Evolution and Economic Selection

Volume II — Foundations: The Paradigm and the Science

THE PLATE

A watercolour of an open book with a garden growing out of its pages: succulents, a flowering stem and new leaves. — Plate II.09The Nursery Bed.Every row in this plot is a guess. The plot is not the guess. The plot is the apparatus that lets a guess be wrong cheaply, and lets a right one be carried into next year by somebody who was not here.

THE LETTER

Economics borrowed the word evolution early and has been careless with it ever since. It usually arrives as a flourish — markets evolve, firms adapt, the fittest survive — and it does no work, because a metaphor that explains every outcome after the fact explains none of them before.

This chapter treats selection as what it actually is: a piece of arithmetic that tells you how fast the average of a population moves, given how much its members differ, how much that difference affects who persists, and how faithfully the difference is passed on. It is exact, it is indifferent to what it is applied to — genes, routines, recipes, cost codes, firms — and once you can compute a response to selection, three things follow that the metaphor cannot give you. You can say how long a better way of working will take to become the ordinary way. You can say when it will not happen at all. And you can say — this is the part the chapter is really about — which trait the economy is actually selecting on, which is never the trait anyone intended, and is always whatever the accounting happens to measure.

A population does not move toward what is good. It moves toward what is measured, at a rate set by how much the population varies and how well the measured thing is inherited. That sentence is either a warning or a design brief, depending on who is holding the pen on the measurement.

We are going to hand you the pen.

— The Editors

DISCOVERY

What is already working

Selection reasoning, done properly, has a good century behind it. Four cases, each of which is already working, and each of which supplies one part of the machinery.

Nelson and Winter, and the routine as the unit of inheritance. In An Evolutionary Theory of Economic Change (1982), Richard Nelson and Sidney Winter proposed that the thing which persists in an economy is not the firm and not the plan but the routine — the repeatable, largely tacit pattern by which a firm does what it does. Routines are inherited, imperfectly: staff move, suppliers copy, spin-outs carry the parent's habits. They vary, because search is noisy. And they are selected, because firms carrying some routines grow and firms carrying others shrink.

The appreciative fact is not the theory but what it did. Nelson and Winter's simulated economy — firms following routines, searching locally, never maximising anything — reproduced the aggregate growth series that neoclassical growth accounting had been fitted to, without a single optimising agent in it. A model with no maximisation matched the data that maximisation was invoked to explain. That is the licence for everything below.

The Illinois long-term selection experiment, running since 1896. A plot of maize at the University of Illinois has been selected for kernel oil content, one generation a year, for more than a century — the longest continuous selection experiment in biology. Starting at 4.70 percent oil, the high line reached about 22.00 percent by generation 100: a rise of 17.30 percentage points, 0.173 points a generation, sustained for a hundred years, ending at 4.68 times where it began. The low line fell to about 0.50 percent — 10.6 percent of the starting value — and selection was eventually discontinued because there was not enough oil left in the kernel to measure reliably.

Read it twice. Slow, relentless pressure on one measured number moved that number nearly fivefold, in a species that reproduces once a year. Firms copy each other faster than that.

William Muir's hens, and the unit of selection. In the 1990s the geneticist William Muir worked on a problem that commercial poultry breeding had been fighting for decades. Hens housed in multiple-bird cages were selected the obvious way: keep the offspring of the most productive individual birds. The result was a line of superb individual layers who were also efficient aggressors. Mortality in those cages ran at about 68 percent.

Muir changed one thing. He kept the offspring of the most productive cages, not the most productive birds. Six generations later, mortality in the group-selected line was 8.8 percent — 59.2 points lower, 7.73 times lower — and eggs per hen housed had risen from 91 to 237, an increase of 160.4 percent, or 24.3 eggs per hen per generation.

Nothing was done to the birds. The birds were the same birds. What changed was the boundary of the thing being counted.

Toyota's supplier network, and inheritance built on purpose. Jeffrey Dyer and Kentaro Nobeoka's study of Toyota's supplier association documented what the evolutionary frame predicts and most firms never build: a deliberate, high-bandwidth transmission mechanism. Consulting teams placed inside suppliers, study groups on a schedule, staff seconded between firms, and an explicit norm that a routine discovered in one supplier belonged to the network.

Variation is cheap and everybody has it. Selection is automatic and nobody can switch it off. Inheritance is the leg that has to be built, and Toyota built it. That is the whole reason the network compounds.

Four cases, one pattern: variation, selection, inheritance — and a choice, usually unexamined, about what unit the selection is applied to. The arithmetic below is how you take that choice deliberately.

THE ARITHMETIC

What works, what does not, and where the line sits

The Price equation. In 1970 George Price wrote down the exact accounting of how the mean of any trait changes in any population, under any selection whatsoever. It has two terms and it is an identity — it is true by construction, which is both its power and, as we will see, its trap.

                Cov(w , z )       E(w · Δz )
       Δz̄  =  ------------  +  ------------
                    w̄                 w̄

  z   the trait, per member          w   offspring per parent
  Δz  the change in the trait        w̄   mean fitness
      passed to the descendant

The first term is selection: how much the trait covaries with who persists. The second is transmission: how faithfully those who persist pass the trait on. Everything that happens to a population's average happens through one of those two, and there is no third place for it to hide.

Take five firms. Let z be capability spend as a share of revenue — training, documentation, maintenance, the things that make a routine survive a departure. Let w be offspring per parent, measured as revenue carried into the next period per unit of this period's revenue.

  firm      z (%)     w      w·z      Δz     w·Δz
  A           2.0    0.80    1.60     0.5    0.400
  B           4.0    0.90    3.60     0.3    0.270
  C           6.0    1.00    6.00     0.1    0.100
  D           8.0    1.30   10.40     0.0    0.000
  E          10.0    1.50   15.00    -0.2   -0.300

  z̄ = 6.000 %        w̄ = 1.100        E(w·z) = 7.320
  Cov(w,z) = 7.320 - (1.100 × 6.000) = 0.720

  selection term      0.720 / 1.100  =  0.6545 pp per generation
  transmission term   0.094 / 1.100  =  0.0855 pp per generation
  Δz̄                                 =  0.7400 pp per generation

Compute the next generation's mean directly and it is 6.74 percent: the identity holds to sixteen decimal places, as identities do. Selection is doing 88.5 percent of the work here and transmission 11.5 percent — and knowing that split tells you where to intervene.

Now change nothing about the firms, and change the fitness function.

Suppose w is not revenue carried forward but this period's operating margin, because that is the number the group scorecard uses and the number the capital committee reads. Capability spend costs margin now and pays later, so the same five firms get a different set of weights.

  firm        z (%)     w'     w'·z
  A             2.0    1.40    2.800
  B             4.0    1.25    5.000
  C             6.0    1.10    6.600
  D             8.0    0.90    7.200
  E            10.0    0.80    8.000

  w̄' = 1.090     E(w'·z) = 5.920     Cov(w',z) = -0.620
  selection term = -0.5688 pp per generation

The same population, in the same year, doing the same things. One measure sends the mean up 0.6545 points a generation. The other sends it down 0.5688. The swing between them is 1.2234 points a generation, and over ten generations of three years — thirty years, one working career — the two economies stand 12.23 percentage points apart: one at +6.55, the other at −5.69.

Here is the cut. Muir did not breed better hens. He changed what a hen's record was kept against, and the population walked to a different place. An economy is in exactly that position, with one difference: the boundary of the thing being counted is not a fact of nature. It is a definition, written down, in a document, with an author. Depreciation schedules, segment reporting, what counts as an asset, where a cost is booked, which entity the return is measured on — those are the fitness function. They are the selection environment. And they are a file that can be edited.

An economy does not get the firms it deserves. It gets the firms its accounting rewards, and the accounting was written by people who mostly thought they were solving a reporting problem.

Multi-level selection, and when it actually operates. Six firms in two clusters of three. Let z be the share of spare capacity a firm lends to a neighbour in a shortage. Lending costs the lender and benefits everyone in the cluster:

  w_i = 1.0 + b·Z_k - c·z_i        b = 0.060    c = 0.030

  cluster 1   z = 6, 8, 10    Z₁ = 8.00    W₁ = 1.2400
  cluster 2   z = 0, 2,  4    Z₂ = 2.00    W₂ = 1.0600

  BETWEEN clusters   Cov(W,Z)/w̄   =  +0.2348 pp per generation
  WITHIN  clusters   E[Cov(w,z)]/w̄ =  -0.0696 pp per generation
                                       -------
  total                                +0.1652 pp per generation

Inside every cluster the lender loses — the within-cluster covariance is −0.0800 in both. Between clusters the lending clusters win, and they win 3.37 times harder than the free-riders win locally. The trait rises.

Wilson and Sober's criteria, stated exactly. Group selection operates when, and only when, all five of these hold — this is the formulation David Sloan Wilson and Elliott Sober set out in Unto Others (1998) and defended in Wilson and Wilson (2007):

The population is structured into groups that persist long enough for the trait to affect group outcomes, and that form and disperse in repeated cycles.
Groups differ in the trait. Between-group variance is strictly greater than zero. Where groups are alike, there is nothing for group selection to act on, whatever the trait does.
The trait affects group fitness — groups contribute differentially to the next generation of groups, as a function of their composition.
Group composition is heritable at the group level. Offspring groups resemble parent groups; otherwise each cycle starts from scratch.
Between-group selection exceeds within-group selection. The trait is at a disadvantage inside its own group — that is what makes it the interesting case — so the between-group advantage must be larger.

Their own summary is the cleanest sentence in the literature: "Selfishness beats altruism within groups. Altruistic groups beat selfish groups. Everything else is commentary."

Criterion five has an exact form in our example. With w_i = base + b·Z_k − c·z_i, the trait rises when (b−c)·V_between > c·V_within, which is:

        b        V_within                  2.6667
       ---  >  1 + --------      =  1 + ---------- = 1.2963
        c        V_between                 9.0000

Here b/c is 2.0000 against a threshold of 1.2963 — critical b* = 0.03889, headroom 1.54 times. Drop b to 0.020, so b/c is 0.6667, and the same six firms in the same two clusters yield −0.1789 points a generation. The trait is selected out. Nothing about any firm changed.

Notice what the condition says about homogeneity. As V_between falls toward zero, the required ratio goes to infinity. An industry whose firms have been made alike cannot be improved by selecting between them at all — which is the cost of consolidation, stated as arithmetic rather than as sentiment.

Replicator dynamics, and the time it takes. Two routines, the better one holding a selection coefficient s. The share x follows ẋ = x(1−x)s, and the time from x₀ to x₁ is ln[(x₁/(1−x₁))·((1−x₀)/x₀)] / s. From 2 percent of a market to half of it, the odds ratio is 49.00 and its logarithm 3.8918:

  s = 0.03 / yr   →  129.7 years to half the market
  s = 0.08 / yr   →   48.6 years
  s = 0.15 / yr   →   25.9 years
  s = 0.30 / yr   →   13.0 years

To nine tenths of the market, the odds ratio is 441.0 and a routine with a 15 percent advantage still needs 40.6 years. A better way of working, left to selection alone, is a multi-generational project. Anything faster than that was not selection; it was somebody deliberately copying.

The honest negative, and it is a large one. Apply the breeder's equation, R = h²·S, to routines. Selection differential S = 4.00 points — the firms you keep are four points above the mean on the trait. Transmission fidelity h² = 0.35, which is an assumption and not a measurement: no published estimate of routine heritability across firms exists at this precision, and the whole calculation is only as good as that number.

  R = 0.35 × 4.00 = 1.40 pp per generation
  generations in a 10-year horizon = 10.0 / 3.0 = 3.33
  total response delivered = 4.67 pp        required = 10.00 pp
  shortfall = 5.33 pp — the programme delivers 46.7% of its mandate

It does not get there. To close it you need R = 3.00 a generation, which means S = 8.57 points, or h² = 0.750, or a generation of 1.40 years. Selection is a slow instrument and a ten-year mandate is usually not enough of it.

And then the sharper negative. Run ten generations where the selection environment reverses sign faster than a generation — the scorecard changes, the sponsor changes, the segment is redefined. The steps are still 1.40 points each, but their expected sum is zero and their standard deviation is 1.40 × √10 = 4.43 points. Two sigma is 8.85.

An industry mean that moved four points in thirty years is inside one standard deviation of no selection at all. Every such move has a story attached, and the stories are all plausible, and none of them are evidence.

That is where evolutionary reasoning in economics stops being science. The Price equation is an identity: it fits every population, always, including a population nothing is selecting. Survivors survived is not a finding, and "the fit survived" is the same sentence when fitness is read off the survival. Edith Penrose said so in 1952 and she was right about the version she was attacking.

The test that separates the real version from the just-so story is prospective and it is about a rate:

The trait is measured before the period, not reconstructed from the winners afterwards.
Fitness is operationalised independently of the outcome it explains — a cost per unit, a defect rate, a rate of repeat purchase — so that fit and survived are two measurements rather than one.
The model states a number and an interval before the period runs. Here: Δz̄ = 0.6545 points, ±0.25, so the claim dies unless the realised change falls in [0.4045, 0.9045]. A realised 0.71 survives. A realised 0.40 or a realised 1.10 falsifies it.

A just-so story cannot take that test, because its fitness variable does not exist until the period is over. If your evolutionary account of an industry cannot be wrong about a number next year, it is not an account. It is a description with a Latin word in it.

DREAM

What becomes ordinary

In an economy that takes selection seriously, the fitness function is a published document.

It has a version number. It states what is being counted, at what boundary, over what period, and it names the body that can change it. Changes carry notice — typically a full generation of whatever is being selected, so that nobody is scored against a rule that arrived after their decisions. This is not bureaucracy. It is the minimum condition under which anyone can rationally invest in a routine whose payback is longer than a year.

Variety is a budget line. Every firm of any size carries a standing allocation for variants it expects to lose money on, because the response to selection is bounded by the variance available. Nobody treats a failed variant as an error: the programme was costed on a kept-to-funded ratio in advance — three of twelve, and the nine were the price of the three. The question asked of a variant programme is never why did nine fail but is the variance wide enough to be worth selecting on.

Inheritance is engineered rather than hoped for. Routines are written down while they are working, not after the person who held them has gone. When a firm fails, its routines do not fail with it — they are deposited, the way a library deposits a book, and the next firm to face the same problem starts from the tenth attempt rather than the first. The failure of a business stops being the destruction of a hundred person-years of learning, which is what it is now, and becomes the release of it.

And the unit of account is chosen on purpose. A firm embedded in a cluster is measured partly on the cluster, because a trait that is costly locally and valuable collectively survives only when the between-group term outweighs the within-group term. It is simply understood that a measure at one boundary produces a population fit for that boundary, and that choosing the boundary is the real decision.

The people inside can say what they are being selected on. That is the whole change, felt from a desk: you know what the number is, you know who can change it, and you know how long the notice is.

DESIGN

The structure that gets there

Selection needs three legs and one choice. Build them explicitly, in this order.

One — the variation budget. Fund a cohort of variants, not a pilot. A single pilot generates no variance and therefore no response to selection; it generates an anecdote. Size the cohort so the spread of outcomes is wider than the noise, and state the kept-to-funded ratio in advance: twelve funded, three kept, an intensity of 0.250. Announcing the intensity before the results is what stops the programme from being retrospectively rewritten around whatever happened to survive.

Garud and Karnøe's comparison of Danish and American wind turbine development is the case here: the Danish path accumulated small variants, tested them in the field and kept what held, while the American path pursued larger leaps with fewer trials. The distributed, high-variance, high-retention path produced the industry. Variation is not the opposite of discipline. It is what discipline is for.

Two — the selection criterion, written down before the period. One number, or a stated combination, with its boundary named. Most of the value of this step is in discovering what you have actually been selecting on, which is rarely what anyone says. If the scorecard rewards this year's margin, capability spend is under negative selection at −0.5688 points a generation regardless of what the strategy document says, and the strategy document loses, every time, because the strategy document is not the fitness function.

Where a trait is costly to the unit and valuable to the group, the criterion must carry a group term large enough to clear b/c > 1 + V_within/V_between. Compute that ratio for your own case. It is usually between 1.1 and 1.5, and it is usually the single most decision-relevant number in the design.

Three — the transmission mechanism. This is the leg everyone skips and it is the one that sets h². Concretely: routines documented while in use, by the people using them, as a condition of the variant being funded. Secondments between units. A standing study group with a schedule rather than an intention. And an explicit rule that a routine discovered in one unit belongs to the network — which is what Toyota's supplier association formalised, and which is why raising h² from 0.35 to 0.75 is a management decision and not a law of nature.

Four — the choice of unit, taken deliberately and revisited. Name the boundary the fitness function is applied to: the individual, the team, the business unit, the firm, the cluster. Write down what the trait costs at that boundary and what it returns at the boundary above. Then check criterion five. Muir's entire contribution was to move one boundary, and it was worth 160.4 percent on output and 59.2 points on mortality.

The sequence. Variation first, because nothing else matters without it. Criterion second, published before results. Transmission third, because it compounds and can be built while the first cohort runs. The unit is revisited annually and changed with notice, never mid-period.

DESTINY

How it holds when nobody is pushing

Three things keep a selection system alive, and each has a named failure.

The fitness function is versioned and amended on a clock. A criterion that can change without notice cannot be invested against; a criterion that can never change becomes a target and stops being a measure. The stable arrangement is a published document, an annual amendment window, and notice of one generation. The failure mode is Goodhart's: once the measure is the target it ceases to be a good measure, and the population optimises the measurement rather than the thing. The defence is not to hide the measure — that destroys the investability — but to hold a second, unpublished, occasionally sampled measure of the same underlying thing, and to amend the published one when the two diverge.

Variance is protected as an asset. The most common quiet death of a selection system is that efficiency drives out variety: the variants are consolidated, the odd unit is brought into line, and within a few cycles V_between approaches zero. From the condition above, the required b/c goes to infinity as that happens — the system becomes unable to respond to selection at all, and it does so without any single decision that looks wrong. The failure mode is monoculture. The defence is a floor: a minimum proportion of capacity held in variants that are not the standard, defended in the budget the way a reserve is defended.

Inheritance outlives its source. Routines are deposited outside the unit that invented them, so that the unit can fail without the learning failing. The failure mode is the one every firm has: a retirement, a reorganisation, an insolvency, and a hundred person-years of routine leaves the building in an afternoon. That is h² falling to near zero, and from R = h²·S the response to selection falls with it, whatever the selection pressure.

And the honest close. If the environment's direction changes faster than a generation, none of this holds — the steps are real and their sum is a random walk of standard deviation 4.43 points, indistinguishable from nothing. Under those conditions the correct decision is not a better selection system. It is to shorten the generation or to stop pretending that selection is the mechanism and start copying deliberately, which is faster and is allowed.

DELIGHT

What it feels like

The pleasure is the nursery bed. Twelve rows, twelve handwritten labels, and the honest expectation that nine of them are wrong. There is a particular lightness in working where being wrong is the budgeted case — where a failed variant is not a reprimand but the thing you paid for, priced at £85,000 and accounted for before it was planted.

Then the second pleasure, quieter and better: the year you find one of your routines running somewhere you never sent it. Somebody's spin-out, somebody's new supplier, a competitor who hired one of your people and kept the habit. It has your fingerprints on it and nobody's name. It stopped being yours and became the ordinary way the thing is done, which is the only form of permanence available to a way of working.

And a third, for the person who keeps the records. Watching a number you have measured every year for six years move 24.3 units a generation in the direction you chose, and knowing exactly why, and being able to show anyone the arithmetic on one page. Selection is the slowest instrument in this book and the least reversible. What it gives back is the rare experience of a change that does not need you to keep pushing it.

OPERATIONALIZE THIS

At the level of finance

The instrument: a variant cohort facility, with a group-level gainshare.

A ring-fenced, multi-year facility that funds a cohort of variants under a published selection criterion, with a defined share of the verified group-level improvement returning to the units that generated the kept variants. It is a breeding programme with a term sheet, and your organisation has probably already approved something structurally identical under the name innovation fund — without the two features that make it work, which are the published criterion and the group-level return.

The mechanics.

Cohort size and cost. Twelve variants at £85,000 each per generation — £1,020,000 a cohort. Three kept, an intensity of 0.250, both published before the cohort runs.
The criterion. One page, versioned, dated, naming the trait, the boundary, the period, the verifier, and the amendment window. Signed by the sponsor and by finance before the first variant is funded. This document is the transaction; everything else is administration of it.
Term. Three generations minimum. At a three-year generation that is nine years, with a hard review after the first cohort is scored.
The gainshare. A fixed share of verified group-level improvement — 25 to 40 percent — to the units that produced the kept variants, paid on the group measure and never on the unit measure. Paying on the unit measure recreates exactly the within-group selection the facility exists to overcome.
Security. The improvement stream, and nothing else. If the cohort produces no verified response, there is nothing to share and the facility says so.

The balance-sheet treatment. The nine variants that are not kept are period expense and should be presented as such, deliberately and without apology: they are the cost of the variance, and variance is the input the response is bought with. The three kept variants, where they create a durable routine, are capitalised as development expenditure and amortised over the routine's expected life at the network boundary, not the unit boundary — because that is where it will actually be used once transmission is working. Talk to your auditors early; this is a conversation about useful economic life and the recognition criteria for development costs, which they have every year.

The counterparty. Internal first — treasury to a cluster of business units, documented in a fortnight. The external version, once one cohort has been scored, is a supplier association or an industry body: the natural counterparty for a group-level measure is whoever already collects industry-wide statistics, because they are the only party with a credible boundary above the firm.

The number that decides it. One line on the front page:

   h² · S · (horizon / generation) · value per point   >   cohort cost · WACC

   0.35 × 4.00 × 3.33 × £310,000  =  £1,446,667
   £1,020,000 × 9%                =  £91,800            →  proceed

At these inputs the facility returns £434,000 a year against a £1,020,000 cohort — 42.5 percent, against a 9 percent WACC, clearing by 33.5 points, with a payback of 2.35 years and a cost of £728,571 per percentage point of response. And the sensitivity that matters: h² is the assumed number. Halve it to 0.175 and the response halves with it. Before you approve the facility, spend a month measuring your actual transmission fidelity, because every figure above is linear in it.

The first ninety days.

Day	Action	Artifact
1–15	Write down what you are currently selecting on. Not the strategy — the scorecard.	The current fitness function, as found
16–30	Compute `Cov(w,z)/w̄` on last year's population for two candidate traits	Two selection terms, signed
31–45	Measure `V_between` and `V_within`; compute the required `b/c`	The threshold, one number
46–60	Draft and sign the criterion, with its amendment window	The signed criterion
61–75	Estimate `h²` from three routines that moved between units last year	A transmission estimate with its method
76–90	Fund the cohort. Publish the predicted Δz̄ and its interval.	The prediction, dated, before the period

Day 90 is the one that matters. A number written down before the period, with an interval, is the difference between an evolutionary economics and a story about one.

APPRECIATIVE QUESTIONS

Twelve, for a room

Discovery — what is already working

Think of a way of working here that spread without anyone mandating it. How did it travel, who carried it, and what made it easy to copy?
Which of our units is most unlike the others — and what has that difference allowed us to learn that uniformity would not have?
When has one of our people taken a routine to another organisation and made it work there? What does it tell us that it survived the move?

Dream — what becomes possible

If everyone here could name, in one sentence, the number they are actually being selected on, what would change on Monday?
Imagine our variant programme three cohorts from now, and one of the kept variants is running in four places. Which variant is it, and who kept it alive at the beginning?
If a routine of ours outlived the unit that invented it, where would it be deposited, and who would go looking for it?

Design — what we build

What is the widest set of genuinely different attempts we could fund this year without anyone needing to call it a strategy?
What group-level number would we be willing to publish before the period, and be scored against afterwards?
Which measure of ours rewards a behaviour inside a team that costs the organisation as a whole — and what would the group term have to be worth for that to reverse?

Destiny — how it holds

What would we have to protect, deliberately and in the budget, to make sure we are still different enough from each other in five years to learn anything?
Who, other than the person who invented it, could run our best routine next quarter — and what would we have to write down this month for that to be true?
What is the first sign we would see that our selection criterion had become a target rather than a measure, and who here would notice it first?

WORKS CITED

Alchian, A. A. (1950). "Uncertainty, Evolution, and Economic Theory." Journal of Political Economy, 58(3), 211–221.

Dudley, J. W. and Lambert, R. J. (2004). "100 Generations of Selection for Oil and Protein in Corn." Plant Breeding Reviews, 24(1), 79–110.

Dyer, J. H. and Nobeoka, K. (2000). "Creating and Managing a High-Performance Knowledge-Sharing Network: The Toyota Case." Strategic Management Journal, 21(3), 345–367.

Falconer, D. S. and Mackay, T. F. C. (1996). Introduction to Quantitative Genetics, 4th edn. Longman.

Friedman, M. (1953). "The Methodology of Positive Economics", in Essays in Positive Economics. University of Chicago Press.

Frank, S. A. (1995). "George Price's Contributions to Evolutionary Genetics." Journal of Theoretical Biology, 175(3), 373–388.

Garud, R. and Karnøe, P. (2003). "Bricolage versus Breakthrough: Distributed and Embedded Agency in Technology Entrepreneurship." Research Policy, 32(2), 277–300.

Hodgson, G. M. and Knudsen, T. (2010). Darwin's Conjecture: The Search for General Principles of Social and Economic Evolution. University of Chicago Press.

Moose, S. P., Dudley, J. W. and Rocheford, T. R. (2004). "Maize Selection Passes the Century Mark: A Unique Resource for 21st Century Genomics." Trends in Plant Science, 9(7), 358–364.

Muir, W. M. (1996). "Group Selection for Adaptation to Multiple-Hen Cages: Selection Program and Direct Responses." Poultry Science, 75(4), 447–458.

Muir, W. M. (2009). "Genetic Selection and Behaviour." Philosophical Transactions of the Royal Society B, 364(1520), 1087–1094.

Nelson, R. R. and Winter, S. G. (1974). "Neoclassical vs. Evolutionary Theories of Economic Growth: Critique and Prospectus." The Economic Journal, 84(336), 886–905.

Nelson, R. R. and Winter, S. G. (1982). An Evolutionary Theory of Economic Change. Harvard University Press.

Penrose, E. T. (1952). "Biological Analogies in the Theory of the Firm." American Economic Review, 42(5), 804–819.

Price, G. R. (1970). "Selection and Covariance." Nature, 227, 520–521.

Price, G. R. (1972). "Extension of Covariance Selection Mathematics." Annals of Human Genetics, 35(4), 485–490.

Sober, E. and Wilson, D. S. (1998). Unto Others: The Evolution and Psychology of Unselfish Behavior. Harvard University Press.

Taylor, P. D. and Jonker, L. B. (1978). "Evolutionarily Stable Strategies and Game Dynamics." Mathematical Biosciences, 40(1–2), 145–156.

Wilson, D. S. and Wilson, E. O. (2007). "Rethinking the Theoretical Foundation of Sociobiology." The Quarterly Review of Biology, 82(4), 327–348.

Winter, S. G. (1964). "Economic 'Natural Selection' and the Theory of the Firm." Yale Economic Essays, 4(1), 225–272.

Note on figures. Every figure in this chapter is computed in lib/verify/II_09.py and printed with its inputs, its units and its source. The maize and poultry figures are as reported in the cited papers. The Price equation examples, the multi-level partition, the group-selection threshold, the replicator times, the breeder's-equation response, the random-walk band and the facility arithmetic are computed there and reproducible. The transmission fidelity h² = 0.35 is an assumption, not a measurement, and is flagged as such wherever it is used.