A History of Thought in Industrial Organization

Industrial organization is an eighty-year argument about a single object: what does concentration mean? A market has a few large sellers rather than many small ones. Is that a fact about power, a fact about efficiency, or no fact at all? The field’s history is the sequence of its answers, and the sequence is a pendulum. Concentration means everything, said Bain in 1951, reading a four-point profit gap across a concentration threshold as oligopoly pricing made visible. It means nothing, answered Demsetz in 1973, reading the same coefficient as efficient firms growing large. It depends, said the game theorists of the 1980s, and then proved the “it depends” so thoroughly that any conduct became somebody’s equilibrium. One market at a time, said the demand estimators of the 1990s, who declared the cross-industry question ill-posed and rebuilt market power as a number attached to a product in a place. Maybe everything again, says the macro-IO literature since 2020, which aggregated those numbers into a single national series and reopened the question the field had spent forty years closing.

What makes this history unusual among the histories of economic ideas is that every answer got enacted. Antitrust is the field’s laboratory: each era’s verdict on concentration is written into a case or a guideline, with numbers, before the next era revokes it, and the courts run systematically behind the journals. The Warren Court blocked mergers at 5 and 7.5 percent shares on the strength of a coefficient the academy was already dismantling. Baxter’s 1982 Merger Guidelines codified the Chicago demolition in the same year game theory began reversing it. The Supreme Court adopted McGee’s predation skepticism in 1993, a decade after Milgrom, Roberts, Kreps, and Wilson refuted it. The 2023 Guidelines restored structural presumptions whose headline statistic had already failed its audit. This lag between revocation in journals and revocation in courts is not an accident of the story. It is one of its findings.

Prologue: Theory Without Measurement (1890-1938)

The era before the argument has no answer to the concentration question, because it has no number. Its claim, stated so it can be contested: by 1938 economics possessed a complete formal theory of firms with market power and no empirical procedure for applying it to a single industry. A defender of the era can reply that Marshall, Chamberlin, and Robinson were building tools, not thermometers, and that demanding concentration ratios of a 1933 price theorist mistakes the genre. The rebuttal is that the era’s own policy arm had already demanded the measurement. The Sherman Act and Standard Oil condemned a firm for controlling the overwhelming majority of refining, with no economics to say what a majority does to a price. Theory and law were both reaching for a number neither could produce, and that shared emptiness, not any positive doctrine, defines the period.

The law came first, and it came empty. The Sherman Act of 1890 declares illegal “every contract, combination… or conspiracy, in restraint of trade” and reaches anyone who shall “monopolize, or attempt to monopolize” (Sherman Act, 1890). It defines neither term. It contains no threshold, no efficiency defense, no reference to prices, output, or shares; its only native numbers are the penalties, a fine of up to $5,000 or a year in prison. It is a blank check whose economic content Congress delegated, whether it knew it or not, to the next 130 years of industrial organization. Every doctrine in this history, the structural presumption, the HHI grid, the recoupment test, is downstream jurisprudence poured into those two undefined words.

The Supreme Court cashed the check in 1911, and the way it did so is the era’s first dispute. Standard Oil dissolved the paradigmatic trust: a holding company controlling on the order of 90 percent of American refining at its peak, broken into some 33 successor companies within six months (Standard Oil, 1911). To get there, Chief Justice White read the statute’s “every restraint” down to every unreasonable restraint, the rule of reason. Steelmanned, this is principled law: a literal reading would outlaw ordinary contracts, and the common-law distinction gave courts a workable standard that promptly delivered the government its landmark victory. Steelmanned the other way, it is law without economics. The opinion infers undueness from a history of acquisitions and an intent to exclude rivals “not as a result of normal methods of industrial development,” but it never models how Standard’s conduct raised prices; it treats the petroleum industry as a whole as the market and measures nothing. Justice Harlan’s partial dissent saw a judicial rewrite that would later excuse “reasonable” combinations. The gap was real enough that in 1958 an economist named John McGee would reread the same trial record, roughly 11,000 pages of it, and find not one instance of the predatory pricing everyone assumed the case was about (McGee, 1958). The canonical monopolization case carried no theory of monopolization. It would take the field half a century to notice.

Economics, meanwhile, had a theory of everything except the middle. Marshall’s Principles posed the problem the whole field inherits: if larger firms produce at lower cost, why doesn’t every industry collapse into one firm? His answer, that firms are mortal “trees of the forest” and that durable scale gains are external to the firm, kept competition alive as a process without ever measuring how much power any firm held (Marshall, 1920). The reconciliation could not survive formalization; Sraffa’s 1926 attack pushed the profession toward the middle ground Marshall had defused. The prologue’s foundational text hands its successors the very problem it claimed to have solved.

The middle ground arrived in 1933, twice. Chamberlin’s Theory of Monopolistic Competition and Robinson’s Economics of Imperfect Competition appeared in the same year, independently, one from Harvard and one from Cambridge. The priority question is real and I will not adjudicate it. Chamberlin had submitted his core argument as a doctoral thesis dated April 1, 1927, six years before publication and before Robinson’s work, a fact he wielded for the rest of his life; he also insisted his construction was distinct, rejecting the very label “imperfect competition” (Chamberlin, 1933). Robinson framed herself as a tool-maker generalizing the Marshallian system out of Cambridge work by Sraffa and Kahn, and she coined “monopsony” along the way (Robinson, 1933). The more interesting fact than who was first is that the same structural idea, every firm faces its own downward-sloping demand curve, so market power is a matter of degree along a continuum, arrived at two universities in one year. And the differences are load-bearing. Chamberlin’s engine is product differentiation: each seller holds a monopoly of its own variety, and free entry drives the firm’s demand curve to tangency with the falling portion of its cost curve, leaving every firm below efficient scale with chronic excess capacity. Robinson’s engine is marginal revenue set against marginal cost, case by case, yielding her durable result that a monopsonist pays labor less than the value of its marginal product, exploitation made rigorous. Different seeds, different harvests: Chamberlin’s differentiation was reborn decades later in demand estimation, Robinson’s monopsony in labor economics.

What neither book contains is a measurement. Chamberlin’s only concrete magnitudes are anecdotes: a New England gasoline retailer who matched posted prices but effectively sold two cents a gallon cheaper through coupons; a Cleveland department store where a five-dollar iron outsold a promoted $3.75 iron by 50 percent (Chamberlin, 1933). Robinson offers no case, no statistic, and explicitly disclaims policy application. One can defend the omission as deliberate, and Robinson did. But the omission is exactly the void the next fifty years fill, and Chamberlin makes it worse in a direction the field took a long time to appreciate: in his theory, even a market with many sellers produces monopoly prices and waste through differentiation alone, so concentration is neither necessary nor sufficient for market power. The 1933 theory tells you market power is everywhere and gives you no way to grade it.

That is where Edward Mason enters, the hinge between diagnosis and instrument. At Harvard in the late 1930s, Mason asked what a science of the firm should look like now that Chamberlin and Robinson had killed the competition-or-monopoly binary, and his answer made structure the starting point: classify markets by their observable structural features, then study how those features govern the price and production policies of the firms inside them (Mason, 1939). The essay contains no number and no table, which is fitting for the era it closes; what it contains is a causal ordering, and orderings are what get inherited. Note the chronology this prologue leaves behind, because it is the era’s verdict on itself. The law that needed a theory of market power was written in 1890, twenty-two years before economics could describe a firm between competition and monopoly. The theory, when it arrived in 1933, attached to no share, no case, and no industry, twenty-two years after the law had already dissolved the paradigmatic trust. Enactment and theory never met. The prologue ends with a theory that says every real seller prices above marginal cost, a law that has acted once on a share it never formally measured, and no way to say which industries are imperfectly competitive or by how much. The loose thread is measurement itself. The next era builds the thermometer, calls its reading concentration, and gives the field’s founding question its first answer: everything.

Era I: The Cross-Section (1939-1968)

The era’s answer, stated as a claim you can contest: structure tells you nearly everything. How concentrated a market is, and how high the walls around it stand, determine how firms behave and what they earn, so market power can be read off a cross-industry correlation without ever observing conduct. Concentration is a cause, a primitive, an input; a number above a critical level is the harm. Every word of that is contestable, and the rest of this history is the contest. But in 1951 the causal arrow points one way, and it points hard.

The arrow was asserted before it was measured. Mason’s 1939 essay is a conceptual charter with no number and no table: a firm’s size relative to total market volume is “one direct indication of the extent of its market control,” and market control governs price and production policy (Mason, 1939). That sentence seeds the ordering that became structure, conduct, performance. It matters that the charter is not itself a finding. Mason licenses reading a coefficient as harm, because in his ordering structure is the exogenous primitive; without that license, any correlation between concentration and profit is just a correlation. Mason asserts the arrow. His student measures its length.

Joe Bain’s 1951 paper is the founding coefficient of the field, and it deserves to be read as what it actually is rather than as the strawman it became. Bain took 42 manufacturing industries, built from the financial statements of roughly 340 firms over 1936 to 1940, measured concentration as the eight-firm share of value of product and performance as the five-year average after-tax return on net worth (Bain, 1951). What he found was not a slope but a step. Industries where the top eight firms supplied 70 percent or more of output earned about 11.8 percent on net worth; industries below the threshold earned about 7.5 percent; and within each group, profit showed essentially no further relation to concentration. A gap of roughly four points, appearing sharply at a break, across dozens of industries.

The step function is the underappreciated part, and it is Bain’s best evidence. Noise does not produce a discontinuity; noise produces scatter. A threshold at CR8 of 70 percent reads naturally as separating two types of oligopolist, those concentrated enough for collusive pricing to hold together and those diffuse enough to approximate competition, which is what oligopoly theory of the time predicted. Bain thought he was supplying the first systematic empirical confirmation of that theory, and turning Mason’s program into measured cross-industry fact. He also thought, and here the era’s method shows its original sin, that the cross-section was the right instrument. Mason had called for close, institutionally rich case studies of firms. The paradigm that bears his name became a body of aggregate statistical correlations he never asked for. The field kept Mason’s unit, the industry, and discarded his method, and the discarded method turns out to be where the era’s fatal question was hiding.

Two objections to the instrument were latent from the start, and it is worth separating them because they have different fates. The first is measurement: an accounting return on equity is a poor proxy for economic profit, a point later made devastating by Fisher and McGowan, and one reason the field eventually migrated to price-cost margins and then away from profitability regressions altogether. The second is identification, and it is the deeper one: the regression’s design decides its interpretation before any data arrive. Bain compared groups of industries; he never compared firms within an industry. Whether that omission was innocuous or fatal depends on which of two structural stories generates the coefficient, and nothing in the grouped comparison can say. Defenders were right that the pattern itself was no accident; the correlation replicated for a quarter century in study after study. What replicated was the coefficient. What never replicated, because it was never tested, was its reading.

Bain spent the next five years answering the obvious follow-up: why do some concentrated industries stay profitable while others do not? Barriers to New Competition reframed the question around potential competition. The “condition of entry,” how far incumbents can hold price above competitive levels without inducing entry, is a second structural variable, and Bain gave it a taxonomy: scale economies, product differentiation, absolute cost advantages, and large capital requirements (Bain, 1956). Ranking about twenty industries by barrier height, the very-high-barrier group earned returns in the high teens against 8 to 12 percent for easy-entry industries, a gap of eight to nine points, twice the concentration gap alone. His most striking judgment was that product differentiation, not scale, was the most potent barrier in consumer goods: automobiles, cigarettes, liquor, fountain pens. Concentration plus barriers now jointly predicted durable market power, and every element of the structure was treated as a fact about technology and taste rather than a choice. That framing, calling scale economies and capital requirements “barriers” when entrants and incumbents bear them alike, would later be attacked as smuggling a normative judgment into a cost fact. But within the era, the two books are one continuous program, and the program owned the field.

Then it walked into the U.S. Reports, and the numbers tell you what enactment did to the idea. Bain located his break at an eight-firm share of 70 percent. The Warren Court blocked mergers at 5. The Justice Department’s 1968 Merger Guidelines completed the administrative enactment in the same spirit, condemning combinations at four-firm shares in the single digits, a share-based structuralism that would stand until Baxter tore it out in 1982.

Brown Shoe (1962) condemned the merger of the fourth-largest shoe manufacturer with the largest independent retail chain on two counts, both trivial by any structural arithmetic (Brown Shoe, 1962). The vertical count: Kinney sold about 8 million pairs of shoes in 1955, roughly 1.6 percent of national nonrubber pairage, and even that was illegal given a documented “trend” toward vertical integration. The horizontal count: a combined retail share exceeding 5 percent in 118 separate cities (women’s shoes topped 20 percent in 32 cities, reaching 57.7 percent in Dodge City). The Court said the quiet part aloud: Congress, in the 1950 Celler-Kefauver amendment, “resolved these competing considerations in favor of decentralization,” accepting that “occasional higher costs and prices might result,” and chose fragmented markets to protect small, locally owned business. On the structuralist premise this is coherent rather than confused. If structure determines performance, preserving competitive structure is preserving competition, full stop. The opinion also built the market-definition machinery, product markets by reasonable interchangeability and cross-elasticity, submarkets by practical indicia, that outlived its permissive threshold by sixty years. The apparatus survived; the 5 percent condemnation did not.

Von’s Grocery (1966) is the doctrinal peak and carries its own obituary inside it. Von’s, the third-largest grocery chain in Los Angeles, acquired the sixth-largest; combined they held 7.5 percent of a $2.5 billion market (Von's Grocery, 1966). There is no analysis of prices, entry, or efficiency anywhere in the majority opinion. The dispositive evidence is a structural time series: single-store grocers in Los Angeles fell from 5,365 in 1950 to 3,818 in 1961 while chains absorbed the independents. For Justice Black, a falling firm count is the harm the statute was written to stop in its incipiency; you do not wait for the monopoly to mature. Justice Stewart’s dissent is the era’s epitaph, written from inside the building: the majority blocked a 7.5 percent merger on conjecture, treating the displacement of corner stores by efficient supermarkets as if it were a finding about competition, and “the sole consistency that I can find is that, in litigation under Section 7, the Government always wins.” That line is not rhetoric. It is a charge of non-falsifiability: if 5 percent and 1.6 percent are enough, no number is a safe harbor, and incipiency has swallowed the statute’s requirement of a probable competitive effect. Note what the Court did to Bain’s number on the way to enacting his logic: it did not adopt the 70 percent threshold, it abolished the threshold. The structural presumption traveled from academy to law by deleting its own limiting principle.

The era closes with its strength and its exposure sitting in the same object, the single reduced-form coefficient. The four-point gap made SCP credible; it also admits a second structural story the cross-section cannot exclude. If collusion raised the industry price, the higher price is a public good, and small and large firms’ returns should rise together. If instead only the largest firms in concentrated industries earn the premium, the pattern traces the superior efficiency of the large, and concentration becomes an outcome of firms getting good and growing rather than a cause of an overcharge. Bain’s design is structurally blind to this distinction, because grouping industries by CR8 averages away exactly the within-industry, by-firm-size variation that would decide it. The loose thread of Era I is not a flaw anyone overlooked; it is the one question the instrument could not ask. The next era asks it, and the field’s answer flips from everything to nothing.

Era II: The Demolition (1968-1982)

The answer flips: industry structure carries almost no information about conduct or performance. The concentration-profit correlation is real but misread; it traces efficient firms growing large, not colluding firms overcharging. Predation, the classic story of the strong crushing the weak, is irrational for a rational monopolist and unobserved in the record. Where entry is free, its mere threat disciplines even a single-firm monopoly. Therefore stop breaking up firms on the strength of a coefficient; the coefficient was never evidence of harm. This claim is contestable in three specific places, and the era’s opponents contested it in exactly those places. The efficiency reading fails if market power, like efficiency, accrues disproportionately to large firms. The predation claim fails if predation can be rationalized on its own terms. The entry claim fails on a single positive sunk cost. Era II wins the war of ideas on all three planks and, remarkably, loses the technical point on all three. That gap between winning the field and winning the argument is the era’s signature.

The demolition’s first charge was loaded before the era began. McGee’s 1958 rereading of the Standard Oil record supplies the premise everything else rests on: the classic exclusionary mechanism does not work (McGee, 1958). The logic is arithmetic. A predator with 75 percent of the market sells roughly three times as much as all rivals combined and therefore loses roughly three times as much on every below-cost sale, while the prey’s shuttered plants and personnel simply reappear when prices recover. Purchase dominates predation: buying a rival at up to the discounted value of future monopoly profit always beats a war whose revenues are lower during the fight and no higher after it. And the record confirms the logic. Reading the full 11,000 pages market by market, McGee found no instance of Standard using below-cost pricing to force a refiner to sell out. Standard bought its monopoly at generous, sometimes lavish, prices; competing refineries actually grew from 38 in 1895 to 123 in 1906. The paradigm case of predation contained no predation. If the strong cannot rationally crush the weak, large size cannot be the product of exclusion, and must be the product of something else.

Stigler then did something stranger than demolition: he built the collusion machine and undermined it in the same motion. His theory of oligopoly set out to rescue the concentration-collusion link by giving it a mechanism, collusion succeeds when secret price-cutting is easy to detect, and detection is a statistical inference problem in the seller’s own sales data (Stigler, 1968). But the model relocates the causal action from seller counts to the number and size of buyers and the observability of transactions; the incentive to cheat falls roughly with the inverse square root of buyers per seller. Concentration survives only as a noisy proxy for detectability. Stigler handed his successors both the microfoundation the structuralists had lacked and the skepticism that the cross-sectional correlation proves nothing, and his successors used the second gift against the first. He also took a knife to Bain’s second structural variable. Redefine a barrier to entry as a cost the entrant must bear that incumbents did not, and two of Bain’s four barriers dissolve on contact: scale economies and capital requirements are borne equally by anyone building a plant, incumbent or entrant, so calling them barriers presupposes the very theory of oligopoly conduct they were supposed to establish. Under Stigler’s definition, only genuine asymmetries survive, some absolute cost advantages and, above all, government restriction, which points the indictment back at the state. Three years later he generalized the skepticism to the state itself: regulation is acquired by the industry and operated for its benefit, with magnitudes attached, oil import quotas costing consumers upward of $5 billion a year, a Civil Aeronautics Board that licensed zero new trunk airlines from 1938 onward (Stigler, 1971). The villain became government-granted entry control, not private structure. Same move, different institution.

Demsetz delivered the decisive counter-reading in nine pages. His 1973 paper poses the question as a public-goods problem: if collusion drove the concentration-profit link, the elevated price is shared by every seller in the industry regardless of size, so small and large firms’ returns should rise with concentration together (Demsetz, 1973). They do not. Disaggregating accounting returns within industries by firm asset size in 1963 IRS and FTC data, Demsetz found that the return-concentration relationship lives almost entirely in the largest size class, several points of return for firms above $50 million in assets, essentially nothing for the smallest firms in the same concentrated industries. Concentration buys higher profits for large firms but not for their small neighbors. His reading: the correlation traces efficiency, low-cost firms grow large and earn high returns, and concentration is the outcome of competition rewarding winners. His policy warning: deconcentration remedies premised on the correlation would punish firms for winning the race, attacking efficiency in the name of attacking monopoly.

Here is the part the textbook version leaves out, and it is the most instructive fact in the era. Demsetz’s test does not actually separate the hypotheses. Market power can also accrue disproportionately to large firms; a dominant-firm price leader earns supracompetitive margins its fringe cannot share, and share-weighted markups generate exactly Demsetz’s pattern. Scherer, Weiss, and eventually Schmalensee’s accounting of the whole literature reached this verdict within fifteen years: the size-disaggregation test is consistent with both stories, and the cross-industry regression is unidentified in either direction. So Demsetz won the framing, SCP became “discredited” and the efficiency reading became doctrine, without proving his positive claim. His negative point, do not infer collusion from the coefficient, held and holds still. His affirmative point, that the coefficient traces efficiency, was never established.

Measure the victory by where it traveled rather than by what it proved. The efficient-structure reading fed directly into Bork and Posner, into the displacement of the Brown Shoe presumption, and into the 1982 Guidelines’ orientation, and it kept governing policy long after IO economists had stopped believing the underlying regression could identify anything. The most consequential victory in the field’s history rests on a test that cannot distinguish the winner from the loser, and the doctrine outlived the evidence by a generation. I take that as the era’s real lesson, and it will matter again in 2020, when the same coefficient returns with the sides reversed and the identification no better.

Peltzman supplied the empirical capstone with a decomposition the prior literature had assumed away. Forty-plus studies had confirmed the concentration-profit correlation by 1977 and read the whole thing as a price effect, implicitly assuming concentration leaves costs untouched. Measure the cost channel and it dominates: for the average industry with rising concentration, the decomposition runs dLnP = +.15 from the collusion channel against -.91 from the cost channel, roughly five to one (Peltzman, 1977). Concentration raises profit not because price rises but because cost falls further than price does. The welfare arithmetic is brutal for the structural program: deconcentrating a typical concentrated industry, pushing its four-firm ratio from about .7 to .5, would raise unit costs on the order of 20 percent, raise prices 10 to 15 percent, and impose a social loss around 13.6 cents per dollar of output, many times any Harberger-style deadweight-loss gain. The paper aged badly on its own terms; the next generation rejected cross-industry accounting-profit regressions as unidentified whoever runs them, Peltzman included. But as ammunition in 1977 it closed the case.

Bork translated the economics into a legal program, and his translation contains a definitional sleight the field took thirty years to unwind. The Antitrust Paradox gave the courts one goal, “consumer welfare,” and argued that most challenged practices raise output, that high shares usually mark superior efficiency, and that horizontal mergers up to roughly 60 to 70 percent share should be permitted (Bork, 1978). The Supreme Court imported the phrase within a year (Reiter v. Sonotone, 1979). But Bork’s consumer welfare was not consumer surplus; his own definition makes it “merely another term for the wealth of the nation,” total surplus. Williamson had already shown the two can move in opposite directions, a merger can raise efficiency and price together, which is precisely the possibility Bork’s synthesis denied. Courts adopted the label while quietly reading it as consumer surplus, producing a governing standard whose meaning was never settled. The ambiguity sat harmlessly in the doctrine for four decades, until a generation arrived that wanted to pry it open.

The era’s theoretical limit case arrived in 1982 and broke almost immediately, which makes it the cleanest experiment in the whole history. Baumol, Panzar, and Willig pushed the severing of structure from performance to its logical extreme: if entry is free and exit costless, a potential entrant can hit and run, undercutting the incumbent and leaving before retaliation, and the perpetual threat forces even a single-firm natural monopoly to price at competitive or Ramsey levels (Baumol et al., 1982). Natural monopoly itself was rigorously redefined as subadditivity of the cost function. A perfectly contestable monopoly prices as if perfectly competitive; the policy lever is guaranteeing free entry, not breaking up firms. Shepherd called it “a kind of fantasy,” and his demolition has worn better than the theory (Shepherd, 1984). Contestability requires a triad that essentially never holds jointly: zero sunk costs, an entry-and-response lag long enough for the entrant to profit and flee before the incumbent cuts price, and entrants with incumbents’ costs. Any positive sunk cost collapses the theorem. The empirical verdict came from the theory’s own flagship application. Airlines, deregulated in 1978 on contestability’s logic, “capital on wings,” consolidated into hub-and-spoke fortresses with fares at concentrated hubs running tens of percent above competitive routes, sustained by frequent-flyer lock-in, reservation-system control, and gate scarcity. Concentration raised prices in the flagship contestable market, exactly the structuralist prediction the theory claimed to overturn. BPW keep the theory round; the theorem is correct in its limit case. Shepherd wins robustness and then the war. The durable residue is a relocation of the whole question: not “how concentrated?” but “how sunk are the costs of entry?”

The demolition became law on June 14, 1982, when William Baxter’s Antitrust Division retired the structuralism of the 1968 Guidelines and replaced it with a price-theoretic pipeline (Merger Guidelines, 1982). Three moves, all in the text. Define the market by the hypothetical-monopolist test, operationalized as a 5 percent price increase and a year’s switching, the SSNIP test that became global orthodoxy. Measure concentration by the Herfindahl-Hirschman Index rather than four-firm ratios, because squaring shares weights the large firms that matter for collusion. Then apply the grid: below HHI 1000, unconcentrated, no challenge; 1000 to 1800, challenge likely above a 100-point increase; above 1800, likely at 100 points, possible at 50. The Department’s own study put HHI 1000 and 1800 at roughly four-firm ratios of 50 and 70 percent, and the worked example is almost a parody of the new precision: a merger of 5 and 10 percent firms raises HHI by 5 times 10 times 2, exactly 100. Two things about the document repay attention. Concentration survives in it only as a proxy for the ease of collusion, Stigler’s detectability logic operationalized; structure is no longer destiny but evidence. And the guidelines hedged the era’s own extreme: entry is a rebuttable factor, not a general theorem, so when Shepherd and the airlines broke contestability, the enforcement framework survived intact. The one overcorrection in the text, efficiencies dismissed as “far easier to allege than to prove,” was reversed by 1997. Otherwise, this is the moment structuralism was formally retired inside enforcement. Von’s Grocery was still good law; no one would ever bring it again.

Era II ends victorious on every plank and technically undefended on each. McGee’s “predation is irrational” is a complete-information, single-market theorem, and the next era answers it on its own turf with asymmetric information. Demsetz and Peltzman established that the coefficient cannot identify collusion, but their efficiency reading is equally unidentified, and the next generation abandons the regression altogether rather than pick a side. And Shepherd’s sunk cost, the thing that broke contestability, becomes the commitment device at the center of what replaces it. Three loose threads, one successor era pulling all of them.

Era III: The Theory Glut (1975-1991)

The answer changes shape: it depends, provably. Feed a noncooperative game the right extensive form, the right information structure, the right order of moves, and you can support almost any outcome as an equilibrium. Predation is irrational in McGee’s complete-information world; add a one-in-a-thousand doubt about the incumbent’s type and it becomes rational. Excess capacity deters entry in Spence; it is never held in equilibrium in Dixit. Concentration falls to zero as markets grow with exogenous sunk costs; it stays bounded away from zero forever with endogenous ones. Each result is airtight; collectively they say that cost and demand functions do not suffice to predict conduct, you must also specify the institutions, and once you specify the institutions you have smuggled in the answer. Whether that is progress or paralysis is the era’s own internal fight, and it was still open when the era closed. Eras III and IV overlap, 1975 to 1991 against 1981 to 2007, because they differ by question, not by calendar. The theorists asked what could rationally happen; the empiricists, often in the same buildings in the same years, asked what was actually happening in a particular market.

The entry-deterrence track opens with a mechanism that outlived its own headline prediction. Spence relocated Bain’s limit-pricing question from price to capacity: an incumbent deters entry by sinking irreversible capacity that lets it credibly flood the market on arrival, and the deterring configuration involves capacity above the cost-minimizing level (Spence, 1977). Dixit then supplied the discipline Spence’s story lacked. Sunk capacity kinks the incumbent’s marginal cost at the installed level, shifting its post-entry reaction curve outward so that an otherwise empty threat becomes a best response; but the credible threat extends only over a bounded range, and holding literal idle capacity is generally not optimal once the post-entry game is Cournot (Dixit, 1980). The signature prediction, persistent visible excess capacity, was walked back almost immediately, and by the literature Spence founded. The device survived: commitment through irreversibility, Shepherd’s fatal sunk cost reborn as a strategic instrument. That inversion, the same object refuting one theory and powering the next, is how this era does business.

The predation reversal is the era’s centerpiece because it is the cleanest case in the field of theory refuting doctrine and doctrine winning anyway. In a single 1982 issue of the Journal of Economic Theory, Milgrom and Roberts and Kreps and Wilson dismantled Selten’s chain-store paradox and with it McGee’s claim that predation is always irrational (Milgrom and Roberts, 1982; Kreps and Wilson, 1982). The move is asymmetric information. Let entrants place even a tiny probability on the incumbent being a committed fighter, and preying on early entrants becomes an investment in reputation that deters later ones. The magnitudes are the striking part. In Milgrom and Roberts’s construction, with the entrant’s prior at one in a thousand, fewer than four entries are ever accommodated; at one in a million, fewer than nine; at one in a billion, fewer than fourteen. The bound grows logarithmically in the doubt and is independent of the number of markets. Kreps and Wilson show the effect survives at a doubt of one in a thousand with roughly ten stages remaining and does not dilute as the horizon lengthens; the limit is discontinuous at zero doubt. Predation is rational precisely where McGee’s single-market present-value arithmetic cannot see it, across markets and time, in beliefs off the equilibrium path. Both papers flag the same coda honestly: the equilibria are fragile and multiple, and predation need never occur on the path, so the theory predicts the possibility of predation while implying it may almost never be observed. Hold that caveat; the law is about to widen it into a wall.

Notice also what the reversal did not deliver, because the omission decides the legal fight to come. Predation and hard competition trace the same price path down. The difference between them lives in intent and in off-path beliefs that data almost never reveal: the deterred entrant who never entered leaves no record, and the fighting incumbent looks exactly like a vigorous competitor. So the theory that vindicated predation as rational conduct simultaneously certified it as nearly unidentifiable in a courtroom. A plaintiff armed with Milgrom-Roberts has a possibility theorem; a defendant armed with McGee has an arithmetic and a clean record. When the two met before judges a decade later, it was no contest, and the reasons were evidentiary, not economic.

Consolidation arrived in 1988 in the form of a graduate textbook, and the textbook performed a quiet burial. Tirole’s Theory of Industrial Organization installed noncooperative game theory as the field’s language and, in one equation, gave Bain’s founding correlation its microfoundation: under Cournot with constant returns, industry profit is proportional to the Herfindahl index, and the ratio of profit to revenue equals the Herfindahl over the demand elasticity (Tirole, 1988). Here, at last, was the theory that made a profit-concentration regression interpretable. And on the facing page Tirole dissolves it: concentration indices are endogenous, correlations cannot be read causally, and Demsetz’s noncollusive reading stands. The book codified the paradigm and buried the empirical program that had motivated the field, in the same chapter.

Then the field paused to ask whether the revolution had produced anything, and the honest answer was uncomfortable. Fisher’s indictment at the 1989 RAND symposium: game-theoretic IO produces “exemplifying” theories, results showing some outcome can occur under some specification, not “generalizing” theories that say what will happen across a broad class (Fisher, 1989). The Folk Theorem is the limiting scandal, any individually rational payoff an equilibrium of the repeated game, and the literature an almost interminable series of special cases, each rigorous, collectively mute. Shapiro’s defense conceded the multiplicity honestly and defended the vocabulary: the two-stage game, an irreversible commitment stage followed by a competition stage; the credibility requirement that separates real threats from empty ones; and the Fudenberg-Tirole taxonomy of strategic postures, top dog, puppy dog, fat cat, lean and hungry, which reduces a sprawling literature to one question, whether competition is in strategic substitutes or complements (Shapiro, 1989). A language is not nothing; before 1975 the field could not even state what made a threat credible. But a language is not a prediction either, and on the narrow point Fisher was conceded rather than refuted.

Sutton’s 1991 book is the honest close, and I read it as the era’s best moment precisely because it refuses both sides of the Fisher-Shapiro fight. If the unobservable details of the game determine the equilibrium, stop predicting points and predict what survives across every admissible game: a bound (Sutton, 1991). The nonconvergence theorem says that with exogenous sunk costs, concentration falls toward zero as market size grows, but with endogenous sunk costs, advertising and quality escalation that scale with the market, concentration stays bounded away from zero forever, no matter how large the market gets and no matter which price-competition subgame you assume. And the bound held in data: across roughly twenty food-and-drink industries in six countries, the estimated asymptotic four-firm floor is 0.06 percent for the low-advertising control group and 19 percent, significantly above zero, for the advertising-intensive group. That is not a possibility theorem. It is a falsifiable prediction that survived contact with data, derived while taking the indeterminacy completely seriously. The era’s answer to Fisher was not to deny the indeterminacy but to change what counts as a prediction.

One paragraph belongs to the parallel track that carried the era’s apparatus into regulation. Laffont and Tirole reframed the problem of a government procuring from a monopolist it cannot monitor: the firm privately knows its cost type and privately chooses cost-reducing effort, observed cost confounds the two, and public funds carry a shadow cost (Laffont and Tirole, 1993). The optimal contract trades rent extraction against effort incentives, gives the most efficient type a fixed-price contract with first-best effort and the less efficient types cost-plus sharing with information rents, and implements the whole thing as a menu firms self-select from. This became the theoretical warrant for the worldwide shift from rate-of-return to incentive regulation, price caps and RPI-X. Its message for the concentration question is procedural rather than structural: a monopoly can be made to perform through well-designed conduct incentives, not deconcentration, though the elegant static optimum is fragile once limited commitment, ratchets, and capture enter.

The era’s policy enactment is Brooke Group (1993), and it is the exhibit for a claim I want to develop rather than mention: the law revokes economics on a lag, and the lag can exceed a decade. Justice Kennedy’s opinion self-consciously imports Chicago industrial organization, treating a predatory scheme as an investment that is rational only if recouped, and codifies a two-prong test: the plaintiff must prove prices below an appropriate measure of cost, and a reasonable prospect, or dangerous probability, of recouping the investment in below-cost prices (Brooke Group, 1993). The recoupment prong is dispositive, and the Court’s arithmetic in the case shows why it is nearly unpassable: Brown & Williamson held 11 to 12 percent of the cigarette market, so roughly eight of every nine dollars of any later oligopoly rents would accrue to rivals it could not exclude, a required recoupment ratio on the order of nine to one. The jury’s $49.6 million verdict, trebled to $148.8 million, was set aside. Predation, the Court said, quoting Matsushita, is “rarely tried, and even more rarely successful.”

Now place the dates side by side. McGee argued predation is irrational in 1958. Milgrom-Roberts and Kreps-Wilson refuted him, on his own turf, in 1982. The Supreme Court codified McGee’s conclusion in 1993, eleven years after the refutation, and the doctrine has governed since, making predation claims close to unwinnable. The chronology is not an embarrassment to be smoothed over; it has a mechanism. Courts do not consume theorems, they consume administrable rules, and the reputation papers’ own honest caveat, equilibria fragile and multiple, predation possibly never observed on the path, is exactly what makes them unusable as a burden of proof and exactly what the Court converted into an evidentiary wall. The technical winner of the predation dispute is the game theorists; the war-of-ideas winner is McGee, whose skepticism outlived the theorem that once justified it and hardened into law after it was overturned. The field’s laboratory runs one full step behind the field. It will do so again.

The loose thread is Fisher’s complaint, accepted and repurposed. If cost and demand functions plus a chosen game form can rationalize any conduct, the missing step is empirical: pin down, with data from a single market, which model and which equilibrium is actually operating. Tirole had already demoted the cross-industry regression to an unidentified reduced form; Sutton had already made concentration a jointly determined equilibrium object rather than a regressor. The next move is to stop asking what concentration means in general and start asking it one market at a time. The strategic revolution proved that everything is possible. The empirical revolution asks which possibility is true, market by market.

Era IV: One Market at a Time (1981-2007)

The fourth answer is the most deflationary: structure means nothing in general and something precise in a single market with a demand system. The cross-industry question, what does concentration do to profits across industries, is not hard. It is ill-posed, and the field abandons it. In its place the era substitutes a demand system estimated brand by brand and a pricing game inverted market by market, from which market power falls out as a number attached to a product in a place at a time: $3,753 on the average 1990 car, 42 percent on a box of cereal. The claim is contestable because it trades reach for rigor. It concedes the entire terrain the first three eras fought over, and it says nothing about whether American industry as a whole is gaining or losing market power. The era treated that silence as a virtue. Its successor would treat it as an evasion.

The charter is six pages long. Bresnahan’s 1982 note settles whether the whole program is even possible: is the oligopoly solution concept identified from data (Bresnahan, 1982)? With demand shifts alone, no; competitive and monopoly pricing trace the same equilibria under suitably chosen costs. But a variable that rotates demand rather than shifting it separates the hypotheses, and the conduct parameter, zero for competition, one for cartel, intermediate for Cournot, becomes an estimable coefficient. Market power is now a measured object, not an assumption. This is the methodological constitution of what came to be called the New Empirical Industrial Organization: refuse to read power off the Herfindahl, estimate it from behavior. (The label’s paternity is itself mildly disputed, as the practice predated the name and several authors have been credited with coining it; I use the term as the field does, for the program this identification result chartered.)

The charter was cashed immediately, and the first application contains the era’s best joke on itself. Porter estimated a switching regression on the Joint Executive Committee railroad cartel, 328 weeks of grain-rate data from 1880 to 1886, and rejected the no-switching null with a likelihood-ratio statistic of 554.1 (Porter, 1983). Cooperative-period prices ran 66 percent above reversionary prices; the implied conduct parameter, 0.336, sits near Cournot, not joint monopoly, even during cooperation. The joke is the reversal Porter himself flagged: he had found a 1980s game-theoretic equilibrium, Green-Porter trigger reversions, apparently operating in 1880s data. The follow-up cut deeper. The price wars were not reliably preceded by demand slumps, as the Green-Porter mechanism has it; unpredictable shifts in individual market shares seem to have triggered reversions, which is Stigler’s 1964 detection story vindicated in the archive. Bresnahan then ran the same play on the 1955 auto price war: with the same demand and cost parameters throughout, Cox tests select collusion as the only surviving model in 1954 and 1956 and competition as the only survivor in 1955 (Bresnahan, 1987). The anomalous year is a one-year change in conduct, not in fundamentals. Conduct had become something you estimate, and something that moves.

Structure itself was next to become an output rather than an input. Bresnahan and Reiss asked how fast competition arrives as firms enter isolated small-town markets, using the ratio of successive per-firm entry thresholds as a scale-free statistic that needs no price data at all (Bresnahan and Reiss, 1991). The ratios fall sharply through the second and third entrant, then flatten; equality of the third through fifth thresholds cannot be rejected for four of five industries, and in the one industry with prices, tires fall about 8 percent from monopoly to quintopoly. The finding entered folklore as “three to five firms is enough for competition,” and the folklore is more confident than the paper: the same tire data show small-town oligopoly prices leveling off well above the competitive urban benchmark, roughly 20 percent above the Bay Area. Conduct stops changing by the third entrant; it does not become competitive. Berry then rebuilt entry with heterogeneous firms and simulation estimators on airline city pairs, and the payoff is a counterfactual that discredits its naive rival: granting every carrier access to both endpoint airports adds about 0.4 carriers per market under the structural model, against about 5 under a probit that ignores strategic interaction (Berry, 1992). A tenfold gap, measuring how much equilibrium logic blunts entry, and, in passing, an unheralded empirical rebuttal of contestability: freeing access does not dissolve concentration when most markets can support only two firms regardless.

The era’s single load-bearing invention arrived in the July 1995 issue of Econometrica, twice. Berry, Levinsohn, and Pakes solved the two problems that had blocked demand estimation for differentiated products for sixty years (Berry et al., 1995). The first is dimensionality: with a hundred products, an unrestricted demand system needs on the order of ten thousand elasticities, and the simple logit that economizes on parameters forces absurd substitution patterns. The random-coefficients logit generates realistic substitution from a handful of taste parameters, because consumers who love horsepower substitute to other fast cars, not to whatever else is cheap. The second is price endogeneity, the differentiated-products version of the oldest problem in econometrics: firms price on quality the econometrician cannot see. The BLP inversion solves for the mean utilities that match observed market shares exactly, which converts the unobserved quality into a structural error that can be instrumented, using the attributes of rival products. Their own diagnostic tells you why the machinery matters: ordinary least squares implies that 1,494 of 2,217 car models face inelastic demand, impossible under profit maximization; instrumenting cuts it to 22. Then the supply side: impose multiproduct Nash-Bertrand pricing, invert the first-order conditions, and a markup falls out for every model sold in America over two decades. The average 1990 markup is $3,753, a Lerner index of 0.239, running from $801 on a Mazda 323 to $10,975 on a BMW 735i, with the twist that the most profitable cars are the mass-market Accord and Taurus, not the high-markup luxury models. The authors flagged their own wart, substitution to the outside good they called “a bit large,” and a refinement literature grew in the crack. But the institutional consequence dwarfs the econometric one. Market power was now routine to measure, product by product, with no natural experiment, no accounting profits, and no concentration ratio anywhere in the pipeline. In the same issue, Goldberg built the demand side from household-level Consumer Expenditure Survey data and ran the kind of counterfactual the machine was built for: remove the 1981 Japanese voluntary export restraint, and only about 54 percent of the sales gains it generated accrue to U.S. firms, the rest leaking to European producers and to Japanese quality upgrading (Goldberg, 1995). Trade policy, evaluated inside an oligopoly equilibrium, off one estimated demand system.

The machine’s cleanest verdict came in cereal, and it retired a presumption the FTC had litigated. Ready-to-eat cereal was the textbook “shared monopoly”: three firms, price-cost margins around 45 percent, heavy advertising, rivalry visibly deflected from price into brand proliferation, and a 1970s FTC complaint that treated the pricing as approximately cooperative. The collusion side owned the prior, and it was not a strawman. Nevo accepted the burden and discharged it (Nevo, 2001). Estimate brand-level demand on 27,862 brand-city-quarter observations, invert the pricing conditions under three conduct hypotheses, and read off the implied margins: 35.8 percent if each brand priced alone, differentiation only; 42.2 percent under the actual multiproduct ownership playing Nash-Bertrand; 72.6 percent under joint monopoly. The observed accounting benchmark is about 46 percent. Competition among heavily differentiated brands, plus firms internalizing substitution across their own portfolios, explains the margins; collusion is rejected and unnecessary. Differentiation contributes roughly 36 points and the portfolio effect another 6. Nevo’s own caveats deserve their place: the test rules out only the extreme joint-monopoly model, not partial cooperation, and the FTC’s actual theory had been brand proliferation as an entry barrier, which his pricing test does not touch. Even so, the canonical inference of Era I, high concentration plus high margins equals market power, is dead in this market. High margins had a competitive explanation, and only a demand system could have found it.

The auction branch ran the same play with bids instead of prices, and deserves its brief mention because it proves the method’s reach. Hendricks and Porter showed that on offshore drainage leases, informed neighbor firms captured about 44 percent of tract value against 23 percent on symmetric wildcat tracts, uninformed bidders earned roughly zero, and the neighbor bid function decreases in the number of neighbor firms, inconsistent with competitive bidding and consistent with neighbors coordinating on one serious bid (Hendricks and Porter, 1988). Tacit coordination, read off the shape of an equilibrium bid function. Guerre, Perrigne, and Vuong then made the recovery nonparametric: each private value equals the bid plus a term in the bid distribution, so latent values are identified from observed bids with no functional form at all (Guerre et al., 2000). Recover unobserved types from equilibrium behavior: it is BLP’s inversion, in a different market institution.

The era’s terminal dispute was staged in 2010, in a single journal issue, on IO’s home ground, and it is the fight the next era inherits. Angrist and Pischke, carrying the credibility revolution’s flag, chose merger prediction as the target because that is where the structural program is most exposed (Angrist and Pischke, 2010). Their narrow case lands. Mergers sit at the center of applied IO, yet a contemporary review found only about 20 studies directly evaluating consummated-merger price effects, against hundreds of union-wage studies in labor. And when simulations met retrospectives, the match was “mixed at best”: structural predictions of post-merger airline prices were poor, design-based cereal estimates differed markedly from the structural ones, and of three structural models checked against a direct estimate, one matched. Against this they set Hastings’s difference-in-differences on the ARCO-Thrifty gasoline merger, a transparent design anyone can read. A good research design, they concluded, beats a good theory. Nevo and Whinston’s rebuttal won the framing by refusing the choice (Nevo and Whinston, 2010). Mergers are heterogeneous and prospective; the merger a regulator must evaluate has not happened, resembles no past merger closely, and therefore requires a model, the only question being whether the model is statistical or economic. Structural simulation needs no merger history, delivers welfare numbers no treatment effect contains, and produces a range under stated assumptions rather than one fragile point. Their challenge to the agencies stands as the era’s best sentence about itself: who should you trust, a model grounded in economic theory and careful measurement, or extrapolation from a handful of past, quite different mergers of questionable exogeneity? The settlement was synthesis: the field kept BLP and adopted the design tradition’s demand for credible variation, and merger retrospectives became the standard tool for validating simulations, exactly the audit Angrist and Pischke had demanded and exactly the use Nevo and Whinston endorsed.

The era’s answer was enacted twice, in two registers. Merger simulation entered agency practice as a working tool: estimate a BLP-style demand system, assume Nash-Bertrand, re-solve the pricing game under merged ownership, and report the predicted price increase for a merger that has not happened, a number no regression on past data can produce. Nevo’s own cereal-merger simulations became the emblem of the pipeline, which is why the design-based critics chose cereal as their target. The concessions the defenders made in print deserve recording, because they are the honest boundary of the enactment: the demand step often leans on less-than-ideal instruments, the simulation holds products, entry, and investment fixed, and it assumes static Nash pricing before and after. Then the tool was audited by the very retrospectives that closed the era, and partly demoted: the 2010 Horizontal Merger Guidelines de-emphasized full simulation in favor of lighter diversion-ratio and upward-pricing-pressure screens that ask less of the superstructure. Enacted, audited, partly revoked, on schedule. The era’s courtroom enactment is Microsoft (2001), and it shows the structural instinct surviving inside effects-based doctrine (Microsoft, 2001). The D.C. Circuit found Microsoft holding more than 95 percent of Intel-compatible PC operating systems, and said, in the era’s voice, that the share is not the violation. What makes it durable is a specified mechanism, the applications barrier to entry: consumers prefer the system with the most applications, Windows had more than 70,000 against OS/2’s peak near 2,500, and developers write for the largest installed base, so applications flow to the incumbent irrespective of quality. Liability attached to the conduct that maintained the position against nascent middleware threats, under a four-step burden-shifting rule of reason; the per se tying holding was reversed, the breakup vacated, and Microsoft’s defense that it spent 17 percent of revenue on R&D was rejected. The opinion cites Bain (1956) and Stigler (1968) by name for its entry-barrier vocabulary and Standard Oil for its rule of reason: the whole eighty-year argument, compressed into one case. Structure is scrutinized only after a conduct mechanism is specified. That is Era IV’s methodological reflex, written into law.

The loose thread is the one the era chose, on principle, not to tie. It built a device that measures market power as a number, the estimated markup, and insisted the number be read one dial at a time. BLP’s markup belongs to the auto market; Nevo’s belongs to the cereal aisle; the era’s identity is the refusal to aggregate them into the cross-industry claim it had declared ill-posed. Nothing in the toolkit enforces that refusal. It was methodological manners, not a theorem. The next era reads every dial at once, calls the result market power, and the eighty-year argument reopens at national scale.

Era V: Scale Returns (2008-present)

The current answer, stated as its advocates would state it: maybe everything again. After four decades in which the profession treated the cross-industry concentration-profit regression as dead, the regression came back at economy scale, estimated from production functions rather than accounting returns, and it says market power has been rising for forty years. The claim is genuinely contestable, and the era’s distinctive feature is that it is contested at the level of measurement, not just interpretation: the skeptics deny that the measured markups reflect market power at all, on grounds none of which has been refuted. And the industries where the stakes are highest, the digital platforms, are precisely the ones where the tools built to adjudicate the dispute, price-cost tests, one-sided market definition, recoupment analysis, stop working. The field reopened its founding question at the moment it lost the instruments to answer it. The era runs as three movements.

Movement 1: the cross-section reopens

De Loecker, Eeckhout, and Unger did the forbidden thing. Their method, the production approach, needs no demand system and no conduct assumption: under cost minimization, a firm’s markup equals the output elasticity of a variable input divided by that input’s share of expenditure in revenue, the share read off financial statements, the elasticity estimated from sector-level production functions (De Loecker et al., 2020). Run on the full Compustat panel from 1955 to 2016 and aggregated with sales weights, the series is roughly flat until 1980 and then climbs from 1.21, 21 percent over marginal cost, to 1.61 by 2016, with the average profit rate rising from about 1 percent to about 8 percent over the same window. The anatomy matters as much as the level. The median firm’s markup barely moves; the rise lives entirely in the upper tail, and market share reallocates within industries from low-markup to high-markup firms. One common cause then organizes a cluster of macroeconomic puzzles: a firm’s labor share is mechanically its labor coefficient over its markup, so rising markups deliver the falling labor share, and the authors extend the logic to the capital share and declining dynamism. They were explicit about why the old instruments could not have seen this: concentration ratios are hostage to market definitions, while the production approach needs only firm accounts. This is Bain’s founding coefficient reborn, forty years of methodological modesty repealed in one series, and it reset the field’s agenda within two years of publication.

Two adjacent facts arrived alongside. Autor, Dorn, Katz, Patterson, and Van Reenen decomposed the fall of the manufacturing labor share, 10.42 percentage points over 1982 to 1997 and 5.65 over 1997 to 2012, and found reallocation between surviving firms accounts for 8.24 and 4.90 points of it, with the within-firm change around 0.4 (Autor et al., 2020). Output flowed to superstar winners: high-productivity, high-markup, low-labor-share firms taking share. The authors named the stakes themselves, in a footnote citing both principals of the founding dispute: this pattern reads as Demsetz’s differential efficiency, competition working, not Bain’s entry barriers and collusion. Gutiérrez and Philippon read the same decade the other way: concentration rose in the United States but not Europe, the concentration-productivity link flipped from positive to zero or negative, industry leaders cut investment, and the cumulative shortfall in the business capital stock reached 5 to 10 percent by 2016, with intangibles explaining only a quarter to a third of it, and three natural experiments, the China trade shock among them, pointing the causal arrow from declining competition to declining investment (Gutiérrez and Philippon, 2017). The same rise in concentration, read as efficiency by one 2020 QJE paper and as power by the NBER paper beside it. Dispute 1 of this history, rerun with firm-level data on both sides.

Then the audits landed, and this is where I would place the era’s center of gravity. Traina’s correction is chronologically prior to the fact it corrects and logically its reply: the DLEU series counts only cost of goods sold as the variable input, and SG&A, selling, general, and administrative expense, grew from about 12 percent of variable costs to about 22 over the sample (Traina, 2018). Count SG&A as variable, and the markup series becomes a shallow U, from roughly 15 percent over cost in 1950 down to under 10 around 1980 and back to about 15 today; the post-1980 rise is no larger than the pre-1980 decline. Reweight Compustat’s selected public firms to the Census size distribution and the aggregate flattens to about 10 percent over cost. DLEU’s counter, that SG&A is partly fixed overhead and does not belong in a flexible-input formula, is the unresolved core of the fight, and it is worth being precise about what kind of fight it is: the same estimator, the same data, one accounting switch, opposite conclusions about forty years of American capitalism. Bond, Hashemi, Kaplan, and Zoch then attacked the estimator itself: when a revenue elasticity stands in for the output elasticity, as it must when firms’ physical output is unobserved, profit maximization forces the ratio estimator to equal exactly one, regardless of the true markup (Bond et al., 2021). The revenue-based series carries zero markup information without output prices or demand structure; the omitted-price problem is not a bias, it is fatal. And Syverson ran the internal-consistency check: a markup of 1.61 with fixed scale elasticity implies a 2015 profit share around 25 percent of revenue, roughly half of value added, more than all capital income combined; allow DLEU’s own estimated rise in scale elasticity and the identity still demands 1.33 where the data deliver 1.14 (Syverson, 2019). His verdict, “not yet dispositive,” is the fairest one-line summary the dispute has.

Berry, Gaynor, and Scott Morton delivered the disciplinary judgment, and it is Era IV speaking in its own defense: there is no well-defined causal effect of concentration on price (Berry et al., 2019). A merger, a Demsetz cost fall, and a superstar reallocation all produce the same concentration-markup correlation with opposite welfare implications; a markup can explode because marginal cost fell toward zero. Concentration is an outcome, not a treatment. And the aggregate premise drew a named objection of its own: national census concentration is not a market, and the national series can rise while competition in local markets intensifies, the diagnosis associated with Rossi-Hansberg, Sarte, and Trachter and lodged against the premise the power and efficiency readings share (Rossi-Hansberg et al., 2021). The point is diagnostic rather than a single settled magnitude: national concentration can rise while the local-market interpretation remains unsettled. Industry-level verdicts, where the one-market-at-a-time toolkit actually applies, differ: hospitals show harm, airlines are mixed, wholesaling is ambiguous. The scoreboard, honestly kept: the critics hold the technical points, on accounting, on identification, on internal consistency, on the causal object; DLEU hold the agenda, because the number reset what the field argues about, and “markups probably rose, magnitude contested, causal meaning unproven” is now the consensus resting place. It is Bain against Demsetz with GMM, and, as in 1973, the war-of-ideas winner and the technical winner are not the same side.

Movement 2: the platforms break the toolkit

The theory that governs the era’s hardest cases was built before the cases existed, for credit cards. Rochet and Tirole showed that a platform serving two interdependent groups sets a total price whose level obeys a Lerner formula in the sum of the two sides’ elasticities, while the allocation across sides follows the ratio of elasticities: the side that is more elastic, or that confers more benefit on the other side, is charged less, possibly nothing, possibly less than nothing (Rochet and Tirole, 2003). Armstrong’s version makes the logic legible in one expression: the equilibrium price to one group equals cost plus the differentiation parameter minus the externality that group exerts on the other side (Armstrong, 2006). The doctrinal consequence is stated in the theory itself: a below-cost price on one side of a platform is not evidence of predation or cross-subsidy; it is the efficient way to internalize a cross-side externality, and one-sided price-cost tests are not merely lenient but conceptually the wrong instrument. Add the limiting case, consumer-facing prices of exactly zero, and the consumer-welfare standard loses its principal gauge: there is no price to raise, so the harm, if there is harm, must be located in quality, attention, data, or foreclosed entry, none of which the machinery was built to price.

Khan’s 2017 article drove into exactly that gap, and its anchor is a number Brooke Group cannot parse (Khan, 2017). Amazon priced bestseller e-books at $9.99, below the wholesale cost it paid publishers, sold the Kindle below manufacturing cost, and held roughly 90 percent of e-book retail by 2009; the Justice Department, in the Apple litigation, read the conduct as benign loss-leading. Under the recoupment test that is the right answer: recoupment, if it comes, comes through Prime memberships, marketplace fees, and hardware, complementary lines the doctrine does not examine, rather than through later price increases on the same good. Khan’s conclusion is that the price-and-output metric goes blind precisely where platform economics predicts below-cost pricing, and that structural dominance itself should be treated as the injury. One can accept her diagnosis of the doctrinal gap, and the two-sided-market theory independently confirms it, without accepting the remedy; the field has mostly done exactly that. The dispute with Bork is thus rejoined at the foundations: his standard was built on the claim that low prices vindicate the firm charging them, and the platforms supply low prices as a strategy of position.

The enforcement revival that carried this movement into policy runs on a number that failed its audit, which makes it this era’s version of the field’s oldest habit. Kwoka’s synthesis of merger retrospectives reported consummated mergers raising prices 7.22 percent on average across 42 mergers, price increases in about 95 percent of mergers leaving six or fewer rivals, and remedies failing at their job, divestiture cases up 7.05 percent and conduct-remedy cases up 16.03 (Kwoka, 2015). The FTC’s economists then audited the scorecard: the averages are unweighted where inverse-variance weighting can flip the sign of a worked example from +8.3 to -1.99, no standard errors accompany them, the remedy conclusions rest on six divestiture and four conduct transactions, two-thirds of the 49 transactions come from airlines and publishing, only seven postdate 2000, and dropping a single miscoded study cuts the divestiture figure from 7.05 to 4.43 percent (Vita and Osinski, 2018). Vita and Osinski do not argue mergers are benign; they argue the data cannot bear Kwoka’s weight. Meanwhile Nocke and Whinston rederived what the screen should be: for unilateral effects the synergy required to keep consumers whole depends on the change in the HHI, not its level, a thousand points of level adding about 0.35 percentage points to the required synergy against 3.15 per hundred points of change, and in estimated beer-market demand a 3 percent synergy tolerates a change of only about 150 to 200 points (Nocke and Whinston, 2022). By that arithmetic the 2010 thresholds were too lax, and the HHI-level presumption, a fixture since Baxter imported Stigler’s collusion logic, has no theoretical grounding in the unilateral-effects cases it actually decides.

The 2023 Merger Guidelines enacted the revival, and the document narrates its own pendulum. A merger is now presumptively unlawful at a post-merger HHI above 1,800 with a change above 100, or at a share above 30 percent with a change above 100 (Merger Guidelines, 2023). Footnote 15 recites the history in the agencies’ own voice: the 1982 guidelines set 1,000 and 1,800 with a 100-point trigger, every revision kept them through 2010, the 2010 guidelines raised them to 1,500 and 2,500 with triggers of 100 and 200, and the agencies “have therefore returned” to the original thresholds. Footnote 16 re-anchors the 30 percent screen on Philadelphia National Bank (1963) and its “undue percentage share” language. Baxter’s machinery, the HHI and the hypothetical monopolist, is retained; the tolerance is reset to pre-Baxter levels, Warren-Court case law quoted approvingly. The pattern deserves to be stated plainly, because it is the third time this history has recorded it: the guideline enacts the war-of-ideas winner, Kwoka’s framing and the neo-Brandeisian energy behind it, while the technical winners, Vita-Osinski on the statistics and Nocke-Whinston on the level-versus-change logic, are reflected only partially, in the direction of stricter change-based screening. Policy consumes framings, not scoreboards.

The live case is United States v. Google (2024), and it shows both how far the revival has gotten and how carefully the courts are still walking. Judge Mehta found Google holding 89.2 percent of general search queries, 94.9 percent on mobile, with Bing near 6; more than $26 billion paid in 2021 for default placement, nearly four times all of Google’s other search-specific costs combined; and roughly half of all U.S. queries flowing through access points locked by exclusive defaults, about 28 percent through Apple and 19.4 through Android (Google Search, 2024). The holding is that Google is a monopolist that acted to maintain its monopoly, in violation of Section 2, under the Microsoft exclusive-dealing framework. The opinion’s care is the point: a high share earned by a superior product is legal, and the 89.2 percent becomes actionable only because the defaults deny rivals the query scale that feeds search quality, scale foreclosure rather than pricing, litigated that way because the pricing toolkit has nothing to say about a product whose price is zero. The remedies phase declined divestiture in favor of behavioral relief, which makes the case a live experiment on the structural-versus-conduct-remedy question Kwoka’s 7.05-versus-16.03 split raised, with the field’s audit of that split unresolved beneath it.

Two older disputes reopened in the guidelines’ shadow, each with its number. Vertical foreclosure had been dormant since Brown Shoe condemned a 1.6 percent national foreclosure on a trend; the Chicago single-monopoly-profit logic then held the field for four decades, embodied in Vertical Merger Guidelines issued in 1984 and never updated, and vindicated in court when the government lost AT&T/Time Warner in 2018, its first vertical-merger litigation in roughly forty years (United States v. AT&T, 2018). Khan’s platform-integration argument and Google’s exclusive-dealing loss have reopened the question without resolving it: a dominant platform that owns the rails its rivals ride is Brown Shoe’s trend logic returned at scale, or it is efficient integration, and the doctrine currently cannot say which. And the killer-acquisition evidence put a magnitude on a harm the reporting thresholds were never built to see: about 6.4 percent of pharmaceutical acquisitions discontinue the target’s overlapping drug projects, deals typically too small for Hart-Scott-Rodino review (Cunningham et al., 2021). The figure is bounded, one industry, a minority of deals, and most small-target acquisitions finance startup exit rather than kill innovation; it fed the 2023 Guidelines’ serial-acquisition framework rather than any general presumption (Merger Guidelines, 2023).

One paragraph on common ownership, a skirmish this history notes because it reached the guidelines, though the battle lives mostly in the finance literature. Azar, Schmalz, and Tecu computed that diversified investors holding all the major airlines push ownership-adjusted concentration, the MHHI delta, to an average around 2,044 by 2014, ten times the merger guidelines’ screening threshold, and estimated fares 3 to 7 percent higher because of it, 10 to 12 percent in a BlackRock-BGI natural experiment (Azar et al., 2018). The rebuttals were forensic: the price correlation comes entirely from the market-share component of the measure rather than the ownership component, placebo tests keep the effect with real shares and random ownership and lose it with the reverse, the natural experiment shows no positive causal effect once instrumented, and recoding the equity of bankrupt carriers, at least one major airline being in bankruptcy in half the sample quarters, attenuates the result to insignificance (Dennis et al., 2022). Economy-wide, point estimates are near zero with bounds tight enough to reject even modest effects (Koch et al., 2021). The theoretical possibility survived into the 2023 Guidelines; the empirical case lost its status as a settled fact. Readers who want the ongoing exchange should look to the finance literature, where it continues.

The question the era must not resolve

Does the economy-wide evidence vindicate Bain, or is the field about to relearn Demsetz’s lesson at scale? On one side: the DLEU series, the investment gap, the retrospective price effects, the Amazon and Google records, all reading rising concentration as rising power. On the other: the SG&A correction, the revenue-elasticity identity, the profit-share arithmetic, the superstar reallocation, the ill-posedness verdict, all reading the same numbers as artifact, efficiency, or category error. These are the same two readings of the same coefficient that split the field in 1951 and 1973, now with production-function econometrics and firm-level panels on both sides, and the honest report is that the dispute is live and undecided. What is new, and what the next era inherits along with the question, is the state of the instruments: the platforms have broken the price-cost test, the one-sided market definition, and the recoupment doctrine at exactly the moment the question those tools were built to answer has returned. The field reopened the concentration question as it lost the means to close it.

Epilogue: The Pendulum and the Coefficient

Run the sequence once more, at speed. No answer, because no number (1890-1938). Everything: a four-point profit gap opening at an eight-firm share of 70 percent, read as oligopoly pricing made visible (1951). Nothing: the same gap, disaggregated by firm size, read as efficiency’s signature, because only the large firms earn it (1973). It depends, provably: any conduct is some game’s equilibrium, so the coefficient is a shadow cast by unmodeled institutions (1975-1991). One market at a time: the coefficient is ill-posed, and market power is $3,753 on a 1990 car and 42 percent on a box of cereal (1981-2007). Maybe everything again: a sales-weighted markup rising from 1.21 to 1.61, the founding coefficient reassembled from firm-level production functions (2020). Each swing believed itself final. None was.

The observation I want to leave with is not that the field changed its mind five times. Fields do that; it is called learning. The observation is that after eighty years, the contested object is still the same regression. Bain’s profit-on-concentration cross-section, Demsetz’s size-disaggregated version of it, Peltzman’s decomposition of it, Tirole’s equation 5.15 microfounding it, BLP’s refusal of it, and the DLEU markup series reviving it are, at bottom, one object: a measured association between how concentrated economic activity is and how profitable it is, awaiting a causal reading. The field never resolved the regression. It changed, era by era, the identification technology brought to bear on it, and with the technology, the standard of proof the same question had to meet. Bain’s grouped cross-section was refuted by within-industry disaggregation; the disaggregation was declared unidentified by equilibrium theory; the theory was declared unfalsifiable by its own critics and rescued as bounds; the bounds gave way to demand systems that identified conduct market by market; the demand systems were audited by quasi-experimental retrospectives; and the production-function estimates that reopened the aggregate question were promptly attacked with an identification argument, the revenue elasticity forcing the estimand to one, that is recognizably the same genre of objection Demsetz raised in 1973, stated now in the language of estimators rather than public goods. The question held still. The proof moved underneath it.

This is also why the technical winner and the war-of-ideas winner diverge so regularly in these pages, five times by my count. Demsetz won the field without a test that could distinguish his story from Bain’s. McGee’s skepticism became law a decade after game theory refuted it. Fisher won the falsifiability point and Sutton converted the loss into a method. Angrist and Pischke won the audit and Nevo and Whinston won the framing. Kwoka’s 7.22 percent reached the guidelines after the audit had cut its headline remedy number by more than a third. A field whose central regression is never decisively identified will be governed by framings, because framings are what remain when scoreboards are contested. That is not a corruption of the science. It is what the science looks like when the standard of proof is itself in motion.

The law, meanwhile, keeps its own time. Every era’s answer was enacted, with numbers: Bain’s logic at 5 and 7.5 percent shares in Brown Shoe and Von’s Grocery, enacted below the 70 percent threshold that generated it, as the academy began revoking it; Demsetz’s reading in Baxter’s 1,000 and 1,800, published in the same year the reputation papers rationalized the predation his school had declared irrational; McGee’s conclusion in Brooke Group’s recoupment test, 1993, eleven years after its refutation; Era IV’s method in merger simulation, adopted, audited, and demoted to screening; and the revival in the 2023 return to 1,800 and 100, with Philadelphia National Bank quoted as if the intervening sixty years were a detour. The lag is structural, and I have argued it has a mechanism: courts convert economics into administrable rules, administrability favors the confident vintage of a theory over the qualified current one, and by the time a doctrine is codified, the journals have usually moved. Anyone who wants to know what antitrust law will believe in 2035 should read what the field stopped believing around 2025. On current form, that would be the strongest versions of both revivals: the aggregate markup series as a measure of power, and the precise zero on common ownership as a measure of its absence.

Where does the pendulum hang now? Not at either pole. The reallocation facts are canonical; the rising-markup fact survives in a softened form, direction probable, magnitude contested; the causal reading of both is exactly where Bain and Demsetz left it, two structural stories inside one coefficient, with the identification honors currently held by the skeptics and the agenda held by the revivalists. The enforcement laboratory is running the era’s experiments live: guidelines reset to 1968-adjacent tolerance, a Section 2 platform verdict resting on scale foreclosure because price foreclosure has no meaning at a zero price, remedies chosen behavioral over structural with the remedy evidence itself under audit. And the instruments are, for the first time since 1951, visibly inadequate to the question being asked of them. The cross-section was reopened by economists who trusted production functions over demand systems; it is being contested by economists who trust neither without output prices; and the industries at the center of the policy argument sell at prices of zero, where every tool this history has described, from Bain’s profit rates to BLP’s inverted first-order conditions, loses its grip at once.

So the eighty-year argument stands open, which is the only honest place to leave it. The field’s founding regression is still on the table, better measured and no better identified than the day Bain published it. What the eighty years produced is not an answer but a discipline: a precise inventory of what the coefficient cannot say, bought one demolition at a time. Whether the current swing ends in a vindicated Bain, a re-vindicated Demsetz, or a third position no one has stated yet, the one prediction this history licenses is that the answer will be enacted before it is settled, and revoked in the journals before it is revoked in the courts. The pendulum is moving toward “everything” again. The coefficient, as always, is waiting.