脏通道的参数。

设一个潜在的防御驱动 S(t)S(t) 经两条各有权重的通路到达两个靶:

H(t)=h0−cH(wCS(t))(心脏观测),G(t)=Gmaxe−kG(wVS(t))(内脏传导度).H(t) = h_0 – c_H\,(w_C S(t)) \quad(\text{心脏观测}), \qquad G(t) = G_{\max}e^{-k_G (w_V S(t))}\quad(\text{内脏传导度}).

HH 只约束乘积 cHwCc_H w_C。kGwVk_G w_V 在 HH 中根本不出现。

推导

# Chapter 6. Structural non-identifiability: the measurement does not pin down
# the claim. In the composite model one latent defensive drive S(t) reaches
# two targets by separate pathways with separate weights:
#     cardiac observable    H(t) = h0 - cH * (wC * S(t))        <- this is
#     what HRV sees visceral conductance  G(t) = Gmax * exp(-kG * (wV * S(t)))
#     <- this is the claim
# H constrains only the PRODUCT cH*wC. Nothing in H constrains kG*wV. The
# program fits four parameter sets that reproduce the SAME cardiac trace to
# within floating-point error, and reports the visceral conductance each one
# implies. If the traces are identical, no amount of cardiac data can choose
# between them. Synthetic data; author's model; illustrative units.
import math

DT, T_END = 0.05, 600.0
EVENTS = [(60, 30, 1.0), (200, 30, 1.5), (360, 30, 0.9), (500, 30, 1.2)]
H0, RS, AS = 1.000, 0.08, 1.0

def appraisal(t):
    for t0, dur, amp in EVENTS:
        if t0 <= t < t0 + dur:
            return amp
    return 0.0

def drive():
    S, out = 0.0, []
    for i in range(int(T_END / DT)):
        t = i * DT
        S += DT * (AS * appraisal(t) - RS * S)
        out.append(S)
    return out

def cardiac(S, cH, wC):
    return [H0 - cH * wC * s for s in S]

def visceral(S, kG, wV, gmax=1.0):
    return [gmax * math.exp(-kG * wV * s) for s in S]

def rmse(a, b):
    return math.sqrt(math.fsum((x - y) ** 2 for x, y in zip(a, b)) / len(a))

if __name__ == "__main__":
    S = drive()
    print("One latent drive S(t), peak {:.3f}, mean {:.3f}"
          .format(max(S), math.fsum(S) / len(S)))

    # the "truth" used to generate the observed cardiac trace
    H_obs = cardiac(S, cH=0.040, wC=1.00)
    print(
        "Observed cardiac index: baseline {:.3f}, minimum {:.3f}, mean"
        " {:.3f}\n".format(
            H_obs[0], min(H_obs), math.fsum(H_obs) / len(H_obs)
        )
    )

    print(
        "Four parameter sets. cH*wC is held at 0.040 in all four, so"
        " every one"
    )
    print("reproduces the observed cardiac trace exactly; kG*wV is free.\n")
    print("   cH     wC    cH*wC   cardiac RMSE   kG     wV   kG*wV"
          "  mean G  min G")
    sets = [(0.040, 1.00, 0.80, 1.00),
            (0.080, 0.50, 0.80, 0.50),
            (0.020, 2.00, 0.80, 2.00),
            (0.040, 1.00, 0.05, 1.00)]
    rows = []
    for cH, wC, kG, wV in sets:
        H = cardiac(S, cH, wC)
        G = visceral(S, kG, wV)
        mg = math.fsum(G) / len(G)
        rows.append((kG * wV, mg, min(G)))
        print(
            "   {:5.3f} {:6.2f} {:7.3f} {:14.2e} {:5.2f} {:5.2f} {:7.2f}"
            " {:7.3f} {:6.3f}".format(
                cH, wC, cH * wC, rmse(H, H_obs), kG, wV, kG * wV, mg, min(G)
            )
        )

    lo = min(r[1] for r in rows)
    hi = max(r[1] for r in rows)
    print(
        "\nEvery cardiac RMSE is at the level of floating-point error, so the"
        " four are"
    )
    print(
        "the SAME observation. The mean visceral conductance they imply"
        " ranges from"
    )
    print(
        "{:.3f} to {:.3f}, a factor of {:.1f}, and the minimum from {:.3f} to"
        " {:.3f}.".format(
            lo, hi, hi / lo, min(r[2] for r in rows), max(r[2] for r in rows)
        )
    )

    print(
        "\nWhat would break the tie: a direct measurement of the visceral"
        " bed, taken"
    )
    print(
        "at the same time as the cardiac one. Without it, 'HRV rose,"
        " therefore"
    )
    print(
        "visceral perfusion was restored' is not an inference; it is the"
        " assumption"
    )
    print("kG*wV = cH*wC, written as if it had been measured.")

    print(
        "\nA second, weaker claim IS identifiable from H alone: the TIMING"
        " of S."
    )
    peak_i = min(range(len(H_obs)), key=lambda i: H_obs[i])
    print("   time of maximum drive from the cardiac trace: {:.1f} s"
          .format(peak_i * DT))
    print("   (true value by construction: {:.1f} s)".format(
        max(range(len(S)), key=lambda i: S[i]) * DT))
    print(
        "   So the cardiac channel can time the response and order two"
        " conditions,"
    )
    print("   which is what the Chapter 8 protocol asks of it, and no more.")
# One latent drive S(t), peak 17.057, mean 2.874
# Observed cardiac index: baseline 1.000, minimum 0.318, mean 0.885
#
# Four parameter sets. cH*wC is held at 0.040 in all four, so every one
# reproduces the observed cardiac trace exactly; kG*wV is free.
#
#    cH     wC    cH*wC   cardiac RMSE   kG     wV   kG*wV  mean G  min G
#    0.040   1.00   0.040       0.00e+00  0.80  1.00    0.80   0.568  0.000
#    0.080   0.50   0.040       0.00e+00  0.80  0.50    0.40   0.636  0.001
#    0.020   2.00   0.040       0.00e+00  0.80  2.00    1.60   0.506  0.000
#    0.040   1.00   0.040       0.00e+00  0.05  1.00    0.05   0.884  0.426
#
# Every cardiac RMSE is at the level of floating-point error, so the four are
# the SAME observation. The mean visceral conductance they imply ranges from
# 0.506 to 0.884, a factor of 1.7, and the minimum from 0.000 to 0.426.
#
# What would break the tie: a direct measurement of the visceral bed, taken
# at the same time as the cardiac one. Without it, 'HRV rose, therefore
# visceral perfusion was restored' is not an inference; it is the assumption
# kG*wV = cH*wC, written as if it had been measured.
#
# A second, weaker claim IS identifiable from H alone: the TIMING of S.
#    time of maximum drive from the cardiac trace: 230.0 s
#    (true value by construction: 230.0 s)
#    So the cardiac channel can time the response and order two conditions,
#    which is what the Chapter 8 protocol asks of it, and no more.

四组参数把 cHwCc_H w_C 固定在 0.040,因此把观测到的心脏曲线重现到浮点误差量级(残差 0.00e+00)——它们是同一个观测。而它们蕴含的平均内脏传导度从 0.506 到 0.884(1.7 倍),最低值从 0.000 到 0.426。

结构不可识别性。左:四组参数产生的心脏观测曲线完全重合(四条线叠在一起,故只看到一条)。右:同样四组参数蕴含的内脏传导度轨迹,彼此相差很远。任何长度、任何精度的心脏数据都无法在这四者之间作出选择。

这不是「数据不够」的问题,也不是「噪声太大」的问题。它是结构性的:无论记录多长、多精确的心脏数据,都无法在这四组之间作出选择,因为它们预测的心脏数据完全相同。所以

「心率变异度上升了,所以内脏灌注恢复了」不是一个推论,而是把假设 kGwV=cHwCk_G w_V = c_H w_C 写成了测量。

程序的最后一段给出心脏通道能识别的东西:反应的时点。从心脏曲线读出的最大驱动时刻(230.0 秒)与构造的真值完全一致。所以心脏通道可以给反应计时、可以给两个条件排序,而这正是第八章对它的全部要求。

注

这不是说心率变异度没用。 它是一个有标准、有参考值、有临床意义的测量 (Task Force of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology 1996; Shaffer and Ginsberg 2017)。本节说的是它不能替代另一个系统的测量。要打破上面的平局,需要与心脏观测同时进行的内脏床区的直接测量;在没有这种测量时,关于内脏的结论就只能以假设的形式陈述,并明确标为假设。

假说 · 尚未验证

作者的假说(尚未确立): 在同一批受试者中同时测量心脏迷走指标与外周灌注指标时,二者在个体内的时间相关性低(例如 |r|<0.3|r| < 0.3),因此前者不能作为后者的代理指标。什么样的结果能证明它是错的: 在事先登记的研究中同时记录两类指标,若个体内相关的汇总估计的 95% 置信区间下限超过 0.5,并在独立样本中重现,则本假说错。

协方差分析与基线

在进入个案证据之前,先把分析方法定下来,因为它影响第八章的全部数值。

证据

为什么用协方差分析。 Vickers 与 Altman 比较了处理有基线与随访测量的受控试验的几种方法,指出以基线为协变量的协方差分析优于比较变化量或比较终点值 (Vickers and Altman 2001)。Senn 重新讨论了「变化量」与协方差分析的关系,并说明了把基线当协变量在随机化试验中的正当性 (Senn 2006)。Barnett、van der Pols 与 Dobson 解释了回归到均值的来源与处理方法 (Barnett, Pols, and Dobson 2005)。

这三点在本书里有一个共同的后果:任何「练习前后比较」的分析都不可解释,因为进入练习的时点与回归到均值的方向相关。下一节把这件事量化。

脱敏个案能支持什么

原说法常以个案支持:某人练习若干周后症状减轻。本节用一个完全没有治疗效应的生成器模拟这种证据形式。生成器里有三样东西,全是结构性的:序列自相关(AR(1),ϕ=0.6\phi = 0.6)、在状态较差时进入练习(选择效应,阈值 +1.2 个标准差)、以及向总体均值的自发漂移。

推导

# Chapter 6. The desensitisation case analysis, and why a single case cannot
# carry it. A synthetic AB design: 14 daily baseline ratings, then 28 ratings
# "on practice". There is NO treatment effect anywhere in the generator. What
# there is: (a) an autocorrelated series (AR(1), phi = 0.6), (b) entry at a
# high point, because people start practising when things are bad -- the
# selection that produces regression to the mean, and (c) a slow spontaneous
# drift toward the population mean. The program reports how often the naive
# comparison of phase means "finds" an improvement, and what the correct tests
# do. Synthetic data; fixed seed; standard library only.
import math
import random
from statistics import NormalDist

N_A, N_B, PHI, SD = 14, 28, 0.6, 1.0
# population mean rating; entry when the level exceeds +1.2 SD
MU, ENTRY = 5.0, 1.2

def one_case(rng, drift_per_day=0.0, true_effect=0.0):
    # burn in until the series is above the entry threshold, then start the
    # baseline
    x = rng.gauss(0, SD)
    for _ in range(200):
        x = PHI * x + rng.gauss(0, SD * math.sqrt(1 - PHI ** 2))
    tries = 0
    while x < ENTRY * SD and tries < 10000:
        x = PHI * x + rng.gauss(0, SD * math.sqrt(1 - PHI ** 2))
        tries += 1
    a, b = [], []
    for i in range(N_A):
        a.append(MU + x)
        x = PHI * x + rng.gauss(0, SD * math.sqrt(1 - PHI ** 2))
    for i in range(N_B):
        b.append(MU + x - true_effect + drift_per_day * i)
        x = PHI * x + rng.gauss(0, SD * math.sqrt(1 - PHI ** 2))
    return a, b

def mean(v):
    return math.fsum(v) / len(v)

def naive_t(a, b):
    """Independent-samples t, the comparison a case report usually makes."""
    na, nb = len(a), len(b)
    ma, mb = mean(a), mean(b)
    va = math.fsum((x - ma) ** 2 for x in a) / (na - 1)
    vb = math.fsum((x - mb) ** 2 for x in b) / (nb - 1)
    sp = ((na - 1) * va + (nb - 1) * vb) / (na + nb - 2)
    se = math.sqrt(sp * (1 / na + 1 / nb))
    return (ma - mb) / se if se > 0 else 0.0

def eff_n(n, phi):
    """Effective sample size of an AR(1) mean; the naive t ignores this."""
    return n * (1 - phi) / (1 + phi)

if __name__ == "__main__":
    rng = random.Random(1979)
    REPS = 4000
    T_CRIT = 2.021                                 # t, 40 df, two-sided 0.05
    hits = drops = 0
    diffs = []
    for _ in range(REPS):
        a, b = one_case(rng)
        d = mean(a) - mean(b)
        diffs.append(d)
        t = naive_t(a, b)
        hits += t > T_CRIT
        drops += d > 0
    print(
        "NO treatment effect in the generator; AR(1) phi = {:.1f}; entry"
        " above +{:.1f} SD.".format(PHI, ENTRY)
    )
    print(
        "   phase-A mean minus phase-B mean, average over {} cases:"
        " {:+.3f} SD".format(REPS, mean(diffs))
    )
    print(
        "   cases showing ANY improvement:            {:.1%}".format(
            drops / REPS
        )
    )
    print(
        "   cases 'significant' by the naive t test:  {:.1%}  (nominal 2.5%)"
        .format(hits / REPS)
    )

    print(
        "\nThe apparent improvement is {:+.3f} SD on average with no effect"
        " present,".format(mean(diffs))
    )
    print(
        "and the naive test declares a benefit in {:.0%} of cases. Two"
        " causes, both".format(hits / REPS)
    )
    print(
        "structural: entry at a high point guarantees the baseline is"
        " above the"
    )
    print(
        "person's own mean, and autocorrelation means the {} baseline"
        " points are".format(N_A)
    )
    print(
        "worth only about {:.1f} independent observations (and the {}"
        " practice points".format(eff_n(N_A, PHI), N_B)
    )
    print("about {:.1f}), so the naive standard error is far too small."
          .format(eff_n(N_B, PHI)))

    print("\nWhat the same design finds when there IS an effect of 0.5 SD:")
    rng2 = random.Random(1980)
    hits2 = sum(naive_t(*one_case(rng2, true_effect=0.5)) > T_CRIT
                for _ in range(REPS))
    tpr, fpr = hits2 / REPS, hits / REPS
    print(
        "   'significant' in {:.1%} of cases, against {:.1%} with no effect"
        " at all.".format(tpr, fpr)
    )
    print(
        "   So a 'significant' single case is about {:.1f} times more likely"
        " when a".format(tpr / fpr)
    )
    print(
        "   0.5 SD effect is present than when nothing is: a likelihood"
        " ratio of"
    )
    print(
        "   about {:.1f}, which on any reading is weak evidence. The design"
        " does not".format(tpr / fpr)
    )
    print(
        "   separate a real half-SD effect from the structural artefacts"
        " above."
    )

    print(
        "\nWhat a single case CAN establish, and what the Chapter 8 checklist"
        " asks"
    )
    print(
        "for instead: a randomised multiple-baseline or ABAB design with the"
        " phase"
    )
    print(
        "change times decided in advance by someone else, a pre-specified"
        " outcome,"
    )
    print(
        "and a randomisation test that respects the autocorrelation. The"
        " program"
    )
    print("below runs that randomisation test on one series with no effect.")
    rng3 = random.Random(1981)
    a, b = one_case(rng3)
    series = a + b
    obs = mean(a) - mean(b)
    # randomisation test over admissible change points, 8 .. 34
    stats = []
    for cp in range(8, len(series) - 7):
        stats.append(mean(series[:cp]) - mean(series[cp:]))
    p = sum(1 for s in stats if s >= obs) / len(stats)
    print(
        "   observed A-minus-B difference: {:+.3f}; randomisation p = {:.3f}"
        .format(obs, p)
    )
    print(
        "   over {} admissible change points. A p value near 1/{} is the"
        " floor this".format(len(stats), len(stats))
    )
    print(
        "   test can reach, which is itself a limit on what one case can"
        " show."
    )
# NO treatment effect in the generator; AR(1) phi = 0.6; entry above +1.2 SD.
#    phase-A mean minus phase-B mean, average over 4000 cases: +0.294 SD
#    cases showing ANY improvement:            69.4%
#    cases 'significant' by the naive t test:  28.5%  (nominal 2.5%)
#
# The apparent improvement is +0.294 SD on average with no effect present,
# and the naive test declares a benefit in 28% of cases. Two causes, both
# structural: entry at a high point guarantees the baseline is above the
# person's own mean, and autocorrelation means the 14 baseline points are
# worth only about 3.5 independent observations (and the 28 practice points
# about 7.0), so the naive standard error is far too small.
#
# What the same design finds when there IS an effect of 0.5 SD:
#    'significant' in 61.6% of cases, against 28.5% with no effect at all.
#    So a 'significant' single case is about 2.2 times more likely when a
#    0.5 SD effect is present than when nothing is: a likelihood ratio of
#    about 2.2, which on any reading is weak evidence. The design does not
#    separate a real half-SD effect from the structural artefacts above.
#
# What a single case CAN establish, and what the Chapter 8 checklist asks
# for instead: a randomised multiple-baseline or ABAB design with the phase
# change times decided in advance by someone else, a pre-specified outcome,
# and a randomisation test that respects the autocorrelation. The program
# below runs that randomisation test on one series with no effect.
#    observed A-minus-B difference: +0.913; randomisation p = 0.704
#    over 27 admissible change points. A p value near 1/27 is the floor this
#    test can reach, which is itself a limit on what one case can show.

结果是:在没有任何治疗效应的情况下,基线相位均值减去练习相位均值平均为 +0.294 个标准差,69.4% 的个案出现「改善」,而朴素的两样本 t 检验在 28.5% 的个案中宣称获益(名义水平 2.5%)。原因有两个,都是结构性的:在高点进入保证了基线高于本人的均值;而自相关使 14 个基线点只值约 3.5 个独立观测、28 个练习点只值约 7.0 个。

那么当真的有效应时呢?真实效应 0.5 个标准差时,该比例为 61.6%。所以「显著的个案」在有半个标准差效应时出现的概率只是无效应时的约 2.2 倍——似然比 2.2,按任何标准都是弱证据。

注

个案能做什么。 这不是说个案无用。Kazdin 对单个案设计有系统的论述 (Kazdin 2011),Edgington 与 Onghena 给出了随机化检验的方法 (Edgington and Onghena 2007)。可用的做法是:随机化的多基线或 ABAB 设计,相位变更的时点由他人事先决定,结局事先指定,用尊重自相关的随机化检验分析。程序最后对一个无效应序列跑了这样的随机化检验:观测到的 A 减 B 差为 +0.913,而随机化 p=0.704p = 0.704(27 个可容许的变更点)。注意 1/271/27 是这个检验能达到的下限,这本身就是一个人的数据所能给出的证据的上界。

本章小结

把链条接成三状态模型后,模型内部「降低高阶精度降低防御驱动、收回传导度」成立,但这是算术——π\pi 乘在唯一抬高 SS 的项上——不是关于身体的发现。该模型有三处已知缺陷,其中最要紧的是它无法表示第一章的情形;承认这一点的变体给出一个可检验的边界条件:常驻警觉先验超过约 0.3 时,降低 π\pi 由有益转为有害。结构可识别性给出本章的主要结果:四组参数把同一条心脏观测曲线重现到浮点误差量级,而蕴含的平均内脏传导度从 0.506 到 0.884、最低值从 0.000 到 0.426,所以「心率变异度上升故内脏灌注恢复」是把假设写成了测量;心脏通道能识别的只是反应的时点。最后,在完全无效应的生成器下,AB 个案设计由于进入时点的选择与自相关,平均给出 +0.294 个标准差的「改善」,朴素检验在 28.5% 的个案中宣称获益,而真实 0.5 个标准差效应下为 61.6%,似然比约 2.2。

参考文献

Barnett, Adrian G., Jolieke C. van der Pols, and Annette J. Dobson. 2005. “Regression to the Mean: What It Is and How to Deal with It.” International Journal of Epidemiology 34 (1): 215–20. https://doi.org/10.1093/ije/dyh299.
Edgington, Eugene S., and Patrick Onghena. 2007. Randomization Tests. 4th ed. Boca Raton: Chapman; Hall/CRC.
Kazdin, Alan E. 2011. Single-Case Research Designs: Methods for Clinical and Applied Settings. 2nd ed. New York: Oxford University Press.
Morrison, Shaun F. 2001. “Differential Control of Sympathetic Outflow.” American Journal of Physiology: Regulatory, Integrative and Comparative Physiology 281 (3): R683–98. https://doi.org/10.1152/ajpregu.2001.281.3.R683.
Senn, Stephen. 2006. “Change from Baseline and Analysis of Covariance Revisited.” Statistics in Medicine 25 (24): 4334–44. https://doi.org/10.1002/sim.2682.
Shaffer, Fred, and J. P. Ginsberg. 2017. “An Overview of Heart Rate Variability Metrics and Norms.” Frontiers in Public Health 5: 258. https://doi.org/10.3389/fpubh.2017.00258.
Task Force of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology. 1996. “Heart Rate Variability: Standards of Measurement, Physiological Interpretation, and Clinical Use.” Circulation 93 (5): 1043–65. https://doi.org/10.1161/01.CIR.93.5.1043.
Vickers, Andrew J., and Douglas G. Altman. 2001. “Analysing Controlled Trials with Baseline and Follow up Measurements.” BMJ 323 (7321): 1123–24. https://doi.org/10.1136/bmj.323.7321.1123.

Fan, Y. (2026). Interoceptive Decoupling, Salience-Network Switching and Autonomic Visceral Restoration: From Precision-Weighting in the Default Mode Network and the Redistribution of Vagal Tone to the Neurophenomenology of a Non-Judgmental Imagery Stream, with a Textual Examination of neiguan in the Huangdi Neijing [内感受解离、突显网络切换与内脏自主神经修复] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24020418 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).