What does Drift and Noise on Long Pages conclude, in numbers, about “Estimating Drift and Noise from Dwell and Payment: EZ-Diffusion and Identifiability”?

Answer: EZ-diffusion turns three numbers, the share of payers and the mean and variance of dwell, into three parameters, drift, barrier separation and non-decision time, in closed form. Program p51 shows that drift is the hardest parameter to estimate: with one hundred decided visitors per version its error exceeds three tenths, and only around sixteen hundred does it fall below a tenth. Program p52 shows that when drift, barrier separation and noise are scaled together nothing observable changes, so noise can be estimated only relative to an anchor; this book's anchor is that visitors' caution does not change with a redesign. The next chapter uses this estimation perspective to discuss the trade-off between speed and accuracy, and the evidence that points against this book.

Scope of the numbers: Chapter 4 pushes a page forward section by section: deleting the feature table of an illustrative page gains 0.0081, and the best and worst orders of sections differ by 0.1764. Marked boxes separate published evidence, runnable Python calculations (20 programs), classical sources and the author's hypotheses.

中文版

Data table

The numbers below are printed verbatim by the chapter's Python program and can be reproduced by running it. standard-library Python only

nv meanv sda meana sdTer meanTer sd
1000.29630.09062.00300.09380.39770.0740
4000.23610.04222.02340.06070.40580.0436
16000.25200.02312.00060.02980.40310.0292
true0.25002.00000.4000
population pay share among the decided: 0.6225
Show the full program
# Chapter 5. p51: recovering drift, barrier separation and non-decision
# time from three summary numbers with the EZ-diffusion equations
# (Wagenmakers, van der Maas and Grasman 2007), noise scale s fixed at 1.
# 输入:已决定访客中的付款比例 Pc、决定时长的均值 MRT 与方差 VRT。
#   L = logit(Pc),x = L (L Pc^2 - L Pc + Pc - 1/2) / VRT,
#   v = sign(Pc - 1/2) s x^(1/4),a = s^2 L / v,
#   MDT = (a / 2v) (1 - exp(-va/s^2)) / (1 + exp(-va/s^2)),Ter = MRT - MDT。
# 真值 v = 0.25、a = 2、Ter = 0.4(加载与扫视时长),起点 a/2;示意。
# 模拟用欧拉步 0.01 与 p11 的连续性校正,每档样本量重复 10 次。
import math
import random
import statistics

V, AA, TER, DT = 0.25, 2.0, 0.4, 0.01


def visitor(rng):
    sq = math.sqrt(DT)
    lo, hi = 0.5826 * sq, AA - 0.5826 * sq
    x, k = AA / 2, 0
    while lo < x < hi:
        x = x + V * DT + sq * rng.gauss(0.0, 1.0)
        k += 1
    return x >= hi, TER + k * DT


def ez(pc, mrt, vrt, s=1.0):
    lg = math.log(pc / (1 - pc))
    x = lg * (lg * pc * pc - lg * pc + pc - 0.5) / vrt
    v = math.copysign(s * x ** 0.25, pc - 0.5)
    a = s * s * lg / v
    y = -v * a / (s * s)
    mdt = (a / (2 * v)) * (1 - math.exp(y)) / (1 + math.exp(y))
    return v, a, mrt - mdt


rng = random.Random(20261113)
print("n      v mean  v sd    a mean  a sd    Ter mean  Ter sd")
for n in (100, 400, 1600):
    est = []
    for _ in range(10):
        data = [visitor(rng) for _ in range(n)]
        pc = len([1 for yes, _ in data if yes]) / n
        ts = [t for _, t in data]
        est.append(ez(pc, statistics.fmean(ts), statistics.variance(ts)))
    cols = list(zip(*est))
    print("%-5d  %.4f  %.4f  %.4f  %.4f  %.4f    %.4f"
          % (n, statistics.fmean(cols[0]), statistics.stdev(cols[0]),
             statistics.fmean(cols[1]), statistics.stdev(cols[1]),
             statistics.fmean(cols[2]), statistics.stdev(cols[2])))
print("true   %.4f          %.4f          %.4f" % (V, AA, TER))
pc0 = 1 / (1 + math.exp(-V * AA))
print("population pay share among the decided: %.4f" % pc0)

Single source and citation

Fan, Y. (2026). Fokker-Planck Drift-Diffusion on Long Sales Pages: Belief-Drift Rates and First-Passage Times to the Payment Barrier (Version 1.0.0), Chapter 5: Estimating Drift and Noise from Dwell and Payment: EZ-Diffusion and Identifiability. 5D Library. DOI pending

DOI pending · monograph page goes live after publication · 5DL-B-240 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026240,
  author    = {Fan, Yafu},
  title     = {Fokker-Planck Drift-Diffusion on Long Sales Pages: Belief-Drift Rates and First-Passage Times to the Payment Barrier},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

Other chapter answers in this book