5D Library › 诊学一体转化漏斗、掌握度门控课程与持牌人网络经济学 · English

第五章 及格线是质量过滤器:静态最优与声誉动态

牌照卖的是什么

执业牌照的经济学有两种经典解释。Leland 把牌照看作最低质量标准:当客户无法事先判断从业者的水平时,劣质从业者会把整个市场拖下去,最低标准可以阻止这一点,但也会减少供给、抬高价格 (Leland 1979)。Shapiro 则指出,高质量产品的溢价是对声誉的回报:卖家之所以维持高质量,是因为声誉一旦受损,未来的溢价就没了 (Shapiro 1983)。Kleiner 与 Krueger 的调查显示,职业牌照在美国覆盖了相当大比例的劳动力,并与更高的工资相关 (Kleiner and Krueger 2013)。

对于一张由企业而不是政府颁发的牌照,这两种解释给出两个可以计算的问题:及格线定在哪里,持牌人群体的质量与规模如何变化?品牌溢价能否补偿被拒之门外的牌照费?

静态的一年

每年 1,000 位候选人走完课程,掌握度为基础掌握度加 0.08(第四章的课程提升),考一次第三章的试,及格线 TT 从 10 分到 20 分。客户为持牌执业者的每次服务支付带品牌溢价的价格 80exp⁡(4(Q−0.75))80\exp(4(Q – 0.75)),其中 QQ 是持牌群体的平均掌握度;每位持牌人每年服务 100 次;企业收 600 的牌照费外加 10% 的分成。一次服务出问题的概率设为 0.1(1−m)20.1(1-m)^2。

推导

# Chapter 5. [p09] The pass mark as a quality filter: a static optimum.
# 1,000 candidates a year arrive with mastery m = theta + 0.08 (theta ~
# Beta(6, 3), the post-course lift of p08), capped at 0.97. They sit
# the exam of p05 once; the pass mark T runs from 10 to 20 points.
# Certified pool: size N(T) and mean mastery Q(T), integrated on a grid.
# Clients pay a certified practitioner a session price that carries a
# brand premium, price = 80 exp(4 (Q - 0.75)); each practitioner holds
# 100 sessions a year. The firm earns a licence fee of 600 plus a 10%
# royalty. Clients' harm: a session goes badly with chance 0.1 (1-m)^2.
import math

K = 6.0

def bb_pmf(n, a, b):
    lb = lambda x, y: math.lgamma(x) + math.lgamma(y) - math.lgamma(x + y)
    return [math.exp(math.lgamma(n + 1) - math.lgamma(j + 1)
                     - math.lgamma(n - j + 1)
                     + lb(j + a, n - j + b) - lb(a, b)) for j in range(n + 1)]

def total_pmf(m):
    one = bb_pmf(4, K * m, K * (1 - m))
    tot = [1.0]
    for _ in range(5):
        nxt = [0.0] * (len(tot) + 4)
        for i, x in enumerate(tot):
            for j, y in enumerate(one):
                nxt[i + j] += x * y
        tot = nxt
    return tot

def pool(T, grid=400):
    n = q = harm = 0.0
    for i in range(grid):
        th = (i + 0.5) / grid
        w = 168 * th ** 5 * (1 - th) ** 2 / grid
        m = min(0.97, th + 0.08)
        pp = sum(total_pmf(m)[T:])
        n += w * pp
        q += w * pp * m
        harm += w * pp * 0.1 * (1 - m) ** 2
    return 1000 * n, q / n, harm / n

if __name__ == "__main__":
    print("    T   certified  mean m  price   bad/100   firm profit")
    best = None
    for T in range(10, 21):
        n, Q, h = pool(T)
        price = 80 * math.exp(4 * (Q - 0.75))
        prof = n * (600 + 0.10 * 100 * price)
        best = max(best or (prof, T), (prof, T))
        print("   {:2d} {:11.1f} {:7.4f} {:6.2f} {:9.3f} {:13.0f}".format(
            T, n, Q, price, 100 * h, prof))
    print("   profit-maximising pass mark: T = {}".format(best[1]))
    print("\nReading: in a static year the firm earns most by passing")
    print("almost everyone: the premium a stricter mark buys is too")
    print("small to offset the licences it forgoes. A static model")
    print("cannot justify the 16/20 mark; only reputation over time")
    print("can (p10).")
#     T   certified  mean m  price   bad/100   firm profit
#    10       909.4  0.7678  85.90     0.707       1326784
#    11       868.0  0.7762  88.85     0.657       1291977
#    12       815.2  0.7863  92.52     0.599       1243348
#    13       750.5  0.7981  96.99     0.537       1178226
#    14       674.0  0.8116 102.36     0.471       1094334
#    15       586.7  0.8268 108.76     0.402        990152
#    16       490.7  0.8436 116.33     0.333        865282
#    17       389.0  0.8621 125.26     0.265        720647
#    18       285.5  0.8822 135.75     0.199        558862
#    19       184.5  0.9038 147.99     0.138        383736
#    20        89.4  0.9264 162.01     0.084        198401
#    profit-maximising pass mark: T = 10
#
# Reading: in a static year the firm earns most by passing
# almost everyone: the premium a stricter mark buys is too
# small to offset the licences it forgoes. A static model
# cannot justify the 16/20 mark; only reputation over time
# can (p10).

及格线从 10 提到 16,持牌人从 909.4 人降到 490.7 人,平均掌握度从 0.7678 升到 0.8436,服务价格从 85.90 升到 116.33,每百次服务出问题的次数从 0.707 降到 0.333。但企业的年利润从 1326784 降到 865282。静态的一年里,利润最高的及格线是 10:更严格的及格线带来的溢价,抵不上它拒掉的牌照费。

这个结论与作者的提纲相反,但它是诚实的:只要客户的支付意愿对质量的反应不够强,单看一年,企业的利益就是放宽及格线。静态模型无法为 16/20 辩护。

五年的声誉

静态模型漏掉了一样东西:出问题的服务会传出去。程序让每次出问题的服务以概率 ss 变成公开的负面事件,每个事件使品牌声誉 RR 下降 0.004;声誉每年向持牌群体的平均掌握度靠拢三分之一;服务价格与下一年的候选人数都随声誉变化。持牌人每年以 0.8 的概率续期。按 10% 贴现,计算五年的净现值(NPV),每个及格线模拟 50 次取平均。

推导

# Chapter 5. [p10] Reputation over five years: the pass mark and NPV.
# Year by year: candidates arrive (1,000 exp(3 (R - 0.75))), each with
# mastery m = min(0.97, theta + 0.08), theta ~ Beta(6, 3), and pass the
# p05 exam at mark T. Licensees renew with probability 0.8. Each holds
# 100 sessions; a session goes badly with chance 0.1 (1 - m)^2, and each
# bad session becomes a public scandal with chance s = 0.02. The brand's
# reputation R (in mastery units, R0 = 0.75) moves a third of the way
# toward the pool's mean mastery and loses 0.004 per scandal. Session
# price 80 exp(4 (R - 0.75)); the firm takes 600 per licence plus 10%.
# NPV at 10% over 5 years, mean of 50 runs; also R in year 5.
import math
import random

K, YEARS, RUNS = 6.0, 5, 50

def pass_prob(m, T, memo={}):
    key = (round(m, 3), T)
    if key not in memo:
        a, b = K * m, K * (1 - m)
        lb = lambda x, y: (math.lgamma(x) + math.lgamma(y)
                           - math.lgamma(x + y))
        one = [math.exp(math.lgamma(5) - math.lgamma(j + 1)
                        - math.lgamma(5 - j) + lb(j + a, 4 - j + b)
                        - lb(a, b)) for j in range(5)]
        tot = [1.0]
        for _ in range(5):
            nxt = [0.0] * (len(tot) + 4)
            for i, x in enumerate(tot):
                for j, y in enumerate(one):
                    nxt[i + j] += x * y
            tot = nxt
        memo[key] = sum(tot[T:])
    return memo[key]

def run(rng, T, s=0.02):
    R, pool, npv = 0.75, [], 0.0
    for y in range(YEARS):
        for _ in range(int(1000 * math.exp(3 * (R - 0.75)))):
            m = min(0.97, rng.betavariate(6, 3) + 0.08)
            if rng.random() < pass_prob(m, T):
                pool.append(m)
        price = 80 * math.exp(4 * (R - 0.75))
        npv += len(pool) * (600 + 10 * price) / 1.1 ** y
        scandals = 0
        for m in pool:
            lam = 100 * 0.1 * (1 - m) ** 2 * s
            scandals += rng.random() < 1 - math.exp(-lam)
        qbar = sum(pool) / len(pool) if pool else R
        R = R + (qbar - R) / 3 - 0.004 * scandals
        pool = [m for m in pool if rng.random() < 0.8]
    return npv, R

if __name__ == "__main__":
    rng = random.Random(74010)
    print("    T   NPV (mean)   R year 5   NPV with s = 0.05")
    res = {}
    for T in range(10, 21, 2):
        a = [run(rng, T) for _ in range(RUNS)]
        b = [run(rng, T, 0.05) for _ in range(RUNS)]
        nv = sum(x for x, _ in a) / len(a)
        rr = sum(r for _, r in a) / len(a)
        nb = sum(x for x, _ in b) / len(b)
        res[T] = (nv, nb)
        print("   {:2d} {:12.0f} {:10.4f} {:19.0f}".format(T, nv, rr, nb))
    print("   best T (s = 0.02): {}   best T (s = 0.05): {}".format(
        max(res, key=lambda t: res[t][0]), max(res, key=lambda t: res[t][1])))
    print("\nReading: once scandals feed back into reputation, a lax")
    print("mark erodes the brand that pays every licensee, and the")
    print("best mark moves up; how far depends on how often a bad")
    print("session goes public, which no model can supply. Within five")
    print("years 16/20 is not the NPV optimum; it is the lowest mark")
    print("shown that keeps reputation near where it started.")
#     T   NPV (mean)   R year 5   NPV with s = 0.05
#    10      7586147     0.4953             5365145
#    12      7693477     0.5609             5463101
#    14      7489253     0.6415             5584471
#    16      6625999     0.7392             5391765
#    18      4818147     0.8308             4375957
#    20      1792803     0.8974             1782396
#    best T (s = 0.02): 12   best T (s = 0.05): 14
#
# Reading: once scandals feed back into reputation, a lax
# mark erodes the brand that pays every licensee, and the
# best mark moves up; how far depends on how often a bad
# session goes public, which no model can supply. Within five
# years 16/20 is not the NPV optimum; it is the lowest mark
# shown that keeps reputation near where it started.

当 s=0.02s = 0.02 时,五年 NPV 最高的及格线是 12(7693477),及格线 10 的第五年声誉已跌到 0.4953;当负面事件更容易传开(s=0.05s = 0.05)时,最优移到 14(5584471)。及格线 16 在五年内不是 NPV 最优,但它是表中能让声誉大致保持在起点 0.75 附近的最低及格线(第五年 0.7392)。

所以,声誉反馈确实把最优及格线往上推,推多远取决于负面事件传开的概率 ss,而这个概率没有任何模型能代为给出。若企业打算经营十年以上,或者 ss 比本书设想的更大,16/20 会更接近最优;第六章会看到,加入客户一侧之后,16/20 几乎就是最优。

假说 · 尚未验证

作者的诠释(尚未验证): 在作者的业务中,持牌执业者的一次失误会通过社交媒体迅速传开(ss 高于 0.05),因此及格线应当定在保持声誉的水平,而不是单年利润最高的水平。什么会证明这一判断错:若真实投诉与评价数据显示,持牌执业者的公开负面事件与其考试分数无关,或者品牌评分对负面事件不敏感,那么声誉机制不成立,及格线的选择应回到静态模型。

小结

把及格线从 10 提到 16,持牌群体变小、变好,服务价格上升,但单年利润下降;静态最优是 10。加入声誉反馈后,五年最优移到 12—14,取决于负面事件传开的概率;16/20 是能保住声誉的最低及格线。及格线的选择,本质上是对时间跨度与声誉风险的选择。

参考文献

Kleiner, Morris M., and Alan B. Krueger. 2013. “Analyzing the Extent and Influence of Occupational Licensing on the Labor Market.” Journal of Labor Economics 31 (S1): S173–202.
Leland, Hayne E. 1979. “Quacks, Lemons, and Licensing: A Theory of Minimum Quality Standards.” Journal of Political Economy 87 (6): 1328–46.
Shapiro, Carl. 1983. “Premiums for High Quality Products as Returns to Reputations.” Quarterly Journal of Economics 98 (4): 659–80.

Fan, Y. (2026). Diagnostic-to-Pedagogy Conversion Funnels, Mastery-Gated Curriculum Architecture and Certification Network Economics: Self-Cannibalization Bounds, Spaced-Retrieval Curves and Two-Sided Practitioner Licensing [诊学一体转化漏斗、掌握度门控课程与持牌人网络经济学] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090197 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).