What does Causal Graphs, Suspense Entropy and Foreshadowing conclude, in numbers, about “The Survivors Who Keep Reading: Reader Heterogeneity, Survivor Bias and Censoring”?

Answer: Chapter 6 shows that reader heterogeneity makes late hooks look spectacular in naive comparisons (a ratio of 0.194 for a true effect of zero), and that a naive ratio puts retention at chapter 50 at 0.0918 where the product-limit estimate gives 0.1656.

Scope of the numbers: Its serials, readers and parameters are illustrative models; all calculations are generated by its programs; it is not legal, financial or investment advice, and it keeps one red line: no false suspense to win clicks. Marked boxes separate published evidence, runnable Python calculations (21 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 onlyprinted values that match the answer: 0.0918, 0.1656, 50

readers 5,000; still reading on day 120 (censored): 1,038
chaptertrue S(t)naive shareKaplan-Meier
50.61210.58720.6063
100.45260.40280.4372
250.26600.20740.2629
500.16550.09180.1656
800.11740.03720.1162
naive 'dropout rate' (stopped / joined): 0.792
Show the full program
# Chapter 6. p63: readers still reading are censored, not lost.
# 5,000 readers join a serial on a day drawn uniformly from 0 to 119 and
# read one chapter a day from chapter 1, each leaving after a chapter
# with a constant theta ~ Beta(0.8, 6) (the p61 model). The data are
# read on day 120, so a reader who joined on day d can have read at most
# 120 - d chapters: if still reading, that reader is right-censored.
# "Naive" counts everyone who reached chapter t over all who joined; the
# Kaplan and Meier (1958) product-limit estimate uses only those still
# at risk. Fixed-seed simulation; illustrative parameters.
import math
import random

A, B, N, DAY = 0.8, 6.0, 5000, 120
rng = random.Random(20261112)
obs = []                                   # (chapters read, still reading)
for _ in range(N):
    d = rng.randrange(DAY)
    th = rng.betavariate(A, B)
    cap, t = DAY - d, 0
    while t < cap:
        t += 1
        if rng.random() < th:
            obs.append((t, False))
            break
    else:
        obs.append((cap, True))


def true_s(t):
    lb = (math.lgamma(B + t) - math.lgamma(A + B + t)
          - math.lgamma(B) + math.lgamma(A + B))
    return math.exp(lb)


def km(t):
    s = 1.0
    for c in range(1, t + 1):
        at_risk = len([1 for x, _ in obs if x >= c])
        quit_here = len([1 for x, cen in obs if x == c and not cen])
        if at_risk:
            s *= 1 - quit_here / at_risk
    return s


cen = len([1 for _, c in obs if c])
print("readers {:,}; still reading on day {} (censored): {:,}".format(
    N, DAY, cen))
print("chapter  true S(t)  naive share  Kaplan-Meier")
for t in (5, 10, 25, 50, 80):
    naive = len([1 for x, c in obs if x > t or (x == t and c)]) / N
    print("{:>7}  {:>9.4f}  {:>11.4f}  {:>12.4f}".format(
        t, true_s(t), naive, km(t)))
print("naive 'dropout rate' (stopped / joined): {:.3f}".format(
    (N - cen) / N))

Single source and citation

Fan, Y. (2026). Causal-DAG Validation of Long-Form Narrative, Sawtooth Scheduling of Chapter Suspense Entropy and Forgetting-Aware Foreshadowing Recovery (Version 1.0.0), Chapter 6: The Survivors Who Keep Reading: Reader Heterogeneity, Survivor Bias and Censoring. 5D Library. DOI pending

DOI pending · monograph page goes live after publication · 5DL-B-178 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026178,
  author    = {Fan, Yafu},
  title     = {Causal-DAG Validation of Long-Form Narrative, Sawtooth Scheduling of Chapter Suspense Entropy and Forgetting-Aware Foreshadowing Recovery},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

Other chapter answers in this book