5D Library › 亚知觉色胺药代动力学、5-HT2A 受体脱敏与神经可塑性窗口的生物数学模型 · English

第八章 实践与预注册检验

注

本章是研究方案与文献阅读方法,不是使用指南。 核查表帮助读者在读到一项说法时把问题问清楚。试验方案是供独立研究机构参考的设计草案,只能在取得所需监管批准的司法辖区内、由有资质的团队实施,并须经伦理委员会审查与试验注册。本章不给出任何剂量、来源、途径或疗程;方案中的暴露水平一律写为「由该司法辖区已获批准的研究方案所规定的水平」,由实施机构按其监管批准填入。本书不建议任何人使用它所讨论的任何物质。

第一部分 四步核查表

前七章可以浓缩为一份四步核查表,用于阅读任何「某种亚知觉暴露带来某种好处」的说法。

  1. 链条与方向。 说法有几个环节?有没有哪一环的方向与它所依据的数学相反(第二章的饱和削平峰值、第四章的累计效应上升)?「亚知觉」指的是主观阈、测量检出限,还是一个假定的效应阈(第一章)?
  2. 参数。 说法所依赖的参数在人体内被测量过吗,还是只有动物高暴露的数据(第二、四章)?模型的结论是对参数的某个取值成立,还是对整个合理范围成立?
  3. 测量。 所用的指标与所主张的结论是同一件事吗?指标对噪声的敏感度与对所求效应的敏感度相比如何(第五章:1 比特的熵升高只需 4.6% 的额外噪声方差)?窗口长度与数据质量在条件之间匹配吗?
  4. 设盲与预期。 有安慰剂对照吗?设盲完整性被测量了吗?若未测量,在合理的猜中率下,纯预期能解释多大的观测差别(第三章:40% 猜中率加 0.5 个标准差的预期效应即可解释 0.20 个标准差的差别)?

把核查表用于「亚知觉微剂量」的通行说法,结果如下表。

步骤 通行说法的情况 能支持的结论
链条与方向 六环;第二环与第四环方向相反;「亚知觉」三义混用 作为机制陈述不成立;作为「值得做一次像样的试验」的理由成立
参数 关键参数(乙醯基前药的 KmK_m、人体 5-HT2A 恢复半衰期 τR\tau_R、ke0k_{e0})均未在人体内测量 模型只能给出条件式结论;具体的给药日程无法从药理学推出
测量 连接熵与噪声不可区分;短窗口偏差与效应同量级 熵类指标只能作探索性结局,且须匹配窗口与回归头动
设盲与预期 安慰剂对照研究大体零结果;开放标签调查报告大量获益 现有的正面证据与纯预期解释相容;需要测量设盲的试验

第二部分 预注册的双盲随机对照试验

设计选择

六点,每一点来自前面某一章。

第一,只设一个主要比较:活性条件对匹配安慰剂,双盲。暴露水平由实施机构按其监管批准填入,本书不指定。

第二,只设一个主要结局:第 6 周预先指定的量表分数,以基线分数与中心为协变量作协方差分析 (Vickers and Altman 2001; Senn 2006)。连接熵类指标只作探索性结局,理由见第五章。不作任何药代动力学主张:本方案不测浓度,因此不报告动力学参数(第二、三章)。

第三,设盲完整性必须被测量并供能:用 Bang 设盲指数 (Bang, Ni, and Davis 2004) 与可信度/预期问卷 (Devilly and Borkovec 2000)。缺乏参与者设盲会造成偏倚,这有系统综述的证据 (Hróbjartsson et al. 2014);本方案的做法是测量它。

第四,预先指定「按信念分层」的对照:在相信自己接受了活性条件的参与者中,比较实际接受与实际未接受者。第三章与程序 p11 说明这是把药理与预期分开的分析。

第五,不承诺为这个分层对照供能。这是一个诚实的设计让步:程序 p14 说明,设盲一旦破坏,这个分析需要两倍以上的样本。把它列为预先指定的关键次要分析、报告其区间,比假装它被供能更可靠。

第六,由独立机构在合法辖区实施:方案、分析代码与判定规则在第一例入组前公开登记;作者不参与招募、干预、评估与分析。

方案

要素 规定
法律前提 只在该物质可合法用于此类研究的司法辖区实施,并取得所需的全部监管批准与管制物质许可。缺少该批准即构成本章判定规则中的前置条件失败,结果写「不可解释」
实施与登记 由与作者无关的研究团队实施;伦理委员会批准;公开注册;按 SPIRIT/CONSORT 撰写与报告 (Chan et al. 2013; Schulz, Altman, and Moher 2010);不从作者的社群、学员或客户中招募
暴露 由实施机构按其监管批准规定的水平与制备方式。本书不指定、不建议、不换算任何数值
参与者 成年、健康志愿者;排除精神病性障碍或双相障碍的个人史与一级亲属家族史、当前重性精神障碍、心血管疾病、妊娠或哺乳、以及相互作用风险的合并用药;由中心的精神科与内科医师判断
随机化与设盲 中央随机系统,1:1;安慰剂在外观、气味与包装上匹配;参与者、实施者、结局评估者与统计分析人员均设盲
主要结局 第 6 周预先指定量表分数(第 6 与第 7 周两次独立评估的平均);协方差分析,协变量为基线分数与中心
设盲完整性 第 1 周与第 6 周的分组猜测与信心,计算 Bang 指数;可信度/预期问卷
关键次要分析 按信念分层的对照(预先指定,不供能,报告区间)
探索性结局 认知任务;睡眠;连接熵类指标,须(一)匹配窗口长度、(二)报告并回归头动与生理噪声 (Power et al. 2012; Ciric et al. 2017)、(三)预先指定估计量
安全性 主动采集不良事件,特别包括焦虑、惊恐、血压心率变化、睡眠紊乱与持续性知觉改变;独立的数据与安全监查委员会
等效性判读 若主要效应落在 ±SESOI 之内,按等效性框架报告为「没有值得据此改变做法的差别」(Lakens, Scheel, and Isager 2018)

自盲法设计能产生多大的表观效应

在给出样本量之前,先量化本领域现有正面证据的一个替代解释。

推导

# Chapter 8. The self-blinding design, and how much of an apparent effect it
# can create without any pharmacological effect at all. The design (Szigeti et
# al. 2021) has participants prepare their own capsules and draw them blind,
# then rate themselves. The generator here has NO drug effect: the only thing
# that moves a rating is what the participant BELIEVES they took. Belief is
# correct more often than chance whenever any fraction of people notice
# something (programme p05).
#   rating = person effect + belief * expectancy + noise
# The program reports the apparent between-arm difference, and then inverts:
# given an observed difference, what combination of correct-guess rate and
# expectancy explains it? Synthetic data; fixed seed; standard library only.
import math
import random
from statistics import NormalDist

Z = NormalDist().inv_cdf

def trial(n, p_correct, expectancy, true_effect=0.0, rng=None):
    """Half allocated to active, half to placebo; no carry-over."""
    rng = rng or random.Random(0)
    act, pla = [], []
    for i in range(n):
        for arm, bucket in ((1, act), (0, pla)):
            # belief: correct with probability p_correct, else the opposite
            belief = arm if rng.random() < p_correct else 1 - arm
            y = (true_effect * arm + expectancy * belief
                 + rng.gauss(0.0, 1.0))
            bucket.append(y)
    ma = math.fsum(act) / len(act)
    mp = math.fsum(pla) / len(pla)
    return ma - mp

if __name__ == "__main__":
    rng = random.Random(5211)
    N, REPS = 90, 400
    print(
        "Self-blinding design, {} per arm, NO pharmacological effect"
        " anywhere.".format(N)
    )
    print("Ratings in SD units. Only belief moves a rating.\n")
    print(
        "   P(correct guess)   expectancy 0.3   expectancy 0.6  "
        " expectancy 1.0"
    )
    for pc in (0.50, 0.60, 0.70, 0.80, 0.90):
        row = []
        for ex in (0.3, 0.6, 1.0):
            d = (
                math.fsum(trial(N, pc, ex, 0.0, rng) for _ in range(REPS))
                / REPS
            )
            row.append(d)
        print("   {:16.2f} {:16.3f} {:16.3f} {:16.3f}".format(pc, *row))

    print(
        "\nThe expected apparent difference is (2 p - 1) x expectancy: at"
        " chance"
    )
    print(
        "guessing it is zero, and at 70 % correct guessing with an"
        " expectancy of"
    )
    print(
        "0.6 SD it is about {:.2f} SD - the size of effect that microdosing"
        " surveys".format(0.4 * 0.6)
    )
    print("routinely report. No drug effect is needed to produce it.")

    print(
        "\nInversion: an observed difference of 0.25 SD. Which (p,"
        " expectancy)"
    )
    print("pairs explain it with no drug effect?")
    print("   P(correct guess)   required expectancy (SD)")
    for pc in (0.55, 0.60, 0.70, 0.80, 0.90, 1.00):
        req = 0.25 / (2 * pc - 1)
        print("   {:16.2f} {:26.2f}{}".format(pc, req,
              "   (implausibly large)" if req > 1.5 else ""))

    print(
        "\nWhat this means for a design. The quantity that has to be"
        " measured is"
    )
    print(
        "P(correct guess), and the trial has to be powered to DETECT a"
        " departure"
    )
    print(
        "from chance, not merely to report one. Sample size for the blind"
        " check:"
    )
    print("   true P(correct)   n per arm for 90 % power to reject P = 0.5")
    for pc in (0.60, 0.65, 0.70, 0.80):
        # one-sample proportion test against 0.5, two-sided 0.05
        n = ((Z(0.975) * 0.5 + Z(0.90) * math.sqrt(pc * (1 - pc))) /
             (pc - 0.5)) ** 2
        print("   {:15.2f} {:38d}".format(pc, math.ceil(n)))

    print(
        "\nAnd the comparison that actually isolates a pharmacological"
        " effect:"
    )
    print(
        "stratify the analysis BY BELIEF. Within the group who believed they"
        " took"
    )
    print(
        "the active capsule, compare those who actually did against those"
        " who did"
    )
    print("not. With no drug effect that contrast is zero by construction:")
    print(
        "   expectancy   P(correct)   belief-stratified contrast (true"
        " effect 0)"
    )
    for ex in (0.3, 0.6, 1.0):
        for pc in (0.7,):
            vals = []
            for _ in range(REPS):
                act_b, pla_b = [], []
                for i in range(N):
                    for arm, bucket_pair in (
                        (1, (act_b, pla_b)),
                        (0, (act_b, pla_b)),
                    ):
                        belief = arm if rng.random() < pc else 1 - arm
                        y = ex * belief + rng.gauss(0.0, 1.0)
                        if belief == 1:
                            (act_b if arm == 1 else pla_b).append(y)
                if act_b and pla_b:
                    vals.append(math.fsum(act_b) / len(act_b)
                                - math.fsum(pla_b) / len(pla_b))
            print("   {:10.1f} {:12.2f} {:38.3f}"
                  .format(ex, pc, math.fsum(vals) / len(vals)))
    print(
        "\n   The belief-stratified contrast is near zero at every"
        " expectancy, which"
    )
    print(
        "   is what makes it the informative analysis: it is immune to the"
        " expectancy"
    )
    print(
        "   term by construction. Its cost is power, because it uses only"
        " part of the"
    )
    print("   sample, and Chapter 8 budgets for that.")
# Self-blinding design, 90 per arm, NO pharmacological effect anywhere.
# Ratings in SD units. Only belief moves a rating.
#
#    P(correct guess)   expectancy 0.3   expectancy 0.6   expectancy 1.0
#                0.50           -0.002           -0.008            0.012
#                0.60            0.059            0.120            0.196
#                0.70            0.111            0.245            0.404
#                0.80            0.190            0.363            0.595
#                0.90            0.225            0.485            0.793
#
# The expected apparent difference is (2 p - 1) x expectancy: at chance
# guessing it is zero, and at 70 % correct guessing with an expectancy of
# 0.6 SD it is about 0.24 SD - the size of effect that microdosing surveys
# routinely report. No drug effect is needed to produce it.
#
# Inversion: an observed difference of 0.25 SD. Which (p, expectancy)
# pairs explain it with no drug effect?
#    P(correct guess)   required expectancy (SD)
#                0.55                       2.50   (implausibly large)
#                0.60                       1.25
#                0.70                       0.63
#                0.80                       0.42
#                0.90                       0.31
#                1.00                       0.25
#
# What this means for a design. The quantity that has to be measured is
# P(correct guess), and the trial has to be powered to DETECT a departure
# from chance, not merely to report one. Sample size for the blind check:
#    true P(correct)   n per arm for 90 % power to reject P = 0.5
#               0.60                                    259
#               0.65                                    113
#               0.70                                     62
#               0.80                                     25
#
# And the comparison that actually isolates a pharmacological effect:
# stratify the analysis BY BELIEF. Within the group who believed they took
# the active capsule, compare those who actually did against those who did
# not. With no drug effect that contrast is zero by construction:
#    expectancy   P(correct)   belief-stratified contrast (true effect 0)
#           0.3         0.70                                  0.003
#           0.6         0.70                                  0.013
#           1.0         0.70                                  0.008
#
#    The belief-stratified contrast is near zero at every expectancy, which
#    is what makes it the informative analysis: it is immune to the expectancy
#    term by construction. Its cost is power, because it uses only part of the
#    sample, and Chapter 8 budgets for that.

期望的表观差别约为 (2p−1)×(2p-1)\times 预期效应。在 70% 猜中率与 0.6 个标准差的预期效应下,表观差别约 0.24 个标准差——正是微剂量调查常报告的量级,而不需要任何药理效应。反向计算给出:要用纯预期解释 0.25 个标准差的观测差别,70% 猜中率下只需 0.63 个标准差的预期效应。

程序的最后一段给出解决办法:按信念分层的对照在任何预期水平下都接近零(0.003、0.013、0.008),因为它按构造免于预期项。这就是把它列为关键次要分析的理由。

注

这不是对 Szigeti 等人研究的批评。 恰恰相反:自盲法设计之所以能被这样分析,正因为它记录了参与者的猜测,而这是该研究的长处。Szigeti 等人自己报告了参与者在相信服用了活性物时评分更高 (Szigeti et al. 2021)。本节是把那个观察形式化,并说明一个后续试验应当据此怎样设计。

把握度预算与样本量

推导

# Chapter 8. The power budget for a trial that has to answer three questions
# at once: (1) the headline between-arm comparison, (2) whether the blind
# held, and (3) the belief-stratified contrast of p11, which is the analysis
# that isolates pharmacology from expectancy. The third is the expensive one,
# because it uses only the participants who believed they were on the active
# arm. The program lays out what each question costs so the trial is sized for
# the question it is meant to answer. Exact formulas plus one simulation;
# standard library only.
import math
import random
from statistics import NormalDist

Z = NormalDist().inv_cdf

def n_ancova(d, rho=0.55, power=0.90, alpha=0.05):
    za, zb = Z(1 - alpha / 2), Z(power)
    return 2 * (za + zb) ** 2 * (1 - rho ** 2) / d ** 2 + za ** 2 / 4

def n_blind_check(p_true, power=0.90, alpha=0.05):
    """One-sample proportion against 0.5, per arm."""
    return (
        (Z(1 - alpha / 2) * 0.5 + Z(power) * math.sqrt(p_true * (1 - p_true)))
        / (p_true - 0.5)
    ) ** 2

if __name__ == "__main__":
    SESOI = 0.30
    print(
        "Smallest effect of interest {:.2f} SD; baseline-outcome correlation"
        " 0.55;".format(SESOI)
    )
    print("two-sided 0.05; 90 % power unless stated.\n")

    print("Question 1 - headline between-arm comparison (ANCOVA):")
    n1 = n_ancova(SESOI)
    print("   n per arm = {:.1f} -> {}".format(n1, math.ceil(n1)))
    print("   at SESOI 0.20 SD: {};  at 0.40 SD: {}"
          .format(math.ceil(n_ancova(0.20)), math.ceil(n_ancova(0.40))))

    print(
        "\nQuestion 2 - did the blind hold? Power to detect a departure from"
        " chance"
    )
    print("guessing (per arm):")
    print("   true P(correct guess)   n per arm")
    for p in (0.60, 0.65, 0.70, 0.75, 0.80):
        print("   {:21.2f} {:11d}".format(p, math.ceil(n_blind_check(p))))
    print(
        "   With the {} per arm that question 1 requires, the smallest"
        " departure".format(math.ceil(n1))
    )
    lo, hi = 0.5001, 0.95
    for _ in range(60):
        mid = 0.5 * (lo + hi)
        if n_blind_check(mid) > n1:
            lo = mid
        else:
            hi = mid
    print("   detectable at 90 % power is P(correct) = {:.3f}.".format(hi))

    print(
        "\nQuestion 3 - the belief-stratified contrast. Only participants who"
    )
    print(
        "believed they were on the active arm enter it, and they split"
        " unevenly"
    )
    print("between the arms according to how well the blind held.")
    print(
        "\n   P(correct)   cells (active / placebo)   smaller cell  "
        " detectable"
    )
    print(
        "                                             as % of arm    effect"
        " (SD)"
    )
    N_PER_ARM = math.ceil(n1)
    for p in (0.50, 0.60, 0.70, 0.80):
        # of the active arm, a fraction p believe active; of the placebo arm,
        # (1-p) do
        n_act = N_PER_ARM * p
        n_pla = N_PER_ARM * (1 - p)
        # detectable difference at 90 % power with unequal cells
        se_mult = math.sqrt(1.0 / n_act + 1.0 / n_pla)
        d_det = (Z(0.975) + Z(0.90)) * se_mult * math.sqrt(1 - 0.55 ** 2)
        print("   {:10.2f} {:14.1f} / {:<8.1f} {:15.0%} {:14.3f}"
              .format(p, n_act, n_pla, min(n_act, n_pla) / N_PER_ARM, d_det))
    print(
        "   (both cells come out of a total of {} per arm; the stratified"
        " contrast".format(N_PER_ARM)
    )
    print(
        "    uses all of them, but its precision is set by the SMALLER cell)"
    )

    print(
        "\n   The cost is severe and asymmetric. A blind that holds perfectly"
    )
    print(
        "   (P = 0.50) gives the largest usable cells and the best"
        " precision; a"
    )
    print(
        "   blind that fails badly (P = 0.80) leaves only {:.0f} participants"
        " in the".format(N_PER_ARM * 0.2)
    )
    print(
        "   placebo-but-believed-active cell and inflates the detectable"
        " effect to"
    )
    print(
        "   {:.2f} SD. So a broken blind damages the very analysis meant to"
        " repair".format(
            (Z(0.975) + Z(0.90))
            * math.sqrt(1 / (N_PER_ARM * 0.8) + 1 / (N_PER_ARM * 0.2))
            * math.sqrt(1 - 0.55**2)
        )
    )
    print(
        "   it. This is the central design tension of the chapter, and it is"
        " why the"
    )
    print(
        "   protocol inflates the sample rather than relying on the"
        " stratified"
    )
    print("   analysis to rescue a failed blind.")

    print(
        "\nInflation needed so that the belief-stratified contrast can"
        " detect the"
    )
    print(
        "SESOI of {:.2f} SD, as a function of how well the blind holds:"
        .format(SESOI)
    )
    print("   P(correct)   n per arm required   inflation over question 1")
    for p in (0.55, 0.60, 0.65, 0.70):
        # solve (z+z) sqrt(1/(Np)+1/(N(1-p))) sqrt(1-rho^2) = SESOI
        k = ((Z(0.975) + Z(0.90)) * math.sqrt(1 - 0.55 ** 2) / SESOI) ** 2
        need = k * (1.0 / p + 1.0 / (1.0 - p))
        print("   {:10.2f} {:19d} {:24.2f} x"
              .format(p, math.ceil(need), need / n1))
    print(
        "\n   Even a mild departure from chance guessing multiplies the"
        " required"
    )
    print(
        "   sample several-fold. The protocol therefore treats the"
        " belief-stratified"
    )
    print(
        "   contrast as a pre-specified KEY SECONDARY analysis, reported"
        " with its"
    )
    print("   interval, and does not promise to power it.")
# Smallest effect of interest 0.30 SD; baseline-outcome correlation 0.55;
# two-sided 0.05; 90 % power unless stated.
#
# Question 1 - headline between-arm comparison (ANCOVA):
#    n per arm = 163.8 -> 164
#    at SESOI 0.20 SD: 368;  at 0.40 SD: 93
#
# Question 2 - did the blind hold? Power to detect a departure from chance
# guessing (per arm):
#    true P(correct guess)   n per arm
#                     0.60         259
#                     0.65         113
#                     0.70          62
#                     0.75          38
#                     0.80          25
#    With the 164 per arm that question 1 requires, the smallest departure
#    detectable at 90 % power is P(correct) = 0.625.
#
# Question 3 - the belief-stratified contrast. Only participants who
# believed they were on the active arm enter it, and they split unevenly
# between the arms according to how well the blind held.
#
#    P(correct)   cells (active / placebo)   smaller cell   detectable
#                                              as % of arm    effect (SD)
#          0.50           82.0 / 82.0                 50%          0.423
#          0.60           98.4 / 65.6                 40%          0.432
#          0.70          114.8 / 49.2                 30%          0.461
#          0.80          131.2 / 32.8                 20%          0.528
#    (both cells come out of a total of 164 per arm; the stratified contrast
#     uses all of them, but its precision is set by the SMALLER cell)
#
#    The cost is severe and asymmetric. A blind that holds perfectly
#    (P = 0.50) gives the largest usable cells and the best precision; a
#    blind that fails badly (P = 0.80) leaves only 33 participants in the
#    placebo-but-believed-active cell and inflates the detectable effect to
#    0.53 SD. So a broken blind damages the very analysis meant to repair
#    it. This is the central design tension of the chapter, and it is why the
#    protocol inflates the sample rather than relying on the stratified
#    analysis to rescue a failed blind.
#
# Inflation needed so that the belief-stratified contrast can detect the
# SESOI of 0.30 SD, as a function of how well the blind holds:
#    P(correct)   n per arm required   inflation over question 1
#          0.55                 330                     2.01 x
#          0.60                 340                     2.07 x
#          0.65                 358                     2.18 x
#          0.70                 388                     2.37 x
#
#    Even a mild departure from chance guessing multiplies the required
#    sample several-fold. The protocol therefore treats the belief-stratified
#    contrast as a pre-specified KEY SECONDARY analysis, reported with its
#    interval, and does not promise to power it.

三个问题的代价不同。主要比较在 0.30 个标准差的最小关注效应下需每臂 164 例;设盲检查在该样本下只能检出 0.625 以上的猜中率;而按信念分层的对照在设盲破坏时精度由较小的格决定——80% 猜中率时安慰剂—但相信活性的格只剩 33 人,可检出效应膨胀到 0.53 个标准差。要让分层对照也能检出 0.30 个标准差,即使只有 0.55 的猜中率也需约 2.01 倍的样本。

推导

# Chapter 8. Sample size for the pre-registered trial. Single primary
# comparison: the active condition against a matched placebo, double blind.
# Single primary outcome: the pre-specified wellbeing/affect score at week 6
# by ANCOVA (baseline, site). Planning
# assumptions: SESOI 0.30 SD, baseline-outcome correlation 0.55,
# single-occasion reliability 0.80, outcome = mean of 2 occasions, two-sided
# 0.05, 90 % power. Normal theory first, then a simulation with integer
# scoring, a ceiling, and the drop-out pattern that long daily-dosing
# protocols actually produce.
import math
import random
from statistics import NormalDist

Z = NormalDist().inv_cdf
D, RHO, REL1, K_OCC = 0.30, 0.55, 0.80, 2
ITEMS, ITEM_MAX = 14, 5           # 14 items scored 0-5: range 0-70
CEILING = ITEMS * ITEM_MAX
TRUE_MEAN, TRUE_SD = 42.0, 9.5

def reliability(k):
    return k * REL1 / (1 + (k - 1) * REL1)

def n_formula(d=D, rho_xx=1.0, rho=RHO, power=0.90, alpha=0.05):
    d_obs = d * math.sqrt(rho_xx)
    za, zb = Z(1 - alpha / 2), Z(power)
    return 2 * (za + zb) ** 2 * (1 - rho ** 2) / d_obs ** 2 + za ** 2 / 4

def invert3(S):
    a, b, c = S[0]
    d, e, f = S[1]
    g, h, i = S[2]
    det = a * (e * i - f * h) - b * (d * i - f * g) + c * (d * h - e * g)
    return [
        [(e * i - f * h) / det, (c * h - b * i) / det, (b * f - c * e) / det],
        [(f * g - d * i) / det, (a * i - c * g) / det, (c * d - a * f) / det],
        [(d * h - e * g) / det, (b * g - a * h) / det, (a * e - b * d) / det],
    ]

def one_trial(rng, n, delta_sd):
    rows = []
    delta = delta_sd * TRUE_SD
    err_sd = TRUE_SD * math.sqrt(1 - RHO ** 2)
    meas_sd = TRUE_SD * math.sqrt((1 - REL1) / REL1 / K_OCC)
    for arm in (1, 0):
        for _ in range(n):
            lat_b = rng.gauss(TRUE_MEAN, TRUE_SD)
            b = min(CEILING, max(0, round(lat_b + rng.gauss(0, meas_sd))))
            lat_y = (TRUE_MEAN + RHO * (lat_b - TRUE_MEAN) + delta * arm
                     + rng.gauss(0, err_sd))
            y = min(CEILING, max(0, round(lat_y + rng.gauss(0, meas_sd))))
            rows.append((arm, b, y))
    n2 = len(rows)
    mb = math.fsum(r[1] for r in rows) / n2
    X = [(1.0, float(r[0]), r[1] - mb) for r in rows]
    Y = [float(r[2]) for r in rows]
    S = [
        [math.fsum(x[i] * x[j] for x in X) for j in range(3)]
        for i in range(3)
    ]
    v = [math.fsum(x[i] * y for x, y in zip(X, Y)) for i in range(3)]
    inv = invert3(S)
    beta = [math.fsum(inv[i][j] * v[j] for j in range(3)) for i in range(3)]
    res = [
        y - math.fsum(bb * xx for bb, xx in zip(beta, x))
        for x, y in zip(X, Y)
    ]
    s2 = math.fsum(r * r for r in res) / (n2 - 3)
    df = n2 - 3
    z = Z(0.975)
    tcrit = z + (z ** 3 + z) / (4 * df)
    at_ceiling = sum(1 for r in rows if r[0] == 1 and r[2] == CEILING) / n
    return beta[1] / math.sqrt(s2 * inv[1][1]) > tcrit, at_ceiling

if __name__ == "__main__":
    rel = reliability(K_OCC)
    print(
        "outcome = mean of {} occasions; single-occasion reliability {:.2f}"
        " -> {:.3f}".format(K_OCC, REL1, rel)
    )
    n0 = n_formula(rho_xx=rel)
    print("normal theory: n = {:.3f} -> {} per arm (attenuated d = {:.3f})"
          .format(n0, math.ceil(n0), D * math.sqrt(rel)))

    rng = random.Random(5215)
    print(
        "\nsimulation with integer scores and a ceiling at {}:".format(
            CEILING
        )
    )
    chosen = None
    for n in sorted({math.ceil(n0), 195, 210, 225}):
        res = [one_trial(rng, n, D) for _ in range(1500)]
        pw = sum(r[0] for r in res) / len(res)
        se = math.sqrt(pw * (1 - pw) / len(res))
        ce = math.fsum(r[1] for r in res) / len(res)
        print(
            "   n = {:3d}/arm: power {:.3f} (s.e. {:.3f}); active arm at"
            " ceiling {:.1%}".format(n, pw, se, ce)
        )
        if chosen is None and pw - 2 * se >= 0.90:
            chosen = n
    if chosen is None:
        chosen = 210
    t1 = sum(one_trial(rng, chosen, 0.0)[0] for _ in range(1500)) / 1500
    print("   smallest n with power - 2 s.e. >= 0.90: {}".format(chosen))
    print(
        "   one-sided error rate with no effect: {:.3f} (nominal 0.025)"
        .format(t1)
    )

    print(
        "\nDrop-out in a six-week daily-dosing protocol is higher than in a"
    )
    print("questionnaire-only study; the plan allows 20 %:")
    plan = math.ceil(chosen / 0.80)
    print(
        "   {} analysable per arm; 20 % loss -> {} randomised per arm, {}"
        " in all".format(chosen, plan, 2 * plan)
    )

    print(
        "\nsensitivity, analysable per arm (normal theory, reliability"
        " {:.2f}):".format(rel)
    )
    print("   SESOI \\ rho        0.35    0.55    0.75")
    for d in (0.20, 0.30, 0.40):
        print("   {:4.2f} SD        ".format(d) + "".join(
            "{:8d}".format(math.ceil(n_formula(d=d, rho_xx=rel, rho=r)))
            for r in (0.35, 0.55, 0.75)))

    print(
        "\nThe trial is powered for the PRIMARY comparison only."
        " Programme p14"
    )
    print(
        "showed that the belief-stratified contrast, which is the analysis"
        " that"
    )
    print(
        "separates pharmacology from expectancy, would need several times"
        " this"
    )
    print(
        "sample if the blind does not hold; it is therefore a"
        " pre-specified key"
    )
    print(
        "secondary reported with its interval and is not powered. The blind-"
    )
    print(
        "integrity check at {} per arm can detect a correct-guess rate of"
        " about".format(chosen)
    )
    lo, hi = 0.5001, 0.95
    for _ in range(60):
        mid = 0.5 * (lo + hi)
        k = (
            (Z(0.975) * 0.5 + Z(0.90) * math.sqrt(mid * (1 - mid)))
            / (mid - 0.5)
        ) ** 2
        if k > chosen:
            lo = mid
        else:
            hi = mid
    print("   {:.3f}, which is the resolution the trial has on its own blind."
          .format(hi))
# outcome = mean of 2 occasions; single-occasion reliability 0.80 -> 0.889
# normal theory: n = 184.184 -> 185 per arm (attenuated d = 0.283)
#
# simulation with integer scores and a ceiling at 70:
#    n = 185/arm: power 0.881 (s.e. 0.008); active arm at ceiling 0.7%
#    n = 195/arm: power 0.892 (s.e. 0.008); active arm at ceiling 0.7%
#    n = 210/arm: power 0.913 (s.e. 0.007); active arm at ceiling 0.7%
#    n = 225/arm: power 0.928 (s.e. 0.007); active arm at ceiling 0.7%
#    smallest n with power - 2 s.e. >= 0.90: 225
#    one-sided error rate with no effect: 0.025 (nominal 0.025)
#
# Drop-out in a six-week daily-dosing protocol is higher than in a
# questionnaire-only study; the plan allows 20 %:
#    225 analysable per arm; 20 % loss -> 282 randomised per arm, 564 in all
#
# sensitivity, analysable per arm (normal theory, reliability 0.89):
#    SESOI \ rho        0.35    0.55    0.75
#    0.20 SD             520     414     260
#    0.30 SD             232     185     116
#    0.40 SD             131     105      66
#
# The trial is powered for the PRIMARY comparison only. Programme p14
# showed that the belief-stratified contrast, which is the analysis that
# separates pharmacology from expectancy, would need several times this
# sample if the blind does not hold; it is therefore a pre-specified key
# secondary reported with its interval and is not powered. The blind-
# integrity check at 225 per arm can detect a correct-guess rate of about
#    0.607, which is the resolution the trial has on its own blind.

正态理论给出每臂 185 例;模拟加入整数计分与天花板后,按「把握度减两个标准误不低于 0.90」取每臂 225 例可分析。六周每日给药方案的失访高于问卷研究,按 20% 计,每臂随机 282 例,两臂共 564 例。

按第 33 本第六章的结果,「90% 把握度」指的是在真实效应恰好等于 0.30 个标准差时,得出「活性条件较好」这类结论的概率约为 90%;最强的结论「至少高出 0.30 个标准差」只有约 2.5% 的机会出现。

判定规则

推导

# Chapter 8. Decision rule, fixed before unblinding. Primary comparison:
# active against matched placebo, double blind. D = adjusted difference in the
# week-6 primary score in SD units (ANCOVA: baseline, site), with its 95 % CI.
# POSITIVE D favours the active condition. SESOI 0.30 SD.
# Gates, checked first: more than 20 % of randomised participants without a
# primary outcome; allocation revealed to outcome assessors before database
# lock; the trial not conducted under the required regulatory authorisation.
# Pre-specified qualifiers: the blind-integrity check (a correct-guess rate
# above the pre-set bound means expectancy is a live explanation, and
# programme p11 quantifies how much of D it could account for) and the
# belief-stratified contrast of p11, which is reported whenever the blind did
# not hold. Any adverse-event excess is reported first.
SESOI = 0.30

def expectancy_share(d, p_correct, expectancy_sd=0.6):
    """Share of D that a correct-guess rate p could account for, per programme
    p11:
    an expectancy effect contributes about (2p - 1) * expectancy_sd."""
    if d == 0:
        return 0.0
    return max(0.0, (2.0 * p_correct - 1.0) * expectancy_sd) / abs(d)

def verdict(d, lo, hi, missing=0.10, p_correct=0.52, belief_contrast=None,
            revealed=False, unauthorised=False, ae_excess=False):
    pre = "ADVERSE-EVENT SIGNAL REPORTED FIRST;\n" if ae_excess else ""
    for bad, why in (
        (unauthorised, "conducted without the required authorisation"),
        (
            missing > 0.20,
            "{:.0%} without primary outcome (limit 20 %)".format(missing),
        ),
        (revealed, "allocation revealed to assessors before lock"),
    ):
        if bad:
            return "{}D = {:+.2f}: NOT INTERPRETABLE - {}".format(pre, d, why)
    if lo >= SESOI:
        v = "benefit of at least the SESOI"
    elif lo > 0:
        v = "benefit, but not shown to reach the SESOI"
    elif hi < 0:
        v = "placebo did better"
    elif lo >= -SESOI and hi <= SESOI:
        v = "no difference worth acting on"
    else:
        v = "unclear: the interval is too wide to decide"
    notes = []
    if p_correct > 0.60:
        share = expectancy_share(d, p_correct)
        notes.append(
            "blind did not hold (correct guesses {:.2f});"
            "\n      expectancy alone could account for about {:.0%} of D"
            .format(p_correct, min(share, 1.0))
        )
        if belief_contrast is None:
            notes.append("belief-stratified contrast NOT reported, so the"
                         "\n      pharmacological effect is not isolated")
        else:
            notes.append("belief-stratified contrast {:+.2f} SD"
                         .format(belief_contrast))
    out = "{}D = {:+.2f} ({:+.2f} to {:+.2f}): {}".format(pre, d, lo, hi, v)
    for n in notes:
        out += "\n    qualifier: " + n
    return out

CASES = [
    # clear benefit past SESOI
    (0.44, 0.31, 0.57),
    # benefit below SESOI
    (0.18, 0.04, 0.32),
    # nothing worth acting on
    (0.03, -0.11, 0.17),
    (-0.24, -0.38, -0.10),                                 # placebo better
    (0.15, -0.25, 0.55),                                   # too wide
    # blind broke, no stratum
    (0.40, 0.26, 0.54, 0.10, 0.74),
    # blind broke, stratum ~0
    (0.40, 0.26, 0.54, 0.10, 0.74, 0.05),
    # blind broke, stratum holds
    (0.40, 0.26, 0.54, 0.10, 0.74, 0.33),
    (0.50, 0.36, 0.64, 0.28),                              # too much missing
    (0.30, 0.16, 0.44, 0.10, 0.52, None, False, True),     # unauthorised
    (0.05, -0.09, 0.19, 0.10, 0.52, None, False, False, True),   # AE signal
]

if __name__ == "__main__":
    for c in CASES:
        print(verdict(*c))
    print(
        "\nNote the three 'blind broke' cases. They share an identical"
        " primary"
    )
    print(
        "result and reach different conclusions, because what distinguishes"
        " them is"
    )
    print(
        "the belief-stratified contrast: near zero means expectancy"
        " explains the"
    )
    print(
        "primary difference, while a contrast close to D means it does not."
        " This is"
    )
    print(
        "the analysis programme p11 was built to justify and programme p14"
        " priced."
    )
    print(
        "\nSecondary outcomes (the connectivity-entropy measures of Chapter 5"
        " with"
    )
    print(
        "their motion regressors and matched windows, cognitive tasks,"
        " sleep) are"
    )
    print(
        "reported with intervals and never change the verdict. No"
        " pharmacokinetic"
    )
    print("claim is made at all: per Chapters 2 and 3 the trial measures no")
    print("concentrations and therefore reports no kinetics.")
    print(
        "\nThe example numbers above are invented, to show what each branch"
        " prints."
    )
# D = +0.44 (+0.31 to +0.57): benefit of at least the SESOI
# D = +0.18 (+0.04 to +0.32): benefit, but not shown to reach the SESOI
# D = +0.03 (-0.11 to +0.17): no difference worth acting on
# D = -0.24 (-0.38 to -0.10): placebo did better
# D = +0.15 (-0.25 to +0.55): unclear: the interval is too wide to decide
# D = +0.40 (+0.26 to +0.54): benefit, but not shown to reach the SESOI
#     qualifier: blind did not hold (correct guesses 0.74);
#       expectancy alone could account for about 72% of D
#     qualifier: belief-stratified contrast NOT reported, so the
#       pharmacological effect is not isolated
# D = +0.40 (+0.26 to +0.54): benefit, but not shown to reach the SESOI
#     qualifier: blind did not hold (correct guesses 0.74);
#       expectancy alone could account for about 72% of D
#     qualifier: belief-stratified contrast +0.05 SD
# D = +0.40 (+0.26 to +0.54): benefit, but not shown to reach the SESOI
#     qualifier: blind did not hold (correct guesses 0.74);
#       expectancy alone could account for about 72% of D
#     qualifier: belief-stratified contrast +0.33 SD
# D = +0.50: NOT INTERPRETABLE - 28% without primary outcome (limit 20 %)
# D = +0.30: NOT INTERPRETABLE - conducted without the required authorisation
# ADVERSE-EVENT SIGNAL REPORTED FIRST;
# D = +0.05 (-0.09 to +0.19): no difference worth acting on
#
# Note the three 'blind broke' cases. They share an identical primary
# result and reach different conclusions, because what distinguishes them is
# the belief-stratified contrast: near zero means expectancy explains the
# primary difference, while a contrast close to D means it does not. This is
# the analysis programme p11 was built to justify and programme p14 priced.
#
# Secondary outcomes (the connectivity-entropy measures of Chapter 5 with
# their motion regressors and matched windows, cognitive tasks, sleep) are
# reported with intervals and never change the verdict. No pharmacokinetic
# claim is made at all: per Chapters 2 and 3 the trial measures no
# concentrations and therefore reports no kinetics.
#
# The example numbers above are invented, to show what each branch prints.

前置条件有三项:缺少所需监管批准、超过 20% 的随机参与者没有主要结局、结果锁定前分组向评估者泄露。任何一项不满足,无论主要结局如何都写「不可解释」。注意三个「设盲破坏」的例子:它们的主要结果完全相同而结论不同,区别在于按信念分层的对照——接近零意味着预期解释了主要差别,接近 DD 意味着没有。这正是程序 p11 所论证、p14 所定价的那个分析。

这项试验不能说明什么

第一,它只检验一个暴露水平、一种日程、六周的安排。第二,它不测浓度,因此对第二、三章的药代动力学问题一无所答——那需要单独的、取得批准的药代动力学研究。第三,它不检验第六章的「窗口需要输入」假说,那需要本书第六章末所述的 2×2 因子设计。第四,参与者有相当比例会猜中分组(第三章),所以预期的贡献只能被部分识别。第五,结论不能推及被排除的人群,而被排除的理由正是安全性。第六,如果结果是「没有值得据此改变做法的差别」,那就是本试验的结论,不能用探索性结局或亚组替代。

伦理与利益冲突

这类研究的风险包括焦虑与惊恐反应、血压与心率变化、以及在易感者中诱发精神病性症状,因此排除标准与安全监查不是形式。无论结果正负,方案、统计分析计划、分析代码与汇总结果都将公开;作者不参与招募、干预、评估与分析。

利益披露

作者以 5D Library 的著作(包括本书)作为其社群课程、升维牌卡(5D Ascension Cards,Dimensional Shift Card Decks)与 Yafu AI 知识库的方法论基础,因此与本书主题有商业利益关系。集团另运营 ketohackers.com 与 nadplusnadplus.com 两个健康科普站点,与本书的代谢与可塑性主题相邻;本书不链接这两个站点,也不提及任何产品。本书不推销任何课程、牌卡、物质或服务,不给出任何剂量、来源、途径或疗程,并明确不建议任何人使用本书讨论的任何物质;第八章的试验设计为由独立于作者企业的研究人员在取得所需监管批准的司法辖区内实施和分析,不从作者的社群、学员或客户中招募;书中所有代码与研究方案均公开,供任何人检验。作者不提供医疗服务,不销售任何物质、器械、补充剂或检测。

本章小结

四步核查表从链条与方向、参数、测量、设盲与预期四个方面阅读亚知觉暴露的说法;用于通行说法时,它显示第二环与第四环方向相反、三个关键参数都未在人体内测量、熵类指标与噪声不可区分、而现有的正面证据与纯预期解释相容。预注册的双盲试验以活性条件对匹配安慰剂为唯一主要比较,以第 6 周预先指定量表分数的协方差分析为唯一主要结局,最小关注效应 0.30 个标准差(取半个标准差这一常被引用的最小重要差别界值 (Norman, Sloan, and Wyrwich 2003) 之下的保守值),每臂 225 例可分析、282 例随机,共 564 例;暴露水平由实施机构按监管批准填入,本书不指定。设盲完整性被测量并写入判定规则,而把药理与预期分开的「按信念分层」对照被列为预先指定的关键次要分析且明确不承诺把握度,因为设盲破坏时它需要两倍以上的样本。该试验不测浓度,因此不回答药代动力学问题,书中写明这一点。

参考文献

Bang, Heejung, Liyun Ni, and Clarence E. Davis. 2004. “Assessment of Blinding in Clinical Trials.” Controlled Clinical Trials 25 (2): 143–56. https://doi.org/10.1016/j.cct.2003.10.016.
Chan, An-Wen, Jennifer M. Tetzlaff, Douglas G. Altman, Andreas Laupacis, Peter C. Gøtzsche, Karmela Krleža-Jerić, Asbjørn Hróbjartsson, et al. 2013. “SPIRIT 2013 Statement: Defining Standard Protocol Items for Clinical Trials.” Annals of Internal Medicine 158 (3): 200–207. https://doi.org/10.7326/0003-4819-158-3-201302050-00583.
Ciric, Rastko, Daniel H. Wolf, Jonathan D. Power, David R. Roalf, Graham L. Baum, Kosha Ruparel, Russell T. Shinohara, et al. 2017. “Benchmarking of Participant-Level Confound Regression Strategies for the Control of Motion Artifact in Studies of Functional Connectivity.” NeuroImage 154: 174–87. https://doi.org/10.1016/j.neuroimage.2017.03.020.
Devilly, Grant J., and Thomas D. Borkovec. 2000. “Psychometric Properties of the Credibility/Expectancy Questionnaire.” Journal of Behavior Therapy and Experimental Psychiatry 31 (2): 73–86. https://doi.org/10.1016/S0005-7916(00)00012-4.
Hróbjartsson, Asbjørn, Frida Emanuelsson, Ann Sofia Skou Thomsen, Jørgen Hilden, and Stig Brorson. 2014. “Bias Due to Lack of Patient Blinding in Clinical Trials: A Systematic Review of Trials Randomizing Patients to Blind and Nonblind Sub-Studies.” International Journal of Epidemiology 43 (4): 1272–83. https://doi.org/10.1093/ije/dyu115.
Lakens, Daniël, Anne M. Scheel, and Peder M. Isager. 2018. “Equivalence Testing for Psychological Research: A Tutorial.” Advances in Methods and Practices in Psychological Science 1 (2): 259–69. https://doi.org/10.1177/2515245918770963.
Norman, Geoffrey R., Jeff A. Sloan, and Kathleen W. Wyrwich. 2003. “Interpretation of Changes in Health-Related Quality of Life: The Remarkable Universality of Half a Standard Deviation.” Medical Care 41 (5): 582–92. https://doi.org/10.1097/01.MLR.0000062554.74615.4C.
Power, Jonathan D., Kelly A. Barnes, Abraham Z. Snyder, Bradley L. Schlaggar, and Steven E. Petersen. 2012. “Spurious but Systematic Correlations in Functional Connectivity MRI Networks Arise from Subject Motion.” NeuroImage 59 (3): 2142–54. https://doi.org/10.1016/j.neuroimage.2011.10.018.
Schulz, Kenneth F., Douglas G. Altman, and David Moher. 2010. “CONSORT 2010 Statement: Updated Guidelines for Reporting Parallel Group Randomised Trials.” BMJ 340: c332. https://doi.org/10.1136/bmj.c332.
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.
Szigeti, Balázs, Laura Kartner, Allan Blemings, Fernando Rosas, Amanda Feilding, David J. Nutt, Robin L. Carhart-Harris, and David Erritzoe. 2021. “Self-Blinding Citizen Science to Explore Psychedelic Microdosing.” eLife 10: e62878. https://doi.org/10.7554/eLife.62878.
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). Biomathematical Models of Sub-Perceptual Tryptamine Pharmacokinetics, 5-HT2A Receptor Desensitisation and Neuroplastic Window Dynamics: From Michaelis–Menten Kinetics of Prodrug Deacetylation and Whole-Brain Connectivity Entropy to the Microdose Tolerance Ratchet, with a Textual Examination of the Dose Boundary in the Classical Materia Medica [亚知觉色胺药代动力学、5-HT2A 受体脱敏与神经可塑性窗口的生物数学模型] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24019347 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).