• 1 For economists’ statements on realism in assumptions, see Friedman (1953), Gibbard and Varian (1978 (...)
• 2 Gabaix and Laibson (2008) consider tractability as one of the “key properties of useful models,” al (...)

1How do economists make modeling choices? In this paper, I argue that answering this question requires a deeper look beyond those assumptions and properties of models that economists have deemed critical, and that philosophers and historians have dissected: profit or welfare maximization, stable preferences, (im)perfect information, rational expectations; equilibrium unicity, convergence, or identification.1 We should also consider those model characteristics that economists only discuss quickly or do not address at all because they view them as unsatisfactory, frustrating, temporary, mundane, unimportant, or conventional. In this article, I thus propose to investigate these choices that economists make for the sake of “tractability.” The term frequently appears, even in article titles, but its definition is seldom discussed. It is ubiquitous, yet invisible. Some article titles advertise “A Tractable Model of Reciprocity and Fairness” (Cox et al., 2007), “Functional Forms for Tractable Economic Models” (Fabinger and Weyl, 2018), “A Tractable Equilibrium Search Model of Wage Dynamics” (Bagger et al., 2014) or “Tractable Likelihood-Based Estimation of Non-Linear DSGE Models” (Kollmann, 2017). Public economist Stefania Stantcheva (2020, 2) praises the sufficient statistics approach in taxation economics for being “very tractable and empirically applicable.” “Tractable” is also a term almost systematically associated with famous “workhorse” models, such as the Dixit-Stiglitz model (Brakman and Heijdra, 2004) or the iceberg transport cost model (Krugman, 1995). It is a term used as a casual synonym for expedience, convenience, simplicity, clarity, solvability or manipulability, but without a clear definition or justification. It is considered sometimes a vice, sometimes a virtue of models.2 It evokes a shared meaning, sometimes a characteristic, a property, a requirement, a constraint of a model; it is often accompanied by or interpreted as an instruction to gloss over certain assumptions that are considered epistemically innocuous. Tractability characterizes whole models, but it is usually attached to specific assumptions about the number of variables in a model, representative agent, single firm, full information, zero profit, or specific functional forms (linearity, homogeneity, quadratic form, separability, convexity, symmetry).

• 3 Economists have used various synonyms, as listed above, to discuss the manipulability and solvabili (...)

2The term “tractability” is pervasive enough to prompt the question of how to define it and how it has shaped individual and collective modeling choices in economics. While this paper does not provide a definitive answer, it aims to encourage historians of economics to explore this question, as some methodologists have done in the past fifteen years. To do so, I first survey the economic methodology literature on tractability, one that grew out of methodologists’ attempts to understand why economists make unrealistic assumptions. I then compare these accounts with the few instances where 20^th century economists have explicitly discussed this notion, and show that such historical investigation, despite its limitations, sheds light on features of tractability that require further exploration. Rather than provide a consistent definition of tractability, I trace the changing and ambiguous ways in which economists have used the term, so as to outline research questions for historians.3 I loosely define tractability as a property of an economic model that makes it easy to manipulate and solve. I argue that documenting the collective dynamics at work when tractability motives are invoked would help to understand the endurance or displacement of specific types of models. I suggest that economists moved from viewing (in)tractability as a constraint of both the world and the model to seeing it as a virtue of economic models. But as even those economists who have made tractability a virtue have pointed out, those modeling strategies meant to be transitory sometimes become conventional and hinder change, progress, diversity. This highlights the existence of “tractability traps,” in need of further investigation. I explain how distinguishing between theoretical, empirical and computational tractability allows to ask why some manipulations and types of solutions are considered valid while other remain on the margins of the discipline. This leads to consider the existence of “tractability standards” that differ across time and fields.

• 4 Hall needed to ensure that the correlation of productivity growth with aggregate demand could be co (...)

5It was however Frank Hindriks (2006) who offered a comprehensive case for the inclusion of “tractability assumptions” in methodologists’ classifications. He defined a tractability assumption as “a statement about the tractability of a problem ... [which] would be intractable if it were not for a particular first-order assumption” (Hindriks, 2006, 412). Development in science, for instance mathematical or econometric techniques, new data, or new instruments, help turn previously deemed intractable problems into tractable ones, he explained. His primary case study examines how economists make tractability assumptions to measure some of the variables in their models. By focusing on measurement, he departed with hitherto more theory-based discussions of assumptions. Hindriks (2005) documented how macroeconomist Robert Hall assumed constant productivity growth when he set out to explain why total factor productivity appeared as procyclical in the 1980s. Hall clarified that the Solow Residual captured not only productivity, but also the effects of firms’ increasing returns and market power. To demonstrate this, measuring the mark-up of price over marginal cost—that Hall assumed was a component of the Solow residual—was necessary. This, in turn, required assuming constant productivity growth, despite opposing empirical evidence.4

• 5 I am not asserting that Robinson was the first economist to discuss tractability. For instance, Jea (...)

7 Ironically, one of the earliest and more forthcoming proponents of writing “tractable” models later recanted her advocacy. It is Joan Robinson.5 As a young theorist completing her book on imperfect competition (Robinson, 1933), she penned a 15-pages essay to defend her formal approach to firm behavior and markets. Reflecting on the methodological controversies she observed at Cambridge and throughout Europe, she elucidated how theorists all faced a tradeoff between the tractability and realism of their assumptions:

• 6 “If the pessimistic economist prefers sitting at the apex of a pyramid of completely self-consisten (...)

8The choice between the “manageable” and the “realistic” set of assumptions was one of “temperament … not one of opinion,” she continued: “the optimistic, analytical, English, economist will choose the manageable set, and the pessimistic, methodological, Continental economist will choose the realistic set” (Robinson, 1932, 6-7). Counting herself as an optimistic analytical economist favoring tractable assumptions, she expressed her hope to move from a two-dimensional to a n-dimensional technique in the future.6

• 7 For histories of the Cambridge and British intellectual milieu and controversies of the 1930s and 1 (...)

9In her introduction, Robinson (1933, 1) again acknowledged the “arid” character of her analytical approach, one that would undoubtedly fuel “the impatience of the political, the businessman and the statistical investigator.” She contemplated the “agonizing sense of shame” of the “analytical economist” confronted to a “practical man” (Robinson, 1932, 2). She again apologized for the “adoption of very severe simplifying assumptions” and explained that the book was “presented to the analytical economist as a box of tools,” in need of being perfected up to the point of becoming useful to the practical man (Robinson, 1932, 2). The adoption of more “tractable” yet less “realistic” assumptions associated with Cambridge was criticized by Australian economist Ronald Walker (1943, 58) in his book From Economic Theory to Policy: “It is not that the complexity of realistic assumptions makes them intractable but that they are in conflict with the assumptions used in some of the established theories and, if admitted to the economists’ system, must destroy these theories,” he wrote, before condemning what he saw as a growing “theoretical blight.” A reviewer interpreted the target of Walker’s ire as “the great Cambridge tradition” (Spiegel, 1945, 56). They both defined “tractable” in relation to the available “analytical techniques.”7

• 9 See Medema (1994) for a thorough discussion of Coase’s work and epistemology.

11Coase thus explicitly tied tractability (or manageability) to the available instruments of economic analysis.9 In the 1930s and 1940s, it was primarily those British economists who shared an interest in microeconomic analytical tools, in particular monopolistic competition, who casually discussed assumptions by using the words “(in)tractable” and manageable interchangeably: Cambridge’s Roy Harrod (1934, 336) solved a duopoly problem with linear demand functions but noted that “if more particular demand functions are allowed, the mathematics becomes less tractable”; John Hicks (1936) challenged the existence of tractable monopoly assumptions, while Fritz Machlup (1940, 296) argued that “simple monopoly is still a tractable assumption within a general equilibrium system.”

• 10 See Boianovsky (2020) for a detailed analysis of Samuelson’s use of Medawar’s epistemological writi (...)

12Over the following decades, while Robinson began to question her previous “scholastic” approach (Robinson, 1953), many economists repeated her defense of tractability assumptions, though in a less apologetic manner. In 1970, MIT theorist Paul Samuelson (1970, 1373) reflected on his younger own impatience with “restrictive” assumptions, and his lack of awareness of the truth in Nobel biologist, Peter Medawar’s maxim that “science must deal with that which can be managed, eschewing the intractable.”10 He would use the term time and again, for instance in a 1986 edition of his collected papers, to remind the reader that the language of science requires to provide “prosaic, tractable, describable, testable and understandable model or paradigm” (Samuelson, [1982] 1986, 859). As for Robinson, while she clearly referred to modeling assumptions in her text, in other pages she explained that “economics is still in infancy, more because of the intractable nature of the subject than because of the low mental caliber of economists.” (Robinson, 1932, 12) Robinson and Samuelson thus both ambiguously discuss (in)tractability as a characteristic of the natural and social world, and of the models themselves. The slow but steady rise of uses of the term “tractable” and the ambiguity surrounding its meaning is supported by anecdotal bibliometric evidence. A search the terms “tractable” or “tractability” or “intractable” among the “articles” classified as “economics” in the JSTOR database suggest that the terms were used around 40 times in the 1930s, 75 in the 1940s and 150 in the 1950s. Most of these occurrences were adjectives, indicating that “tractability” had not yet stabilized as a property of economic models. Consistent with the ambiguities found in Robinson and Samuelson’s work, the terms tractable and intractable were sometimes used to discuss war mobilization, book printing, clothes stock reallocation (Rostow, 1942), “economic and social problems” (Menderhausen, 1949, 672), employment and investment (Austin Robinson, 1941, 481; see also Dowdell, 1940, 24), or “business cycle forces” (Bratt, 1953, 20).

• 11 “Modeling trick” is a term that economists often use, whether casually (see for instance Stiglitz, (...)

13In the 1960s, Samuelson defended tractable models in the context of ongoing debates on the state of development economics. He criticized the voluminous, repetitive, and intractable nature of research published by Paul Rosenstein-Rodan, Arthur Lewis or Hollis Chenery (Boianovsky, 2020). Comparing this with how Paul Krugman criticized the same institutionalist development literature twenty years afterwards helps pinpoint the stabilization of what economists meant by tractability. Like Samuelson, Krugman (1995, chap. 1) complained about Albert Hirschman’s unwillingness to endorse “tightly specified models.” He assumed that Hirschman favored institutionalism due to impatience “with the narrowness and seeming silliness of the economics enterprise.” Krugman proceeded to defend those “silly assumptions,” in particular in his own field, trade (Krugman, 1993). He contrasted “big untrue assumptions—constant returns, perfect competition” with new trade theory models that “avoid these big lies but make many small ones along the way to keep matters tractable; the theorist can never forget the degree of falsification involved” (Krugman, 1994, 15). He likely had in mind his own iceberg transport cost model of trade, which treats transportation costs not as a separate variable, but as the melting of the quantity of transported goods. This modeling trick he took from Samuelson.11 Krugman insisted that he came to realize that these simplifying assumptions made for the sake of tractability “were virtues, not vices” and that they “added up to a program that could lead to years of productive research” (Krugman, 1993).

• 12 I have presented the numbers in absolute terms rather than as a percentage of a growing body of art (...)
• 13 See for instance Romer (1986, 664) on banking models or Holmstrom and Costa (1986) on tractable inf (...)

14The use of the term “tractability” in economics rose sharply from the 1970s, coinciding with Krugman’s emphasis on model tractability (defined as its ability to be solved analytically). The number of JSTOR articles in “economics” that used the terms “tractable,” “intractable” or “tractability” increased from around 300 in the 1960s to some 800 in the 1970s and more than 1500 in the 1980s, with the term “tractability”, and more specifically “analytic(al) tractability,” experiencing a significant raise (from 0 in the 1960s to 9 in the 1970s, 49 in the 1980s, 84 in the 1990s).12 By the end of the 1960s, John Gould (1968, 49) and Henry Grabowski (1970, 220), then two young industrial organization specialists, were among the few to discuss the “mathematical tractability” of cost functions and the “analytical (in)tractability” of alternative functional specification of duopoly models. About ten years later, by the 1980s, discussing the tractability of functional forms had become commonplace.13 The growing use of the noun may indicate the stabilization of the understanding as a property of economic models. Indeed, uses of these terms to characterize the social world, rather than the model, gradually become marginal.

• 14 What economists have usually called “workhorse” models is different from what philosophers of scien (...)

15 Krugman was part of this small group of economists associated with MIT who exhibited “a knack for making the kinds of simplifying assumptions that rendered his models tractable but not trivial” (Warsh, 2006, 205). Other members included Samuelson, Robert Solow, or Joseph Stiglitz, all of whom produced “benchmark” or “workhorse” models that were renowned for their “tractable” nature and applicability to various real-world problems.14 The term appears multiple times in a book edited by Steven Brackman and Ben J. Heijdra (2004) on the history and legacy of the monopolistic competition model published by Avinash Dixit and Stiglitz in 1977. This model aimed to analyze the equilibrium and welfare implications of the trade-off that firms face between consumers’ taste for diversity (an incentive to produce several goods) and the existence of economies of scale (an incentive to produce less goods in greater quantities). These phenomena introduced non-convexities in the mathematical analysis, making it difficult to solve models of monopolistic competition. To simplify the analysis, Dixit and Stiglitz used various tricks. They modeled consumer goods’ variety by using a Constant Elasticity of Substitution utility function where the elasticity of substitution between varieties is constant and various products enter symmetrically in the bundle. They used fixed costs to model increasing returns at the firm level and assumed free entry. They modeled firms as making decisions on the basis of the price of their good and of an aggregate price index summarizing all other goods’ prices. They derived an optimum in which each firm produces one good priced with a constant mark-up on marginal cost. The model proved applicable to a wide range of issues in trade, economic geography, growth and macroeconomics. Dixit later reflected that “Joe and I knew that we were doing something new in building a tractable general equilibrium model with imperfect competition, but we didn’t recognize that it would have so many uses” (quoted in Warsh, 2006, 208). He thus links tractability with typicality or versatility, that is, the possibility to apply the model to various research questions and teaching settings.

• 15 See for instance Couix (2020) on the Georgescu-Roegen/Daly vs Solow/Stiglitz controversy on growth (...)

16All these economists found themselves defending their “tractability vs. realism” tradeoff against alternative methodologies.15 However, they also had disagreements among themselves about the best approach to writing models that can be easily manipulated and solved, as well as how to apply them. “Elegant error is often preferred to messy truth. Theoretical tractability is often preferred to empirical relevance,” Richard Lipsey (2001, 169) complained as he assessed the legacy of the Dixit-Stiglitz model. Lipsey and B. Curtis Eaton had proposed an alternative model in multi-characteristic spaces, but it did not receive the same recognition because it relied on less aggregated behavior. Also reflecting on the Dixit-Stiglitz model, Peter Neary (2004, 180) emphasized “the price of tractability”, that is, a reliance on very special functional forms, identical, atomistic firms, free entry and homotheticity. James Tobin (1986, 351) complained that microfoundations had become the only game in town in macroeconomics, and “even the individual’s optimization is simplified and specialized in the interest of analytic tractability. Utility and production functions take parametric forms. By conventions, equations are linear or log linear.” A decade earlier, evolutionary economics pioneers Richard Nelson and Sidney Winter (1977, 272) had emphasized that a “major advantage” of the use of simulation in industrial organization research was the “freedom from tractability constraints of available analytical techniques.” Referring to the use of maximization and equilibrium model in macro in an earlier article, they noted that “the principal virtue of those tools is the gain in analytic tractability and logical coherence that has been obtained precisely by abstracting from all that diversity and change … the question is, how high is the price?” (Nelson and Winter, 1974, 903) Heavy, they answered, since the “neoclassical orthodoxy” cannot properly discuss the diversity rewards and penalties that firms, sectors and countries face as a consequence of technological progress.

• 16 According to Sergi (2017, chap. 2), the shift towards rational expectations initially made macroeco (...)

17Furthermore, what appears as a virtuous hypothesis that allows to tackle new questions in one decade might become a vice later, an assumption that keeps doors closed. The representative agent made macroeconomic models with rational expectations tractable, but it precluded the examination of how macroeconomic shocks affect income and wealth distributions;16 Dixit-Stiglitz models allowed economists to study of why firms engage in trade and how they choose their location, but they could not say anything about strategic behaviors such as commitment or price discrimination. In other words, the collective spread of tractability assumptions creates what I call “tractability traps.” As Robinson and others have highlighted, assumptions made for the sake of tractability are meant to be tentative assumptions; they are waiting to be lifted as mathematical techniques—and computer capabilities advance. In their analysis of how and why economists build models, theorists Alan Gibbard and Hal Varian acknowledged that tractability constraints are dynamic, but not the sole driving force behind modeling:

• 17 This definition overlaps with the one proposed by Fumagalli (2010, 622): “a cluster concept resembl (...)
• 18 I recovered only a few instances where economists discussed empirical tractability before the 1980s (...)
• 19 These parameters are elasticities of capital and labor supply with respect to taxes.

20To better understand the importance of documenting the collective dynamics of tractability in economic modelling, we need to refine our definitions. In line with how most economists have debated tractability in the recent decades, I define it as a set of properties of models that generally includes manipulability and solvability.17 Most of the above discussions by methodologists and economists however focus on one specific type of tractability, the ability to solve and manipulate theoretical models. One exception is Hindriks (2005; 2006), who documents how Hall made tractability assumptions so as to measure the mark-up of price over marginal cost.18 This ability to write empirical tractable models allowing the observation, measurement, identification and estimation of key economic variables has been at the core of the debates on the Lucas critique, structural vs. quasi-experimental models, or the transformation of public economics since the 1980s. In the 1990s and 2000s, optimal taxation economists sought a “tractable” middle-ground between those who advocated for a structural approach where all variables are deep behavioral parameters, and those who endorse a more rigorous empirical reduced-form approach. They developed what Raj Chetty later called the “sufficient statistics” approach. As explained by Piketty and Saez (2012, 2), “by tractable, we mean that optimal tax formulas should be expressed in terms of estimable parameters and should quantify the various trade-offs in a simple and plausible way.”19

• 20 For instance, macroeconomist Fabio Ghironi (2018,1) defines tractability as “the requirement that I (...)

21Some reactions to Krugman’s defense of tractability as a virtue of economic models point to a third type of tractability. Macroeconomist Robert Lucas, for instance, opened his review of Krugman and Helpman’s work on trade with the statement that “the useful development of an economic idea depends critically on one’s ability to formalize it accurately and tractably” (Lucas 1990, 664). He conceived tractability as the ability to manipulate and solve a model, but not merely at the theoretical and empirical levels.20 His writings suggest that many of his choices of (linear and quadratic) functional forms were a response to computational affordances: “there are no theoretical reasons that most applied work has used linear models, only compelling technical reasons given today’s computer technology,” he and Tom Sargent explained in 1979 (Lucas and Sargent, 1979, 13, emphasis in the original). They emphasized that many of their assumptions were for the sake of computational tractability, that is, to bring the model to data given the contemporary state of “computer technology”:

• 21 This is the “certainty equivalence” property demonstrated by Herbert Simon and generalized by Henry (...)

23Second, some modeling choices can be seen as attempts to make models both theoretically and empirically tractable. A typical case is the ascent of the Cobb-Douglas production function. A major argument in favor of using aggregate productions functions, in particular Cobb-Douglas than CES functions in growth and capital theory, for instance, was their mathematical and empirical tractability (respectively the ability to solve theoretical models, and to estimate the parameters of these functions; see Biddle, 2020). Another example is the widespread use of quadratic function in macroeconomics, for instance the quadratic loss function used since the 1960s to represent the central bank’s problem in macroeconomics. As explained by Pedro Duarte (2009), such functions were widely used in engineering and management since the 1950s because of their computational tractability. In monetary economics, this translated into the possibility to run simulations to assess alternative policy rules. But, as monetary economist William Poole remarked, “the linear structure with quadratic loss function is highly tractable. It is generally easy to obtain analytic solution” (quoted by Duarte, 2009, 4). This is because the decision maker only needs to know the expected value of the stochastic variables, not their full distribution.21

• 22 See Backhouse and Cherrier (2017) for further references of historical work documenting shifts in c (...)

24Third, modeling choices made for tractability purpose are largely endogenous, and depends on what economists in a given field at a given time consider crucial assumptions (for instance, rational expectations). What counts as a tractable model also depends on the available technologies, including mathematical and computational tools, algorithms, hardware and software, which are also endogenous to economists’ research agenda. In response to technological limits, economists have sought more powerful computers in other institutions, or borrowed analytical and numerical methods from other sciences, which they have subsequently standardized and spread through the development of software.22

• 23 Around that time, Barro and Grossman’s abandoned the disequilibrium models of involuntary unemploym (...)

28The presence of multiple types of tractability and changing hierarchies raise additional questions for historians regarding the dissemination, stabilization and overturning of modeling choices aimed at making models tractable. One crucial question is whether “tractability standards” that dictate which types of solution are deemed valid by economists can be identified in the history of economics, and if they tighten or loosen over time? For instance, the history of macroeconomics in the 1970s can be reconstructed as a process of tightening tractability standards. From the late 1930s to the 1970s, economists built large-scale macroeconometric models that often consisted of over a hundred equations and had to be estimated and simulated on slow and hard-to-access mainframe computers. Any approach that allowed them to derive numbers from the IBM 360 computer after 20 hours was acceptable, including taking down whole blocks of equations, or mixing limited information maximum likelihood, instrumental variable techniques, and recursive block estimation. Which models macroeconomists were allowed to solve and how they were allowed to solve them considerably narrowed in the 1970s. Empirical strategies had to be adapted to fit the requirement that agents’ decision rules exhibit rational expectations in a stochastic setting.23 This often entailed relinquishing analytical solutions and writing down models that could be “solved” with numerical methods. These methods enabled approximating the true unknown solution through linearization, value function iteration, then perturbation and projection methods. Some macroeconomists, led by Finn Kydland and Edward Prescott, abandonned estimation for calibration methods, until the implementation of new computational methods such as Markov Chain Monte Carlo methods fueled the development of Bayesian estimation in the 2000s.