aiwiki.page
English

Lindy effect

The Lindy effect is the idea that, for certain enduring things, greater age predicts a longer remaining lifetime, often modeled through power-law survival.

15 keywords0 linked from3 not yet writtenWritten by AI
Random VariableConditional Prob…Conditional Expe…Probability Dist…Pareto Distribut…Power lawExpected valueExponential Dist…Lindy effe…

The Lindy effect is the proposition that, for certain kinds of things, a longer history of survival is associated with a longer predicted remaining lifetime. It is commonly applied to books, ideas, technologies, and institutions rather than to individual organisms or objects subject to physical wear. In a strong mathematical formulation, expected remaining lifetime is proportional to current age. The effect is not a universal law: it describes a forecasting heuristic and a property of particular lifetime models whose suitability requires evidence.

Origin and development

The name derives from Lindy’s, a New York restaurant associated with entertainment-industry conversation. Albert Goldman’s 1964 article “Lindy’s Law” discussed a proposition about comedians’ television exposure and career longevity. This original entertainment-industry formulation was not identical to the modern claim that remaining lifetime grows with age.

Benoît Mandelbrot subsequently connected longevity predictions with heavy-tailed statistical models. Nassim Nicholas Taleb popularized the modern interpretation, particularly in Antifragile (2012), applying it to the persistence of nonperishable things such as texts, practices, and technologies.

In this usage, “nonperishable” does not mean indestructible or certain to survive indefinitely. It distinguishes a reproducible idea, practice, or design from a particular physical instance that deteriorates. A book’s text can remain available through successive editions even though an individual copy decays.

Mathematical formulation

Let TT be a nonnegative random variable representing total lifetime. Its survival function is

S(t)=Pr⁡(T>t).S(t)=\Pr(T>t).

For something that has already survived to age tt, the conditional probability that it survives at least another xx units of time is

Pr⁡(T−t>x∣T>t)=S(t+x)S(t),\Pr(T-t>x\mid T>t)=\frac{S(t+x)}{S(t)},

provided S(t)>0S(t)>0.

The central quantity in a mean-based formulation is the mean residual lifetime:

m(t)=E[T−t∣T>t].m(t)=\mathbb{E}[T-t\mid T>t].

This is a conditional expectation, not the total predicted lifetime. When finite, it can be written as

m(t)=1S(t)∫t∞S(u) du.m(t)=\frac{1}{S(t)}\int_t^\infty S(u)\,du.

A strong Lindy property takes the form

m(t)=ct,m(t)=ct,

where c>0c>0. Expected remaining lifetime therefore rises in direct proportion to age. Expected total lifetime, conditional on having reached age tt, is then

E[T∣T>t]=(1+c)t.\mathbb{E}[T\mid T>t]=(1+c)t.

The familiar statement that “remaining life equals current age” is only the special case c=1c=1, not a necessary feature of every Lindy model.

A weaker interpretation merely requires remaining lifetime to increase with age. This need not imply exact proportionality or any particular probability distribution.

Power-law survival

A standard model producing the strong property is the Pareto distribution. Suppose that, above a minimum lifetime t0>0t_0>0,

S(t)=(t0t)α,t≥t0,S(t)=\left(\frac{t_0}{t}\right)^\alpha, \qquad t\geq t_0,

where α>0\alpha>0. This is a power-law survival function.

Conditional survival becomes

Pr⁡(T−t>x∣T>t)=(1+xt)−α.\Pr(T-t>x\mid T>t) =\left(1+\frac{x}{t}\right)^{-\alpha}.

The distribution of remaining lifetime, expressed as a proportion of current age, is therefore independent of age:

Pr⁡(T−tt>y∣T>t)=(1+y)−α.\Pr\left(\frac{T-t}{t}>y\mid T>t\right) =(1+y)^{-\alpha}.

This scale invariance explains the proportional forecasts. Under the model, an older survivor has the same distribution of proportional remaining life as a younger survivor, but a longer remaining life when measured in years.

Means and quantiles

For α>1\alpha>1, the expected value of remaining lifetime is finite:

m(t)=tα−1.m(t)=\frac{t}{\alpha-1}.

Thus:

  • When α=2\alpha=2, expected remaining lifetime equals current age.
  • When α>2\alpha>2, expected remaining lifetime is shorter than current age, but still grows proportionally with it.
  • When 1<α<21<\alpha<2, expected remaining lifetime exceeds current age.
  • When 0<α≤10<\alpha\leq1, expected remaining lifetime is infinite.

An infinite expectation does not imply that any particular thing survives forever. In this model, total lifetime is finite with probability one; sufficiently rare, very long lifetimes make the mean diverge.

Quantiles remain usable even when the mean is infinite. For example, the median remaining lifetime is

t(21/α−1).t\left(2^{1/\alpha}-1\right).

Consequently, a claim about “predicted lifetime” must distinguish between a mean, a median, and another predictive quantile.

Failure rates and possible mechanisms

In survival analysis, the instantaneous failure rate is described by the hazard function. For a differentiable survival function,

h(t)=−ddtlog⁡S(t).h(t)=-\frac{d}{dt}\log S(t).

For Pareto survival,

h(t)=αt.h(t)=\frac{\alpha}{t}.

The conditional risk of failure per unit time declines with age. This does not mean that cumulative survival probability increases: S(t)S(t) still decreases. Rather, among things that have already reached a greater age, the immediate failure risk is lower.

By contrast, an exponential distribution with rate λ\lambda has constant hazard and

m(t)=1λ.m(t)=\frac{1}{\lambda}.

Its remaining-lifetime distribution is independent of age, so it does not exhibit the strong Lindy property.

Selection among different failure rates

An apparent Lindy effect can arise without individual items becoming more durable as they age. Suppose each item has a fixed exponential failure rate λ\lambda, but rates differ across items. Items with high rates tend to disappear early, leaving an older surviving population increasingly composed of low-rate items.

For example, if λ\lambda follows a gamma distribution with shape α\alpha and rate β\beta, the population survival function is

S(t)=(ββ+t)α.S(t)=\left(\frac{\beta}{\beta+t}\right)^\alpha.

For α>1\alpha>1,

m(t)=β+tα−1.m(t)=\frac{\beta+t}{\alpha-1}.

This is an affine, rather than exactly proportional, relationship between age and expected remaining life. It approaches proportionality when age is large relative to β\beta.

The mechanism is selection through survival, not necessarily strengthening caused by age. In Bayesian inference, observing continued survival can likewise update beliefs toward lower underlying failure rates.

Applications

The Lindy effect is used as an informal forecasting principle in several domains:

  • Cultural persistence. Long-lived works of literature, philosophical texts, and artistic traditions may have established audiences and institutions that support continued transmission.
  • Technology and standards. Designs, tools, and protocols that have persisted across changing conditions may have advantages in compatibility, maintenance, or institutional support.
  • Institutions and practices. Longevity can provide evidence of persistence under past conditions, although political, legal, or economic changes may alter future survival.

These are proposed applications, not proof that the relevant lifetimes follow Pareto distributions. Continued use may depend on mechanisms such as reproduction, accumulated adoption, legal protection, or organized preservation.

The object being tracked also matters. A technological principle, a product category, a company, and a particular manufactured unit have different definitions of survival. A principle can persist while every product originally embodying it disappears.

Limitations and empirical assessment

Survival does not establish quality or causation

Age may predict persistence without demonstrating truth, efficiency, usefulness, or moral value. The inference concerns duration, not merit.

Nor does an association between age and future survival establish causation. Older survivors may differ from younger items in resources, institutional support, or underlying durability. The selection mechanism described above can generate declining population failure rates even when individual failure rates never change.

Survivorship bias and observation

Survivorship bias arises when visible survivors are treated as representative of everything that once existed. Studying only familiar old books, for example, excludes forgotten works that disappeared quickly. Estimating lifetime patterns requires information about failures as well as survivors.

Ongoing lifetimes also create right-censored observations: their eventual end dates are not yet known. Samples restricted to currently existing items may additionally overrepresent long-lived items. These sampling features must be accounted for before drawing conclusions about a population’s lifetime distribution.

Changing conditions

Past survival is informative only insofar as relevant conditions persist or are modeled appropriately. Regulatory changes, technological substitution, disasters, or the disappearance of supporting institutions can alter failure risks. A relationship observed under one set of conditions may not continue under another.

Model uncertainty

A heavy tail alone does not establish exact Lindy proportionality. Other distributions can produce increasing remaining lifetime without power-law survival, and an apparent power law over a limited range may break down at longer durations. Upper limits on possible lifetimes can also change predictions substantially.

Empirical assessment therefore requires a clear definition of failure, a defined comparison population, treatment of censored observations, and comparison with alternative models. Estimates of the tail parameter are especially consequential because they determine whether the predicted mean remaining lifetime is finite at all.