5D Library › 关系断裂生存分析、断联期记忆衰减与复联博弈窗口 · English

第一章 把复合当作生存分析问题

两个事件互相竞争

分手之后,一对前任可能发生两件事:复合,或者一方、双方另有新生活,例如进入新的关系。一旦有人另有新生活,复合就不再是原来的问题。关于分分合合恋爱关系的研究发现复合相当常见,并不罕见 (Dailey et al. 2009);这里的问题是怎样计数。生存分析是自然的工具,但两个事件互相竞争时必须小心使用。

p01 跟踪一个合成队列:2,000 对前任,跟踪一年。十分之三的伴侣态度摇摆,如果没有别的事发生就会很快复合;其余的已经心意已定。每一对伴侣另有新生活的机会都一样。

推导

# Chapter 1. p01: Who gets back together, and how to count it. A
# synthetic cohort of 2,000 former couples is followed for 365 days.
# Two events compete: reconciliation, and "moving on" (a new partner,
# after which reconciliation is treated as no longer possible). 30% of
# couples are ambivalent (reconciliation hazard 1/60 a day); the rest
# are settled (1/1500 a day). Moving on has hazard 1/200 a day for all.
# Kaplan-Meier treats moving on as censoring and so overstates the
# chance of reconciliation; the cumulative incidence does not. Seed 92.
import random

N, DAYS, P_AMB, H_AMB, H_SET, H_MOVE = 2000, 365, 0.30, 1 / 60, \
    1 / 1500, 1 / 200

def cohort(rng):
    out = []
    for _ in range(N):
        h = H_AMB if rng.random() < P_AMB else H_SET
        tr, tm = rng.expovariate(h), rng.expovariate(H_MOVE)
        if min(tr, tm) > DAYS:
            out.append((DAYS, "censored"))
        elif tr < tm:
            out.append((tr, "reconciled"))
        else:
            out.append((tm, "moved on"))
    return out

def km_and_ci(data, t_end):
    data = sorted(data)
    at_risk, s_km, s_all, ci = len(data), 1.0, 1.0, 0.0
    for t, kind in data:
        if t > t_end:
            break
        if kind == "reconciled":
            ci += s_all / at_risk      # Aalen-Johansen increment
            s_km *= 1 - 1 / at_risk    # KM: moving on as censoring
        if kind != "censored":
            s_all *= 1 - 1 / at_risk   # survival from either event
        at_risk -= 1
    return 1 - s_km, ci

if __name__ == "__main__":
    rng = random.Random(92)
    data = cohort(rng)
    print("   by day   Kaplan-Meier   cumulative incidence   observed share")
    for t in (30, 90, 180, 365):
        km, ci = km_and_ci(data, t)
        obs = sum(1 for d, k in data if k == "reconciled" and d <= t) / N
        print("   {:6d}   {:12.3f}   {:20.3f}   {:14.3f}".format(
            t, km, ci, obs))
    rec = [d for d, k in data if k == "reconciled"]
    print("   reconciliations in the first 90 days: {:.3f} of all".format(
        sum(1 for d in rec if d <= 90) / len(rec)))
    print("   couples who moved on within the year: {:.3f}".format(
        sum(1 for d, k in data if k == "moved on") / N))
#    by day   Kaplan-Meier   cumulative incidence   observed share
#        30          0.137                  0.128            0.128
#        90          0.286                  0.241            0.240
#       180          0.379                  0.290            0.289
#       365          0.456                  0.311            0.310
#    reconciliations in the first 90 days: 0.775 of all
#    couples who moved on within the year: 0.597

常用的 Kaplan-Meier 估计 (Kaplan and Meier 1958) 把另有新生活的伴侣当作「删失」,好像他们以后仍可能复合。它报告第 30 天 0.137、第 90 天 0.286、年底 0.456。但真正复合的比例只有 0.128、0.240 与 0.310。累积发生率 (Aalen and Johansen 1978) 把另有新生活当作使这对伴侣退出问题的事件,得到 0.128、0.241、0.290 与 0.311,与实际情况相符。差距随时间扩大,因为一年内有 0.597 的伴侣另有了新生活。

同一个队列也说明「复合大多发生在早期」为什么既对又容易误导:全部复合中有 0.775 发生在前 90 天。这说的是谁会复合(主要是态度摇摆、很快复合的那些),而不是对任何一个具体的人保证等待没有用。

窗口会关闭吗?

第二个流行的说法是复合的机会在几个月后就会「关闭」。p02 构造一个不可能关闭的世界:每一对伴侣在整整一年里每天复合的机会都不变,只是彼此大小不同。

推导

# Chapter 1. p02: Does the "window" close, or do the likely couples
# simply leave first? Every couple keeps a constant daily hazard of
# reconciling for the whole year (no couple's chance ever declines),
# but couples differ: hazards are drawn from a gamma law with mean
# 1/250 a day and shape k. The hazard seen in the surviving group still
# falls month by month, because the couples most likely to reconcile
# leave the group first. For k = 0.5 the decline is steep. Seed 92.
import random

def life_table(rng, k, n=20000, days=360):
    hs = [rng.gammavariate(k, (1 / 250) / k) for _ in range(n)]
    ts = [rng.expovariate(h) for h in hs]
    rows = []
    for a in range(0, days, 60):
        risk = [t for t in ts if t >= a]
        ev = sum(1 for t in risk if t < a + 60)
        rows.append((a, ev / len(risk) / 60 * 1000))
    return rows

if __name__ == "__main__":
    rng = random.Random(92)
    print("   observed hazard per 1,000 couples per day, by window")
    print("   window (days)   all alike   shape 2.0   shape 0.5")
    tabs = [life_table(rng, k) for k in (1e6, 2.0, 0.5)]
    for i in range(6):
        a = tabs[0][i][0]
        print("   {:3d} to {:3d}      {:9.2f}   {:9.2f}   {:9.2f}".format(
            a, a + 59, tabs[0][i][1], tabs[1][i][1], tabs[2][i][1]))
    f = tabs[2][-1][1] / tabs[2][0][1]
    print("   shape 0.5: last window / first window = {:.3f}".format(f))
#    observed hazard per 1,000 couples per day, by window
#    window (days)   all alike   shape 2.0   shape 0.5
#      0 to  59           3.62        3.38        2.96
#     60 to 119           3.56        3.11        2.14
#    120 to 179           3.57        2.83        1.77
#    180 to 239           3.55        2.65        1.40
#    240 to 299           3.55        2.48        1.19
#    300 to 359           3.58        2.17        1.05
#    shape 0.5: last window / first window = 0.355

所有伴侣都一样时,每个时间段观察到的风险率都在每天每 1,000 对 3.55 至 3.62 之间。伴侣彼此不同时,群体中看到的风险率仍然逐月下降:在差异很大的人群里(形状参数 0.5)从 2.96 开始,降到 1.05,最后一段只有第一段的 0.355。没有任何一对伴侣自己的机会改变过。最可能复合的伴侣先复合、先离开样本,于是仍然分开的伴侣平均而言是较不可能复合的那些。图上一条下降的曲线,因此不能证明时间本身会关上窗口;它可能只是显示了谁留了下来。

假说 · 尚未验证

作者的诠释(尚未验证): 分手后几个月里复合率的下降,大部分反映的是伴侣之间的差异,而不是每一对伴侣自身机会的下降。什么会证明它错: 在一个对每对伴侣反复测量双方联系与复合意愿的队列中,把这些测量固定之后,伴侣自身的复合机会仍随时间下降,下降幅度与群体风险率的下降相当。

小结

计算复合必须把另有新生活当作竞争事件,否则会高估复合的机会;在 p01 里,到年底高估的部分几乎是真实比例的一半。随时间下降的风险率可以完全来自筛选。读到「有多少伴侣复合」的统计并拿来套用在自己身上时,这两点都很重要。

参考文献

Aalen, Odd O., and Søren Johansen. 1978. “An Empirical Transition Matrix for Non-Homogeneous Markov Chains Based on Censored Observations.” Scandinavian Journal of Statistics 5 (3): 141 to 150.
Dailey, René M., Abigail Pfiester, Borae Jin, Gary Beck, and Glenn Clark. 2009. “On-Again/off-Again Dating Relationships: How Are They Different from Other Dating Relationships?” Personal Relationships 16 (1): 23 to 47.
Kaplan, E. L., and Paul Meier. 1958. “Nonparametric Estimation from Incomplete Observations.” Journal of the American Statistical Association 53 (282): 457 to 481.

Fan, Y. (2026). Relationship Rupture Survival Analysis, No-Contact Memory Decay and Reconciliation Game Windows: Negative-Affect Dissipation Rates, Dual-Process Nostalgia Peaks and Signaling Equilibria [关系断裂生存分析、断联期记忆衰减与复联博弈窗口] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090241 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).