5D Library › 节律声刺激、听觉稳态反应与双耳等时调制工程 · English

第六章 证据:检出力、安慰剂与分岔路径

文献说了什么

证据

已发表的证据。 Schwarz 与 Taylor 比较了双耳拍与单耳拍诱发的稳态反应,发现在 40 Hz 处双耳拍可以诱发可测的反应,但比单耳拍弱 (Schwarz and Taylor 2005)。一篇系统综述检查了 14 项以脑电或脑磁测量双耳拍「夹带」的研究,其中 5 项支持、8 项与之相反、1 项结果混合,作者认为现有证据尚不能确立夹带的说法 (Ingendoh et al. 2023)。一篇关于认知、焦虑与疼痛结局的元分析报告了总体上的正面效应,同时指出研究之间差异很大 (Garcia-Argibay et al. 2019)。

这些研究的分歧并不奇怪。本章用三段程序说明,即使刺激本身完全相同,研究设计也足以决定结果。

单次记录能否看见反应

程序为每位模拟听者生成刺激频率处的一个复数频点(信号加噪声)与十个只有噪声的相邻频点,以「目标频点功率 ÷ 相邻频点平均功率」为检验统计量,用假刺激听者校准 0.05 的阈值,再计算各种刺激下超过阈值的听者比例。

推导

# Chapter 6. p11: Can a single recording show the response at all?
# For each simulated listener the EEG spectrum at the stimulation rate
# is one complex DFT bin: signal s plus complex Gaussian noise, and ten
# neighbouring bins carry noise only. The test statistic is the power
# in the target bin over the mean power of the neighbours; the 0.05
# threshold is calibrated on sham listeners (no drive). Signal power
# per second follows p04: (a^2 / 2) / P(f), with between-listener
# amplitude spread exp(N(0, 0.4)), times the drive: isochronic 1,
# binaural 0.13116, sham 0. Detection rate = share of listeners above
# the sham threshold. 4,000 listeners per cell, fixed seed.
import math
import random

def snr_per_s(fm):
    a = 0.15 + 0.35 * math.exp(-((fm - 40.0) / 10.0) ** 2)
    return (a * a / 2.0) / (4.0 / fm)

def stat(rng, snr):
    amp = math.sqrt(snr)
    ph = rng.uniform(0, 2 * math.pi)
    re = amp * math.cos(ph) + rng.gauss(0, math.sqrt(0.5))
    im = amp * math.sin(ph) + rng.gauss(0, math.sqrt(0.5))
    nb = sum(rng.gauss(0, math.sqrt(0.5)) ** 2 +
             rng.gauss(0, math.sqrt(0.5)) ** 2 for _ in range(10)) / 10
    return (re * re + im * im) / nb

def rate(rng, fm, drive, T, thr, n=4000):
    hit = 0
    for _ in range(n):
        g = math.exp(rng.gauss(0, 0.4))
        if stat(rng, snr_per_s(fm) * T * (g * drive) ** 2) > thr:
            hit += 1
    return hit / n

if __name__ == "__main__":
    rng = random.Random(2026)
    sham = sorted(stat(rng, 0.0) for _ in range(20000))
    thr = sham[int(0.95 * len(sham))]
    print("   sham 95th percentile of the statistic: {:.2f}".format(thr))
    print("\n   rate  T (s)   isochronic   binaural   sham")
    for fm in (10.0, 40.0):
        for T in (10, 60, 300):
            r = [rate(rng, fm, d, T, thr) for d in (1.0, 0.13116, 0.0)]
            print("   {:4.0f} {:6d} {:12.3f} {:10.3f} {:6.3f}".format(
                fm, T, *r))
    print("\nReading: an isochronic 40 Hz tone shows up in nearly every")
    print("listener within a minute. The same rate as a binaural beat")
    print("is barely above the 0.05 false alarm level after ten")
    print("seconds and needs minutes of recording; at 10 Hz the")
    print("binaural beat stays near chance for the whole five minutes.")
#    sham 95th percentile of the statistic: 3.54
#
#    rate  T (s)   isochronic   binaural   sham
#      10     10        0.107      0.049  0.050
#      10     60        0.355      0.053  0.048
#      10    300        0.831      0.078  0.050
#      40     10        0.903      0.093  0.046
#      40     60        0.999      0.298  0.048
#      40    300        1.000      0.774  0.051
#
# Reading: an isochronic 40 Hz tone shows up in nearly every
# listener within a minute. The same rate as a binaural beat
# is barely above the 0.05 false alarm level after ten
# seconds and needs minutes of recording; at 10 Hz the
# binaural beat stays near chance for the whole five minutes.

假刺激的阈值为 3.54。40 Hz 等时音在 10 s 时已能在 0.903 的听者身上看到反应,60 s 时为 0.999;同频率的双耳拍 10 s 时只有 0.093,60 s 时 0.298,300 s 时 0.774。10 Hz 时,等时音 300 s 才达到 0.831,双耳拍在五分钟内始终接近机会水平(0.078)。因此,一项以 10 Hz 双耳拍、每人记录几分钟的研究,得到「无夹带」几乎是必然的;而以 40 Hz、记录较长的研究可以得到「有反应」。两者并不矛盾。

40 Hz 与 10 Hz 下等时音与双耳拍的单人检出率随记录时长的变化。虚线为 0.05。

期望效应与样本量

自评的放松或专注同时包含真实效应与期望效应。只做前后对比、没有对照的设计看到的是两者之和;双盲对照假刺激音轨的设计只看到真实效应。

推导

# Chapter 6. p12: Expectation, sham control and sample size.
# Self-reported calm after a session mixes a true effect (d_true) with
# an expectation effect (d_exp) in standard deviation units. Designs:
#   A. before / after with no control: sees d_true + d_exp;
#   B. blinded track vs sham track (same carriers, no beat or no
#      gating): sees d_true only.
# Normal approximations, two-sided alpha 0.05:
#   power of A with n listeners: Phi(d sqrt(n) - 1.96);
#   n per arm for B at power 0.80: 2 (1.96 + 0.8416)^2 / d^2.
import math

def phi(x):
    return 0.5 * (1 + math.erf(x / math.sqrt(2)))

if __name__ == "__main__":
    d_exp = 0.4
    print("   design A (no control), expectation d = {}, true d = 0:"
          .format(d_exp))
    for n in (15, 30, 60, 120):
        print("     n = {:3d}   chance of a 'significant' result {:.3f}"
              .format(n, phi(d_exp * math.sqrt(n) - 1.96)))
    k = 2 * (1.96 + 0.8416) ** 2
    print("\n   design B (blinded vs sham), listeners per arm for 0.80:")
    for d in (0.1, 0.2, 0.3, 0.5):
        print("     true d = {:.1f}   n per arm = {:5d}   total = {:5d}"
              .format(d, math.ceil(k / d ** 2), 2 * math.ceil(k / d ** 2)))
    print("\n   if 30 % of listeners can tell sham from active, and their")
    print("   expectation d = 0.4 leaks into B, the apparent effect is")
    print("   0.30 x 0.4 = {:.2f} even when the true effect is zero."
          .format(0.30 * 0.4))
    print("\nReading: an uncontrolled before and after survey of 60")
    print("people will find a 'significant' benefit most of the time")
    print("with no true effect at all. A blinded comparison with a sham")
    print("track needs hundreds of listeners for a small true effect,")
    print("and still needs a check that the blinding held.")
#    design A (no control), expectation d = 0.4, true d = 0:
#      n =  15   chance of a 'significant' result 0.341
#      n =  30   chance of a 'significant' result 0.591
#      n =  60   chance of a 'significant' result 0.873
#      n = 120   chance of a 'significant' result 0.992
#
#    design B (blinded vs sham), listeners per arm for 0.80:
#      true d = 0.1   n per arm =  1570   total =  3140
#      true d = 0.2   n per arm =   393   total =   786
#      true d = 0.3   n per arm =   175   total =   350
#      true d = 0.5   n per arm =    63   total =   126
#
#    if 30 % of listeners can tell sham from active, and their
#    expectation d = 0.4 leaks into B, the apparent effect is
#    0.30 x 0.4 = 0.12 even when the true effect is zero.
#
# Reading: an uncontrolled before and after survey of 60
# people will find a 'significant' benefit most of the time
# with no true effect at all. A blinded comparison with a sham
# track needs hundreds of listeners for a small true effect,
# and still needs a check that the blinding held.

期望效应为 0.4 个标准差、真实效应为 0 时,60 人的前后对比有 0.873 的机会得到「显著」结果,120 人时为 0.992。双盲对照要以 0.80 的检验力发现 0.2 的真实效应,每组需要 393 人,共 786 人;0.1 的效应需要每组 1570 人。若有 30% 的受试者能分辨真假音轨,期望效应会以 0.12 的大小漏进双盲比较。因此还必须检查盲法是否成立。

分岔路径

脑电研究的分析自由度很大:5 个频段、19 个电极、3 个时间窗,共 285 次检验。若事先不规定比较哪一个,研究者总能找到「显著」的格子 (Gelman and Loken 2014)。

推导

# Chapter 6. p13: Five bands, nineteen electrodes, three windows.
# A null EEG study (no effect anywhere) tests every combination of
# 5 frequency bands x 19 electrodes x 3 time windows = 285 tests at
# two-sided 0.05 (|z| > 1.96). Electrodes within one band and window
# share a common component with correlation rho; bands and windows
# are independent. We simulate 4,000 null studies and report:
#   the family-wise error rate (FWER) without correction;
#   the expected number of false "significant" cells;
#   the FWER under Bonferroni, |z| > z(0.05 / 285);
#   the max-statistic threshold that holds the FWER at exactly 0.05.
import math
import random

def z_two_sided(alpha):
    lo, hi = 0.0, 10.0
    for _ in range(100):
        mid = (lo + hi) / 2
        if math.erfc(mid / math.sqrt(2)) > alpha:
            lo = mid
        else:
            hi = mid
    return (lo + hi) / 2

def study(rng, rho):
    zs = []
    for _ in range(15):
        c = rng.gauss(0, 1)
        for _ in range(19):
            zs.append(math.sqrt(rho) * c +
                      math.sqrt(1 - rho) * rng.gauss(0, 1))
    return zs

if __name__ == "__main__":
    m = 5 * 19 * 3
    zb = z_two_sided(0.05 / m)
    print("   tests: {}   Bonferroni |z| > {:.3f}".format(m, zb))
    print("   independent tests: no false hit with chance 0.95^{},"
          .format(m))
    print("   that is 1 in {:,.0f}".format(1 / 0.95 ** m))
    rng = random.Random(99)
    print("\n   rho   FWER raw   mean false hits   FWER Bonf   max-stat z")
    for rho in (0.0, 0.5, 0.8):
        raw = bon = hits = 0
        mx = []
        for _ in range(4000):
            zs = [abs(z) for z in study(rng, rho)]
            h = sum(1 for z in zs if z > 1.96)
            hits += h
            raw += h > 0
            bon += max(zs) > zb
            mx.append(max(zs))
        mx.sort()
        print("   {:3.1f} {:10.3f} {:17.2f} {:11.3f} {:12.3f}".format(
            rho, raw / 4000, hits / 4000, bon / 4000, mx[int(0.95 * 4000)]))
    print("\nReading: a study that looks everywhere finds a 'significant'")
    print("cell almost surely, about fourteen of them on average, even")
    print("when nothing happens. Correlation between electrodes lowers")
    print("the FWER but not the expected count. Bonferroni or a")
    print("max-statistic threshold fixed in advance restores 0.05.")
#    tests: 285   Bonferroni |z| > 3.752
#    independent tests: no false hit with chance 0.95^285,
#    that is 1 in 2,232,402
#
#    rho   FWER raw   mean false hits   FWER Bonf   max-stat z
#    0.0      1.000             14.27       0.050        3.752
#    0.5      1.000             14.32       0.043        3.715
#    0.8      0.977             14.25       0.022        3.520
#
# Reading: a study that looks everywhere finds a 'significant'
# cell almost surely, about fourteen of them on average, even
# when nothing happens. Correlation between electrodes lowers
# the FWER but not the expected count. Bonferroni or a
# max-statistic threshold fixed in advance restores 0.05.

在完全没有效应的情况下,独立检验全部不出现假阳性的机会只有 2,232,402 分之一;平均每项研究有 14.27 个假「显著」格子。电极之间的相关不改变假阳性的平均个数(相关 0.5 时为 14.32),只稍微降低至少出现一个的概率(相关 0.8 时为 0.977)。Bonferroni 阈值 3.752 把族错误率压回 0.050 及以下;最大统计量阈值在相关 0.8 时为 3.520,略为宽松而同样有效。

小结

已发表的研究之所以互相矛盾,很大程度上是设计所致:频率、记录时长、对照与分析自由度足以决定结论。一项可信的研究需要假刺激对照、盲法检查、事先登记的单一主要结局与多重比较校正。

参考文献

Garcia-Argibay, Miguel, Miguel A. Santed, and José M. Reales. 2019. “Efficacy of Binaural Auditory Beats in Cognition, Anxiety, and Pain Perception: A Meta-Analysis.” Psychological Research 83 (2): 357–72.
Gelman, Andrew, and Eric Loken. 2014. “The Statistical Crisis in Science.” American Scientist 102 (6): 460–65.
Ingendoh, Ruth Maria, Ella Sophie Posny, and Angela Heine. 2023. “Binaural Beats to Entrain the Brain? A Systematic Review of the Effects of Binaural Beat Stimulation on Brain Oscillatory Activity, and the Implications for Psychological Research and Intervention.” PLOS ONE 18 (5): e0286023.
Schwarz, D. W. F., and P. Taylor. 2005. “Human Auditory Steady State Responses to Binaural and Monaural Beats.” Clinical Neurophysiology 116 (3): 658–68.

Fan, Y. (2026). Rhythmic Acoustic Stimulation, Auditory Steady-State Responses and Modulation Signal Engineering: Binaural versus Isochronic Envelope Depth, 40-Hz Phase-Locking Bounds and Pink-Noise Spectral Masking [节律声刺激、听觉稳态反应与双耳等时调制工程] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090211 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).