1 Introduction

Across much of social science, the general linear model (GLM) holds sway. However, it has significant limitations. This article makes the case for studying social processes as co-evolving systems, involving non-linear dynamics. Co-evolution of species in the natural world is a blind process. In the social world in contrast, purposeful interventions by social actors are omnipresent, in their struggles for positional advantage. The article brings together co-evolving networks and purposeful social action in the “Contingent Historical Model.”

We seek to apply this model in ways that engage with both scholarly and policy concerns. If such investigations are to be fruitful, they must not only be elaborated theoretically, they must also be applied to empirical datasets. This article considers how this can be done, with what sorts of data sets and what forms of data analysis. It takes as its specific example the international datasets on patents, as revealing processes and patterns of technological innovation.

Section 2 examines the assumptions that underlie the GLM, but also the weaknesses that are exposed, when those assumptions are not met. These weaknesses reveal the desiderata that any alternative must fulfill. In light of this, Section 3 examines models of co-evolving systems, as a form of non-linear dynamics. Section 4 elaborates the corresponding mathematics of autocatalytic sets, arguing that these must become a normal part of the social scientist’s toolkit.

Co-evolution was first noticed by Darwinian biologists, as a blind process of the natural world. The complex dynamics of the social world are likewise in varying degree blind and emergent, without collective purpose. Nevertheless, purposeful interventions by social actors are omnipresent; this entails struggles for positional advantage and the exercise of power. Section 5 brings this together in the “Contingent Historical Model” (CHM) which we juxtapose to the GLM. This we seek to apply empirically and in ways that engage with both scholarly and policy concerns.

To do this involves three steps. Section 6 considers how the mathematics of co-evolving systems can be applied to empirical datasets and the requirements those datasets must fulfill. Section 7 applies these to databases of patents, as a vehicle for studying technological innovation. Section 8 examines how far an empirical network of co-evolving technologies can thereby be constructed, as a contingent historical process.

Section 9 brings together the results of this enquiry and the implications for future work.

2 The General Linear Model¹

The task of social science is to explain social phenomena. This, it is commonly asserted, should involve measuring the effects of different “independent” variables on some “dependent” variable of interest. This can be presented diagrammatically. In Figure 1, the independent variables x₁, x₂ and x₃ shape the dependent variable y (but with some effects exerted via the intermediate variable z).

This vision or ontology of the social world can be presented as an equation:

Quantitative social science applies this vision using regression analysis. It commonly casts the problem of explanation in terms of a set of linear equations; this is why it is often described as the “General Linear Model” (Abbott, 2001, Ch. 1). The GLM looks for the straight line that best estimates — and therefore “explains” — the dependent variable y as the additive outcome of a number of independent variables x₁ … x_n plus a random error term u:

The b coefficients measure the rate at which changes in the independent variables x produce changes in y. The error term u is a measure of how closely our equation captures the empirical data — and how much “noise” remains around its predicted values of y.

There are a number of assumptions involved here however, which should be made explicit.

1. The GLM assumes that the separate causal effects of the independent variables can be isolated — there are no significant interactions among them. When interactions are substantial, equation (1) cannot in general be solved mathematically; and statistical models such as equation (2) may produce biased estimators and reduced significance for each of the variables in question.²

2. The GLM assumes that the influence of the dependent variables x on the dependent y is downstream and one-way. There are no significant feedback processes allowing y to influence x. And yet of course, in the real world such feedback processes are common. This may mean that instead of a downstream and uni-directional determination of y by x, the directions of influence run upstream as well (Figure 2). Such self-reinforcing processes are well-recognized across the social sciences, including for example the economics of “cumulative causation” in Myrdal and Kaldor (Toner, 1999), system dynamics (Checkland & Scholes, 1990) and the social policy literature on the dynamics of social exclusion (Room, 1995). Nevertheless, across much of social science the mathematical modelling of such cumulative feedback is under-developed.³

3. The GLM assumes that changes in the independent variables will produce broadly proportional changes in the dependent variable, across the entire range of observable values, within a rather simple “timescape.” In the presence of the afore-mentioned feedback processes however, timescapes are often more complex. Some effects of the independent variables are short-term and immediate, while others are long-term and delayed (Abbott, 2001, Ch. 1). Many social processes are replete with time lags, ratchets and path dependencies (Lieberson, 1987, Ch. 4). Pierson (2004) points to the consequences of long-term and often slow changes in background social and economic conditions (pp. 74–77). There may be long periods of stasis, and then thresholds at which sudden avalanches of reconfiguration occur (as sometimes discussed in terms of “punctuated equilibria” (Baumgartner & Jones, 1993).

   It is possible for the GLM to handle some of these temporal complexities, using more sophisticated statistical techniques. Abbott’s critics claim that he overlooks the advanced econometric techniques available to quantitative sociologists such as structural equation modelling (SEM), time series analysis and hierarchical linear models (Stolzenberg, 2003, p. 422). Nevertheless, taken together, the GLM struggles, when applied to social sub-systems and processes with complex inter-connections.⁴

4. The GLM assumes, finally, that the independent and dependent variables are given and that the relationships among them are fixed. This may well be appropriate, at least in the short term.⁵ Sooner or later however, the relationships among the “variables” may change and the variables themselves amalgamate, divide or disappear and new ones emerge. For understanding these dynamics of wholesale reconfiguration, the techniques of the GLM are of little use.⁶

It is hardly surprising that the GLM exerts such a powerful sway across many disciplines. It is readily visualized; it can be formalized in terms of simple equations; it is convenient and tractable.

While the social world is not in general linear, linear methods can often be used to good effect, if they are understood as “local” approximations. Nevertheless, where the foregoing assumptions fail to hold, the general applicability of the GLM is put in question. Among natural scientists, there are clear and well-recognized limits on applying linear models to particular phenomena. Social scientists need similarly to have alternative strategies available, for when linear models are not appropriate (Jervis, 1997, pp. 34ff).

The weaknesses we have identified provide clues that point directly towards the alternative we require. It is after all precisely by confronting inconsistencies in this way that scientific understanding and practice are able to advance (Kuhn, 1970; Popper, 1994, Ch. 5; Tavory & Timmermans, 2014). The most appropriate alternative, we now argue, is to adopt a model of co-evolving systems, as a form of non-linear dynamics.

3 Co-Evolutionary Dynamics

The GLM finds its inspiration in classical physics. Evolutionary biology provides a quite different source of conceptual and methodological inspiration — albeit one which social scientists have interpreted and applied in a diversity of ways.⁷

The strengths of an evolutionary model align closely with the limitations of the GLM discussed in the previous section. With evolutionary science as our starting point, we now therefore consider co-evolutionary dynamics, as an alternative to the GLM.⁸

In his account of the diversification of species, Darwin was centrally concerned with processes of adaptation to different habitats. He depicted this visually as a “Tree of Life” (Darwin, 1859/1998). This offered successively sprouting branches and sub-branches, as particular “variations” adapted to and exploited different habitats, over many millions of years.

Imagine looking down on the tree of life from above and viewing the top-most branches. Displayed there are the various species that are alive today (or we could consider any other horizontal “cut,” representing the species that lived at another chosen period in the Earth’s history). Across each of these cross-sections, the various branches (species) are connected in ecosystems of interdependence, involving dynamic synergies and arms races of co-evolution (Kauffman, 1993; Maynard Smith & Szathmary, 2000). These ecosystems typically involve populations far removed from each other across the evolutionary tree — for example, flowers and insects. They powerfully influence which species thrive and which are extinguished.

Figure 1 provided a visual representation of the GLM in its most basic terms. Such diagrams can provide powerful images that organize and direct our thinking about a given phenomenon. We now therefore consider a correspondingly parsimonious representation of evolutionary dynamics (Figure 3). This brings centre-stage variables that emerge, divide and disappear, as their interrelationships unfold. In subsequent sections we apply this to the social world.

A and B (at the bottom of the diagram) are two mutually adapted entities in the world of today. They might for example be the populations of two species such as bees and flowering plants, each benefitting as the other thrives.

We then pose the question: how did this mutual adaptation arise? We decline to treat it as a causal correlation, with the population of flowers “causing” the population of bees within a timeless environment. Instead, we seek to unpick the intricate and messy history of successive contingencies that has led to the mutual adaptations of today.

The upper part of the diagram reveals those historical contingencies. A1 and B1 were the ancestors of today’s bees and flowers. As we know from Darwin, in each generation, variations are produced. In general however, as long as the environment remains stable (Period -2) they are unlikely to displace A1 and B1.

It is when some environmental change occurs, at the start of Period -1, that we may expect some of the variations (A21, A22, B21, B22) to be adopted as superior to A1 and B1, which now become extinct. However, which of the variations A21 and A22 becomes preponderant depends in part on the new biotic environment constituted by the arrival of B21 and B22; and vice versa. In short, what is crucial is which of the four sets of interactions between A21 and A22 on the one hand, B21 and B22 on the other, is of greatest mutual benefit.

In the diagram, we show the relationship of A22 with B21 as being this favoured pairing, this synergy or “elective affinity.”⁹ Each will now accelerate the flourishing of the other. Their flourishing will in turn deny resources to A21 and B22, which in the “struggle for existence” become extinct. Hence we arrive at the bottom row of the diagram, Period 0, where A and B dominate. Here, by virtue of their domination, the environment is quite different from that in Period -2 or even in Period -1. And indeed, A and B are themselves may be quite different in their capacities from their respective forebears A1 and B1: perhaps barely recognizable (Shubin, 2008). Nevertheless, the domination of A and B is unlikely to last for ever; further rounds of interaction with the wider eco-system will eventually destabilize it, as new rounds of variation and selection are set in motion.

The dynamic synergies among particular elements have, as their obverse and corollary, the progressive disruption of other connections and elements of the eco-system — and the incorporation of those elements into the “empire” of the favoured elements (see Figure 4). The A22–B21 axis becomes a vector of cumulative change, around which the wider eco-system is progressively reordered and re-configured. This also makes it a non-linear system with strong path dependency, where instead of the additive effects that are central to the GLM, change is multiplicative and self-reinforcing.

The dynamic synergy cannot however continue without limit; nor can the concomitant disruption and recycling of other elements. Some parts of the wider eco-system are too resilient and robust to be unpicked and re-worked; they constitute an “evolutionarily stable state” (Maynard Smith, 1982). On the other hand, the disruptive change that is driven by the elective affinity of A22–B21 is forever opening up new possibilities for other elements of the ecosystem, other dynamic synergies that may eventually match, surpass or disrupt A22–B21.

In short therefore: the GLM finds its inspiration in Newtonian mechanics, with stable entities having causal effects on other stable entities. The process of co-evolutionary change elucidated by Darwin and his successors is quite different. Here the populations of different species share a common space, and interact with each other in a variety of mutual synergies and antagonistic threats. Through successive generations, the populations of these different species successively re-constitute each other, for the next round of the competitive struggle. These selection dynamics can shift; they will not necessarily be incremental and gradual; they can instead involve punctuated equilibria and tipping points and core shifts (Kauffman, 1993).

There is a large scientific literature concerned with these co-evolutionary dynamics — conceptually, methodologically and empirically — and with their application beyond the biological realm. This literature has been central to the larger discussion of complex systems that has developed over recent decades.¹⁰

The sections that follow apply these insights to the modelling and empirical analysis of social and technological change.

4 The Maths of Co-Evolving Systems

We now require an empirical methodology, appropriate to the dynamics of co-evolving systems and applicable to the social world. In developing this methodology, we continue to take inspiration where appropriate from evolutionary biology, as well from social scientists who have sought to capture these dynamic processes in their empirical research.

In the GLM, the variables and their boundaries are fixed. So are their interrelationships. The practitioner of GLM seeks to distinguish and measure the effects of the independent variables on the dependent variable. Much energy has been devoted to clever ways of doing this, even under apparently unpropitious circumstances.

Co-evolutionary dynamics involve a quite different notion of causation. Here are phenomena whose configuration and frequency are driven by a dynamic that is synergistic not additive, as depicted parsimoniously in Figure 3. It is the “elective affinities” among particular elements that lead to their progressive domination of the system in question and the corresponding reconfiguration of the variables.

4.1 Qualitative System Dynamics

Powell has developed qualitative system dynamics (QSD) for the analysis of organizational change (Powell, 1992; Powell & Bradford, 1998); he builds on the work on system dynamics and causal loops by such writers as Checkland and Coyle (Checkland & Scholes, 1990; Coyle, 1996). He first maps the organizations of interest and the connections of interdependence among them. He then labels each line of interdependence, to indicate its direction, but also whether the relationship is direct or inverse — whether, in other words, an increase in some property or activity of the “upstream” node causes a change in the “downstream” node that is positive or negative.

Within this map, Powell proceeds to identify those cycles whose links are all positive. These are cycles which loop back on themselves in self-reinforcing circles. When any one element starts increasing, the whole sub-system experiences explosive growth; when any starts decreasing, the sub-system experiences implosive collapse. Powell refers to these as “runaway loops.” In Figure 5 the runaway loop is marked as a dotted line. Each element in this loop accelerates the flourishing of its downstream neighbour, just as happened in Figure 3 between the elements A22 and B21. Meanwhile other cycles of interdependence loop back on themselves, in ways that dampen down change and stabilize the system as it presently exists.

Having identified the runaway loops, the next step is to assess how strong is the self-reinforcing dynamic. This will determine the speed at which the cycle “runs away.” Second, we need to know how well-connected the sub-system or loop is to the system as a whole, so that its runaway loops have wider influence. There may be particular threshold effects: beyond a certain point, the runaway subsystem triggers other sub-systems, which in their turn begin also to “run away.” The particular dynamics that emerge will be heavily dependent on the way that elements are connected to each other.

Nevertheless, Powell’s QSD assumes that the configuration of elements — in Figure 5 for example — remains fixed. There are no novelties and no extinctions. He does not allow for the sub-system that runs away to change its configuration or to re-shape the larger system in which it is embedded. Powell advances our quest: but he does not provide us with a co-evolutionary dynamic.

4.2 Autocatalytic Sets

Jain and Krishna (2003) are interested in autocatalytic sets (ACS) and the key role these appear to have played in the origins of life. An ACS comprises a set of simple molecular organisms, none able individually to self-replicate, but each providing a catalyst for each of its fellows — a process of symbiotic and co-evolutionary “boot-strapping” for collective self-replication (see also Kauffman, 1993, Ch. 7). This is of central relevance to any Darwinian account of the origin of species.

Like Powell, Jain and Krishna employ a methodology of directed graphs (networks where the direction of the connection matters).¹¹ They model a population located at each of the different nodes, dependent on the growth of population at a number of other nodes. Their first step is to watch which nodes thrive and which do not — and the significance of the ACSs for this variation in fortune. They notice (Section 3) that the population of a node enjoys particularly rapid growth if it is part of an ACS. An ACS thus plays a role in Jain and Krishna similar to that of “runaway loops” in Powell.

So far, this analysis by Jain and Krishna — like that of Powell — does not allow for any reconfiguration of the system or any transformation of its elements. This we will refer to as the “fast” dynamic, as the nodes within a given configuration of the network thrive to different extents, their populations eventually reaching an equilibrium. Jain and Krishna then however apply a Darwinian rule, to effect a re-configuration of the network (Sections 4–5). This we will refer to as the “slow” dynamic, occurring only once the fast dynamic has reached an equilibrium. From among the nodes that have flourished the least, a randomly chosen few are extinguished; random new nodes and connections are then added, mimicking the mutations in Darwin. Such reconfigurations are of course central to any evolutionary framework.

This formal analytic of the fast and the slow dynamic thus captures co-evolutionary dynamics in elegant and parsimonious fashion. The fast dynamic sees the various nodes or species thriving differentially, by virtue of their interactions with their fellows. It is those nodes associated with ACSs that thrive the most; just as in Figure 3, it was elements A22 and B21 that thrived by virtue of their mutual affinity. As with Darwin, the slow dynamic sees those nodes or species that fail to thrive being extinguished, allowing new nodes or mutations of species to discover a niche for themselves.

Jain and Krishna use computational models to simulate what can happen in such a dynamic system, depending on the parameters and algorithms adopted. For one illustrative moment in these successive transformations, see Figure 6. One ACS has developed, centred on the five nodes coloured red, but also benefitting a range of blue nodes, which thrive on their accelerated development, while not themselves contributing to the self-reinforcing feedback loops. The white nodes however are left without benefit from these elective affinities; they will therefore provide the low fitness candidates for extinction. The subsequent reconfigurations associated with the slow dynamic may occasionally enable new ACSs to form, undermining those which currently dominate, and producing dramatic shifts or collapses within the connective geometry of the system.

Computational models of this sort can serve as ideal types, illuminating the range of empirical dynamics to be found in the real world (Gilbert & Troitzsch, 2005). To apply such models empirically is likely to be rather demanding in terms of data, depending as it does on details of connections across a complex system, collected repeatedly and in timely fashion. This is an obvious example of where “big data” analysis may play a key role — on the one hand drawing on the large-scale administrative data that are routinely amassed and regularly updated; on the other hand, deploying modern computational capacity to scan such data sets, for the “runaway loops” and the ACSs they reveal (Mayer-Schonberger & Cukier, 2013).

5 Contingency and the Arts of Civilisation

Co-evolution was recognized by Darwinian biologists as a blind process of the natural world. It was also a contingent process, as Gould (1991) has conclusively argued; if the biological “tape of history” were to be “replayed” many times, altering some apparently minor detail in that chain of contingencies, each would have produced a quite different result. Similarly, when Jain and Krishna run successive simulations of their model, small shifts in the algorithms of extinction and mutation can produce quite different trajectories. The fruit of any such enquiry is not therefore a set of “universal truths,” of the sort to which the GLM aspires, but instead some “timebounded truths” about the contingent dynamics of change (Brown & Langer, 2011).

Few of the “replays” of Gould’s biological “tape of history” would have produced anything like homo sapiens. That however made all the difference. Homo sapiens now purposefully intervenes in the contingent dynamics of biological co-evolution, shifting the algorithms of extinction and mutation and the terms on which the tape of history is played. Thus in some degree, our species makes its own history, not least in regards to the biological world of which Darwin wrote and where human interventions are producing new mass extinctions.

This article however is concerned with the coevolutionary dynamics of social rather than biological life. The complex dynamics of the social world are in some degree as blind and emergent as the dynamics of natural world, without collective purpose. Nevertheless, purposeful interventions by social actors are omnipresent; and this entails struggles for positional advantage and the exercise of power. It is necessary therefore to consider how to conceptualize this purposeful agency in relation to the co-evolutionary dynamics of the social world.

Those who champion evolutionary ideas and want to apply them to the social world disagree as to what it is that evolves. For Dawkins, it is a matter of understanding social dynamics by reference to the demands of biological evolution. It is for example by reference to the “selfish gene,” that we should understand the evolution of cooperation and altruism (Dawkins, 1976). Sloan Wilson (2008) likewise retains a strong focus on the biological substrate of human behaviour.

Evolutionary economists in contrast leave no place for biological selection.¹³ In Darwinism, it is the genetic legacy of a species that is re-worked; in evolutionary economics, it is the technological and institutional legacy of a society (Potts, 2000; Hodgson, 2002; Crouch, 2005; Beinhocker, 2007). New “variations” in technology emerge from the “animal spirits” and inventiveness of entrepreneurs, in what Schumpeter described as “swarms of innovation.”

Just as the pigeon breeders and horticulturalists described by Darwin (1859/1998, Ch. 1) looked out for novel characteristics in the offspring of each new generation, entrepreneurs are forever on the lookout for new technologies whose coevolutionary dynamics can open new markets and yield disproportionate returns. Such dynamics may entail co-evolution between different technologies; between new technologies and new markets; between new forms of industrial organization and new systems of public regulation, etc. To discover and nurture such dynamics is central to what we may call the arts of civilization (Bronowski, 1981: Chs. 2–4).¹⁴

How a given technological innovation will then fare — and how it may interact with other technologies and institutions — can never be entirely foreseen. Entrepreneurial ingenuity may propose new variants: but it is processes of differential selection through the market that dispose; and these can seem just as collectively “blind” and devoid of overall intent as the processes of natural selection that drive speciation in the wild. To understand the economy in such evolutionary terms — and more generally as a complex system with emergent features — has attracted growing numbers of heterodox economists.¹⁵

Even as those collectively blind processes unfold however, strategically purposeful social actors will attempt to modify them to their own ends — depending on the resources and power at their disposal. This is a struggle for positional advantage. Attempts to apply evolutionary models to the social world have in general neglected such exercise of power. We however bring power centre-stage — all set within the hierarchical relationships of domination and dependence which characterize our human societies.¹⁶

If Figure 3 — our generic model of coevolutionary dynamics — is to inspire our social enquiry, it must be seen as playing out within those relationships of power and domination — and the institutional rules and regimes in which they are embodied. That is for example why, when we come to examine the co-evolution of technologies, we will also engage with the multi-level institutional processes within which new patents establish their intellectual property rights. In the biological world, the simple selective dynamics summarized in Figure 3 play out across complex multi-level food webs; in the social world the selective dynamics, through which technologies and capabilities develop, play out across contested multi-level institutional webs of positional advantage.¹⁷

We will henceforth speak of the Contingent Historical Model (CHM) as our alternative to the General Linear Model: integrating blind co-evolutionary dynamics with purposeful agency and the struggle for positional advantage (Room, 2011; 2016). It is this model that we seek to apply empirically, in ways that engage with scholarly concerns. We will also assess the practical value of Jain and Krishna for policy actors, in detecting early or “weak” signals of impending change, and applying the arts ofcivilizationn to the policy world as it unfolds.¹⁸

6 The Investigation of Empirical Dynamics

We now consider how the mathematics of co-evolving systems — in particular the model of Jain and Krishna — can be applied to empirical datasets — and the desiderata those datasets must fulfill.

In the present article, we focus on databases of patents, and their adequacy for studying the empirical processes by which new technologies co-evolve. This empirical project is the initial focus of the DCICSS research group.¹⁹ Case studies in other empirical areas will follow. Depending on their distinctive features, we may need to modify the model; and there are always likely to be trade-offs between the realism of the model and its mathematical tractability.

We construe the technologies of the modern world as a connected network of loci or nodes. The level of activity on each node is affected by that at the other nodes to which it is connected. This is the “fast dynamic” of Jain and Krishna. The pattern of outcomes in such a vast connected system cannot however be predicted in advance, as the simple aggregation of activity at the micro-level — it is blind and emergent and the outcome may be counter-intuitive. This is a general feature of complex systems (see for example Schelling, 1978; Squazzoni, 2012).

The “slow dynamic” is different. It involves the extinction of nodes with low levels of activity and the introduction of new nodes and connections. In Jain and Krishna, these are random novelties. It can also however be effected by purposeful selection by social actors. This may happen at the local level, for example when inventors produce new technological devices (Koenig, Battiston, & Schweitzer, 2008), a few of which find elective affinities that enable them to thrive disproportionately. It can happen on a larger scale, when the big actors of government and the corporate world consider the emergent macro-outcomes of the fast dynamic and purposefully intervene, intending to steer them in new directions.²⁰

Against this conceptual background, Table 1 summarizes the desiderata for the datasets we will require, when we study non-linear dynamics within co-evolving systems, using the Jain and Krishna model, and the tasks that will be involved. We will be able to judge, by reference to this Table, whether the datasets available to us, for any empirical case study, sufficiently meet these requirements.

As we shall see, the process of mapping data from empirical datasets into our model of co-evolving systems is itself highly demanding. We confront, first, the challenges facing any researcher who uses large databases, in checking the reliability and coverage of the data and their consistency between countries. More than this however, we are mapping these data not into the rather simple architecture of the GLM, as captured in equation (1), but into the tangled societal counterpart of the “tree of life,” as captured in equation (3), re-shaped over multiple periods of co-evolutionary dynamics and purposeful agency, within a changing institutional architecture and political economy.

Section 7 will apply these desiderata to databases of patents, as a vehicle for studying technological innovation. Section 8 examines how far an empirical network of co-evolving technologies can thereby be constructed, as a contingent historical process. Section 9 will bring together our findings.

7 Patents as Technological Innovations

7.1 Introduction

Technological innovation is a key feature of the modern world, driving all else, as celebrated by such diverse luminaries as Adam Smith, Marx and Schumpeter. How might we study the dynamics of technological innovation, as innovators combine existing and new technologies in novel ways? What sort of datasets might allow us to track these changes over time and better anticipate their direction?

Databases of patents have developed over the last century and are harmonized internationally. Many scholars have used them for studying different aspects of technological innovation; there is a rich and self-critical literature on which we can build (Napolitano et al., 2018).

The World Intellectual Property Organisation (WIPO) is the forum charged with overall governance of the international patent system; it includes the major national patent offices and the European Patent Office (EPO). WIPO is responsible for the global patent classification system and its annual updates; EPO is responsible for publishing the global PATSTAT database, set within the latest version of the classification system.

A patent can be viewed from two standpoints (Strumsky, Lobo, & Leeuw, 2012). First, it constitutes a new capability, a force of production. It combines and applies knowledge and technologies in new ways, whether incremental or more radical. This combinatorial ontology is central to much of the innovation literature, including Schumpeter and Hayek, and more recently Teece (2009) and Potts (2000).²¹

Second, a patent is a claim to novelty — and one whose commercial potential the inventor wishes to protect. A patent application acknowledges the “prior art” on which it builds, its intellectual debts, but it also makes clear what is new. A patent thus constitutes a claim to intellectual property from which others will be excluded — part therefore of the relations of production. Once a patent is granted, the inventor may develop it commercially. It may alternatively be sold, with the new owners either exploiting or shelving it, for a period at least, to avoid it threatening their existing product lines and markets.

As both forces and relations of production, patents remind us that technological change takes place within particular legal and institutional settings. These vary between countries and over time; they are socially and politically constructed and contested (Polanyi, 1944). Relevant here is the literature on “varieties of capitalism” (Hall & Soskice, 2001) and “national innovation systems” (Lundvall, Intarakumnerd, & Vang, 2006). A study of patents may serve to illuminate the co-evolution of technologies and institutions in different political economies. This does not mean that these are mutually insulated national domains. On the contrary, the patent regimes of individual countries can have consequences for innovation elsewhere, not least through the interconnections woven by international companies and their production chains.

It follows that a study of patents should serve to illuminate major theoretical and empirical questions that are central to sociological enquiry. It should allow us to lay bare alternative trajectories of development and the scope for intervention by policy makers and corporate actors, in pursuit of those alternatives. In thus demonstrating the value of models of co-evolving dynamics for the study of patents, we hope to show that such models deserve greater attention from the social science community at large.

8 Patents as a Contingent Historical Process

This conceptual and methodological challenge nevertheless opens up an important opportunity for our analysis of innovation. Far from the selection of patents as our case study being ill-advised, it will prove to have been particularly apt.

We have thus far been interested in how technologies co-evolve with each other, in autocatalytic processes as modelled by Jain and Krishna. Nevertheless, we grasp technologies through the institutions by which we organize them. The pace and direction of innovation depends, indeed, as much on the new institutions that are being invented — laws on e-commerce for example and on new forms of intellectual property — as on the new technologies themselves. The designers and inventors of these institutions (in government, in business, etc) are re-shaping the world, no less than the designers and inventors of technologies.

In this way, institutions and technologies interact and co-evolve with each other. Many new technologies emerge, only to find themselves on an institutional terrain that stifles them. In some other country, the institutional terrain may be more supportive. Some technologies and institutional forms may thus discover an “elective affinity” that enables both of them to thrive and together to dominate their socio-economic ecosystem.²⁵ These elective affinities between technologies and institutions are therefore as potentially significant, as those between different technologies, in generating autocatalytic effects and in driving processes of innovation.

The international patent system is one such institutional domain; its officers are one group of institutional inventors. How then shall we think of the autocatalytic dynamics involved in the development of new technologies and their registration as patents?

Consider Figure 7. We start from the north-east quadrant and proceed clockwise.

A new year has started. WIPO has updated the classification system it published 12 months earlier, involving a dendrogram of classes and sub-classes. EPO has published the latest version of the PATSTAT database, locating all patents, both recent and not so recent, within this most recent version of the classification system. It thus maps the technological capabilities of the society. It also records the knowledge claims that are officially recognized, and hence the flows of know-how on which successive generations of patents relied — all organized in as parsimonious, searchable and up-to-date a form as the international patent officers consider possible. This mapping will now provide the stable framework that inventors can use, to search the “prior art” and to register their new patents over the coming months. It will also allow patent officers to evaluate those new applications.

We will ask: Q1: What technological capabilities do these patent classes capture — and what flows of know-how, from established patent classes (and sub-classes) to new patents, did they involve? We construct an empirical network, with patent classes as nodes and knowledge flows as edges.

We return to this NE quadrant, after working our way through the other three.

This moment of publication is the starting gun for the registration of new patents. Within the south-east quadrant, inventors start registering new patents and populating the classes of the newly-updated classification system. They cite a variety of existing patents (and thus patent classes) as the antecedents on which they have drawn in novel ways.

This is the continuous present of the technological fast dynamic. Within this quadrant, the classes are given and fixed, just as with the nodes in Jain and Krishna. Inventors anchor their claims to novelty within this stable substrate, the prior art; it is only thanks to the parsimonious clarity and searchability of the classification system, that the novelty of their own invention can be rigorously demonstrated.

We will ask: Q2: How do the various nodes of the network (in this case, the classes and sub-classes) thrive differentially? What is their resulting fitness — the number of patents registered in each class or sub-class during the year in question?

Even as the inventors register their grounding in the prior art, they also document the novelty of what they offer. Some of these novelties will however in due course, and taken together, abrade against and undermine the classification system, even though it has only recently been updated. Here therefore is a stable classification system, but one across which there progressively spreads a mass of prospective instabilities. In this way, the continuous present unfolds on the stable ground of the past, but also defies and challenges it.

In places, the sheer numbers of new patent applications threaten to overwhelm the categories’ capacity to describe and contain the promised novelties, without blurring or losing their distinctiveness. In other places, the novelties defy the categorical boundaries and the distinctions they draw. This makes it increasingly difficult for the national patent officers to apply the classification system as it stands. We might indeed describe the pressures for such revisions, and their distribution across the array of patent classes and sub-classes, as the rate of activity — the fertility — of different parts of the classification system. This is the institutional fast dynamic, located in the south-west quadrant.

We might ask: Q3: What are the mounting pressures for revision to the classification system and how are they distributed across the different classes and sub-classes? What limitations and inconsistencies in the classification system do these pressures expose?

Nevertheless, these micro-processes are largely invisible; we cannot view them directly within the patent publications.²⁶

This is a fertile, creative and expert activity, a practical craft on the part of the national patent officers, as they consider what adjustments would be needed to provide all these new patent applications with an appropriate niche, capturing and displaying its distinctive novelty. As in all scientific work, such classification is an essential tool in building conceptual links for understanding and progressing the architecture of knowledge (Bowker & Star, 2002: Ch. 9). It goes beyond mere categorization and search; for it helps ensure that each patent comes to the notice of other relevant strands of invention, enabling cross-pollination and the discovery of further elective affinities.

That task is taken further at year end by their WIPO colleagues at international level, with their annual updates of the international classification system. Now, in the north-west quadrant, we move from the fast to the slow dynamic; from the video of the continuous present, to looking back from the end of each year, a careful snapshot, a pause for constructive reflection. WIPO reads the reports and recommendations from the national officers. It takes stock of the pressures they have encountered for adjustment of the classification system and the conceptual tangles these have produced. It considers how to simplify and untangle the web of citations as economically as possible. In information theory, this parsimony appears as a counterpart to the export of entropy in physics. Elements of the classification system serving no useful purpose are eliminated, as surely as the least fit species in an ecosystem.

We might ask: Q4: What rules and processes of sorting and re-working of the classification system do the national and international patent officers follow, in their efforts to maintain a parsimonious, searchable and consistent system?

It is useful to distinguish between new patents registered in a given year, using the classification system published at the start of that year, and patents registered in earlier periods, but now re-classified by reference to the current system. The ad hoc modifications made during each year, by the national patent officers within the south-west quadrant, are driven primarily by the former. The latter also matter however; and when the international office considers what annual revisions to make, it will need to weigh up how easily those earlier patents can be mapped into the revised classificatory matrix that the newest patents suggest, while still maintaining a parsimonious and searchable scheme.²⁷

This is the institutional slow dynamic — with WIPO looking back and checking the consistency and elegance of the various ad hoc adjustments that their national counterparts have proposed, on the basis of the previous twelve months. The new classification of technologies that WIPO now publishes — and the new version of the PATSTAT database that the EPO then publishes — is the result of this reflection. The slow dynamic is however not just a look back at the preceding year, simplifying and pruning the tangles of the classification system, after the myriad local adjustments it has suffered over the past months. It also looks forward and establishes the new classificatory terrain for the year ahead. This is the north-east quadrant, the slow dynamic in relation to technological forms. The new classification system will provide the stable framework that inventors will use, to register their new patents and their debts to the prior art, over the coming months.

Our progress around the four quadrants has thus left us with four empirical research questions. Q1 and Q2, concerned with the technology slow and fast dynamics, can be addressed directly using the PATSTAT database. At the end of each year, we have the newly published WIPO classification scheme and the latest PATSTAT database, as well as their predecessors of 12 months earlier. Some of these patents will have been registered only during the last 12 months; others are of longer-standing. This enables us annually to construct an empirical technology network, showing patent classes and sub-classes as the nodes, the citations and knowledge flows as the edges, the numbers of patents as indicating the levels of activity, as viewed from the standpoint of the new and updated classification system. We can identify ex post the growth, consolidation or decline of autocatalytic sets through the year. This is the core of the empirical analysis in which we engage.

Meanwhile however Questions Q3 and Q4 concern the evolution of the classification system but cannot be directly addressed through the PATSTAT database. The latest PATSTAT database leaves in the background the classification system published 12 months ago, which nevertheless during the year served as the immediate point of reference for inventors. Those inventions are instead now displayed by reference to the revised classification system that WIPO produced during the final weeks of the year. That revision was driven by the stresses and strains which those inventors placed on the classification scheme, obliging the national and international patent officers to bring their collective wisdom, experience and ingenuity to bear — the arts of civilization. Thus inventors benefitted from the legacy of classifications transmitted from the past; but they also then exposed the limitations in that legacy, laid bare by the very novelties they sought to register.

Of that journey through the south-west and the north-west quadrants we have no direct knowledge, unless we mount a fieldwork enquiry of our own, surveying the national and international patent officers. We can however, albeit not without some effort, track how patents registered in previous years have over subsequent years been reallocated to new positions, as the classification system has been revised. Within the DCICSS research group, Napolitano and Pugliese (2017) have created just such a unified database, mapping each patent against all the codes it had in different classifications in different years, so as to track not only changes in the classification system but also the detailed flow of patents through that changing system.²⁸ This could be used to study indirectly the pressures exerted on the classification system through each 12 months by the latest cohort of inventions.

These are the selective dynamics through which technologies and capabilities develop, played out across contested multi-level institutional terrains, in a process which we earlier referred to as contingent historical development.

9 Conclusion

We now bring together the results of this study and its preliminary findings on technology networks.

The previous section displayed the co-evolution of technologies and institutions, patents and classifications, in a contingent historical process. Centre-stage are zones where a proliferation of new technologies is manifest in patents that challenge the established classification system. It is the dynamics of this zone of intense innovation, in both technologies and institutions, that we seek to capture, using our model of co-evolving networks taken from Jain and Krishna.

Notice the relationship between Figure 7 and the earlier Figure 3. Figure 3 offered a parsimonious representation of co-evolutionary dynamics. It was the starting point for our presentation of the Contingent Historical Model — a radical alternative to the General Linear Model, which has dominated much of social science. It looked back from the ecosystems of today to the elective affinities that had developed, among the populations of different species, during earlier periods. Those affinities enabled some species to thrive, while others failed and became extinct. Only by revisiting those earlier periods, was it possible to make sense of the complex food webs that we see today.

Figure 7 has taken us on an analogous journey through successive time periods, in order to make sense of the patent system we see today, with its wide range of interrelated technologies, classified to clear and parsimonious effect. The journey alternated between the slow and the fast dynamics, and between innovations in technologies and in classifications, as we took transverse cuts through successive annual data sets. Those cuts involve alternating snapshots and videos. The fast dynamic tells of a continuing present; the slow dynamic looks back from the end of each year, a pause for both reflection and foresight. We thus disassemble the co-evolutionary dynamic into four moments, each of which we study by appropriate interrogation of the PATSTAT databases for successive years. With the data laid bare, from this succession of vantage points, it becomes possible to construct the empirical network of co-evolving technologies and to investigate the autocatalytics of change.

The process of mapping data from empirical datasets into our model of co-evolving systems is thus non-trivial. Instead of the rather simple architecture of the GLM, we face the complex process of technological and classificatory co-evolution, by which the patent database is regularly re-woven. There was no certainty that the PATSTAT database could be bent to our research purposes. In the event however, the selection of this database for our empirical study has proved particularly apt.

This therefore is the principal empirical finding of the present article: not yet a systematic analysis of the autocatalytic empirics of technological development, but a necessary first step, in establishing the fitness-for purpose of this particular empirical database, and its exploitability in the service of the CHM. To this extent, the present article reports a conceptual and analytical approach and assesses the feasibility of a specific database for an empirical application of that approach.

Beyond this however, the DCICSS project has meanwhile published preliminary findings on the autocatalytic empirics of technological development, using the PATSTAT data (Napolitano et al., 2018).²⁹ This study incorporates many elements of the conceptual framework described here. It demonstrates that the evolution of the technology network does indeed involve a growing autocatalytic structure. It also confirms that the technology fields in the core of the autocatalytic set display greater fitness, in terms of a greater number of patents. Finally, it reveals core shifts, whereby different groups of technology fields come to dominate the autocatalytic structure, only then to be overthrown. This points to radical innovation, with new combinations developing among distant as well as closely related technology fields. These are just the sorts of dynamic change that Jain and Krishna explore through their analysis of eigenvectors and eigenvalues. This empirical analysis will be extended in the next phase of the DCICSS work.

This demonstrates the promising possibilities for this form of analysis. The research also has potential benefit for policy makers, providing indicators — “weak signals” — of impending change, so that they can better steer the dynamics of autocatalytic innovation. It will need to be set within a wider institutional analysis, including the literature on national innovation systems (Lundvall et al., 2006) and varieties of capitalism (Hall & Soskice, 2001). It will also need to integrate co-evolutionary dynamics with a more thorough analysis of power and political economy (Room, 2016).

Beyond the present programme of work, the DCICSS project expects to apply the Jain and Krishna model, mutatis mutandis, to other fields of social, economic and technological change. It may of course be that in many potential areas of study, there are no appropriate databases immediately available. Nevertheless, this article has at least established the conceptual and methodological coherence and viability of such an empirical enterprise.