5D Library › Scaling Laws, Allometric Growth and the Statistical Physics of Organisational Metabolism · 中文

Chapter 4 The Mortality of Companies: Constant Hazard, Half-Lives and Administrative Overhead

Do companies “die of old age”?

In his 2017 book Scale West argued that the socioeconomic output of cities grows superlinearly, so that cities seldom perish, whereas many indicators of companies scale roughly linearly or even sublinearly, so that companies resemble the organisms of Chapter 2: maintenance costs eventually swallow growth, and companies stagnate and “die” (West 2017). The analogy is suggestive, but it contains two propositions that need to be tested separately. Does the risk that a listed company exits change with its age? And do administrative and coordination costs grow superlinearly with size and drag large companies down?

Evidence

The mortality of listed companies. Daepp, Hamilton, West and Bettencourt (2015) used the records of more than 25,000 North American listed companies in the Compustat database from 1950 to 2009 to define corporate “birth” (entry into the database) and “death” (disappearance from it), and found that the mortality of listed companies is approximately a constant hazard over long periods of observation, independent of the company’s age; the typical half-life is about a decade, similar across sectors. They stressed that the most common cause of corporate “death” is disappearance through mergers and acquisitions, not bankruptcy or liquidation (Daepp et al. 2015). In Chapter 9 of Scale, “Toward a Science of Companies”, West gives the figures of 28,853 companies traded on US markets from 1950 to 2009 and a half-life of about 10.5 years (West 2017).

This differs from a classic finding of organisational sociology, the “liability of newness”: Freeman, Carroll and Hannan (1983) found in samples of labour unions, newspapers and semiconductor firms that organisational death rates fall with age (Freeman et al. 1983). The samples differ greatly: listed companies have already survived their start-up years, and their “deaths” include acquisitions, whereas studies of the liability of newness follow organisations from the day they are founded. The program below will also show that even when no company’s risk changes with age, the risk of a population can fall with age.

Evidence

Size and volatility. Stanley and colleagues (1996), analysing US listed manufacturing companies, found that the standard deviation of a company’s annual growth rate falls with its size as a power law with exponent about 0.15: large companies grow more steadily (Stanley et al. 1996). Axtell (2001), using census data on all US firms, showed that the distribution of firm sizes is close to a Zipf distribution (Axtell 2001). Volume 20 discussed where power-law distributions come from; they are only cited here.

The tools of survival analysis

Calculation

Let TT be the time a company survives after listing. The survival function is S(t)=P(T>t)S(t) = P(T > t), and the hazard function h(t)h(t) is the rate at which companies still alive at tt exit in the next instant. If h(t)=λh(t) = \lambda is constant, S(t)=e−λtS(t) = e^{-\lambda t} and the half-life is ln⁡2/λ\ln 2/\lambda; age does not change the risk, which is the mathematical meaning of “not dying of old age”. The Weibull distribution S(t)=e−(t/θ)kS(t) = e^{-(t/\theta)^k} generalises this case (Weibull 1951): a shape parameter k>1k > 1 means a hazard that rises with age (ageing), k<1k < 1 one that falls, and k=1k = 1 a constant hazard. For companies still alive when observation ends we know only that they have survived “at least this long”; this is right-censoring, and the Kaplan–Meier estimator Ŝ(t)=∏ti≤t(1−di/ni)\hat S(t) = \prod_{t_i \le t} (1 – d_i/n_i) handles it correctly, where nin_i is the number of companies still under observation just before tit_i and did_i the number exiting at tit_i (Kaplan and Meier 1958).

The program generates 20,000 synthetic listed companies (not Compustat data), listing at times spread uniformly over 1950–2009 and observed until the end of 2009; the combined hazard of all exits is constant at ln⁡2/10.5\ln 2 / 10.5, and 60 % of exits are acquisitions (an illustrative share). It then estimates the half-life, the hazard by age band and the Weibull shape parameter; looks at what happens when acquisitions are treated as censored; and finally mixes two kinds of company, each with its own constant hazard (fragile ones at 0.15 a year, robust ones at 0.03).

# Survival analysis of SYNTHETIC listed firms (not Compustat data). 20,000 firms list at
# random times in 1950-2009 and are observed until the end of 2009 (right-censoring).
# Every exit (acquisition, merger, failure, other) has the same constant hazard
# ln 2 / 10.5 per year. The share of exits that are acquisitions (60 %) is ILLUSTRATIVE.
import math, random
rng = random.Random(2015)
LAM = math.log(2) / 10.5

def kaplan_meier(data):
    """data: list of (time, event 1/0). Returns list of (time, S(t)) at event times."""
    data = sorted(data); n, s, out, i = len(data), 1.0, [], 0
    while i < len(data):
        t, d, c = data[i][0], 0, 0
        while i < len(data) and data[i][0] == t:
            d += data[i][1]; c += 1; i += 1
        if d: s *= 1 - d / n; out.append((t, s))
        n -= c
    return out

def median(km):
    return next(t for t, s in km if s <= 0.5)

def weibull_shape(data):
    """Censored-data MLE of the Weibull shape k (profile likelihood, bisection)."""
    ev = [t for t, d in data if d]; D = len(ev); sl = sum(math.log(t) for t in ev)
    def score(k):
        tk = [t ** k for t, _ in data]
        s = sum(a * math.log(t) for a, (t, _) in zip(tk, data))
        return D / k + sl - D * s / sum(tk)
    lo, hi = 0.2, 5.0
    for _ in range(50):
        mid = (lo + hi) / 2
        lo, hi = (mid, hi) if score(mid) > 0 else (lo, mid)
    return (lo + hi) / 2

def firms(hazards):
    out = []
    for _ in range(20000):
        lam = rng.choice(hazards); entry = rng.uniform(1950, 2010)
        life, window = rng.expovariate(lam), 2010 - entry
        out.append((min(life, window), int(life <= window), life <= window
                    and rng.random() < 0.6, entry))
    return out

def hazard_by_age(data, bands=((0, 5), (5, 10), (10, 20), (20, 40))):
    res = []
    for lo, hi in bands:
        exposure = sum(min(t, hi) - lo for t, _ in data if t > lo)
        events = sum(d for t, d in data if lo < t <= hi)
        res.append(events / exposure)
    return res

if __name__ == "__main__":
    f = firms([LAM]); data = [(t, d) for t, d, _, _ in f]
    km = kaplan_meier(data)
    print(f"constant hazard: {sum(d for _, d in data)} exits among 20,000 firms;"
          f" KM half-life {median(km):.1f} years (true 10.5)")
    print("   hazard by age 0-5, 5-10, 10-20, 20-40 years: "
          + ", ".join(f"{h:.3f}" for h in hazard_by_age(data))
          + f"  (true {LAM:.3f})")
    print(f"   Weibull shape k = {weibull_shape(data):.2f} (k = 1: no ageing)")
    fail = [(t, 1 if d and not acq else 0) for t, d, acq, _ in f]
    print(f"if acquisitions are treated as censored, 'failure half-life' ="
          f" {median(kaplan_meier(fail)):.1f} years")
    old = [x for x in f if x[3] <= 1990]               # firms with a full 20-year window
    fl = sum(1 for t, d, acq, _ in old if d and not acq and t <= 20) / len(old)
    ex = sum(1 for t, d, _, _ in old if d and t <= 20) / len(old)
    print(f"   firms listed by 1990: within 20 years {ex:.1%} exited, {fl:.1%} failed")
    mix = firms([0.15, 0.03]); dm = [(t, d) for t, d, _, _ in mix]
    print("mixture of fragile (0.15/yr) and robust (0.03/yr) firms, each constant:")
    print("   pooled hazard by age: " + ", ".join(f"{h:.3f}" for h in hazard_by_age(dm))
          + f";  Weibull k = {weibull_shape(dm):.2f}")
# constant hazard: 15072 exits among 20,000 firms; KM half-life 10.5 years (true 10.5)
#    hazard by age 0-5, 5-10, 10-20, 20-40 years: 0.067, 0.065, 0.067, 0.067  (true 0.066)
#    Weibull shape k = 0.99 (k = 1: no ageing)
# if acquisitions are treated as censored, 'failure half-life' = 26.0 years
#    firms listed by 1990: within 20 years 73.8% exited, 28.9% failed
# mixture of fragile (0.15/yr) and robust (0.03/yr) firms, each constant:
#    pooled hazard by age: 0.081, 0.064, 0.048, 0.035;  Weibull k = 0.81

Under censoring, the Kaplan–Meier estimate correctly recovers the half-life of 10.5 years; the hazard in all four age bands lies between 0.065 and 0.067, and the Weibull k=0.99k = 0.99. This is the pattern reported by Daepp and colleagues.

The next two lines show the difference between “exit” and “death”. If acquired companies are treated as censored and only bankruptcies and liquidations are counted, the “bankruptcy half-life” is 26.0 years; but this number describes an imaginary world without acquisitions. Among companies listed before 1990, which have a full 20-year window of observation, 73.8 % exited within 20 years, but only 28.9 % actually failed. So “the half-life of a listed company is about a decade” must not be turned into “half of all companies go bust within ten years”: the businesses, employees and brands of acquired companies often live on.

The last line is a statistical trap. Neither kind of company has a hazard that changes with age, but together the hazard of the population falls from 0.081 to 0.035, with Weibull k=0.81k = 0.81, which looks like “the older, the safer”. The reason is that fragile companies exit first, and those that remain are increasingly the robust ones. A hazard that falls with age may therefore reflect learning by organisations, or merely selection.

Administrative overhead and survival: a schematic model

Calculation

The following is a schematic model, not an estimate. A company of NN people has a profit margin of 20 % before administrative and coordination costs; these costs take a share s(N)=0.05(N/10)δs(N) = 0.05 (N/10)^{\delta} of revenue, so that their total grows as N1+δN^{1+\delta}. Each year the margin also receives a random shock with standard deviation 0.5N−0.150.5 N^{-0.15} (large companies are steadier; the exponent is borrowed from Stanley and colleagues’ result on growth rates, and applying it to profits is an assumption). The company exits when cumulative losses exceed half a year’s revenue. The program compares 20-year survival across sizes and values of δ\delta.

# A SCHEMATIC model (not an estimate) of firm size, administrative overhead and survival.
# A firm of N people earns a margin of 20 % before overhead; overhead takes a share
# s(N) = 0.05 (N/10)^delta of revenue, i.e. total overhead grows as N^(1 + delta). Each
# year the profit rate also gets a shock with SD 0.5 N^(-0.15) (larger firms are steadier;
# the exponent is borrowed from the growth-rate study of Stanley et al. 1996). The firm
# exits when cumulative losses exceed a buffer of half a year's revenue. 20 years.
import random
rng = random.Random(1996)

def survival(N, delta, years=20, runs=4000, buffer=0.5):
    margin = 0.20 - 0.05 * (N / 10) ** delta
    sd = 0.5 * N ** -0.15
    alive = 0
    for _ in range(runs):
        e = buffer
        for _ in range(years):
            e += margin + rng.gauss(0, sd)
            if e < 0: break
        else:
            alive += 1
    return margin, alive / runs

if __name__ == "__main__":
    sizes = (10, 100, 1000, 10000, 100000)
    print("overhead exponent delta | 20-year survival for N = "
          + ", ".join(f"{n:,}" for n in sizes))
    for delta in (0.0, 0.10, 0.15, 0.20):
        row = [survival(n, delta) for n in sizes]
        print(f"   {delta:.2f}  survival " + "  ".join(f"{s:6.1%}" for _, s in row))
        print("         margin   " + "  ".join(f"{m:+6.2f}" for m, _ in row))
# overhead exponent delta | 20-year survival for N = 10, 100, 1,000, 10,000, 100,000
#    0.00  survival  82.3%   95.5%   99.6%  100.0%  100.0%
#          margin    +0.15   +0.15   +0.15   +0.15   +0.15
#    0.10  survival  83.1%   93.8%   99.1%  100.0%  100.0%
#          margin    +0.15   +0.14   +0.12   +0.10   +0.07
#    0.15  survival  82.5%   92.7%   97.9%   98.9%   83.9%
#          margin    +0.15   +0.13   +0.10   +0.06   +0.00
#    0.20  survival  81.0%   92.0%   95.0%   70.7%    0.0%
#          margin    +0.15   +0.12   +0.07   +0.00   -0.12

With δ=0\delta = 0, administrative costs are proportional to headcount and size brings only benefits: a 10-person company survives 20 years with probability 82.3 %, companies of 1,000 or more almost surely, and small companies die mainly of volatility. With δ=0.15\delta = 0.15, administrative costs eat the margin of a 100,000-person company down to about 0, and survival falls from 98.9 % at 10,000 people to 83.9 %; with δ=0.20\delta = 0.20, a 10,000-person company is already down to 70.7 %, and a 100,000-person company has a margin of −0.12 and always exits. Survival thus first rises and then falls with size: small companies die of volatility, large ones of coordination costs. Where the turning point lies depends entirely on δ\delta, and this book has no reliable estimate of δ\delta for real companies.

Hypothesis · not established

The superlinear coordination-cost hypothesis (not yet established). The author conjectures that within an industry, spending on management, administration and internal coordination grows with headcount with an exponent greater than 1 (δ>0\delta > 0), and that companies with larger δ\delta are at higher risk of exiting through financial distress (excluding acquisitions at a premium) over the following decade. If, in pre-registered company panel data, the upper limit of the 95 % confidence interval of this exponent does not exceed 1, or δ\delta is unrelated to subsequent exit through distress, the hypothesis is refuted.

Summary

The exit hazard of listed companies is approximately constant and independent of age, with a half-life of about a decade; but most “exits” are acquisitions and mergers, not failures. Survival analysis must handle censoring and competing risks correctly, and heterogeneity within a population can create the illusion that “the older, the safer”. The schematic model shows that if administrative and coordination costs grow superlinearly with size, survival first rises and then falls with size, with a turning point set entirely by that exponent. The next chapter narrows the view from firms to teams: the more people, the less each produces?

References

Axtell, Robert L. 2001. “Zipf Distribution of U.S. Firm Sizes.” Science 293 (5536): 1818–20. https://doi.org/10.1126/science.1062081.
Daepp, Madeleine I. G., Marcus J. Hamilton, Geoffrey B. West, and Luı́s M. A. Bettencourt. 2015. “The Mortality of Companies.” Journal of the Royal Society Interface 12 (106): 20150120. https://doi.org/10.1098/rsif.2015.0120.
Freeman, John, Glenn R. Carroll, and Michael T. Hannan. 1983. “The Liability of Newness: Age Dependence in Organizational Death Rates.” American Sociological Review 48 (5): 692–710. https://doi.org/10.2307/2094928.
Kaplan, Edward L., and Paul Meier. 1958. “Nonparametric Estimation from Incomplete Observations.” Journal of the American Statistical Association 53 (282): 457–81. https://doi.org/10.1080/01621459.1958.10501452.
Stanley, Michael H. R., Luı́s A. N. Amaral, Sergey V. Buldyrev, et al. 1996. “Scaling Behaviour in the Growth of Companies.” Nature 379 (6568): 804–6. https://doi.org/10.1038/379804a0.
Weibull, Waloddi. 1951. “A Statistical Distribution Function of Wide Applicability.” Journal of Applied Mechanics 18 (3): 293–97. https://doi.org/10.1115/1.4010337.
West, Geoffrey B. 2017. Scale: The Universal Laws of Growth, Innovation, Sustainability, and the Pace of Life in Organisms, Cities, Economies, and Companies. Penguin Press.

Fan, Y. (2026). Scaling Laws, Allometric Growth and the Statistical Physics of Organisational Metabolism: From Kleiber's Three-Quarter Power Law and Fractal Supply Networks to the Mortality of Firms, Read Beside Classical Chinese Organisational Thought [规模法则、异速生长与组织新陈代谢的统计物理] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23788553 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).