J Evol Econ (2002) 12: 539–562

c© Springer-Verlag 2002

Evolutionary economics and the counterfactualthreat: on the nature and role of counterfactualhistory as an empirical tool in economics�

Robin Cowan1 and Dominique Foray2

1 MERIT, University of Maastricht, PB 616, 6200 MD Maastricht, The Netherlands(e-mail: r.cowan@merit.unimaas.nl)

2 IMRI, Universite Dauphine, 75775 Paris Cedex, France(e-mail: dominique.foray@dauphine.fr)

Abstract. Counterfactual conditional statements are ubiquitous in any scientificendeavour. This paper contains an analysis of the nature of counterfactual condi-tionals and the conditions under which they are considered assertable by scientists.The paper then applies this analysis to the use of counterfactuals in evolutionaryeconomics, arguing that because evolutionary economics is inherently concernedwith historical processes it cannot avoid the use of counterfactual history as one ofits tools of empirical analysis. We discuss the strengths and pitfalls of counterfactualhistory. We argue that because evolutionary economics starts from the foundationthat randomness may be inherent in any economic system, the very aspects of evo-lutionary economics that make counterfactual history a desirable empirical toolalso make that tool difficult to employ.

1 Introduction

This paper is concerned with a methodological question in economics. In particularit examines the role and nature of counterfactual history as an empirical tool inevolutionary analysis of economic phenomena. As a method, counterfactual historyhas not met with universal approbation, and many economists look upon it withconsiderable dis-favour. In 1983 the American Economic Review published a paperby Preston McAfee in which he asks about American economic growth underthe counter-factual proposition that Columbus had not discovered the new world.McAfee clearly means to ridicule the use of counter-factual history. Part of thegoal of the present paper is to explain both why counterfactual history is appealing

� We thank the participants of the International Seminar on Evolutionary Economics as a ResearchProgramme in Stockholm, May 1997, for many helpful comments. We also thank Lorri Baier for manyhelpful substantive and textual comments.Correspondence to: R. Cowan

540 R. Cowan and D. Foray

as a tool for certain approaches to economic analysis, and why it may appearunappealing to someone deeply schooled in an axiomatic, deductive tradition.

Empirical analysis in any research field is entwined in theoretical analysis. Thatis, empirical work depends on theory for concepts, definitions and hypotheses, all ofwhich are used as foundations for empirical investigation. Thus if theoretical work ina field evolves and becomes more complex, the range of tools employed in empiricalanalysis is also likely to evolve, and in general we would expect the toolbox to grow.One of the changes in theoretical economics in recent years has been an increase inthe number of models producing multiple equilibria. Multiple equilibria have beenpresent in analysis almost since the birth of the discipline, but for many years, partof the methodology of economics seemed to be to design models that had uniqueequilibria. This is certainly sensible in many ways: it simplifies comparative staticsanalysis, and it removes the need for detailed examination of the (difficult) dynamicprocesses involved in equilibrium selection. But the growth of the literature of pathdependence, evolutionary modelling, game theoretic models and much modernmacro-economics has brought with it multiple equilibria, which in these fields aregenerally accepted as a normal and untroubling outcome. The pervasiveness of thisresult cannot help but affect methodology. For instance, when an equilibrium isunique, the welfare question is restricted to whether this equilibrium can be madebetter. This is often answered through a comparative statics exercise, which inprinciple involves small adjustments to parameters, resulting in small changes toequilibrium values: we can adjust a tax rate to move the current equilibrium to apoint very near and ask which of the two is preferred. But when multiple equilibriaare present, a second question is possible: would a different equilibrium be or havebeen, better? This cannot necessarily be answered with a simple comparative staticsexercise.

Kaldor (1934) points out that there are two general types of multiple equilib-ria models, but that analysis of both will depend on analysis of the path to theequilibrium. First, the (possibly single) equilibrium moves in response to eventsalong the path. Fisher (1983) for example, shows that if dis-equilibrium trades takeplace, the final equilibrium value, even if there is enough convexity in the systemto guarantee that only one exists, cannot be predicted from the initial data. Simply,in an Edgeworth Box, equilibrium lies somewhere in the lens formed by the initialendowments and indifference curves. An out of equilibrium trade will create a new“endowment” and thereby a new core. An out of equilibrium trade at a differentnon-equilibrium price will create a different new “endowment” and thereby a dif-ferent new core. Thus the set of possible equilibria will change if out-of-equilibriumtrades take place, and will change differently depending on the (almost certainlyunpredictable) prices at which those trades take place. Notice that for this view thatequilibria move in response to out-of equilibrium phenomena to make sense, timemust pass before the equilibrium is reached.1

Kaldor’s second reference is to cases in which there are several equilibria whichdo not move in response to the path of adjustment. This is simply the case of multipleequilibria for a given set of parameter values. Given the initial data, the analyst is

1 For a similar discussion of the same issue see David (1997).

Evolutionary economics and the counterfactual threat 541

able to predict, but can only narrow his predictions to some set of equilibria – themarket will evolve to one of a set of known equilibria, but which of them it willbe is not determined. In a static model, this exists when there are appropriate non-convexities – perhaps some form of increasing returns. The “in principle” difficultythis poses is how one of these equilibria is chosen – there must be some selectionmechanism. It is in the selection of the equilibrium that the path dependence emergesor becomes important. With some exceptions, of course, it seems again that time isgoing to play a central role here; the selection of an equilibrium is something thatwill take place as a process evolves over time.

When multiple equilibria are present, there is no guarantee that they are welfareequivalent. This raises the possibility of potential regret (see Strotz, 1956). Whetheror not potential regret has a place in a model will depend on the nature of theselection mechanism. How do we get one equilibrium rather than another? Forsome economists the answer to this question is “History.” Selection mechanisms areinherently historical. Historical processes can have the feature of path dependence,and this is commonly present in evolutionary models. Here is the moment at whichthe Potential Regret result arises. Had history taken a different course in the past,we would now be at a different, and better, position. The historicity present hereimplies that explanations which address the issue of potential regret will necessarilybe historical. The questions “Is potential regret possible?”, or “Is regret actual in thiscase?” effectively ask what would have been the case had history taken a differentcourse. This is a counterfactual question, equivalent in this regard to the question“What would be the effect on employment if the tax rate were reduced by x%?”This paper is concerned with the former type of counterfactual, and in particularhow historical concerns and the introduction of path dependence into economicmodels changes the nature of counterfactuals.

The paper is organized as follows: We begin by addressing the role of coun-terfactual conditionals in science in general, and then turn to their relationshipwith views of causation. This is followed by a discussion of two distinct views ofcounterfactual conditionals, the branching view, and the non-actual-possible-worldview. The following sections bring these abstract considerations back to economics,addressing both the difficulties raised for counterfactual analysis, and suggestionsregarding how to strengthen it.2

2 There is a difficulty of nomenclature here. It would be convenient to have a phrase that captures thebody of analysis for which historical processes are central. This includes work concerned with selectionmechanisms, evolutionary game theory, adjustment processes, path dependence, what is known tightlyas evolutionary economics, significant parts of modern macro-economics, and no doubt a variety ofother work. With apologies to the members of the Schumpeter Society, we will refer to this collectionas evolutionary. The reason is that much of economics that lies outside “Evolutionary Economics” hasevolutionary features – concern with processes, positive feedbacks, variety among the agents and so on(see Nelson, 1995 for a characterisation of “Evolutionary Economics”), and it is the presence of thesefeatures in the theory that demand the kind of historical counterfactual analysis that we are discussinghere. When we refer to Evolutionary Economics per se we will capitalize it.

542 R. Cowan and D. Foray

2 Counterfactuals in science

The prototypical theoretical result in mainstream economics is the comparative stat-ics result: the number of patents filed will increase by 10% for every 25% increasein R&D spending. These results are usually presented as theorems or propositionsderived from axioms. Similar results are presented in most sciences, though the de-gree of precision and extent of generality vary, both within and across disciplines.What is key here is that these statements can all be re-cast explicitly as equiv-alent counter-factual conditionals, that is, conditionals in which the antecedentsare false.3 Indeed, Nelson Goodman (1983) argues convincingly that six types ofstatements are equivalent, and that any statement of one type can be re-cast as astatement of any of the other five types. The types are counter-factual conditionals,scientific laws, dispositional statements (p.3), factual conditionals (p.4), statementsregarding possible worlds (p.34), and causal statements (Goodman is not explicitabout the last category but it is implicit in what he writes on pages 37 and 45). Sothe following statements are equivalent: If R&D spending were 25% higher thanit is, patenting would be 10% higher. The ratio of the change in R&D spending tothe change in patents is 2.5:1. Patenting has the tendency to increase (in the ratio1/2.5) as R&D spending increases. Since patenting increased by 10%, R&D spend-ing must have increased by 25%. In a world like ours but with 25% more R&Dspending, we would observe 10% more patenting. Changes in patenting activityare caused by (inter alia) changes in R&D spending.4

Counterfactual conditionals are ubiquitous in science. They can be, and are, usedto express causal relations. Further, they are part of the link between theoretical andempirical work. Empirical work is successfully characterized as the investigation ofcounterfactual conditional claims. Do the data show that if R&D spending increasesthen patenting increases in turn?

3 Causation

Counterfactuals also appear in another place that is germane to this discussion.Philosophical discussions of causation often employ counterfactual conditionals asthe key concept. Indeed for many years there were vigorous attempts to unpack cau-sation formally in those terms. The appeal is obvious: “C causes E.” is very muchlike the statement “If C occurs then E occurs.” The problems with this approach

3 This idea is not new in economics: Fogel (1973) fends off an attack on the new economic historymade by Genovese (1962) by presenting a long list of “hidden” counterfactuals in Genovese’s ownwork.

4 McCloskey (1989, p.150) has a similar list regarding money supply and inflation; Cowan and Foray(1997) illustrate using minimum wages and unemployment.A claim that appears impossible to render sensibly in this way might be “If R&D spending goes up25% next year, patenting will increase by 10%.” This appears not to lend itself to exposition as acounterfactual, since the future has, as yet, no fact. The counterfactual arises when one notes that theclaim is implicitly contrasting a world in which R&D spending does not increase by 25% next year.Now, it is clear that (at least) one of the two contrasting cases will be contrary to fact. The counterfactualcan be formed prospectively, even though we cannot know which of the cases will be fact, and whichcounter to fact.

Evolutionary economics and the counterfactual threat 543

turned out to be insuperable when addressed purely in terms of material condi-tionals, so that endeavour has effectively been given up. Part of the difficulty wasa tension between logical analysis and physical analysis. To most non-Humeans,causation does seem like a real feature of the world. If one thing did not cause an-other there would be no reason for the world not to be totally chaotic, and thereforeno reason, beyond observations of Humean constant conjunction, to believe that itmight not suddenly become so. This is what logical analysis has so much difficultycapturing. So we can note that understanding the statement “If C occurs then Eoccurs” as a causal statement involves understanding more than simply the formallogic involved. It involves understanding things about C and E, about what linksthem and how C “necessitates” the occurrence of E.

We can make this point a slightly different way. All counterfactual condition-als, when viewed as material conditionals or read purely truth-functionally, aretrue. Logically, they can be written as ∼C & (C → E).5 Such a statement has thetruth value true. Thus, truth-functionally, “If R&D spending were higher than it ispatenting would be higher than it is.” and “If R&D spending were higher than itis patenting would be lower than it is.” are both true, since R&D spending is nothigher than it is. This makes it sound as if the logical conclusion of scientific en-deavour is “Anything can happen.” (which of course is not inconsistent with manyof the beliefs of the spiritual movement of the 1990s). But this is clearly not wherescience is headed or wants to end up. How do we distinguish between these twostatements, accepting one and rejecting the other?

The first, brief part of the answer is that it cannot be done purely with logicalanalysis, and no modern treatment of either counter-factuals or causation attemptsto do so. Peter Gallison (1987) shows convincingly just how much this is the casefor experimental physics. Experimenters must decide when they have arrived atthe correct result, and when errors introduced by faulty equipment, extraneousforces, observation error and so on are not affecting the results in a significant way.Gallison shows in a series of cases that while formal logical analysis is important,it is probably not the most important factor in making that decision.

The longer answer has had two strands in the literature. A meta-linguistic anal-ysis, (associated with writers such as Chisolm, Goodman, and Mackie) gives upon truth and falsity and asks instead about “assertability”. We are willing to assertthat an increase in patenting follows an increase in R&D spending, but we areunwilling to assert that a decrease in patenting would follow an increase in R&Dspending. This removes the onerous burden of logical proof, but leaves open thequestion of what determines assertability. The other strand, following Stalnaker(1968) or Lewis (1973 a,b) for example, retains truth and falsity, but attempts amore sophisticated analysis of what determines it. For our purposes, it is not nec-essary to choose between these approaches, for the main problem is over-arching:in Goodman’s words, it is “to define the circumstances under which a given coun-terfactual holds while the opposing conditional with the contradictory consequentfails to hold.” (p.4).6 Meta-linguistic analysis is relatively clear that assertability

5 This should be read as “C is false, and (C implies E)”.6 There may be exceptional cases in which both hold, but for our purposes we need not be concerned

with those cases. Lewis’ (1973b) theory allows for such a possibility.

544 R. Cowan and D. Foray

follows from being able to infer the consequent from the antecedent plus someadditional premises. On the theories of Lewis and Stalnaker, truth or falsity of acounterfactual follows from the same considerations as truth and falsity of a factualconditional, but to evaluate the former we consider a non-actual possible world.The two approaches have similar problems: which additional premises are consid-ered; which possible world is examined.7 But having determined which premises orwhich world, there is the issue of how to use the set-up to evaluate the conditional.The consequent must somehow be linked to the antecedent. Purely formal linkswill almost certainly fail to provide the answer in general, just as they do in evalu-ating many factual conditionals. Goodman is explicit that this link must be causal;Lewis is less so. For Lewis a counterfactual conditional is true if the factual condi-tional is true in a closest possible world. How do we evaluate factual conditionals?Gallison’s work shows that the answer to this question is complicated, but at thevery least non-contradiction of known causal laws is one of the strongest criteria.Laws are included in descriptions of possible worlds, and a conditional which wasshown to be an instantiation of a causal law (or a system of them) would certainlybe judged true. Even if we do not go as far as Goodman (or Elster, see below) inasserting that the consequent must follow from the antecedent by a causal law, ourideas about causation will be central in determining how we evaluate counterfactualconditionals.

Emphasizing the place of causation conforms well with the intuitions whichmake counterfactual claims relatively common in everyday life. Asserting a coun-terfactual conditional is in essence making a claim about the regularity and coher-ence of the world. Causal forces connect events and provide the coherence that weseek. The reliability of a counterfactual arises from the same source, namely thecoherent, consistent structure of events, even under Hume’s minimalist, “constantconjunction” view of causation. To defend a counterfactual conditional is to expli-cate the causal connections among the events involved.8 If the causal account isjudged acceptable, the conditional is judged assertable. This brings us to the natureof causation in economics.

7 Which additional premises to consider is determined by “co-tenability”. Lewis (1994, pp. 57 and69) provides a definition which he claims captures the intent of metalinguistic theorists and also makesthe two theories identical.

8 Accounts of counterfactuals vary in their explicitness regarding the use of ideas about causationin our evaluation of counterfactuals. Simon and Rescher (1966) are explicit that the acceptability of acounterfactual is determined by the causal models we employ, through their version of modal orderings.Stalnaker (1968), and similarly Lewis (1973a, 1973b), are not explicit that causal analysis is involved,though Lewis makes an explicitly tight connection between causation and counterfactual conditionals inthat he seeks to provide a counterfactual analysis of causation. For both authors, though, the acceptabilityof a counterfactual conditional is determined by asking whether the consequent is true in a world inwhich the antecedent is true but which is minimally different from the actual world. Jackson’s (1977)suggestion is to posit a world in which the antecedent is true and inter alia in which “causal laws . . . areidentical to ours” (p.7). Bennett (1984, p. 73), has a similar approach, though speaks of “laws” ratherthan causal laws explicitly. Goodman’s exposition is very explicit regarding the causal under-pinningsof counterfactuals, and it is clear that many authors feel there is a tight connection between the two.

Evolutionary economics and the counterfactual threat 545

1.1 Causation in economics

Within economics as a whole we can see that there are implicitly two notions ofcausation. Traditionally, the fundamental goal in economics has been to explainphenomena in terms of equilibria.9 Note that in this context, an equilibrium canbe dynamic, involving change, growth or even cycles in the variables of interest.Something is explained when it is shown to be held in place by countervailingforces. The economy (firm, market) exhibits behaviour X because there is no in-centive to deviate from that behaviour, and in fact deviations are punished. Theexplanation of X consists in finding the forces that impinge on X and showing thatthey are in balance. How these forces came to impinge on X , or how they came intobalance is not generally considered part of the explanation. It is assumed that someprocess would produce this result, but that process is rarely spelled out. Further, “inabstracting from the actual course of adjustment to the initial shock one is assum-ing that the adjustment process has little effect on the final outcome.” (Hausman,1990, p. 171), thus ruling out the processes described by Fisher (1983), mentionedabove. This implicit assumption connects well with uniqueness of equilibria. Thefinal result of an analysis is to find the effects of some independent variable on adependent one, R&D spending on patenting for example. Uniqueness implies thatthe path to the equilibrium is in a very real sense unimportant. If this equilibriumis the only place we can end up, and if we are in or close to equilibrium most of thetime (otherwise its usefulness as an explanatory tool would necessarily be calledinto serious question) what difference does it make how we get there? It is not eventhat interesting as a curiosity, since the path to equilibrium must be relatively short.

Causation, when seen this way, can be called “sustaining causation”, and fromthis point of view understanding a phenomenon involves knowing what holds thatphenomenon in place, and why there is no deviation from it. There is another typeof causation, though, which we can call “originating causation”.10

The process by which an event, state or phenomenon comes to exist is centralto explanation in Evolutionary Economics, and must at the very least be consideredwhen a mechanism to select among equilibria is at play. The general goal ofexplaining the movement of something over time, or to explain the existence of aphenomenon in terms of how it came to be is common to all evolutionary sciences(Nelson, 1995). Explanations of, for example, the location of an industry wouldinvolve a dynamic model of location choice, in which choices made by firms thisperiod depend on the location choices of previous actors. The process might evolveto a stationary distribution, but also might never do so, if the choices of later actorssufficiently upset the choices of earlier actors that they provide incentives to change

9 See Reder (1982, p. 12) for example, who describes a dominant, Chicago, view as assuming that“one may treat observed prices and quantities as good approximations to the long-run competitiveequilibrium values”.

10 See Cowan and Rizzo (1996) for a discussion of causation in economics and in particular on thedifference between sustaining and originating causes.

546 R. Cowan and D. Foray

location.11 The cause of the current distribution of activity is seen as historical. Wemust look back at decisions made in the past, and trace their effects on followingdecisions, as the choice of one actor affects the incentives of other actors. To un-derstand a phenomenon, then, we must understand the process by which it came tobe. There is a sequence of events leading to this outcome, and we must be able totrace this sequence, and the causal links between the events in it, before we “reallyunderstand” why this phenomenon exists.

This is all part and parcel of the possible existence of multiple equilibria. Whenmultiple equilibria exist, the selection mechanism gains prominence. Selectionmechanisms (in the world) are inherently historical. If there are two possible out-comes, one cannot simply look at sustaining causes to explain which is selected –by definition both are sustainable. One needs to address origins.

This discussion of causation impinges on counterfactual conditionals in a cru-cial way. We argued above that unpacking a counterfactual conditional involvesexamining causal connections. This is the process by which a counterfactual isjudged – if the causal story is judged acceptable, or credible, the counterfactualis deemed assertable. Clearly, then, the nature of causation will determine whatis unpacked during this enterprise. In a science in which causation is ahistorical,examination of the counterfactual claim will not involve history. Contrarily, if cau-sation is viewed as inherently historical, then examining a counterfactual claim willnecessarily involve history in one way or another.

We can make the same argument in a slightly different way. We pointed outabove that a counterfactual claim is implicitly a causal claim, but this is deeper thansimply a semantic identification. A claim about a counterfactual condition is aboutthe same thing as a causal claim. Thus the counterfactual will contain within it thenature of causation. What this means is that if unique equilibria are derived fromaccepted axioms, then simple counterfactual conditionals will be the norm. But ifmultiple equilibria exist, and events on the path are important in determining whichis selected, and perhaps where it is located, then counterfactuals involving historywill be common. In the style of economics we have been calling evolutionary,empirical work will involve historical excursus: hence the counterfactual history.

This discussion of causation makes contact with the solution of Conrad andMeyer (1964) to the problem of counterfactual conditionals. As noted above, thedifficulty arises from logical problems with the counterfactual statements them-selves. The solution, Conrad and Meyer suggest, is to look for statements that areco-tenable with the counterfactual statement. Consider, for example, the statement,“If there had been no Civil War, slavery in the southern US would have been dis-mantled peacefully.” There is no way to prove this analytically, nor is there a directempirical “proof”. But, following Conrad and Meyer’s exposition, this statementwould only be true if slavery were unprofitable at the time of the war. Thus we canimplicitly test this statement by testing the profitability of slavery. Notice, however,that this approach makes several assumptions about the causal forces at work. It

11 The lack of an equilibrium is not necessarily considered either a failure or a problem to be addressed.The role of equilibrium in economic analysis is something occasionally debated within EvolutionaryEconomics. We hope, within the course of the argument presented in this paper, to remain agnostic onthat issue.

Evolutionary economics and the counterfactual threat 547

assumes that profitability is key to the continuance of an institution, and, moretellingly, it assumes that the non-occurrence of the Civil War would not have af-fected the profitability of slavery, nor would it have set in play any other events thatwould have over-whelmed the effects of the (non)profitability of slavery. Both ofthese assumptions may be valid in this case, but by making them explicit we seethat the co-tenability solution runs the risk of being too static. It fits easily with theview of counterfactuals as describing a world as similar as possible to the actualworld at the time under examination, but it may not fit so well if one’s view ismore historical, namely that the counterfactual world must be the outcome of somefeasible path from some point in the history of the actual world.

We turn now to a discussion of these two views of counterfactual statements.

4 Two views of counterfactuals

The philosophical literature contains (at least) two views of counterfactuals: coun-terfactuals as non-actual possible worlds; and counterfactuals as untaken branchesin the tree of history.

4.1 Possible worlds

David Lewis (1973 a,b) sees a counterfactual conditional as describing a possiblebut not an actual world. There are many worlds which in principle could exist,the actually existing world being only one of them. A counterfactual conditionalis merely describing another. One might be a world just like ours, but in whichR&D spending is 25% higher than it is in ours. Lewis’s view is that to do scienceeffectively, which of course involves examining counterfactual conditionals, thisalternative world should be as much like ours as is possible, and in fact he describesa formal metric. His exposition is largely ahistorical since within it is no constraintthat it be possible to get to the alternative world from some point in our own actualhistory. It is simply a world running parallel to our own in some alternative universe,and which we examine when doing so might tell us something interesting aboutour own.12

12 Though he does not develop it particularly, Lewis allows for a more historical version of his theory.In effect, it involves putting non-simultaneous dates into the counterfactual: “If the roulette ball hadlanded on red instead of black last week, then today . . . ”. Since roulette wheels are deterministic, theantecedent involves a minor miracle if we are to keep the world the same up until the time at which theball landed. But following the landing of the ball, it seems that many major miracles could be necessary tokeep that possible world “just like” ours. So, Lewis (1974, p. 77) concludes, following the counterfactuallanding of the ball, we invoke existing causal (and other) laws, to trace out the consequences. What isless clear is what to do if it would have taken a major miracle to change the colour. Should we go furtherback in time, (to when the croupier got out of bed perhaps) to find a (counterfactual) historical path thatinvolves only a minor miracle? If this is the case, then this theory becomes very like a branching theory.Other authors, (see Bennett, 1984, or Jackson, 1987 for example) develop possible worlds theories whichexplicitly include temporal notions and thus share some of the features of explicit branching theories.For a discussion of possible worlds theories, see Sanford (1989, Ch. 7, 9 and 10).

548 R. Cowan and D. Foray

4.2 Branching

There is a second view of counterfactuals, defended by Jon Elster (1978, Ch. 6),which involves a branching view of history. History is seen as a tree, in whicheach decision taken represents a branching point. We proceed up the tree, movinghigher and higher, never descending, as decisions are made and history unfolds.A counterfactual analysis involves moving back down the tree to some branchingpoint, and examining a branch not taken. Sometime last year a decision was takenabout this year’s R&D budget. The analyst returns to the point at which that decisionwas taken and asks “If at that point a different decision had been taken, how wouldhistory have unfolded?”. Bring this alternative history up to today’s date and askhow the alternative present differs from our own.

Given the historical nature of evolutionary economics, and the view that smallcauses can beget large effects, it seems that the branching view of counterfactualsis more apt.

But in its most general form, a counterfactual analysis amounts to postulatingthat something is different than in fact it is, and examining what follows from thisdifference. Two difficulties arise. What is it feasible to postulate as a difference?How do we tell what follows? The key to addressing both of these difficultieslies in there being theory that can be employed in the analysis. Theory confinesa counterfactual analysis in two ways. First, it restricts what can be postulatedas the counterfactual antecedent. Second, it determines what follows from thosepostulates.13

4.3 Branching and the role of theory in counter-factual construction

The branching approach to counterfactual analysis consists in choosing a branchingpoint in our actual history, and tracing an alternative history that would followfrom a different branch having been taken. The branching view is common amongeconomic historians. When discussing the “new economic history”, Fogel (1973,p. 139) states, “In order to determine what would have happened in the absence of agiven institution, the economic historian needs a set of general statements that will

13 The introduction of theory in this way runs the risk of introducing circularity of two types. Ifthe aim of the empirical work is to test the theory, in the strictest version of testing, then using thetheory to guide the counterfactual construction of history runs the risk of assuming what is to be tested.This risk is not always as severe as it is made out to be. Often the basis of scientific judgement hasto do with internal consistency, which is not affected by this issue. The second source of circularityis more abstract. It would undercut the very starting point of counterfactual analysis. Theories can beand often are expressed as law-like statements. As Goodman has pointed out, any law-like statement isequivalent to a counterfactual conditional. Thus any counterfactual analysis which employs theory toguide and restrain it rests on further counterfactual claims (since whatever form the theory is expressedin, it can be re-written as a set of counterfactuals). So to argue that counterfactual analysis in general isacceptable because any counterfactual analysis must be grounded in theory fails as a defence of this formof endeavour. It is equivalent to saying that counterfactual analysis in general is acceptable since anycounterfactual analysis must be grounded on a set of counterfactual conditionals. This is problematicif one is concerned with defending the technique itself. We leave aside this concern, assuming that is itbetter dealt with, and indeed has been dealt with, in the philosophical literature.

Evolutionary economics and the counterfactual threat 549

allow him to deduce a counterfactual situation from institutions and relationshipsthat actually existed.” Elster (1978, p.181) is more explicit, claiming that thereare three conditions that must be met to make any such counterfactual analysiscompelling. First, the antecedent of the counterfactual conditional must be feasiblein a static sense; that is, it must describe a state that is internally consistent.14 Theclaim that had railroads not existed, GDP would only be slightly smaller, needs,as one necessary condition, that a world without railroads does not contain logicalcontradictions. Second, the antecedent, “railroads do not exist in 1890” must alsobe “insertable”; there must be some feasible historical path from a real historicalstate that would produce the state described in the counterfactual antecedent. Giventhe actual history of the world to some point, 1840 say, it must be possible to trace afeasible history which includes no logical or physical impossibilities from that pointto the one in which the antecedent – railroads do not exist in 1890 – obtains. Third,the consequent, “GDP in 1890 is virtually the same as it is in our world containingrailroads”, must be linked to the antecedent by a (dynamic) theory. Theory enterstwice. It enters into insertability conditions – tracing out a history from a real point,obeying all known laws, to reach the antecedent. It can also enter in the link betweenantecedent and consequent if they are separated in time. It is clear that all three ofthese conditions are governed by theoretical considerations: internal consistencyis a theoretical notion; the feasibility of a hypothetical history implies that it doesnot contradict held theories of how the world works; and the connection betweenantecedent and consequent is, if it is to be compelling, based on accepted causaltheories.

The importance of the role of theory here introduces what Elster (p. 184) con-siders to be the basic paradox of counterfactual analysis. The stronger is the theory(that is to say the more it restricts what can happen) the better grounded is the“deduction” of the consequent from the antecedent, and the tighter is the connec-tion between the (actual) world at the point when counter-facts are inserted, andthe counterfactual state under discussion; the tighter the connection, for examplebetween the actual world of 1840 and the 1890-world-without-railroads. But onthe other hand, the stronger is the theory, the more restrictive are the conditions ofinsertability, that is, the smaller is the set of antecedents that are consistent with“what we know about the world”. Any counterfactual analysis must avoid bothrisks of vagueness and risks of absurdity. An analysis is vague when the modelconnecting consequent to antecedent is too loose, or not specified completely – itallows too many things to happen. Absurdity arises when the antecedent runs awryof a tight theory, that is, a theory that is highly specific about what can happen,and so when used to examine the counterfactual antecedent, reveals a state that isinternally incoherent or absurd.15

The branching view of counterfactuals implies a stronger notion of consistencythan the possible worlds view. To logical and physical consistency the branching

14 This consistency condition is common to many theories of counterfactuals, Stalnaker (1968) beingperhaps most explicit about it.

15 One potential way around this paradox, which cannot always be employed, however, is to observethat the theory determining whether the antecedent is insertable may be different from the theory usedto deduce the consequent, provided the theories are not inconsistent.

550 R. Cowan and D. Foray

view adds historical (or originary causal) consistency as a requirement. Thus toask what would be the effect on GDP if railroads did not exist in 1890 is notto ask about the world in the instance that railroads were vaporised on the firstof January 1890, since some of the remaining physical and institutional structureswould be historically incongruous. Historical consistency rules that out as an avenueof inquiry, and demands instead that we consider the possibility that they werenot introduced in 1840, nor between then and 1890. Notice that this permits oraccommodates some looseness in the statement of the counterfactual antecedent.Whether the antecedent is stated as “railroads do not exist in 1890”, “railroads werenot introduced in 1840, or any time up to 1890”, or “railroads were never invented”,all demand the same analysis. The same theories apply in each case, assuming thecounterfactual consequent in each case refers to the state of GDP in 1890.

5 Counterfactuals in economics

One of the strengths of the axiomatic, deductive aspect of neo-classical economics,and something which makes the counterfactual analysis used there seem innocuous,is that the theory is very strong. Its strength places severe restrictions on what canbe assumed as antecedents. The primitives of the theory, utility functions and pro-duction functions, are tightly constrained by the analytical tradition and demandsof rigour: a small number of functional forms is acceptable, and all have similarregularity properties. Thus to a great extent the only possible counterfactual an-tecedents have to do with changing parameter values in the models. In any modelwith a unique equilibrium consequents follow very tightly from counterfactual an-tecedents. In fact the theory has been specifically constructed so that consequentsfollow antecedents by the laws of formal logic or mathematics. Tighter connectionsare not available even in principle.16 Further, when equilibrium is the focus of theanalysis the accepted form of causation is sustaining causation, from which it fol-lows that comparative statics analysis tells a complete story (again see Hausman,1990). We can place this in the context of Elster’s three conditions for counterfac-tual analysis. Because changes in parameter values are by definition exogenous tothe model, they lie outside the scope of economics, and so many changes are ac-ceptable. The only constraints are internal, logical consistency, and that the changenot create a violation of the functional form regularity conditions that most modelscontain. The “course of history” that produces the parameter changes is exogenousto the analysis, and so an economist has little, if anything, to say about it. Further,since the equilibria are typically unique, all possible histories will lead to the newequilibrium. The link between consequent and antecedent is deductive, which isthe strongest kind of link we know. So if the tenets of the theory are accepted,counterfactual analysis will be a very strong tool.

It is clear that any deductive science will be able to make very strong counter-factual claims, since theory provides a strong link between consequent and an-tecedent. If one agrees with the premises of the basic theory, one cannot argue with

16 This should not be read to be saying that therefore the theory is correct. It is simply to point outthat if the premises of the theory are accepted, then antecedents logically imply their consequents.

Evolutionary economics and the counterfactual threat 551

the counterfactual claims that are drawn from it. Some of evolutionary economicsfits this mould. Specifically, it is often possible to derive, using an evolutionarymodel, results about patterns of behaviour. Technology choice models, for exam-ple, derive results about proportions of agents using different technologies.17 Andevolutionary game theory models typically have this feature.

6 Counterfactuals and evolutionary economics

Not all of evolutionary economics can be characterized this way though. Many ofthe models yield probabilistic results. This is not surprising given the underlyingview that many processes are contingent (see Nelson, 1995). Probabilistic resultsare not a difficulty when two conditions are satisfied. The first is that there is adeducible relationship between the probability distribution and some exogenousvariable (a form of a comparative statics result); and second, that the phenomenonunder investigation is manifest in a way that permits estimation of the entire dis-tribution. When both of these are fulfilled, the counterfactual conditional relatingchanges in the parameter to changes in the probability distribution can be examinedsince both are observable. It is common, though, that the second condition fails, andthat we cannot observe distributions, but rather a very small number of realisationsof the distributed variables. This is typically the case with potential regret results.In the technology choice literature for example, a typical result which is provableanalytically is that with non-zero probability, the system will standardise on aninferior technology. When turned into a counterfactual conditional referring to aparticular instance we get: “Had we standardized on gas-graphite nuclear power re-actors, we would now be better off.” This claim seems much stronger than anythingpresent in the theory which only claims that if there are enough cases of technologycompetitions, some of them will have the result that an inferior technology wins.It makes no claims about nuclear power, nor about this particular realisation of itsdevelopment. If we were to confront the theory with the nuclear power case, whatwould be involved?

On the “possible worlds” view of counterfactuals, we need to imagine a worldin which we simply replace all the light water reactors with gas-graphite reactors,and ask whether it would be a better world. 18 This is the thought experiment ofa comparative statics analysis, and makes perfect sense if history does not matter.A (formal, analytic) map from technology-type to net-benefits-from-use will be

17 See David and Greenstein (1990) or Foray (1989) for surveys of this literature. See Cowan andCowan (1998) or Bassanini and Dosi (1998) for more recent contributions that derive results deductively.

18 The answer would almost certainly be that it would not, since the current state of gas-graphitereactors is not a particularly good one. Notice in general that if this is the view taken of the counterfactualconditional, it will be rare to find examples of dominant inferior technologies. If there were one, whywould users not switch? Apart from co-ordination problems, which can in principle be solved by makinginformation public, there is no reason for rational users not to create the other possible world. Thusthere cannot be dominant inferior technologies. This is the approach taken by strident opponents ofpath dependence and in our view involves applying a static possible worlds approach when a branching(or dynamic possible worlds) approach is appropriate. (See for example the writings of Liebowitz andMargolis, 1995 e.g.)

552 R. Cowan and D. Foray

enough to tell us whether gas-graphite would have been better. An ahistorical maprenders the possible worlds approach sensible.

On the branching view of counterfactuals, we must go back in history, prob-ably to about 1960, and change several of the decisions made at that point, andthen re-write the history to the present. (See Cowan, 1990, for a detailed history ofthis example). This is the approach taken if historical events do make a difference.And this is the approach appropriate to evolutionary economics in a great manyinstances. This follows from the presence of positive feedbacks in evolutionarymodels. Changes in decisions can be magnified, and re-enforced by the system.Again, though, if an analytic map from technology-type to benefits-from-use ex-isted, the counterfactual conditional would be strong again, even if history wereinvolved in that mapping. What renders this attempt infeasible, from the point ofview of many evolutionary economists, is the importance of randomness, coupledwith positive feedbacks. When these are both present, at the level of analysis typi-cally conducted by economists any outcome is under-determined by the data, evenin an equilibrium model. A central tenet of Evolutionary Economics (and commonto much of evolutionary economics) is that there are many sources of indetermi-nacy in any economy: the identities of interacting agents, which will necessarilybe beneath the vision of the analyst; the process of learning, which is often unpre-dictable; the presence or details of innovation which are, again, unpredictable, areall examples. This indeterminacy implies that model outcomes (at least the detailsof them) are under-determined, and further, that the actual determination of themis historical. This indeterminacy weakens the link between antecedent and conse-quent in the counterfactual. Evaluating it is no longer just a matter of formal logic,it now involves judgement.

It is important to notice that under-determination notwithstanding, history isnot unconstrained—that would make investigation impossible. But history is onlyloosely constrained, since even in a world governed by causal laws, the presence ofrandomness (or even simply ignorance on the part of the analyst, which means hecannot predict everything) implies that within those laws many things are possible.

We argued above that law-like statements can be re-expressed as counterfactu-als, and that thereby the investigation of law-like statements will be, implicitly atleast, an investigation of counterfactuals. Analysts investigate, therefore, the causallink between antecedent and consequent of the counterfactual conditional. If causa-tion is viewed historically, so will be the counterfactual and so the investigation ofit. But if history is only loosely constrained, many things can happen, which meansthat in all probability there is some link between consequent and antecedent. But“many things can happen” does not mean that everything happens with equal prob-ability. Nonetheless, the fact that many things have non-trivial probabilities impliesthat tracing a tight link between an event (or counter-event) and its consequencesmay be problematic, the more so as the two are removed from each other in time.History unfolds by choosing among alternative possibilities. That is, history makesa choice about what today is like. With every possible today, there is associated a setof possible tomorrows. If agents are learning (often in an unpredictable way) andagents are innovating (in an unpredictable way) and interacting with other agents(in an unpredictable way), the choice of today’s state does not place strong restric-

Evolutionary economics and the counterfactual threat 553

tions on the set of tomorrow’s possible states. This means that tracing the effectsof a (counterfactual) event becomes less secure as the tracing occurs over longertimes.

The most compelling counterfactual analyses exist when the theory on whichthey are based is tightest and most restrictive about what can take place. One ofthe intuitive strengths of evolutionary economics is the premise that in a systemas complicated and complex as any modern economy must be, any theory thattoo tightly constrains what might happen will leave out very important aspects ofeconomic activity. But this makes counterfactual analysis very difficult. If historymatters, then counterfactual analysis will often be historical. Without severe theo-retical constraints on what history can produce, though, the analysis runs the riskof being far from compelling.

We can summarise this section briefly. To accept a counterfactual conditionalwe demand an argument that it is possible to “infer” the counterfactual consequentfrom the counterfactual antecedent plus auxiliary conditions. This inference mustbe causal, so causes and the causal structure must be understood (Goodman, 1978).Hausman points out that for a comparative statics analysis to be coherent, theeconomist must believe that there is a process that would connect the states beforeand after the change in the exogenous variable. He argues that for a variety of reasonsthis process is not of particular interest in general. But, as Tunzelman (1990, p. 296)argues, counterfactual histories “represent an empirical working out of comparativestatics”, that is, they are an explicit rendering of Hausman’s process. Evolutionaryeconomics contains a general belief that the process between initial and final statesis important, due to the under-determined nature of outcomes, and so counterfactualhistory will be an important part of explanation. To come to the same conclusionfrom a different direction, we note that virtually all sciences embrace some formof causality, and use that notion as part of their explanations and to support theirpredictions. For Evolutionary Economics, since it is generally accepted that “cause”refers to originating or genetic causes and that the causal structure is temporal orhistorical, the argument supporting the counterfactual conditional will necessarilybe temporal or historical. But one of the underlying beliefs of many evolutionaryeconomists is that outcomes are under-determined; in general any basic data presentmore than one possible outcome. Thus theory does not tell us what will happen, itonly restricts us to a set of possibilities.

7 Evolutionary economics and counterfactual history

The discussion in the previous section seems to be driving toward the conclusionthat since empirical counterfactual analysis is so difficult to perform convincinglyin evolutionary economics, this branch of economics is fated to be either non-empirical or mere story-telling and therefore highly speculative, and presumablyof little practical value. In this section we argue that we need not be driven to thisconclusion. We present four different routes away from it – three “in practice”, andone “in principle”.

554 R. Cowan and D. Foray

7.1 Weak counterfactuals and appreciative theorizing

Counterfactual analysis can be used to buttress theory. The aim of appreciativetheorizing is less to formalise general laws, which could be used in predictingthe future, than to use general, well-understood theoretical concepts to understandeither the past or the present. 19 Evolutionary economics takes almost as a preceptthe idea that understanding is necessarily historical. Thus to understand the presentwe must know and understand the history that brought us here. As time passes,the economy is continually eliminating future branches, and thereby producingthe present. Understanding this process involves showing either that there were nobranching points, and thus the present state is inevitable given the starting point, oridentifying the branching points (those at which there was a non-trivial probabilityof taking a different route), and showing why the economy followed the route itdid.20

We refer to this exercise as one of weak counterfactual history, since it involvesminimal knowledge or assertions about the non-actual branches. In this regard itis only necessary to argue that the branches not taken would lead to a differentpresent.21 Even more generally, weak counterfactual analysis is restricted to under-standing the major events and chains of decisions which, coupled with processesthat magnified rather than damped their effects, can be considered as having playeda role in disconnecting some sub-regions of the tree from the branches followed byactual history.

Advantages of a weak counterfactual argument are clear. Accepting its goalsrenders Elster’s conditions less important, and thus permits economists to treat alarge number of problems in those terms. Drawbacks also seem clear however.Weak counterfactual analysis does not give welfare results, and in a sense can onlyargue that the actual outcome is not the only possible one. This does permit theclaim of potential regret, but cannot assert more; it will not make a case for actualregret (see Foray, 1997). For that we need something stronger.

7.2 Strong counterfactual analysis

Appreciative theorising, or appreciative analysis more generally, is sometimes con-sidered not enough. In particular, economists are often concerned about makingwelfare comparisons between two different states. This exercise involves knowingsomething about the nature of both states. It is difficult enough to know about the ac-tual present. But to compare it to a counter-factual present can be extremely tricky.If, following the arguments given above, we adopt the branching view, rather thanthe non-actual possible worlds view of counterfactuals, this comparison is made

19 For an example see Malerba et al. (1999).20 It is immediately clear that such an approach excludes one view of counterfactuals, namely coun-

terfactuals as non-actual possible worlds. The point of the exercise is to show that certain historicalevents contributed to forming the present state, and that had different branches been taken at some pointin the past, we would be at a different present.

21 It is possible that this argument may be difficult in certain cases, but in general to have identified abranching point is to have identified a moment when one future is eliminated.

Evolutionary economics and the counterfactual threat 555

more difficult as we must compare our state not simply with another state havingminimal variations from it, but rather with the state that would exist now followingsome counterfactual change at some point in the past.

The obvious difficulty of the exercise is that strong counterfactual analysis con-sists of (re)-vitalising something that did not happen.22 In contrast to weak coun-terfactual argument, the restrictions discussed above must be given much moreconsideration in the analysis. The theory or principles employed in strong counter-factual argument must, therefore, i) constrain the world of potential developments;and ii) use strong laws on the basis of which both to construct and to develop thealternative paths “selected” from the set of possible paths.

The first principle is relatively easy to illustrate. Consider the hypothesis thatenvironmental problems would be less severe if the internal combustion enginedid not dominate personal mobility technology. There are, currently, four seriouspossible automobile propulsion sources: internal combustion; lead-acid, or moreadvanced batteries; fuel cells; and hybrid battery-internal combustion technology.An argument that internal combustion is an inferior technology (on the environ-mental criterion) could only be built using as competing technologies steam andlead-acid batteries. This follows from the fact that the important events determin-ing the outcome of the competition took place at the turn of the century, and thefuel cell technology did not then exist as an automobile propulsion source. Thusthe important branching point in the history was a three-way branching in whichthe internal combustion branch was taken. This is related to Elster’s insertabilitycriterion. The fuel cell option is not insertable at the crucial point in time.23 But thisis, rather than a weakness, a strength of the analysis. A small set of options impliesthat a small number of branches need be traced to address the issue of whether thepresent state is an optimal one. Bifurcation points with few branches make the casefor regret stronger, by deriving it from a point with relatively few options.

The second condition, that counterfactual history be based on strong laws, ismore difficult to illustrate. A causal story is simplest and most compelling whenthere is a single causal factor that clearly dominates all other considerations. Inthis case a relatively tight link between antecedent and consequent can be drawn,because the construction of the chain of events is tightly controlled and constrained.

22 It is worth pointing out that facts about paths not chosen are often available. When this is thecase, analysis is considerably facilitated. In many cases of apparent total technological standardization,for example, small islands of the non-standard technology can be found. These islands can be used assources of information about how the technology might have developed had it been more extensivelyused. Another case might be a particular industry that had different trajectories in different countries orregions: the cotton industry in Texas and Mexico, which were very similar to each other at one point,adopted different strategies for pest control with dramatically different results (Cowan and Gunby,1996); for a similar case in ferrous casting in France and Germany, see Foray and Grubler (1989). Thuscounterfactual analysis is not necessarily “analysis without facts”.

23 It is possible that on a different history the fuel cell could have been available in 1900. But to makethis counterfactual analysis would involve going back yet further in time to find a point at which researchon fuel cells could have been taken up. Elster uses this notion to create a measure of difference – howfar back in history do we have to go to trace a feasible path to the counterfactual present we wish toexamine?

556 R. Cowan and D. Foray

The law must be strong in two senses: it must be general (that is to say, robust)and important. Generality is straightforward. The law must cope with a variety ofspatial and temporal contexts. Many macro-and micro-economic laws are generalin this sense.24

The other dimension, importance, is less simple. We need a law “importantenough” to cause the counterfactual construction we base on it to be, in effect,isolated from the influence of other laws. It must be over-ridingly strong. It clearlyfollows that the strength of a law cannot be dissociated from the context of theenquiry. Network effects, for example, can create forces which underlie a stronglaw in the field of telecommunications technologies. But these same forces willnot be “important enough” to build a counterfactual history about, for example, re-gional polarisation, since it is not plausible that the dynamics generated by networkexternalities could be isolated from the effects of transportation costs, infrastruc-ture externalities, local labour markets and so on. In this dimension, “importance”implies that the ceteris paribus clause can be plausibly invoked.

There is another problem which must be faced, namely the risk of circularity.Any economic system, agent or phenomenon will have many forces acting in or onit. The discussion immediately above had to do with identifying and examining theeffects of the strongest forces. This implies that forces can be identified as primaryor secondary. Concentrating attention on primary forces (of which one hopes thereare few) connected to strong laws will not generate misleading results. The iden-tification of primary and secondary forces is one typically done in the abstract, ortheoretically, and here lies the possibility of circularity, or perhaps assumption ofthe result. Evolutionary theory, in common with other theories of complex systems,typically has the feature that positive feedbacks, and the interconnections betweendifferent parts of a system, can magnify what appear to be secondary considerationsinto primary ones. Unfortunately, determining whether or not this is a relevant con-sideration can only be done using counterfactual history. So again the very featurethat makes counterfactual history a vital tool makes it a tricky one to use. We can seethis in the discussion of Robert Fogel’s (1964) work on US railroads. One centralissue of Paul David’s (1975) criticism is that Fogel ignored this possibility. Fogelmade a priori judgements about which forces were primary and which secondary,and then showed that when the primary ones were considered, US railroads had acertain effect on GNP. One interpretation of David’s arguments is that if the forcesconsidered secondary were put into a complete counterfactual history, they wouldhave been acknowledged to be as important as the primary ones.

7.3 Stronger theory to contstrain counterfactual analysis

We have just argued that whether or not a law is strong depends on context. Thismight suggest that a general change in context could have the effect of increasingthe strength of laws generally. In other words, it may be possible to make strongcounterfactual history more compelling by looking at theory slightly differently,reconsidering what questions are feasible, what count as respectable answers, or

24 On the issue of robustness of economic laws, see Kindleberger (1989).

Evolutionary economics and the counterfactual threat 557

what phenomena it is appropriate to analyse, if these have the effect of making thetheory in general stronger. If, for example, the theory is seen to be more deductive,and in particular more deterministic, this reduces the “anything can happen” prob-lem. But because of the underlying tenets of evolutionary economics, to adopt sucha strategy implies thinking about economics problems in a slightly different way.

Neo-classical economics attempts to make predictions of the behaviour of eco-nomic agents, either as individuals or as aggregations. In models of perfect compe-tition all agents are infinitesimally small, and thus as individuals have no influenceon outcomes, so deducing market outcomes is facilitated by there being few “pub-lic” effects of individual actions. At another extreme, we observe models with fewenough agents that all can be traced and the non-anonymity is tractable. Evolution-ary models, though, often have a large but finite number of heterogeneous agentsin them, and these agents are affected by each others’ actions. When this is true,a detailed prediction of agent behaviour is simply not possible.25 It is this largepopulation of heterogeneous agents, interacting with and affecting each other, thatcauses the serious under-determination of evolutionary models. Even if there is noontological uncertainty, epistemological uncertainty can be extreme.

Whether or not under-determination is important, though, is crucially dependenton the level of analysis. Where in the world it will rain today is certainly under-determined by my knowledge of the weather system. That it will rain somewhere isnot. And with a small amount of research one could almost certainly predict fairlyaccurately how much rain will fall today. Thus we can reduce considerably thedegree of under-determination simply by changing the level or nature of analysis.While the details of the actions of individual agents may be under-determined,general patterns of behaviour may not be. This was an idea present in the work ofHayek (1967, for example), and based on roughly the same arguments – there is toomuch information present in the economy to be processed by any analyst. Any agentwill have more or different information than any analyst, thus the goal of predictingdetailed agent-level behaviour is un-reachable. There will be general tendencies,however, which are predictable, due, in effect, to the central limit theorem.

Consider models of firm location. A firm chooses a location based in part onthe locations of other firms. Locating near another firm on the one hand permitsencroaching on its market, but on the other reduces one’s monopoly power. In earlymodels of such a system, (Eaton and Lipsey 1975, is a good example) the questionwas where firms would locate in a certain space. It quickly became apparent, though,that in many models attempts to expand beyond very small numbers of firms, orbeyond one dimension, created intractable problems, partly because of the economicprocess involved. (Note that the goal was not to ask where a specific firm, FirmA, would locate but rather a weaker one, whether there would be any firm at all atsome location X). The question was un-answerable, possibly even in principle. Adifferent question, though, and one which involves a different level of analysis, iswhether firms will cluster or spread out. This is of course related to the central issuein Hotelling’s (1929) original paper. It is also possible to ask how large clusters will

25 It is inconceivable that the analyst could get enough information, and even if he could get theinformation, he would not be able to process it.

558 R. Cowan and D. Foray

be, and how close together they will be. Here the questions do not refer specificallyto firm behaviour, but rather to the properties of the system as a whole, and inparticular the pattern or distribution of location. It is possible to address thesequestions analytically. By changing the unit of analysis we have recovered strongtheory and the analytical results that are produced using deductive tools, yet themodels underlying these results retain the evolutionary processes inherent in theeconomic decisions made by firms.

A second approach to making evolutionary theory tighter is to make the anal-ysis statistical. The “interacting agents literature” models large numbers of (het-erogeneous) agents who interact both within and outside the market. Populationproperties evolve as agents respond both to their own past actions and to the actionsof other agents. Here, theory seeks to make predictions about statistical propertiesof populations rather than about the properties of individual agents (this literature isreviewed in Kirman, 1996, and Durlauf, 1996). Schelling (1978) refers to analysiswhich explores the relationships between behavioural characteristics of individualsand those of the aggregate. Developing counterfactual arguments is dependent onour ability to calculate aggregate data as determined by individual actions. Thisis indeed possible for some classes of social and economic interactions. Schellinguses a simple example to illustrate: “if we know that every driver, on his own, turnshis lights on at sundown, we can guess from our helicopter we shall see all thecar lights in a local area going on at about the same time.” (Schelling, p. 13). Inmore complex cases, however, agents’ behaviour or agents’ choices depend on thebehaviour or the choices of other agents; and those situations often do not permitany simple summation or extrapolation to the aggregates. Then counterfactual con-structions must look at the system of interactions between individuals and theirneighbourhoods.

Both approaches, by changing the focus of analysis, produce strong theoryand analytical results. They both reduce the historicity present in the abstract re-search, and permit counterfactual analysis which is more like comparative staticsin appearance. Tight theory and tight counterfactual conditionals are possible.

7.4 A different criterion by which to judge the success of an analysis?

There are occasions in which strengthening the theory to make its predictions moreprecise in one of these ways is the right approach. But there may be phenomenafor which it is not appropriate, and some looseness in the theory, and consequentlooseness in the counterfactual empirical claims is either desirable or necessary.When this is the case, the arguments given above indicate a problem for empiricalevolutionary research. It may fall prey to the fairly common view that while thetheory is pretty good, empirical evidence is frequently weak. But this may betoo hasty. One can only judge empirical endeavours of an approach by standardsappropriate to them; standards to judge success must be consistent with the taskundertaken.

Intuitive arguments against counterfactual history by and large suppose that ittends to be based on theory which is so loose that it allows anything to happen. If onechanges the past one of course can change the present. Any counterfactual claim

Evolutionary economics and the counterfactual threat 559

can be asserted given the weakness of the theory involved. Therefore, because theresults are “loose” and are based heavily on the judgement of the analyst, they areof little (or no) value. But this argument is to say that “this apple makes bad orangejuice.” Much of economics has opted for an axiomatic approach involving deductivetheory. Counterfactual conditionals are deduced under standard assumptions, andempirical investigation involves checking the values of parameters (traditionally,checking only that they are of the right sign). Predictions about behaviour arearrived at deductively, and are supported by finding an elasticity in the right range.Is the measured elasticity of patenting with respect to R&D spending large or smallor negative? Answering this question can corroborate the theory with which thequestion was generated. The argument is of the form, “If assumptions 1 through nhold, then the elasticity of patenting with respect to R&D is large.” Measuring thevalue of the elasticity can provide very strong indications regarding the theory andwhether the theory is valuable in understanding innovation.26

When counterfactual conditionals are not (formally) deduced from axioms orassumptions, or when the analysis generates multiple or possibly infinite equilibria,this simple argument form is not available. Knowing an elasticity neither provesnor disproves the theory, nor does it permit one to predict the future. The latterfollows from the under-determination typical of the approach we are considering.

If one opts for a non-deductive theory, empirical analysis will be of a differentnature, and it follows that the criteria by which it is judged will also be different. Inparticular, the assertability of a counterfactual necessarily rests on a different typeof argument. A simple statistical test will not, in general, provide the empiricalarguments that buttress a theory, simply because under the tenets of evolutionarytheory, the counterfactuals that emerge are necessarily historical. To reject evolu-tionary theory on the grounds that it provides no identifiable statistical tests butrather merely the opportunity to “tell stories”, involves first justifying the criteriaby which it is being judged.

Loosely, what empirical work does is to confront theory with “the facts”. Buteven Popper admits that the notion of a brute fact, that is, a fact that exists inde-pendent of some held theory, is not tenable (see Hollis, 1994, p. 76; or Feyerabend,1988, p. 155). All facts or observations are theory-dependent; what counts as orconstitutes a fact will depend crucially on the theories one holds. This implies thatthe criteria we use to judge a theory on empirical grounds will also be theory-dependent.27

To argue that the theory is weak because it provides no testable hypotheses,where the definition of “testable hypotheses” follows from the milieu of deductivetheory, demands first a justification of the use of deductive theory. This has yet to be

26 Recall the stricture that in fact knowing the elasticity does not prove the theory. The only strongstatement we can make is that if the elasticity is not large the theory is wrong. This suggests that even onits own criteria, if strictly applied, the empirical evidence in support of much economic theory is fairlyweak.

27 We are not arguing here that it is therefore impossible to compare theories (see Kuhn, 1962), ratherwe are arguing that assertions about the strength or weakness of empirical evidence for a theory musttake into account what it is possible to produce as evidence, and how that sort of evidence ought to bejudged.

560 R. Cowan and D. Foray

provided in economics (or, arguably, anywhere else for that matter; see Feyerabend,1988). To reject particular instances of the analysis as having failed to pass the testof empirical evidence involves accepting the standards of evidence that are impliedby the theory itself. If you accept the theory, then you must accept its definition ofa testable hypothesis. The idea that you can separate the two was given up with thedeath of the theory-observation distinction.

Arrow (1995, p. 21), suggests that “The theory is pretty goood, the empiricalevidence may be by definition pretty hard to come by, not just as a practical matter.”Indeed, the argument above suggests that creating a robust empirical case for anevolutionary model will be a challenge due to the very nature of the theory. Thismay make the empirical evidence for the theory appear weak. And perhaps by thestandards of axiomatic, deductive analysis the empirical evidence is weak. But ifyou admit that the “theory is pretty good”, you must accept that the theory itselfimplies that empirical evidence will take a certain form, from which it follows thatcertain standards apply when judging that evidence. The standards developed tojudge empirical evidence when the underlying theory is axiomatic and deductive donot necessarily apply to other types of theory. Different standards do. In the contextof a path-dependent, evolutionary economics, whether or not a counterfactual con-ditional is accepted as empirical evidence depends not on the statistical significanceof parameter estimates but rather on whether an historical, causal explanation iscompelling, and thus whether the counterfactual is assertable. The evidence looksmuch stronger when judged on its own grounds.

References

Arrow K (1995) Economics as it is and as it is developing: a very rapid survey. In: Albach H, RosenkransR (eds) Intellectual property rights and global competition. Sigma, Berlin

Arrow K, Skitovsky T (eds) (1969) Readings in welfare economics. Allen and Unwin, LondonBassanini A, Dosi G (1998) Competing technologies, international diffusion and the rate of convergence

to a stable market structure. IIASA Interim Report IR 98–012Bennett J (1983) Counterfactuals and temporal direction. Philosophical Review 9: 57–91Bunge M (1979) Causality and modern science, 3rd edn. Dover, New YorkChisolm R (1946) The contrary-to-fact conditional. Mind 55: 289–307Conrad AH, Meyer JR (1964) The economics of slavery and other studies in economic history. Aldine,

ChicagoCowan R (1990) Nuclear power reactors: a study of technological lock-in. Journal of Economic History

50: 541–566Cowan R, Cowan W (1998) Technological standardization with and without borders in an interacting

agents model. MERIT Research Memorandum, # 2/98–018Cowan R, Foray D (1998) Counterfactual history and evolutionary economics. MERIT Research Mem-

orandum, # 97–012Cowan R, Gunby P (1996) Sprayed to death: path dependence, lock-in and agricultural pest control.

Economic Journal 106: 521–542Cowan R, Rizzo M (1996) Genetic causation and modern economic history. Kyklos 49(3): 273–317David PA (1975) Transport innovations and economic growth: Professor Fogel on and off the rails. In:

Technical choice, innovation and economic growth: essays on American and British experience inthe nineteenth century. Cambridge University Press, Cambridge

David PA (1988) Path dependence: putting the past into the future of economics. Technical Report#533, Institute for Mathematical Studies in the Social Sciences, The Economics Series, StanfordUniversity

Evolutionary economics and the counterfactual threat 561

David PA (1997) Path dependence and the quest for historical economics: one more chorus in the balladof QWERTY. Oxford Discussion Papers in Economic and Social History, #20

David PA, Greenstein S (1990) The economics of compatibility standards: an introduction to recentresearch. Economics of Innovation and New Technology 1: 3–41

Durlauf S (1996) Statistical mechanics approaches to socioeconomic behavior. Santa Fe Institute Work-ing Paper 96-08-069

Eaton C, Lipsey R (1975) The principle of minimum differentiation reconsidered: some new develop-ments in the theory of spatial competition. Review of Economic Studies 42: 27–49

Elster J (1978) Logic and society: contradictions and possible worlds. Wiley, TorontoFeyerabend P (1988) Against method. Verso Books, New YorkFisher FM (1983) Disequilibrium foundations of equilibrium economics. Cambridge University Press,

CambridgeFogel R (1964) Railroads and American economic growth: essays in econometric history. Johns Hopkins

Press, BaltimoreFogel R (1973) The specification problem in economic history. In: Temin P (ed) The new economic

history, Penguin. Reprinted from (1967) Journal of Economic History 27: 283–308Foray D (1989) Les Modeles de competition technologique: une revue de la litterature. Revue

d’Economie Industrielle 48: 16–43Foray D (1997) The dynamic implications of increasing returns: technological change and path depen-

dent inefficiency. International Journal of Industrial Organization 15: 733–752Foray D, Grubler A (1990) Morphological analysis, diffusion and lock-out of technologies. Research

Policy 19(6): 535–550Gallison P (1987) How experiments end. University of Chicago Press, ChicagoGenovese E (1962) The Significance of the Slave Plantation for Southern Economic Development.

Journal of Southern Economic History 28: 422–437Goodman N (1983) Fact, fiction and forecast, 4th edn. Harvard University Press, CambridgeGoodman N (1978) Ways of worldmaking. Hackett, IndianapolisHollis M (1994) The philosophy of social science: an introduction. Cambridge University Press, Cam-

bridgeHausman D (1990) Supply and demand explanations and their ceteris paribus clauses. Review of Political

Economy 2: 168–187Hayek F von (1967) Studies in philosophy, politics and economics. University of Chicago Press, ChicagoHirsch W (1952) Manufacturing progress functions. The Review of Economics and Statistics 34: 143–

155Hotelling H (1929) Stability in competition. Economic Journal 39: 41–57Jackson F (1977) A causal theory of counterfactuals. Australasian Journal of Philosophy 55: 3–21Kaldor N (1934) A classificatory note on the determinateness of equilibrium. Review of Economic

Studies 1: 122–136Kindleberger C (1989) Economic laws and economic history. Cambridge University Press, CambridgeKirman A (1996) Economies with interacting agents. European University Institute, FlorenceKuhn TS (1962) The structure of scientific revolutions. University of Chicago Press, ChicagoLewis D (1973a) Causation. Journal of Philosophy 70: 556–567Lewis D (1973b) Counterfactuals. Harvard University Press, Cambridge, MALiebowitz S, Margolis S (1995) Path dependence, lock-in and history. Journal of Law, Economics and

Organization April: 206–226Mackie J (1962) Counterfactuals and causal laws. In: Butler RJ (ed) Analytical philosophy. Basil

Blackwell, Oxford 66–80Malerba F, Nelson R, Orsenigo L, Winter S (1999) ‘History-friendly’ models of industry evolution: the

computer industry. Industrial and Corporate Change 8(1): 3–40McAfee RP (1983) American economic growth and the voyage of Columbus. American Economic

Review 73(4): 735–740McCloskey D (1989) Counterfactuals. In: Eatwell J, Milgate M, Newman P (eds.) The new palgrave: a

dictionary of economics. Norton, New YorkMill JS (1967) A system of logic. Longman, LondonNelson R (1995) Recent evolutionary theorizing about economic change. Journal of Economic Literature

33: 48–70

562 R. Cowan and D. Foray

Putnam H (1987) The many faces of realism. Oper Court, La Salle, ILReder MW (1982) Chicago economics: permanence and change. Journal of Economic Literature 20:

1–38Sanford DH (1989) If P then Q: conditionals and the foundations of reasoning. Routledge, LondonSchelling T (1978) Micromotives and macrobehaviors. Norton, New YorkSimon HA, Rescher N (1966) Cause and counterfactual. Philosophy of Science 33: 323–240Stalnaker R (1968) A theory of conditionals. In: Rescher N (ed) Studies in logical theory. Basil Blackwell,

OxfordStrotz R (1956) Myopia and inconsistency in dynamic utility maximisation. Review of Economic Studies

23: 165–180Tunzelman GN von (1990) Cliometrics and technology. Structural Change and Economics Dynamics

1(2): 291–310Wright TP (1932) Factors affecting the cost of airplanes. The Journal of the Aeronautical Sciences 3:

122–128