5D Library › Collective Dynamics, Coupled-Oscillator Synchrony and Distributed Consensus · 中文

Chapter 8 Practice and a Pre-Registered Trial

Note

The checklist in this chapter is a thinking tool. It is not management consulting, makes no political claim of any kind, and promises no result. The example is a fictional 12-member committee; all members, networks and estimates are invented to demonstrate the calculations and come from no real organisation, community or client.

Part 1 A four-step checklist against echo chambers and for calibrating synchrony

The models of the first six chapters can be organised into four questions. They apply to any situation in which a group of people exchange views repeatedly and then give a joint judgement — for example a committee estimating how many people will attend an event.

  1. Measure influence. Draw the network of “who listens to whom” and compute each person’s influence πi\pi_i and the effective number Neff=1/∑πi2N_{\mathrm{eff}} = 1/\sum\pi_i^2 as in Chapter 5. A largest influence far above 1/N1/N means the conclusion is decided mainly by a few people.
  2. Independence first, exchange second. Before any discussion, have everyone write down and save a private estimate. Use these independent estimates to check how far the consensus would be pulled if the most influential person carried a bias (the πib\pi_i b of Chapter 5).
  3. Look for clusters and bridges. Using the bounded-confidence model of Chapter 6, see into how many groups the independent estimates fall under a reasonable confidence bound, and whether the links between groups all run through one person — if so, that person is both the only bridge and the only filter.
  4. Calibrate synchrony. Fast convergence does not mean a good conclusion (Chapter 5: speed and accuracy are unrelated; Chapter 2: over-synchrony can do harm). Record the disagreement in every round, fix the number of rounds and the stopping condition in advance, and always keep “the simple mean of the independent estimates” as a benchmark.

The checklist applied to the fictional committee. The 12 members’ independent estimates have been written down. Before: the convener (member 0) and everyone listen to each other, and the other members form two small circles {1…6} and {7…11}, each listening only to the two circle neighbours. After: the 12 sit in a ring, each listening to two neighbours on each side, and the convener is an ordinary seat. The program gives the results of the four steps: step 3 first shows into how many groups the independent estimates fall under different confidence bounds; steps 1, 2 and 4 compare the two networks (step 2 assumes the convener has a bias of +15 and individual noise has standard deviation 20, estimating the root-mean-square error of the consensus from 20,000 simulations).

Calculation

# Four-step check for a FICTIONAL 12-member committee that estimates a quantity together
# (all numbers invented). Before: member 0 (the convener) hears and is heard by everyone;
# the rest form two small circles {1..6} and {7..11}, each member hearing only its two
# circle neighbours. After: every member hears 2 neighbours on each side of one ring of 12
# (member 0 becomes an ordinary seat).
import math, random

N = 12
def circle(m):                 # members m[0], m[1], ... in a circle, each hears 2
    return {m[k]: {m[k - 1], m[(k + 1) % len(m)]} for k in range(len(m))}
C = {**circle(list(range(1, 7))), **circle(list(range(7, N)))}
BEFORE = [set(range(1, N))] + [C[i] | {0} for i in range(1, N)]
AFTER = [{(i + s) % N for s in (-2, -1, 1, 2)} for i in range(N)]
START = [92, 120, 118, 125, 131, 115, 122, 78, 84, 70, 88, 81]   # first private estimates

def influence(adj):            # DeGroot with self-loops: pi_i proportional to d_i + 1
    w = [len(a) + 1 for a in adj]; return [x / sum(w) for x in w]

def degroot(adj, x, rounds):
    for _ in range(rounds):
        x = [(x[i] + sum(x[j] for j in a)) / (len(a) + 1) for i, a in enumerate(adj)]
    return x

def hk_groups(x, eps):         # bounded-confidence clusters among the private estimates
    for _ in range(200):
        x = [sum(v for v in x if abs(v - u) <= eps) / sum(abs(v - u) <= eps for v in x)
             for u in x]
    return sorted({round(v, 1) for v in x})

def bias_error(adj, rng, bias=15.0, sd=20.0, draws=20000):
    pi = influence(adj); e = 0.0                 # member 0 carries a bias of +15
    for _ in range(draws):
        e += (sum(p * rng.gauss(0, sd) for p in pi) + pi[0] * bias) ** 2
    return math.sqrt(e / draws)

def cross_links(adj, g):        # links joining the two camps, and members holding them
    links = [(i, j) for i, a in enumerate(adj) for j in a if i < j and g[i] != g[j]]
    top = max(sum(k in e for e in links) for k in range(len(adj)))
    return len(links), top / len(links)

if __name__ == "__main__":
    rng = random.Random(8)
    print(f"private estimates {START}; mean {sum(START) / N:.1f}")
    for eps in (10, 20, 30, 40):
        print(f"   tolerance eps = {eps}: groups end at {hk_groups(START, eps)}")
    side = [int(v >= 100) for v in START]       # the two camps visible at eps = 20
    for name, adj in (("before", BEFORE), ("after", AFTER)):
        pi = influence(adj); top = max(range(N), key=pi.__getitem__)
        x5 = degroot(adj, START, 5)
        print(f"[{name}] 1 max pi = {pi[top]:.3f} (member {top}), N_eff = 1/sum pi^2 ="
              f" {1 / sum(p * p for p in pi):.1f} of {N}")
        eq = math.hypot(20 / N ** .5, 15 / N)
        print(f"         2 RMS error, member 0 biased by +15 (noise SD 20):"
              f" {bias_error(adj, rng):.2f}  [equal weights {eq:.2f}]")
        nl, nm = cross_links(adj, side)
        print(f"         3 links between the camps: {nl} of {sum(map(len, adj)) // 2};"
              f" share via one member {nm:.2f}")
        cons = sum(p * v for p, v in zip(pi, START))
        print(f"         4 spread after 5 rounds {max(x5) - min(x5):5.1f} (start"
              f" {max(START) - min(START)}); consensus {cons:.1f}")
# private estimates [92, 120, 118, 125, 131, 115, 122, 78, 84, 70, 88, 81]; mean 102.0
#    tolerance eps = 10: groups end at [82.6, 121.6]
#    tolerance eps = 20: groups end at [82.2, 121.8]
#    tolerance eps = 30: groups end at [101.7]
#    tolerance eps = 40: groups end at [102.0]
# [before] 1 max pi = 0.214 (member 0), N_eff = 1/sum pi^2 = 9.8 of 12
#          2 RMS error, member 0 biased by +15 (noise SD 20): 7.21  [equal weights 5.91]
#          3 links between the camps: 6 of 22; share via one member 1.00
#          4 spread after 5 rounds  10.0 (start 61); consensus 100.6
# [after] 1 max pi = 0.083 (member 0), N_eff = 1/sum pi^2 = 12.0 of 12
#          2 RMS error, member 0 biased by +15 (noise SD 20): 5.91  [equal weights 5.91]
#          3 links between the camps: 6 of 24; share via one member 0.33
#          4 spread after 5 rounds  12.3 (start 61); consensus 102.0

Step 3: the mean of the 12 independent estimates is 102.0; with a confidence bound of 10 or 20 the opinions gather into two groups, around 82 and around 122, and only with a bound of 30 do they form one group. The committee in fact has two camps.

Step 1: in the network before, the convener’s influence is 0.214 (the average is 0.083) and the effective number 9.8; after, influence is perfectly equal and the effective number 12.0. Step 2: if the convener has a bias of +15, the root-mean-square error of the consensus is 7.21 before and 5.91 after, the same as an equally weighted mean. The second half of step 3: both networks have 6 links between the camps, but before, all of them run through the convener (share 1.00), while after, at most a third run through any one person. Step 4: the network before converges faster — after 5 rounds the disagreement falls from 61 to 10.0, against 12.3 after; the consensus values are 100.6 and 102.0. The network before reached, faster, a conclusion that depends more on the convener — an example of Chapter 5’s “speed and accuracy are unrelated”.

Part 2 A pre-registered cluster-randomised controlled trial

Question

In repeated rounds of continuous numerical estimation, when a core seat in the network carries a known bias, does a “centralised hub topology” (everyone sees the core seat; high largest influence) give a larger error in the group’s final estimate than an “equal-degree decentralised regular ring” (every seat has the same degree and the core seat is an ordinary member)?

Evidence

Existing evidence. In the experiment of Becker and colleagues in 2017, the results of centralised networks depended on the central person: when that person’s initial estimate lay away from the truth, the error of the group’s mean estimate rose by 19 %; when it lay towards the truth, it fell by 43 % (Becker et al. 2017). Lorenz and colleagues in 2011 found that social influence reduced the diversity of estimates without a corresponding gain in accuracy (Lorenz et al. 2011). But the author found no pre-registered cluster-randomised trial in which the bias of the core seat is fixed in advance, the two arms differ only in topology, and the group’s final error is the single primary outcome.

Design

Element Specification
Design Two-arm parallel cluster-randomised controlled trial; the unit of randomisation is a “group” of 20 seats, paired by session and allocated 1:1 within pairs
Participants Adults aged 18 or over, fluent in the task language and answering on a computer, recruited through a public online research-participant platform; not recruited from the author’s community, students or customers
Programmed seat 1 of the 20 seats in each group is played by a program and shows twice the true value for every question (log bias b=ln⁡2b = \ln 2); the consent form states in advance that “some displayed estimates may be generated by a program”
Task 10 quantity-estimation questions (pre-tested in an independent sample; questions are chosen whose initial crowd median is within about 10 % of the truth), 3 rounds each: round 1 answered alone; in rounds 2 and 3 participants revise after seeing the previous-round estimates of 4 seats
Centralised arm Everyone sees the programmed seat and 3 peers (positions −1, +1, +2 on a ring of the 19)
Decentralised arm The 20 seats form a ring and everyone sees two seats on each side; the programmed seat is an ordinary member of the ring, seen by only 4 people
Primary outcome Group level: the mean over the 10 questions of the absolute difference between the mean log estimate of the 19 participants in round 3 and the log of the true value
Secondary outcomes Individual mean absolute log error in round 3; dispersion of round-3 estimates (SD of logs); size of the movement towards the programmed seat; change in group error from round 1 to round 3
Sample size 49 analysable groups per arm, 58 randomised; 116 groups and 2,204 participants in all
Analysis Linear regression at the group level (arm + pair block); see below

The analysis in detail: with the group as the unit of analysis, the primary outcome is regressed on arm and pair block, reporting the mean difference dd (centralised arm minus decentralised arm; positive means centralisation increases the error) with its 95 % confidence interval, two-sided α=0.05\alpha = 0.05. Sensitivity analysis: a mixed model with the group error on each question as the outcome and group as a random effect. Reporting follows the CONSORT extension for cluster trials (Campbell et al. 2012).

In both arms everyone sees 4 estimates in every round; only “whose estimates are seen” differs. There is a single primary outcome, for three reasons: it measures directly the “final consensus error” the framework is concerned with; it is defined at the unit of randomisation (the group), avoiding treating members of the same group as independent; and the log scale suits the right-skewed distributions common in quantity estimates. Participants receive a fixed payment plus a bonus for the accuracy of their own round-3 estimates, the same in both arms.

Why the size of the bias must be fixed. In the DeGroot model, a programmed seat that never updates will, in the long run, pull everyone to its value as long as it is connected to the network; the two topologies differ only in how fast this happens in the first rounds. This trial has only 3 rounds and compares exactly that finite-time difference. The program below also shows that when the bias is small the centralised structure pulls everyone’s estimates together and reduces noise, which can offset the bias.

Sample size

Calculation

# Chapter 8 trial. Part 1: a MODEL of one group (assumptions, not estimates). 20 seats;
# seat 0 is a disclosed scripted seat that always shows truth x e^b (log bias b > 0).
# 19 people answer 10 questions in 3 rounds: round 1 alone, then each sees 4 estimates of
# the previous round and moves a share a_i ~ U(0.1, 0.6) toward their mean.
#   hub arm: everyone sees seat 0 plus 3 peers (offsets -1, +1, +2 on a ring of the 19)
#   ring arm: 20-seat ring, everyone sees seats at -2, -1, +1, +2 (seat 0 is ordinary)
# Log estimate = truth + c_q + e_i, c_q ~ N(0, 0.05) (questions pre-tested to have crowd
# medians near the truth), e_i ~ N(0, 0.60). Group outcome = mean over the 10 questions of
# |mean final log estimate of the 19 people - log truth|.  Part 2: groups per arm.
import math, random
from statistics import NormalDist, mean, stdev

H, Q = 19, 10

def views(arm):
    if arm == "hub":
        return [[None] + [(i + d) % H for d in (-1, 1, 2)] for i in range(H)]
    ring = [[(s + d) % 20 for d in (-2, -1, 1, 2)] for s in range(1, 20)]
    return [[None if v == 0 else v - 1 for v in r] for r in ring]   # None = scripted seat

def group(arm, cq, b, rng):
    vw = views(arm); a = [rng.uniform(0.1, 0.6) for _ in range(H)]; err = 0.0
    for c in cq:
        x = [c + rng.gauss(0, 0.60) for _ in range(H)]
        for _ in range(2):
            x = [(1 - a[i]) * x[i] + a[i] * sum(b if j is None else x[j]
                                                for j in vw[i]) / 4 for i in range(H)]
        err += abs(sum(x) / H)
    return err / Q

def groups_per_arm(sd, diff, power=0.90, alpha=0.05):
    z = NormalDist().inv_cdf
    return math.ceil(2 * (z(1 - alpha / 2) + z(power)) ** 2 * sd ** 2 / diff ** 2) + 1

if __name__ == "__main__":
    rng = random.Random(2017); cq = [rng.gauss(0, 0.05) for _ in range(Q)]
    print("scripted bias b   hub arm (SD)       ring arm (SD)      difference hub - ring")
    for b in (0.40, math.log(2), 1.00):
        hub = [group("hub", cq, b, rng) for _ in range(1000)]
        ring = [group("ring", cq, b, rng) for _ in range(1000)]
        print(f"  {b:.3f} (x{math.exp(b):.2f})  {mean(hub):.4f} ({stdev(hub):.4f})"
              f"    {mean(ring):.4f} ({stdev(ring):.4f})"
              f"     {mean(hub) - mean(ring):+.4f}")
    print("analysable groups per arm, power 90 % (+1 for the t distribution):")
    print("   SD of outcome   SESOI 0.02   SESOI 0.03   SESOI 0.05")
    for sd in (0.025, 0.030, 0.040):
        print(f"       {sd:.3f}      " + "".join(f"    {groups_per_arm(sd, s):5d}    "
                                              for s in (0.02, 0.03, 0.05)))
    n = groups_per_arm(0.030, 0.02); rec = math.ceil(n / 0.85)
    print(f"plan (b = ln 2): SD 0.030, SESOI 0.02, 15 % of groups lost -> {n}"
          " analysable,")
    print(f"   {rec} randomised per arm: {2 * rec} groups of 20 seats,"
          f" {2 * rec * H} people")
    n80 = groups_per_arm(0.030, 0.02, power=0.80)
    print(f"80 % power alternative: {n80} analysable per arm")
# scripted bias b   hub arm (SD)       ring arm (SD)      difference hub - ring
#   0.400 (x1.49)  0.1172 (0.0265)    0.1204 (0.0282)     -0.0032
#   0.693 (x2.00)  0.1432 (0.0299)    0.1213 (0.0284)     +0.0219
#   1.000 (x2.72)  0.1806 (0.0344)    0.1228 (0.0294)     +0.0578
# analysable groups per arm, power 90 % (+1 for the t distribution):
#    SD of outcome   SESOI 0.02   SESOI 0.03   SESOI 0.05
#        0.025             34           16            7    
#        0.030             49           23            9    
#        0.040             86           39           15    
# plan (b = ln 2): SD 0.030, SESOI 0.02, 15 % of groups lost -> 49 analysable,
#    58 randomised per arm: 116 groups of 20 seats, 2204 people
# 80 % power alternative: 37 analysable per arm

The first part of the program is not data but a model of the trial: each person’s round-1 log estimate is the truth plus a question bias (SD 0.05) plus individual noise (SD 0.60), and in each later round they move a share aia_i (uniform between 0.1 and 0.6) of the way towards the mean of the 4 estimates they see. These are assumptions, not estimates from any study. The model gives three results. With a bias of 0.40 for the programmed seat (showing 1.49 times the truth), the centralised arm’s mean error (0.1172) is actually slightly below the decentralised arm’s (0.1204), a difference of −0.0032-0.0032: everyone is pulled towards the same fixed value, the noise in the group mean falls, and this offsets the bias. With a bias of ln⁡2\ln 2 (twice the truth) the difference is +0.0219+0.0219; with a bias of 1.00 (2.72 times) it is +0.0578+0.0578. These differences depend on the question biases cqc_q that happen to be drawn: rerun during checking with 4 other seeds (600 groups per arm), the difference at bias 0.40 lies between −0.006-0.006 and +0.002+0.002, with no consistent sign, and at bias ln⁡2\ln 2 between +0.016+0.016 and +0.029+0.029. The robust conclusion at bias 0.40 is therefore that the two arms hardly differ, not that centralisation is better. So “centralisation always increases error when the core seat is biased” does not hold; the conclusion depends on the size of the bias, and this protocol fixes b=ln⁡2b = \ln 2.

The sample size rests on these assumptions: a smallest effect of interest of 0.02 (in the model the decentralised arm’s error is about 0.12, so 0.02 is a relative increase of about 17 %, of the same order as the 19 % increase Becker and colleagues reported when the central person deviated from the truth); a standard deviation of the group outcome of 0.030 (0.028 to 0.030 in the model); two-sided α=0.05\alpha = 0.05, 90 % power, plus 1 for the tt distribution. Each arm needs 49 analysable groups; allowing about 15 % of groups to become unanalysable through attrition (fewer than 15 completing round 3), 58 groups are randomised per arm, 116 groups and 2,204 participants in all. With a standard deviation of 0.025 or 0.040, each arm needs 34 or 86 groups; with a smallest effect of interest of 0.03 and SD 0.030, 23. If a smaller alternative is wanted, 80 % power needs 37 analysable groups per arm, which should be stated at registration as well. A pilot precedes the main trial, solely to check the program, interface and questions, not to change the sample size.

Note

What “90 % power” means. As in Volume 33: 90 % is the probability of reaching some conclusion that “centralisation increases the error” when the true effect equals the smallest effect of interest (d=0.02d = 0.02); in that case the decision rule reaches the strongest conclusion, “increases it by at least 0.02”, only about 2.5 % of the time. The decision program below confirms this by simulation.

Decision rule

Calculation

# Decision rule for the Chapter 8 trial, fixed before any data are collected.
# d = mean group outcome (mean absolute log error of the final group estimate), hub arm
# minus ring arm, with its 95 % CI; positive d = the hub topology makes groups less
# accurate.
# SESOI = 0.02. Gates: more than 20 % of randomised groups unanalysable (fewer than 15 of
# the 19 people complete all three rounds), or the scripted seat failed to display in more
# than 5 % of group-rounds -> NOT INTERPRETABLE.
import math, random
from statistics import NormalDist
SESOI = 0.02

def verdict(d, lo, hi, lost=0.08, display_fail=0.01):
    if lost > 0.20:
        return (f"d = {d:+.3f}: NOT INTERPRETABLE - {100 * lost:.0f} % of groups lost"
                " (limit 20 %)")
    if display_fail > 0.05:
        return (f"d = {d:+.3f}: NOT INTERPRETABLE - scripted seat missing in"
                f" {100 * display_fail:.0f} % of rounds")
    if lo >= SESOI:      v = "hub topology adds at least 0.02 to the error"
    elif lo > 0:         v = "hub topology adds error; 0.02 not shown"
    elif hi < 0:         v = "hub topology REDUCES the error"
    elif lo >= -SESOI and hi <= SESOI: v = "no difference worth acting on"
    else:                v = "unclear: the interval is too wide to decide"
    return f"d = {d:+.3f} ({lo:+.3f} to {hi:+.3f}): {v}"

SHORT = {"hub topology adds at least 0.02 to the error": "adds >= 0.02",
         "hub topology adds error; 0.02 not shown": "adds",
         "no difference worth acting on": "no difference",
         "hub topology REDUCES the error": "reduces",
         "unclear: the interval is too wide to decide": "unclear"}

if __name__ == "__main__":
    for case in ((0.031, 0.019, 0.043), (0.034, 0.022, 0.046), (0.012, 0.001, 0.024),
                 (0.002, -0.010, 0.014), (-0.015, -0.027, -0.003), (0.010, -0.025, 0.045),
                 (0.030, 0.018, 0.042, 0.24), (0.030, 0.018, 0.042, 0.05, 0.09)):
        print(verdict(*case))
    # Verdict frequencies: 49 analysable groups per arm, SD of the group outcome 0.030,
    # 10,000 simulated trials for each true d (t quantile with 96 df).
    rng = random.Random(44); n, sd, tq = 49, 0.030, 1.9850
    for true in (0.02, 0.03, 0.0):
        tally = {}
        for _ in range(10000):
            a = [rng.gauss(true, sd) for _ in range(n)]
            b = [rng.gauss(0, sd) for _ in range(n)]
            ma, mb = sum(a) / n, sum(b) / n
            ss = sum((x - ma) ** 2 for x in a) + sum((x - mb) ** 2 for x in b)
            d, se = ma - mb, math.sqrt(ss / (2 * n - 2) * 2 / n)
            v = verdict(d, d - tq * se, d + tq * se).split(": ", 1)[1]
            tally[SHORT[v]] = tally.get(SHORT[v], 0) + 1
        ranked = sorted(tally.items(), key=lambda kv: -kv[1])
        print(f"true d = {true:+.2f}: "
              + ", ".join(f"{k} {c / 100:.1f} %" for k, c in ranked))
# d = +0.031 (+0.019 to +0.043): hub topology adds error; 0.02 not shown
# d = +0.034 (+0.022 to +0.046): hub topology adds at least 0.02 to the error
# d = +0.012 (+0.001 to +0.024): hub topology adds error; 0.02 not shown
# d = +0.002 (-0.010 to +0.014): no difference worth acting on
# d = -0.015 (-0.027 to -0.003): hub topology REDUCES the error
# d = +0.010 (-0.025 to +0.045): unclear: the interval is too wide to decide
# d = +0.030: NOT INTERPRETABLE - 24 % of groups lost (limit 20 %)
# d = +0.030: NOT INTERPRETABLE - scripted seat missing in 9 % of rounds
# true d = +0.02: adds 88.0 %, unclear 6.8 %, no difference 2.6 %, adds >= 0.02 2.6 %
# true d = +0.03: adds 62.6 %, adds >= 0.02 37.2 %, unclear 0.2 %, no difference 0.0 %
# true d = +0.00: no difference 80.7 %, unclear 14.0 %, reduces 2.7 %, adds 2.6 %

The first eight lines are invented examples. If more than 20 % of randomised groups are unanalysable, or the programmed seat failed to display in more than 5 % of group-rounds, the report says “not interpretable” whatever the result. Only when the lower limit of the 95 % confidence interval for dd is at least 0.02 does the report say “the centralised structure increases the error by at least 0.02”. Simulation shows that with 49 groups per arm and a true dd of exactly 0.02, in 10,000 simulated trials 88.0 % conclude “increases the error, 0.02 not shown” and 2.6 % reach the strongest conclusion, 90.6 % together; with a true dd of 0.03, 37.2 % reach the strongest conclusion; with no true difference, 80.7 % conclude “no difference worth acting on”, 2.6 % wrongly conclude “increases the error” and 2.7 % “centralisation reduces the error”. Secondary outcomes are reported only as estimates and intervals and are not used to overturn the primary conclusion.

What this trial cannot show

First, the conclusion holds only for the stated bias (the programmed seat shows twice the truth); the model shows that with a smaller bias the direction may reverse, and the size and direction of a real central person’s bias are not fixed. Second, a programmed seat is not a real leader: a real central person gives reasons, has a reputation and can be persuaded; this trial tests only the effect of network position itself. Third, this is an abstract online estimation task and cannot be generalised to real organisations, communities or public opinion, still less used to support or oppose any political claim. Fourth, the trial has only 3 rounds; in the DeGroot model a seat that never updates eventually pulls any connected network to its value, so the trial compares a difference in speed over a finite number of rounds. Fifth, the network structure is imposed by the experimenter, whereas real networks are chosen by people themselves, and the effects may differ. Sixth, participants on online platforms do not represent the general population.

Ethics and conflicts of interest

This is a low-risk study: the task is abstract quantity estimation, involves no health intervention and collects no health information or identifiable personal information. The trial must be reviewed and approved by an independent ethics committee, every participant gives informed consent online, and the protocol and analysis plan are registered before the first group is enrolled (Nosek et al. 2018). The consent form states in advance that “some displayed estimates may be generated by a program” but not which seat; this is a limited withholding of information approved by the ethics committee, and after the task all participants are told the position and bias of the programmed seat and the purpose of the study. Anyone may withdraw at any time and is paid for the part completed. Data are stored anonymously and used only to compute the pre-specified measures. The results will be published whether positive or negative.

Disclosure

The author uses the works of 5D Library, including this book, as the methodological basis of his community courses, the 5D Ascension Cards (Dimensional Shift Card Decks) and the Yafu AI knowledge base, and therefore has a commercial interest in the subject of this book. The book promotes no course, card deck or service; the Chapter 8 trial is designed to be run and analysed by researchers independent of the author’s business and does not recruit from the author’s community, students or customers; all the code and the study protocol are published openly for anyone to check. The author founded and runs the Yafu Global and Asia Influencer X community businesses and has a direct commercial interest in the subject of this book. The book is not management consulting and makes no political claim of any kind.

References

Becker, Joshua, Devon Brackbill, and Damon Centola. 2017. “Network Dynamics of Social Influence in the Wisdom of Crowds.” Proceedings of the National Academy of Sciences 114 (26): E5070–76. https://doi.org/10.1073/pnas.1615978114.
Campbell, Marion K., Gilda Piaggio, Diana R. Elbourne, and Douglas G. Altman. 2012. “CONSORT 2010 Statement: Extension to Cluster Randomised Trials.” BMJ 345: e5661. https://doi.org/10.1136/bmj.e5661.
Lorenz, Jan, Heiko Rauhut, Frank Schweitzer, and Dirk Helbing. 2011. “How Social Influence Can Undermine the Wisdom of Crowd Effect.” Proceedings of the National Academy of Sciences 108 (22): 9020–25. https://doi.org/10.1073/pnas.1008636108.
Nosek, Brian A., Charles R. Ebersole, Alexander C. DeHaven, and David T. Mellor. 2018. “The Preregistration Revolution.” Proceedings of the National Academy of Sciences 115 (11): 2600–2606. https://doi.org/10.1073/pnas.1708274114.

Fan, Y. (2026). Collective Dynamics, Coupled-Oscillator Synchrony and Distributed Consensus: From the Kuramoto Phase Transition and Vicsek Flocking to DeGroot Spectral Convergence, Read Beside Classical Chinese Thought on Harmony and Conformity [集群动力学、耦合振子同步与分布式共识的数学模型] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23828063 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).