11

The Lindy effect is the observation that, contrary to the pattern with perishable things like people, the longevity so far is a good predictor of the future longevity. In other words, technologies that have already survived a long time can be expected to continue to survive for a long time (whereas people who are old are likely to die soon).

Taleb argues thus in a recent Wired article:

For a perishable human, every year that elapses reduces his life expectancy by a little less than a year.

The opposite applies to non-perishables like technology and information. If a book has been in print for 40 years, I can expect it to be in print for at least another 40 years. But – and this is the main difference – if it survives another decade, then it will be expected to be in print another 50 years.

He isn't arguing that this is a perfect rule, just a good statistical estimate (so don't be too quick to provide single examples as answers!)

Given that this is a statistical generalization, is the weight of evidence in favour of the idea?

Community
  • 1
matt_black
  • 56,186
  • 16
  • 175
  • 373
  • 1
    It sounds like it's just a generalization from some things in experience. So it sounds like it's vacuously true of the kinds of things of which it is true. – Mike Dunlavey Mar 06 '13 at 00:39
  • 4
    This is a heuristic estimate based on limited information for items that follow certain probability distributions (e.g. the lengths must spread over several magnitudes). It works as a rough estimate where there is no more available info than longevity. (Computer process longevity is another classic example.) It doesn't work where the probability distribution doesn't follow the assumptions. See also Half-Lives, Pareto, Long Tail, Benford's Law, etc. So what does it mean to have a "weight of evidence" behind it? It is easy to find examples that don't fit, but that doesn't invalidate it. – Oddthinking Mar 06 '13 at 08:13
  • I'm not sure "good statistical estimate" paints the right picture; my understanding is that this kind of estimation is basically the weakest kind you can do, and it has many ways of failing (e.g. one of the grounding assumptions is that there's nothing particularly special about the current point in time - something which which is difficult to be sure of). So, it's perhaps better than nothing, but not necessarily "good". – Daniel B Mar 06 '13 at 09:23
  • Maybe I'm not understanding but, by the Lindsey rule, wouldn't nothing ever perish? – Publius Mar 06 '13 at 09:37
  • 2
    Given that some books do eventually go out of print, the Lindy effect cannot be valid in the way suggested for books. The Lindy effect is correct for the first half of the actual lifetime of an item and wrong for the last half of its lifetime. You never know which half you are in. – RedGrittyBrick Mar 06 '13 at 11:55
  • @Oddthinking I can think of a simple statistical test. Take a sample of things existing at a particular point in history (lets say books in print in 1950) and test whether their longevity at that point was a good predictor of how long they *stayed* in print over the next 60 years. The same sort of test would work in other domains. So realistic meaningful tests are possible for some applications of the principle. – matt_black Mar 06 '13 at 17:41
  • @RedGrittyBrick To clarify what I think the claim is: the idea is not that there is a *perfect* predictor of the future lifetime of an individual thing. Rather, the claim is that, if you had to make a bet on the average lifetime of say 100 such things, the average lifetime so far would be a very good estimate (or number to bet on). – matt_black Mar 06 '13 at 17:45
  • @Avi Nope. If the object has lasted 5 years, for example, you would expect it to last another 5, then die. Whereas, if it has already lasted a few thousand years (like the pyramids) then you would expect it to last a few thousand more at least. – matt_black Mar 06 '13 at 17:47
  • 1
    @matt_black: Thanks for the clarification (I confess I'm still a bit in the dark). Are you sure http://stats.stackexchange.com/ isn't a better home for this question? – RedGrittyBrick Mar 06 '13 at 18:35
  • @Matt: there are several problems: 1) "books in print" is an arbitrary choice. If that fails, you have merely chosen the wrong domain that doesn't match the required prob distribution, not falsified the theory. 2) Whatever domain, there are likely to be better predictors (including actually predicting something useful, like how many copies to stock), so what does "good" mean? This predictor's main benefit is it requires little thought. – Oddthinking Mar 06 '13 at 20:51
  • @Oddthinking Disagree. It is a good test as it is easy to apply and test (and I won't wiggle out by changing the definition). Moreover, *simple* heuristics are often better than complex ones (see Gigerenzer's book http://amzn.to/YdvyOA) – matt_black Mar 06 '13 at 23:20
  • So, the question you are asking is not "Does the Lindy Effect work?" but "Does it work on books-in-print, in particular?" Also, if I come up with a much more accurate heuristic (which I am confident I could do, as soon as we find some data), does that invalidate the theory or not? Should we take this to chat? – Oddthinking Mar 07 '13 at 00:36
  • @matt_black right, but then if it were in print for those 10 years you'd expect it to be in print for 20. And if it lasted that long you'd expect it to be in print for 40, etc. – Publius Mar 07 '13 at 02:54
  • One possible reading of this is that if something has survived much longer than expected, you should start to suspect that the standard life expectancy model doesn't apply to this one exemplar and it's following a different rule. But simple bayesian statistics would support that inference anyway. – octern Apr 26 '15 at 16:40

2 Answers2

6

It's a model or approximation: which can be applied, more or less usefully, to some (a subset of) things.

It's supported by, for example:

  • http://en.wikipedia.org/wiki/Burn-in means that an older/tested component, when it passes its short burn-in period, then has a longer subsequent life expectancy than a young/untested component which has not.

  • http://en.wikipedia.org/wiki/Bathtub_curve suggests that the effect would be applicable to components which "don't wear out" ... or for which, the magnitude of the "wear out failures" is insignificant compared to the magnitude of the "infant mortality" failures.

I haven't read his book (Black Swan) but from the article you cited Mr. Taleb isn't even talking about "things", necessarily: for example it's about software, which doesn't "wear out" in a conventional way.

ChrisW
  • 26,552
  • 5
  • 108
  • 141
  • +1 It's utility (if it has any) applies to collections not individuals. If I have 2000 things of the same type I can divide them into those that are older than the median age and those younger. I can expect that the old group represents robust survivors whose fragile peers have gone. I can predict that at today + median, the ratio of survivors of today's tough old group will exceed survivors of today's mixed young group. This probably applies to light bulbs in an office building. Whether you can benefit from this is moot. Lindy effect = *able veterans are tougher than newborn*. Obvious? – RedGrittyBrick Mar 06 '13 at 15:06
2

I can think of two major flaws in the Lindy effect concerning the longevity of non-perishable things:

  1. We have witnessed a rapid (almost exponential) rate for advances and developments in technology and sciences in the last century (or in even in the last 50 years). This prevents us from suggesting any rules concerning what kind of technology or theory will last more. (Edit: As a reference to the above claim, one can consider Moore's law, proposed by Gordon E. Moore one of the co-founders of Intel, which asserts the advancements in electronic,specifically chip, industry is too fast that the number of chips doubles every two years. It is now so acceptable in the industry that experts use his theory for their forward-looking planning. See also Vinge's theory about exponentially accelerating changes for other evidences for the above point.)
  2. The Lindy effect (and in some extent your question) rely on the assumption that the future can be predicted by the past events. There is certainly no scientific evidence or logical justification for this assumption about general events and consequently for the Lindy effect. Although induction is an accepted logic for scientific theories it is hardly acceptable for predicting future, as a philosopher (C. D. Broad) says "induction is the glory of science and the scandal of philosophy", see also the page Problem of Induction in wikipedia.
  • Please [provide some references](http://meta.skeptics.stackexchange.com/q/5) to support your claims. – Oddthinking Mar 06 '13 at 08:07
  • I'm not sure of the usual scope of the Lindy effect for broad technologies. Its not obviously wrong, though. The idea of the integrated circuit built on doped silicon is from the 1960s-70s and is reasonably likely to last a while. But the 8086 and the specific semiconductor technologies that built it have mostly gone. – matt_black Mar 06 '13 at 18:09
  • As can be seen in the Wikipedia link "advancements in electronic,specifically chip, industry is too fast that the number of chips doubles every two years.", as well as being ungrammatical, is wrong. – Mark Hurd Mar 12 '13 at 05:28
  • No, the Lindy effect do not rely on the assumption that future can be predicted. Quite the opposite: it states that it cannot be predicted, and that our only reliable data is the observed lifespan of an object, and the statistical probability that this objects is at its half-life. Suppose you meet an ET tomorrow, you have no idea about its age, you can only guess it is in the middle of its life, because it is the most probable (the middle of a bell curve). – gb. May 08 '13 at 04:37