Skip to main content
Free

From Good to Bad Concentration? US Industries over the Past 30 Years

New York UniversityNew York UniversityNew York University, CEPR, and NBER

Abstract

We study the evolution of profits, investment, and market shares in US industries over the past 40 years. During the 1990s, and at low levels of initial concentration, we find evidence of efficient concentration driven by tougher price competition, intangible investment, and increasing productivity of leaders. After 2000, however, the evidence suggests inefficient concentration, decreasing competition, and increasing barriers to entry as leaders become more entrenched and concentration is associated with lower investment, higher prices, and lower productivity growth.

We analyze the evolution of concentration in US industries over the past 40 years. Figure 1 summarizes the four stylized facts that motivate our work. Concentration and profits have increased, while the labor share and investment have decreased (fig. 1a–1d, respectively).1 This is true across most US industries as shown by Autor et al. (2017a; labor shares), Gutiérrez and Philippon (2016; investment and profits), and Grullon, Larkin, and Michaely (2019; concentration and profits). Although these stylized facts are well established, we are still far from consensus on what is causing them and what they tell us about the health of the US economy. The most prominent explanations can be organized in two groups:

•  Good concentration: The observed trends may be explained by good sources of concentration, such as increases in the elasticity of substitution (henceforth σ) or technological change leading to increasing returns to scale and intangible capital deepening (henceforth γ). Autor et al. (2017a, 180) argue for σ, noting that concentration reflects “a winner take most feature” explained by the fact that consumers have become more sensitive to price and quality due to greater product market competition. Haskel and Westlake (2017) argue for γ, emphasizing how scalability and synergies of intangible capital can lead to increasing returns to scale. Under σ and γ, concentration is good news: more productive firms expand yet competition remains stable or increases.

•  Bad concentration: Alternatively, the trends may reflect bad sources of concentration, which we summarize as rising barriers to competition (henceforth κ).2 Furman (2015, 12), for example, shows that “the distribution of returns to capital has grown increasingly skewed and the high returns increasingly persistent” and argues that it “potentially reflects the rising influence of economic rents and barriers to competition.”3 According to κ, concentration is bad news: it increases economic rents and decreases innovation.

The goal of this paper is to differentiate between these explanations at the aggregate and industry level. Before discussing our approach and results, however, it is important to clarify three points. First, these hypotheses are not mutually exclusive. Leaders can become more efficient and more entrenched at the same time—which can explain their growth but also the rise of barriers to entry (Crouzet and Eberly 2018). Indeed, a combination of these explanations is often heard in the discussion of internet giants Google, Amazon, Facebook, or Apple.
Fig. 1. 
Fig. 1. 

Evolution of US concentration, profits, labor shares, and investment. (a) Cumulative change in eight-firm concentration ratio (CR8; in %), based on the cumulated sales-weighted average change in CR8. Data from the US Economic Census based on Standard Industrial Classification-4 codes before 1992 and North American Industry Classification System-6 codes after 1997. When multiple tax groups are reported, only taxable firms are included. CR8 equals the market share (by sales) of the eight largest firms in each industry. We include only those industries that are consistently defined over each 5-year period. Change from 1992 to 1997 imputed from Autor et al. (2017b). (b) Profits/value added (VA), (c) labor share, and (d) net investment to net operating surplus, based on quarterly data for the nonfinancial corporate sector from the Financial Accounts of the US, via Federal Reserve Economic Data. Profit rate defined as the ratio of after-tax corporate profits with inventory valuation adjustment and capital consumption adjustment to VA (series W328RC1A027NBEA and NCBGVAA027S, respectively). Labor share defined as the ratio of compensation of employees (NCBCEPQ027S) to gross VA (NCBGVAQ027S). Net investment to net operating surplus defined as the ratio of net investment (gross fixed capital formation minus consumption of fixed capital, series NCBGFCA027N minus NCBCFCA027N) to net operating surplus (series NCBOSNQ027S). Dotted lines show the average of the corresponding series before and after 2002.

Second, intangibles can play a role in all theories. They may increase the elasticity of substitution (e.g., through online price comparison), increase returns to scale (e.g., organizational capital), and also create barriers to entry (e.g., through patents and/or the compilation of Big Data).

Third, these specific patterns are unique to the US. Figure 2a shows that profits margins have increased in the US, but they have remained stable or decreased in Europe, Japan, and South Korea. Figure 2b shows that concentration has increased in the US but it has remained roughly stable in Europe and Asia.4 Last, figure 2c shows that the labor share has declined in the US, but it has remained stable in Europe since 2000.5 Assuming that all advanced economies use similar technologies, the uniqueness of US trends suggests that technology alone cannot explain the trends.

Fig. 2. 
Fig. 2. 

Profits, concentration, and labor shares across advanced economies. (a) Gross Operating Surplus over Production (GOS/PROD) for nonagriculture business sector excluding real estate, from Organization for Economic Co-operation and Development and Structural Analysis (OECD STAN) database. (b) Change in eight-firm concentration ratio (CR8; since 2000) for nonagriculture business sector excluding real estate, based on Compustat but adjusted for coverage using OECD STAN. CR8 for Japan and Korea (JPN + KOR) reported only since 2006 because Compustat coverage increases rapidly beforehand. (c) Change in labor share (since 2000) for market economy from EU capital (K), labor (L), energy (E), materials (M), and service (S; KLEMS). See appendix E (see https://www.nber.org/data-appendix/c14237/appendix.pdf) for details. EU28 = current European Union member states; EU15 = original European Union member states; NA = North America; CP = Compustat.

Approach

We begin by using a sequence of simple models to clarify the theories of good and bad concentration. We derive a broad set of predictions regarding the joint evolution of competition, concentration, productivity, prices, and investment under each theory. We then evaluate these predictions empirically, first at the aggregate level, then at the industry level. Although some of these predictions have been studied by the literature, we contribute new facts/results for each of them. We also clarify several measurement issues and, perhaps more important, we show how the combination of all the facts helps us differentiate good and bad concentration.

Aggregate results

Table 1 summarizes our aggregate results. It contrasts the theoretical prediction of theories of good and bad concentration against the observed evolution of each measure.6 Predictions in the right column are consistent with the data after 2000. Predictions in the middle column are not.

Table 1. 

Summary of Test Measures and Predictions

 DataTheories
“Good”“Bad”
(i) Exit rate+ (σ)
(ii) Corr(ΔCR, ΔTFP)+ to −+
 Corr(ΔCR, ΔP)− to ++
(iii) Aggregate investment rate+
 Leader investment rate+
(iv) Leader turnover+ (σ)/− (γ)

Note. CR = concentration ratio; TFP = total factor productivity.

View Table Image

According to theories of good concentration, the growth of large firms is an efficient response to technological change. Under σ, competition increases as consumers become more price elastic. More productive firms expand to capture a larger share of the market, while less productive firms either shrink or exit. Economic activity reallocates toward more productive firms, increasing industry-level productivity and decreasing prices. Under γ, technological change leads to increasing returns to scale. Large firms again respond by expanding, which increases concentration and productivity while decreasing prices. The productivity gap between small and large firms grows.

If the economy experiences good concentration, we should observe: (i) concentration driven in part by exit; (ii) concentration associated with higher productivity and lower prices; and (iii) stable or increasing investment rates relative to Tobin’s Q—particularly for leaders. If the increase is driven by σ, we should also find higher volatility of market shares as demand responds more strongly to cost shocks. If the increase is driven by γ, however, the prediction could flip: volatility of market shares could fall as leaders’ comparative advantages become (potentially) more persistent (e.g., Aghion et al. 2019).

We already know that σ and γ are important for certain industries during certain periods. For instance, they describe well the evolution of the retail industry from 1990 to 2005 (Basu et al. 2003; Blanchard 2003). The rise of superstores and e-commerce led to more price competition, higher concentration, higher productivity, and the exit of inefficient retailers (Hortacsu and Syverson 2015). The question is whether these theories explain the evolution of the economy as a whole over the past 30 years. We test these predictions in the data and find some support for them during the 1990s. During this period, concentration is correlated with rising productivity, falling prices, and high investment, particularly in intangibles. Since 2000, however, these predictions are rejected by the data. The correlation between concentration and productivity growth has become negative, while the correlation between concentration and price growth has become positive; exit rates have remained stable; investment relative to Q has fallen; and market shares have become more persistent. Estimates of returns to scale based on the methodology of Basu, Fernald, and Kimball (2006) have remained stable, as have other estimates in the recent literature (Ho and Ruzic 2018; Diez, Fan, and Villegas-Sanchez 2019). All these predictions are consistent with the κ theory.

Barriers to competition therefore emerge as the most relevant explanation over the past 15 years. It correctly predicts the evolution of profits, entry, exit, turnover, prices, productivity, and investment in most industries.

Industry results

Aggregate trends are interesting, but the dynamics of individual industries are more informative: σ and γ cannot explain the broad trends but they probably matter for some industries. To obtain a systematic classification of industry-level changes, we perform a principal components analysis (PCA) on a wide range of measures related to competition. We find that the first principal component, PC1, captures the σ and γ theories of good concentration while the second principal component, PC2, captures theories of bad concentration. This distinction is quite stark and allows us to show which industries have experienced good versus bad concentration and compare the importance of each theory over time.

Durable computer manufacturing exhibits the highest loading on PC1. It exhibits high intangible capital intensity but remains relatively competitive, likely as a result of intense foreign competition. By contrast, telecommunications, banking, and airlines are predominantly explained by κ, consistent with the results of Gutiérrez and Philippon (2018). They exhibit high concentration, high profits, and low productivity growth. Interestingly, some industries, such as nondurable chemical manufacturing and information (data), load heavily on both PC1 and PC2. These industries hold large amounts of intangible assets but also exhibit high barriers to entry. They are good examples of intangible assets giving rise to barriers to entry, as emphasized by Crouzet and Eberly (2018). In fact, Crouzet and Eberly (2018) argue that the health-care sector, which includes nondurable chemical manufacturing, is one where market power derived from intangible assets is largest.

Looking at the evolution of loadings over time further emphasizes the transition from good to bad concentration. The average PC1 score (reflecting good concentration) was substantially higher than PC2 in 1997 and increased faster from 1997 to 2002. But PC2 caught up afterward and, by 2012, explained a larger portion of industry dynamics. Our results therefore indicate that the US economy has transitioned from good to bad concentration over the past 30 years.

Related literature

Our paper contributes to a growing literature studying trends in competition and concentration in the US economy. The literature began by (separately) documenting the stylized facts. Haltiwanger, Jarmin, and Miranda (2011, 2) find that “job creation and destruction both exhibit a downward trend over the past few decades.” Decker et al. (2015) argue that, whereas in the 1980s and 1990s declining dynamism was observed in selected sectors (notably retail), the decline was observed across all sectors in the 2000s, including the traditionally high-growth information technology (IT) sector. CEA (2016) and Grullon et al. (2019) document the broad increases in profits and concentration; Elsby, Hobijn, and Sahin (2013) and Karabarbounis and Neiman (2014) document the decline in the labor share; and IMF (2014), Hall (2015), and Fernald et al. (2017) discuss the decline in investment in the context of weak overall growth. Akcigit and Ates (2019) review some of the literature.

Over time, the literature began to connect these facts and propose theories of “good” and “bad” concentration (we use “good” and “bad” for didactic purposes). The most prominent explanations of good concentration include Autor et al. (2017a) and Van Reenen (2018, 1), who argue that rising concentration and declining labor shares are explained by an increase in σ, which results in “winner take most/all” competition, and Alexander and Eberly (2018) and Crouzet and Eberly (2018), who link the rise in concentration and the decline in investment to intangible capital. Bessen (2017) links IT use to industry concentration. Ganapati (2018) links concentration to increasing labor productivity and stable prices. Aghion et al. (2019) and Ridder (2019) develop models where information and communication technologies increase returns to scale, leading to higher concentration and lower labor shares.

Moving to bad concentration, Grullon et al. (2019) show that firms in concentrating industries exhibit higher profits, positive abnormal stock returns, and more profitable merger and acquisition deals. Barkai (2017) documents a rise in economic profits and links it to concentration and labor shares. De Loecker, Eeckhout, and Unger (2019) argue that markups have increased. Gutiérrez and Philippon (2016) link the weakness of investment to rising concentration and market power, while Lee, Shin, and Stulz (2016) find that capital stopped flowing to high Q industries in the late 1990s. Eggertsson, Robbins, and Wold (2018) introduce time-varying market power to a standard neoclassical model to explain several of our stylized facts. Gutiérrez and Philippon (2018), Jones, Gutiérrez, and Philippon (2019), and Gutiérrez and Philippon (2019) argue that domestic competition has declined in many US industries because of increasing entry costs, lax antitrust enforcement, and lobbying.

We would like to note that this debate between good and bad concentration has a direct precedent in the industrial organization literature of the 1970s and 1980s. By then, the discussion was centered on how to interpret the positive correlation between profits and concentration at the industry level, first documented by Bain (1951). While this fact was commonly rationalized as evidence of market power (“bad concentration”), Demsetz (1973) argued that the observed pattern was instead explained by differences in productivity (“good concentration”). This seminal contribution spawned a series of empirical papers evaluating these two hypotheses, reviewed in Schmalensee (1987).

Finally, our paper is also related to the effect of foreign competition, particularly from China (see Bernard et al. 2012 for a review). Bernard, Jensen, and Schott (2006) show that capital-intensive plants and industries are more likely to survive and grow in the wake of import competition. Bloom, Draca, and Van Reenen (2016) argue that Chinese import competition leads to increased technical change within firms and a reallocation of employment toward more technologically advanced firms. Frésard and Valta (2015) find that tariff reductions lead to declines in investment in markets with competition in strategic substitutes and low costs of entry. Within industry, they find that investment declines primarily at financially constrained firms. The decline in investment is negligible for financially stable firms and firms in markets featuring competition in strategic complements. Hombert and Matray (2015) show that research and development (R&D)–intensive firms were better able to cope with Chinese competition than low-R&D firms. They explain this result based on product differentiation, using the Hoberg and Phillips (2017) product similarity index. Autor, Dorn, and Hanson (2013), Pierce and Schott (2016), Autor, Dorn, and Hanson (2016), and Feenstra, Ma, and Xu (2017) study the effects of Chinese import exposure on US manufacturing employment. Feenstra and Weinstein (2017) estimate the impact of globalization on markups, and conclude that markups decreased in industries affected by foreign competition. Some of these papers find a reduction in investment for the “average” firm, which is consistent with our results and highlights the importance of considering industry leaders and laggards separately.

The remainder of this paper is organized as follows. Section I derives theoretical predictions. Section II discusses measurement issues related to common empirical proxies of competition. Section III tests aggregate predictions related to business dynamism, productivity, prices, investment, and returns to scale. Section IV replicates the exercise at the industry level, using PCA. Section V concludes.

I.  Theory

We use a few simple models to derive testable predictions for the various hypotheses. The timing of the models follows the classic model of Hopenhayn (1992): (i) there is a sunk entry cost κ; (ii) firms draw their productivities a (and/or idiosyncratic demand shocks); and (iii) they either produce with a fixed operating cost ϕ or they exit early.

A.  Good Concentration, Bad Concentration

Let us start with the simple case where there is no heterogeneity. Consider, then, an industry with N identical firms with productivity ai=A for all i[0,N], and industry demand Y. Suppose the game among the N firms leads to a markup μ over marginal cost. In other words, firms set the price

p=1+μA
and firm i’s profits are
πi=(p1A)yiϕ=μ1+μpyiϕ.
In a symmetric equilibrium with identical firms, all firms produce
yi=YN for all i[0,N].
So profits are
π=μ1+μpYNϕ.
Under free entry, we have
E[π]r+δκ,
where r is the discount rate, δ is the (exogenous) exit rate, and κ is the sunk entry cost. The free entry condition is then
Nμ1+μpY(r+δ)κ+ϕ.
A simple case is when industry demand is unit elastic (Cobb-Douglas). In that case Y(p)=Y¯/p and we have N(μ/(1+μ))(Y¯/((r+δ)κ+ϕ)). We then have the following proposition.
Proposition 1. 

In response to shocks to ex post markups μ, concentration is positively related to competition. In response to shocks to κ, concentration is negatively related to competition.

This proposition summarizes the fundamental issue with using concentration as a proxy for competition. Concentration is endogenous and can signal either increasing or decreasing degrees of competition. In other words, when looking at concentration measures, it is crucial to take a stand on why concentration is changing, in particular to see whether it is driven by shrinking margins or by higher barriers to entry.

Corollary 1. 

Concentration is a valid measure of market power only when concentration is driven by barriers to entry or by mergers.

Note that it is straightforward to extend the analysis to the case where μ depends on the number of firms. We can write μ/(1+μ)=l¯Nθ, where l¯ is the baseline Lerner index and θ is the elasticity of the markup to concentration. In a standard constant elasticity of substitution (CES) monopolistic competition model, for instance, we have θ=0 and l¯=1/σ. We can then write the free entry condition as N1+θ(lY¯/((r+δ)κ+ϕ)), which shows that our propositions are valid when markups vary with concentration.

B.  Selection and Ex Post Profits

Consider now the case of heterogenous marginal costs. Heterogeneity creates a selection effect and we need to distinguish between the number of firms that enter (N^) and the number of firms that actually produce (N). Formally, consider the following industry entry game:

•  Each entrant pays κ for the right to produce one variety i[0,N^].

•  After entry, each firm draws productivity ai and decides whether to produce with fixed operating cost ϕ and markup μi.

Let NN^ be the number of active producers. We reorder the varieties so that i[0,N] are active while i(N,N^] exit early. The demand system is given by the CES aggregator
Yσ1σ=0Nyiσ1σdi,
where σ>1 is the elasticity of substitution between different firms in the industry. This demand structure implies that there exists an industry price index P1σ0Npi1σdi such that the demand for variety i is
yi=Y(piP)σ.
The firm sets a price pi=(1+μi)/ai and the profits of firm i are now given by πi=(μi/(1+μi)σ)aiσ1PσYϕ. If we assume monopolistic competition, the optimal markup μm=1/(σ1) maximizes μi/(1+μi)σ. But we do not need to consider only this case. We could assume limit pricing at some markup μ<1/(σ1), strategic interactions among firms, and so on. For now we simply keep μ as a parameter.

Firms with productivity ai<a* do not produce, so the active producers are N=(1F(a*))N^, where N^ is the number of firms that pay the entry cost. Similarly, the density of producers’ productivity is dF*(a)=dF(a)/(1F(a*)). Because all the firms draw from the same distribution of productivity, we have

P=1+μA*N1σ1,
where average productivity is
(1)A*(aσ1dF*(a))1σ1.
Equilibrium profits are then
π(ai;a*,PY,N)=μ1+μ(aiA*)σ1PYNϕ.
There is a cutoff a* such that only firms above the cutoff are active producers
π(a*;a*,PY,N)=0.
The productivity cutoff a* solves (μ/(1+μ))(a*)σ1(PY/N)=ϕ(A*)σ1. For simplicity we consider again the log-industry demand case, so PY is exogenous and equal to Y¯. Using the definition of A* in equation (1), and N=(1F(a*))N^ and dF*(a)=dF(a)/(1F(a*)), we find that
μ1+μY¯ =ϕN^a>a*(aa*)σ1dF(a).
The right-hand side is increasing in σ and decreasing in a*, so we have the standard selection effect.
Lemma 1. 

The cutoff a* increases with the demand elasticity σ.

From the free entry condition we have
(r+δ)κ=(1F(a*))×E[π|a>a*].
Because 1F(a*) decreases with σ, it follows that E[π|a>a*] must increase with σ for a given κ.

Proposition 2. 

For a given free entry condition, an increase in σ leads to higher rate of failed entry (early exits) and higher profits for remaining firms (selection effect). An increase in κ, on the other hand, leads to lower entry, lower exit, and higher profits.

This proposition allows us to distinguish the σ hypothesis from the κ hypothesis.

C.  Increasing Returns

Now suppose that firms can choose between two technologies after entry: low fixed cost and low productivity (AL, ϕL) or high fixed cost and high productivity (AH, ϕH). Let us ignore idiosyncratic productivity differences for now. Profits are then

π(a,ϕ)=μ1+μ(aA)σ1PYNϕ.
The choice of technology clearly depends on the size of the market and the elasticity of demand.
Lemma 2. 

Firms are more likely to switch to the high returns to scale technology when σ is high.

Assume that the parameters are such that the firms decide to switch to ai=AH for all i. Equilibrium profits are then π=(μ/(1+μ))(PY/N)ϕH. Free entry then requires π=(r+δ)κ

N=μ1+μPYϕH+(r+δ)κ.
Concentration increases when firms switch to the high returns to scale technology. The behavior of equilibrium profits depends on the selection effect. Without idiosyncratic risk, profits are simply pinned down by free entry. If we take into account idiosyncratic risk, then equilibrium profits increase when firms switch to the high returns to scale technology because the selection effect intensifies.
Proposition 3. 

A switch to increasing returns technology is more likely when demand is more elastic. A higher degree of increasing returns to scale leads to more concentration, higher profits, and higher productivity for the remaining firms.

This proposition connects σ and γ, as often discussed in the literature. Note that we can measure the degree of returns to scale γ as the ratio of average cost (ϕ/y)+(1/A) to marginal cost 1/A:
γ1ϕAy=ϕϕ+(r+δ)κμN1σ1,
which is increasing with ϕ because N is decreasing in ϕ. Therefore, if we were to measure γ under the old and the new technologies, we would indeed find γH>γL.

D.  Dynamics of Market Shares

Consider finally the case where, after entry, firms are subject to demand and productivity shocks. In the general case, we have j[0,1] industries and i[0,Nj] firms in each industry. The output of industry j is aggregated as Yj,t((σj1)/σj)=0Njhi,j,t1/σ(yi,j,t)((σj1)/σj)di, where σj is the elasticity between different firms in the same industry and hi,j,t are firm-level demand shocks. The demand for good (i, j) is given by

yi,j,t=hi,j,tYj,t(pi,j,tPj,t)σj,
where Pj,t is the industry price index. The nominal revenues of firm i are
pi,j,tyi,j,t=pi,j,t1σjhi,j,tPj,tσjYj,t
and the market share of firm i in industry j is
si,j,t=pi,j,tyi,j,tPj,tYj,t=hi,j,tNj((1+μj)ai,j,t(1+μi,j)Aj,t)σj1,
where μj is the industry average markup and Aj,t is the industry average productivity, as defined earlier. If we track the market shares of firms over time, we have the following proposition.
Proposition 4. 

The volatility of log-market shares is

logs2=logh2+(σj1)2loga2,
where loga2 is the volatility of idiosyncratic productivity shocks.

Therefore,
Corollary 2. 

All else equal, an increase in σj leads to an increase in the volatility of market shares in industry j.

In summary, we have established that an increase in σ leads to an increase in concentration, productivity, exit, the volatility of market shares, and investment. Similarly, an increase in γ results in more concentration, higher profits, and higher productivity for surviving firms.7 Finally, an increase in κ leads to an increase in concentration and a decrease in productivity, exit rates, market share volatility, and investment (relative to Q).

II.  Measurement Issues

Before testing our predictions, we discuss two important issues related to the measurement of concentration and markups.

A.  Foreign Competition and Concentration

First, when computing industry concentration, it is important to control for imports. We compute import-adjusted concentration measures (CR8IA) and use them throughout the paper. Figure 3 shows the importance of the correction, focusing on manufacturing industries that are highly exposed to foreign competition. While domestic concentration increased by 6.7 percentage points in these industries, import-adjusted concentration (dotted line) increased by only 1.6 points.8 Foreign competition, therefore, plays an important role in manufacturing. But import-exposed industries only account for about 10% of the private economy, so foreign competition cannot explain the aggregate trends that we have presented earlier.

Fig. 3. 
Fig. 3. 

Domestic versus import-adjusted concentration for high import manufacturing industries. Weighted average absolute change in domestic (solid line) and import-adjusted (dotted line) eight-firm concentration ratio (CR8) across North American Industry Classification System (NAICS)-6 manufacturing industries in the top three quantiles of import shares as of 2012. Imports accounted for 29% of sales plus imports in these industries, on average. Domestic concentration from US Economic Census. Import adjusted concentration defined as CR8jtIA=CR8jt×(salejt/(salejt+impjt))=CR8jt×US Sharejt. NAICS-6 industries are included if they are consistently defined from 1997 to the given year. See appendix (see https://www.nber.org/data-appendix/c14237/appendix.pdf) for details.

B.  Markup Measurement

The second issue relates to measurement of markups. De Loecker et al. (2019) estimate markups using the methodology of De Loecker and Warzynski (2012). The idea is to compare the elasticity of output to a variable input, with the cost share of that input. De Loecker et al. (2019) implement this methodology using cost of goods sold (COGS) as their main measure of variable input. While this approach is promising in theory, the question is whether it provides a reliable measure of market power. There are measurement issues with COGS that we discuss in appendix A (see https://www.nber.org/data-appendix/c14237/appendix.pdf). Our main concern, however, is that technology can change over time in a way that creates challenges for COGS-based markup measures.

Identification: The China Shock

We use the China shock to illustrate this issue, following Autor et al. (2016) and Pierce and Schott (2016). Chinese competition led to a strong replacement effect. Figure 4 shows the normalized number of firms in industries with high and low Chinese import penetration.9 Both groups have the same preexisting trends, including during the dot-com boom, but start to diverge after 2000. In unreported tests, we confirm that this relationship is strongly statistically significant.

Fig. 4. 
Fig. 4. 

Number of firms by Chinese exposure. Annual data. Number of firms from Compustat; import penetration based on data from NBER Center for Economic Studies and Peter Schott. Manufacturing industries only, split into “high” (above-median) and “low” (below-median) import exposure (IE) based on import penetration from 1991 to 2015. See appendix E (see https://www.nber.org/data-appendix/c14237/appendix.pdf) for details.

Realized imports are endogenous so, in the rest of the section, we use the instrument proposed by Pierce and Schott (2016). The instrument exploits changes in barriers to trade following the US granting permanent normal trade relations (PNTR) to China.10 Pierce and Schott (2016) show that industries facing larger NTR gaps experienced a larger increase in Chinese imports and a larger decrease in US employment. We follow Pierce and Schott (2016) and quantify the impact of granting PNTR on industry j as the difference between the non-NTR rate (to which tariffs would have risen if annual renewal had failed) and the NTR rate as of 1999:

NTR Gapj=Non NTR RatejNTR Ratej.
This measure is plausibly exogenous to industry demand and technology after 2001. The vast majority of the variation in NTR gaps is due to variation in non-NTR rates set 70 years prior to passage of PNTR. See Pierce and Schott (2016) for additional discussion.

Profits versus Markups

Figure 5 reports results of the following regressions across firms i in industry j

(2)πi,j,t=y=19912007βt×NTR Gapj+δi+γt+εi,j,t,
where πijt denotes a given outcome variable (profits, etc.). All regressions include firm and year fixed effects and are weighted by firm sales. Standard errors are clustered at the North American Industry Classification System (NAICS)-6 industry level. Consistent with the identification assumption, we see no significant pre-trends before 2000 and strong responses afterward. Consistent with the increase in exits, the operating income of US companies falls upon Chinese accession to the World Trade Organization (WTO; fig. 5a).
Fig. 5. 
Fig. 5. 

Profits, SG&A intensity, and markups around China shock. (a) log(OIADP). (b) log(SG&A/COSTS). (c) SALE/COGS. Firm financials from Compustat. Normal trade relations gap from Pierce and Schott (2016). Figure reports regression results following equation (2), including 95% confidence intervals. Only firms that existed before 1997 are included. SALE/COGS and XSGA/XOPR (SG&A/total operating expenses) are winsorized at the 2% and 98% level, by year. See text for details. SG&A = sales, general, and administrative expenses; OIADP = operating income after depreciation; COSTS = total costs; SALE = sales; COGS = cost of goods sold.

What is more remarkable, however, is the increase in the share of sales, general, and administrative expenses (SG&A) in total costs. SG&A is the second major component of costs and includes all intangible-building activities (e.g., R&D, advertising, and IT staff expenses). US firms react to the increased competition by almost doubling their SG&A intensity (fig. 5b), a result consistent with the shift toward intangible capital documented in table 4, as well as the increased product differentiation documented by Feenstra and Weinstein (2017). The increase in SG&A is precisely the type of technological change that may affect the validity of COGS-based markups. Indeed, figure 5c shows that SALE/COGS (ratio of sales to COGS) appears to increase rather than decrease upon the shock.11 COGS-based markup measures would fail to classify the China shock as an increase in competition, while exit and profit margins do.12

We can also get a broad evaluation of the usefulness of markups by studying the evolution across regions. Figure 6 plots the sales-weighted average ratio of sales to COGS against gross profit rates by region.13 The shift toward intangible expenditures is clearly present across all advanced economies: SALE/COGS rises everywhere as the cost-share of COGS falls. This may suggest a global rise in market power, but profits show us the opposite—especially for the European Union 15 and the United Kingdom. Only in the US do we observe a large increase in profits. In the remaining regions, the decline in COGS is fully offset by a rise in SG&A so that profits remain flat (operating income before depreciation equals sales minus COGS and SG&A). Given the inability of markup estimates to control for technology, we focus on profits and market share dynamics in the rest of the paper.14

Fig. 6. 
Fig. 6. 

Weighted average SALE/COGS versus gross profit rates by region (1995 = 1). (a) USA; (b) EU15; (c) GBR; (d) JPN. SALE/COGS equals the sales-weighted average ratio of sales to cost of goods sold across all Compustat firms in a given region. GOS/PROD based on Organization for Economic Cooperation and Development Structural Analysis database for nonagriculture business sector excluding real estate. EU15 = original European Union member states; GBR = Great Britain; JPN = Japan; SALE = sales; COGS = cost of goods sold.

III.  Aggregate Evidence

A.  Entry, Exit, and Turnover

Having clarified some measurement issues, let us return to the main goal of the paper: differentiating theories of good versus bad concentration. We begin with market share turnover. Industrial organization economists rightly complain about the use of Herfindahl-Hirschman indexes (HHIs) or concentration ratios (CRs) at the broad industry × country level as measures of market power. The limitations of national CRs and HHIs are well understood. NAICS industries and countries are much broader than product markets—and concentration may evolve differently at more granular levels.15 But there is a more fundamental problem: depending on the nature of competition, technology, and supply and demand primitives, concentration may be positively or negatively correlated with competition and markups. In other words, concentration “is a market outcome, not a market primitive” (Syverson 2019, 4).

Leader Turnover

To obtain an alternate measure of market power, we consider turnover of market shares and market leadership. In particular, one can ask: Given that a firm is at the top of its industry now (top 4, top 10% of market value), how likely is it that it will drop out over the next 5 years? Per proposition 4, increases in σ would result in higher leader turnover, while increases in κ would result in lower turnover.

Figure 7 tests this prediction. We define turnover in industry j at time t as the probability of leaving the top 4 firms of the industry over a 5-year period,

TopTurnjt=Pr(zi,j,t+5<zj,t+5#4|zi,j,tzj,t#4),
where zi,j,t denotes either the sales of firm i at time t or its market value of equity, and zj,t#4 is the value of zi,j,t for the fourth largest firm at time t in industry j.16 We then average turnover across all industries in a given year. We focus on the post-1980 period, after the addition of Nasdaq into Compustat. As shown, the likelihood of a leader being replaced was 35% in the 1980s, rose to 40% at the height of the dot-com bubble, and is only 25% today. Appendix A (see https://www.nber.org/data-appendix/c14237/appendix.pdf) presents results by sector.
Fig. 7. 
Fig. 7. 

Turnover of leaders by sale and market value (MV; Compustat North America, following Bureau of Economic Analysis industries). Only industry years with five or more firms are included. See text for details.

Persistence of Market Shares

Leader turnover focuses on the right tail of the distribution. Let us now broaden the sample to include all firms and study the persistence of market shares. We follow proposition 4 and estimate an AR(1) model of the log-market share for firm i that belongs to Standard Industrial Classification (SIC)-3 industry j, using a 5-year rolling window:

logsi,j,t=ρj,tlogsi,j,t1+ϵi,j,t.
Figures 8a and 8b plot the sales-weighted average ρj,t and root mean squared error (RMSE), respectively. In line with the decline in turnover, the persistence of market shares increases after 2000, whereas the RMSE falls.17
Fig. 8. 
Fig. 8. 

Persistence and volatility of market shares (MS). (a) MS persistence. (b) root mean squared error (RMSE). Autocorrelation and RMSE for AR(1) model of firm-level log-market shares, following Standard Industrial Classification-3 industries. Estimates based on a 5-year rolling window. Only industry years with five or more firms and firms with a market share higher than 0.02 are included.

Leaders clearly have less to worry about today than 30 years ago. Their market shares and leadership positions are far more persistent today than even 15 years ago. Why might this be? In Gutiérrez and Philippon (2019), we study competitive pressures directly, focusing on the entry and exit margins. We show that exit rates have remained stable, while the elasticity of entry with respect to Tobin’s Q was positive and significant until the late 1990s but fell close to zero afterward. The behavior of entry, exit, and turnover is inconsistent with σ, but consistent with κ.

B.  Concentration, Productivity, and Prices

According to σ and γ, concentration rises as high productivity leaders expand, increasing industry-level productivity and decreasing prices. If more productive firms have lower labor shares, the aggregate labor share also falls. Autor et al. (2017b) document a reallocation from high- to low-labor-share establishments, while Ganapati (2018) finds that changes in concentration are uncorrelated with changes in prices but positively correlated with changes in productivity. Kehrig and Vincent (2017) and Hsieh and Rossi-Hansberg (2019) make similar arguments for manufacturing and service industries, respectively.

BLS and Compustat

We begin our analysis with relatively aggregated data from the Bureau of Labor Statistics (BLS) multifactor productivity (MFP) Tables. This data set includes total factor productivity (TFP), prices, wages, and labor productivity. We complement it with Compustat-based concentration measures to obtain the same industry classification in left- and right-hand side variables. We assess the joint evolution of productivity, prices, and markups using regressions of the form

Δ5log(Zj,t)=βΔ5log(CR4j,t)+γt+εjt,
where Z is the variable of interest and Δ5 denotes a 5-year change. We consider TFP, prices, and markups of prices over unit labor costs (ULC): Δ5logμ=Δ5logPΔ5logULC, where Δ5log(ULC)Δ5log(W)Δ5log(LPt).

Table 2 summarizes the results. Columns 1, 3, and 5 are based on pre-2000 changes and exhibit correlations in line with σ and γ: positive and significant with TFP and negative (although insignificant) with prices and markups. However, the relationship seems to have collapsed after 2000. The correlation between concentration and TFP turns negative (though insignificant), while the correlation with prices and markups turns positive.

Table 2. 

Concentration, TFP, Prices, and Markups: BLS Industries

 Δ5 log(TFP)Δ5 log(P)Δ5 log(μ)
Pre-00Post-00Pre-00Post-00Pre-00Post-00
(1)(2)(3)(4)(5)(6)
Δ5 log(CR4IA).186*−.044−.093.077−.102*.116+
(.070)(.051)(.069)(.088)(.047)(.064)
Cons.016.025**.074**.097**.048**.045**
(.013)(.009)(.013)(.010)(.012)(.011)
Year FEYYYYYY
R2.12.1.048.07.041.082
Observations941419414194141

Note. Table shows the results of industry-level ordinary least squares regressions of contemporaneous 5-year changes in TFP, prices (P), markups (μ), and import-adjusted concentration over the periods specified. Data include all industries covered in the BLS multifactor tables. CR4 from Compustat. Standard errors in parentheses are clustered at industry level. TFP = total factor productivity; BLS = Bureau of Labor Statistics; CR = concentration ratio; FE = fixed effects.

+ p < .10.

*p < .05.

**p < .01.

View Table Image

To illustrate the transition, figure 9 plots the evolution of markups and concentration for the telecom and air transportation industries. While they exhibit little (or negative) correlation before 2000, both rise sharply afterward. This is consistent with the cross-country analyses of Gutiérrez and Philippon (2018).

Fig. 9. 
Fig. 9. 

Change in markup and concentration since 1991: (a) telecom and (b) airlines (Bureau of Labor Statistics multifactor tables for markups and Compustat for import-adjusted concentration). CR8 = change in eight-firm concentration ratio.

The BLS MFP tables provide several advantages. They cover the full economy, include TFP estimates, and follow a consistent segmentation that can be mapped to other Bureau of Economic Analysis (BEA) data sets. This allows us to include the evolution of prices, unit-labor costs, and markups in the PCA of Section IV. However, using broad industry definitions limits the power of our regressions, hence the previous large confidence intervals. Let us now bring in more granular data.

BEA, NBER, and Census

We roughly follow Ganapati (2018) and combine concentration data from the US Economic Census with price data from the NBER Center for Economic Studies (NBER-CES) database for the manufacturing sector and the BEA’s detailed gross domestic product (GDP) by industry accounts for nonmanufacturing.18 Combined, these data sets allow us to estimate real labor productivity and analyze the evolution of markups using the previous definitions.

We estimate regressions of the following form:

Δ5log(Zjt)=βΔ5log(CR4j,t)+γs,t+εjt,
where j denotes industries and t denotes years. γs,t denotes sector-year fixed effects. Table 3 reports results for prices and markups. Before 2002, the correlation is small and often insignificant, in line with the results of Ganapati (2018). After 2002, however, increases in concentration are systematically correlated with increases in prices. Columns 7–9 show a similar effect but instead of sorting on time (pre-/post-2002), we sort by ending levels of concentration. When ending concentration is low, there is not much correlation between changes in concentration and changes in markups. When concentration reaches a high level, however, the correlation is much stronger, especially in the nonmanufacturing sector. See appendix B (see https://www.nber.org/data-appendix/c14237/appendix.pdf) for additional results, including a decomposition of the correlation between concentration and markups into the underlying components: prices, wages, and labor productivity.
Table 3. 

Concentration versus Prices: Pre- and Post-2002

 Δ5 log(P)Δ5 log(μ)Δ5 log(μ)
AllMfgNonMfgAllMfgNonMfgAllMfgNonMfg
(1)(2)(3)(4)(5)(6)(7)(8)(9)
Δ5 log(CR4jt)−.01.05*−.03.02.10**−.00.12*.12**.12*
(.03)(.02)(.04)(.04)(.03)(.05)(.05)(.04)(.06)
Δ5 log(CR4jt) × 1>2002.17**.20**.17**.23**.13+.26**   
(.04)(.06)(.05)(.06)(.07)(.07)   
Δ5 log(CR4jt) × High CR      .18*.08.41**
      (.09)(.07)(.14)
High CR      .03*.07**−.01
      (.02)(.02)(.02)
Δ5 log(LPjt)−.39**−.37**−.39**      
(.05)(.07)(.08)      
Δ5 log(wjt).58**.74**.48**      
(.14)(.28)(.12)      
Cons.05**.06+.06**.05**.11**.03**.04**.06**.02+
(.02)(.03)(.02)(.01)(.02)(.01)(.01)(.01)(.01)
Sector × Year FEYYYYYYYYY
R2.47.45.5.39.34.36.38.38.35
Observations2,0831,6824012,0831,6824012,0831,682401

Note. Table shows the results of industry-level OLS regressions of contemporaneous 5-year changes in prices, markups, and concentration over the periods specified. P = prices; μ = markups; Mfg = manufacturing industries; NonMfg = nonmanufacturing industries; CR = concentration ratio; FE = fixed effects. Observations are weighted by sales. Standard errors in parentheses are clustered at industry level.

+ p < .10.

*p < .05.

**p < .01.

View Table Image

The joint evolution of concentration, TFP, and prices appears consistent with the σ and γ theories before 2000. Over the past 15 years, however, concentration is correlated with lower TFP and higher prices. The evidence is now more closely aligned with the κ theory.

Our data and correlations are consistent with the ones in Ganapati (2018) but our interpretation is quite different. Regarding prices, we agree that the full sample correlation is small, but as we have shown the correlations after 2000 and at high level of concentration are large and positive. The most important disagreement, however, relates to the correlation with productivity. The existing literature has failed to recognize that, given what we know about firm-level data, we should expect a quasi-mechanical correlation between concentration and productivity at the level of detailed industries (NAICS level 4 or 5, for instance). We know that the firm-size distribution is skewed. At NAICS level 5, the top four firms account for about one-third of output. We also know that firm-level shocks are large. Therefore, changes in industry output at level 5 are strongly affected by idiosyncratic firm-level shocks. If a large firm experiences a positive shock, industry output increases and concentration increases at the same time.

Therefore, in the regressions run by Ganapati (2018) or Autor et al. (2017b), one would expect a mechanical positive correlation between changes in CR4 and changes in output or productivity or both (depending on the details of the shocks). At level 4 the kurtosis of log changes in CR4 is 8.8. Once we move to level 2 or level 3, the law of large number kicks in and these effects are muted. At level 2, for instance, log changes in CR4 have a skewness of 0 and a kurtosis of 2.5. In other words, the changes are basically normal. This has nothing to do with synergies or with the value of concentration per se. It is just fat-tail econometrics. Ganapati (2018) claims that, because changes in concentration and changes in industry productivity are positively correlated on average, we need not worry about the (smaller) impact of concentration on prices.19 The earlier reasoning suggests that this claim is incorrect.

C.  Investment and Profits

Under σ and γ, the increase in concentration is driven by technological change linked to the rise of intangibles. In that case, aggregate investment would remain in line with Q, while intangible investment would increase. However, as shown in figure 10, the growth of the capital stock has fallen across all asset types since 2000, notably including intellectual property assets. Moreover, the decline in investment is not explained by Tobin’s Q, as shown by figure A24 (see https://www.nber.org/data-appendix/c14237/appendix.pdf). In fact, investment is near its historical trough while Q is near its historical peak.

Fig. 10. 
Fig. 10. 

Growth rates of capital stock. Growth rate of private nonresidential fixed assets, based on section 4.2 of the Bureau of Economic Analysis fixed assets tables.

Is the fall in investment pervasive across firms? In table 4, we define leaders by constant shares of market value to ensure comparability over time.20 Capital K includes intangible capital as estimated by Peters and Taylor (2016). As shown, the leaders’ share of investment and capital has decreased, while their profit margins have increased. By contrast, laggards exhibit much more stable investment and profit rates. As shown in figure A25 (see https://www.nber.org/data-appendix/c14237/appendix.pdf), the increase in leader profits is not fully explained by a reallocation effect with higher profit firms becoming leaders: profits increased within firms for leaders and decreased slightly for laggards.

Table 4. 

Investment, Capital, and Profits by Leaders and Laggards

 1980–951996–2017Difference
LeadersMidLaggardsLeadersMidLaggardsLeadersMidLaggards
0–33 (%)33–66 (%)66–100 (%)0–33 (%)33–66 (%)66–100 (%)0–33 (%)33–66 (%)66–100 (%)
Share of OIBDP.36.33.32.35.32.33.00−.01.01
Share of CAPX + R&D.36.32.32.33.30.36−.02−.02.04
Share of PP&E.34.33.33.33.29.37.00−.04.04
Share of K.33.33.33.32.31.36−.01−.02.03
(CAPX + R&D)/OIBDP.59.58.60.43.44.52−.16−.14−.08
OIADP/SALE.13.11.09.16.14.10.03.03.01

Note. Table shows the weighted average value of a broad set of investment, capital, and profitability measures by time period and market value (MV). Leaders (laggards) include the firms with the highest (lowest) MV that combined account for 33% of MV within each industry and year. Annual data from Compustat. See appendix E (see https://www.nber.org/data-appendix/c14237/appendix.pdf) for details. OIBDP = operating income before depreciation; CAPX = capital expenditures; R&D = research and development; PP&E = property, plant, and equipment; K = capital; OIADP = operating income before depreciation; SALE = sales.

View Table Image

Is the decline in investment by leaders linked to concentration? According to σ and γ, leaders should increase investment in concentrating industries, reflecting an escape-competition strategy (σ) or their increasing relative productivity (γ). We test this at the firm level by estimating the following regression for firm i that belongs to BEA industry j:

(3)Δlog(Kijt)=β1Qit1+β2CR8jt1IA×Leadi,j,t+β3CR8jt1IA+β4Leadijt1+β5log(Ageit1)+ηt+δi+εit,
where Kit is firm capital (property, plant, and equipment [PP&E], intangibles, or total), CR8jtIA is the import-adjusted Census-based CR8, and Leadi,j,t is an indicator for a firm having a market value in the top quartile of segment k. We include Qit1 and log(Ageit1) as controls, along with firm and year fixed effects (ηt and δi). β2 is the coefficient of interest. Table 5 shows that with the exception of manufacturing, leaders in more concentrated industries underinvest. This is inconsistent with σ and γ but consistent with κ.
Table 5. 

Investment by Leaders in Concentrating Industries

 AllMfgNonMfg
Δ log(PPE)aΔ log(IntPT)bΔ log(KPT)a + bΔ log(PPE)aΔ log(IntPT)bΔ log(KPT)a + bΔ log(PPE)aΔ log(IntPT)bΔ log(KPT)a + b
(1)(2)(3)(4)(5)(6)(7)(8)(9)
CR8jt1IA−10.98+.58−4.82−17.10+−3.49−3.81−7.0612.13−2.35
(5.96)(6.00)(5.38)(9.21)(8.29)(7.52)(9.19)(10.75)(9.18)
CR8jt1IA×leadit1−11.95*−18.92**−15.14**1.44−1.35−1.15−13.64*−23.92**−17.44**
(4.66)(5.80)(4.51)(7.20)(9.58)(7.25)(6.10)(7.53)(5.90)
log Qit−113.45**11.66**12.90**11.99**9.85**10.66**15.60**14.16**15.96**
(.43)(.37)(.35)(.53)(.42)(.40)(.73)(.67)(.61)
Leadit−14.19**3.83**3.03**3.69**2.392.37+2.47+2.99+1.51
(.99)(1.13)(.91)(1.42)(1.67)(1.32)(1.38)(1.72)(1.38)
log ageit−1−15.11**−18.85**−17.17**−15.50**−18.31**−17.34**−14.36**−19.29**−16.75**
(.78)(.72)(.64)(1.03)(.86)(.83)(1.18)(1.25)(1.00)
Year FEYYYYYYYYY
Firm FEYYYYYYYYY
R2.1.12.15.098.14.16.11.12.15
Observations63,68063,34265,28533,70034,29334,30829,98029,04930,977

Note. Table shows the results of firm-level panel regressions of the log change in the stock of capital (K; deflated to 2009 prices) on import-adjusted concentration ratios (CR), following equation (3). Regression from 1997 to 2012 given the use of Census concentration measures. We consider three measures of capital: property, plant, and equipment (PP&E), intangibles (Int) defined as in Peters and Taylor (2016; abbreviated as PT), and their sum (total). Leaders (Lead) include firms with market value in the top quartile of the corresponding Bureau of Economic Analysis segment j for the given year. Q and log-age included as controls. As shown, leaders decrease investment with concentration rather than increase it. Annual data, primarily sourced from Compustat. Standard errors in parentheses are clustered at the firm level. Mfg = manufacturing industries; NonMfg = nonmanufacturing industries;, FE = fixed effects.

+ p < .10.

*p < .05.

**p < .01.

View Table Image

Case Study: The China Shock Again

Another way of investigating the role of κ for investment is to examine the behavior of leaders and laggards following the China shock. Figure 11 plots the average stock of K across Compustat firms in a given year, split by the 1999 NTR gap (see Section II for details). K includes PP&E as well as intangibles, as estimated by Peters and Taylor (2016). In low exposure industries, leaders and laggards exhibit similar growth rates of capital. By contrast, leaders increase capital much faster than laggards in high exposure industries.

Fig. 11. 
Fig. 11. 

Change in average firm KPT by Chinese exposure (1991 = 1). Annual data from Compustat, Peters and Taylor (2016; abbreviated as P&T), Schott (2008), and Pierce and Schott (2016). Manufacturing industries only, split into high (above-median) and low (below-median) exposures based on the 1999 normal trade relations (NTR) gap. Leaders (Lead) defined as firms with market value in top quartile of the distribution within each North American Industry Classification System level 6 industry, as of 2001. Only firms-year pairs with non-missing KPT included. K = capital; Lag = laggards.

Figure 11 suggests that leaders react to increased competition from China by increasing investment. We confirm this by estimating a generalized difference-in-differences (DiD) regression:

(4)log(Ki,j,t)=β1Post01×NTR Gapj×ΔIPt¯+β2Post01×NTR Gapj×ΔIPt¯×Leaderi,j,0+Xj,tγ+ηt+μi+εit,
where the dependent variable is a given measure of capital for firm i in industry j during year t. ΔIPt¯ captures time-series variation in Chinese competition averaged across all industries.21 The first two terms on the right-hand side are the DiD terms of interest. The first one is an interaction between the NTR gap and ΔIPt¯ for the post-2001 period. The second term adds an indicator for leader firms to capture differences in investment between leaders and laggards. The third term includes several industry-level characteristics as controls, such as capital and skill intensity.22 We include year and firm fixed effects ηt and μi.

Table 6 reports the results. It shows that leaders increase investment in response to an exogenous increase in competition. We consider three different measures of capital: PP&E, intangibles (from Peters and Taylor 2016) and total capital (equal to the sum of PP&E and intangibles).23 Columns 1–3 include all US incorporated manufacturing firms in Compustat over the 1991–2015 period. Columns 4–6 focus on continuing firms (i.e., firms that were in the sample before 1995 and after 2009) and show that leaders invested more than laggards, even when compared to firms that survived the China shock.

Table 6. 

Investment of Leaders and Laggards Following the Accession of China to the WTO

 All FirmsContinuing Firms
log(PPEt)alog(InttPT)blog(ktPT)a+blog(PPEt)alog(InttPT)blog(ktPT)a+b
(1)(2)(3)(4)(5)(6)
Post01 × NTRGap−8.035**−.426−1.884−11.214**−3.284+−4.670**
(2.008)(1.962)(1.578)(2.138)(1.921)(1.534)
Post01 × NTRGap × Lead9.267**6.978**6.643**9.601**8.319**7.998**
(2.005)(1.159)(1.149)(2.457)(1.457)(1.459)
Firm FEYYYYYY
Year FEYYYYYY
Industry controlsYYYYYY
R2.14.52.49.18.57.54
Observations34,71135,04335,07515,90616,01716,034

Note. Table shows the results of firm-level panel regressions of measures of capital on NTR Gapj×ΔIPj,tUS¯, following equation (4). We consider three measures of capital: gross property, plant, and equipment (PPE), intangibles (Int) defined as in Peters and Taylor (2016; abbreviated as PT) and their sum (total). Regression over 1991–2015 period. Leaders (Lead) defined as firms with market value in top quartile of the distribution within each North American Industry Classification System Level 6 industry, as of 2001. All regressions include measures of industry-level production structure as controls (see text for details). Only US-headquartered firms in manufacturing industries with non-missing KPT included. Standard errors in parentheses are clustered at the industry level. WTO = World Trade Organization; NTR = normal trade relations; FE = fixed effects.

+ p < .10.

**p < .01.

View Table Image

Our results are consistent with Frésard and Valta (2015) and Hombert and Matray (2015). Frésard and Valta (2015) find a negative average impact of foreign competition in industries with low entry costs and strategic substitutes. They briefly study within-industry variation and find that investment declines primarily at financially constrained firms. Hombert and Matray (2015) studies within-industry variation with a focus on firm-level R&D intensity. They show that R&D-intensive firms exhibit higher sales growth, profitability, and capital expenditures than low-R&D firms when faced with Chinese competition, consistent with our finding of increased intangible investment. They find evidence of product differentiation using the index of Hoberg and Phillips (2017). In the appendix of Gutiérrez and Philippon (2017), we study the dynamics of employment and find that leaders increase both capital and employment, while laggards decrease both. Employment decreases faster than capital so that K/Emp increases in both groups of firms. Since initial publication of these results in Gutiérrez and Philippon (2017), Pierce and Schott (2018) obtained similar results using Census data to cover the entire sample of US firms.

In summary, leader profit margins increased while investment relative to Q decreased, in line with κ. The falling growth rate of the capital stock—including intangibles—and the decline in leader investment, particularly in concentrated industries, is inconsistent with σ and γ.

D.  Returns to Scale

So far, we have evaluated the different theories indirectly by looking at their predictions about observable measures. In the case of γ, however, we can test the theory directly.

In Gutiérrez and Philippon (2019), we use industry- and firm-level data to estimate returns to scale. Industry-level estimates are based on BLS capital (K), labor (L), energy (E), materials (M), and service (S; KLEMS) data, following the methodology of Basu et al. (2006) while incorporating the instruments of Hall (2018). These estimates have the advantage of relying on well-measured inputs, outputs, and prices while following an established literature and set of instruments. However, the limited data availability implies that we can only estimate long-run average changes, such as an increase from before to after 2000. We perform this estimation and find a small increase in returns to scale, from 0.78 before 2000 to 0.8 afterward.

We complement industry-level estimates with firm-level estimates based on Compustat, roughly following Syverson (2004) and De Loecker et al. (2019). In particular, we estimate

Δlogqit=γ[αVΔlogv+αKΔlogk+αXΔlogx]+ω,
where γ measures the average return to scale across all firms. The variables v, k, and x denote COGS, capital costs, and overhead costs (SG&A), respectively. The equation αV=PVV/(PVV+rK+PXX) denotes the cost share of the COGS (likewise for αK and αX).24 We again find stable estimates since 1970.

The relative stability of returns to scale is consistent with a variety of estimates in the literature, including Ho and Ruzic (2018) for manufacturing in the US, and Salas-Fumás, San Juan, and Vallés (2018) and Diez et al. (2019) across EU industries. Thus, γ cannot explain the aggregate trends, though it likely matters for some industries.

IV.  Industry Evidence

Aggregate trends are interesting, but they obscure the dynamics of individual industries: one size does not fit all. In this section, we perform a PCA on a wide range of variables related to competition (and covering all types of measures in table 1) to obtain a systematic classification of the drivers of industry-level changes. We follow the industry segments in the BLS KLEMS and perform the PCA on the correlation matrix, so all measures contribute equally. Because we include census-concentration ratios, agriculture and mining are excluded from the analysis.

Figure 12 shows the variables included in the analysis and the resulting loadings of the first two principal components. Together, these components explain 34% of the variance. They have an intuitive interpretation. PC1 seems to capture the σ and γ theories of good concentration. It exhibits a positive loading on the level and changes in concentration (cr4_cen) and a high loading on intangible capital intensity (intan_kshare). The corresponding industries face significant import competition (import_share) and exhibit stable or declining profits (profit_margin). TFP increases (dtfp_kl), and unit-labor costs fall (Dlogulc). Prices also fall (Dlogp) but less than unit-labor costs so that markups rise (Dlogmu). Leader turnover falls while the investment gap is close to zero.

Fig. 12. 
Fig. 12. 

Principal component (PC) loadings. See text for details and appendix E (see https://www.nber.org/data-appendix/c14237/appendix.pdf) for definitions of variables. Cen = Census; BEA = Bureau of Economic Analysis industry accounts; PT = Peters and Taylor (2016); CP = Compustat; KL = capital (K), labor (L), energy (E), materials (M), and service (S; KLEMS); mv = market value; ikgap_cp = investment gap_compustat. See text for further details.

PC2, by contrast, seems closely related to the κ theories of bad concentration. It captures a sharp increase in concentration despite limited growth in intangibles and negative import competition. Profits rise and the labor share falls. Markups also rise, but for inefficient reasons: prices rise while productivity and ULC remains largely flat.

Figure 13 contrasts the 2012 loadings on PC1 and PC2 for each industry. We highlight the six industries with the highest score according to PC1 and PC2. Durable computer manufacturing, computer services, and nondurable apparel exhibit high loadings on PC1 and low loadings on PC2. They appear to remain strongly competitive despite increases in intangibles and concentration, likely as a result of foreign competition as shown in figure 14. In fact, figure 14 confirms the importance of foreign competition for domestic concentration and serves as a comforting validation of our PCA.

Fig. 13. 
Fig. 13. 

Principal component (PC) scores, by industry. See text for details and appendix E (see https://www.nber.org/data-appendix/c14237/appendix.pdf) for definitions of variables. Dur = durable; Nondur = nondurable; mgmt = management; Adm = administration; Acc = accommodation; Transp = transport; Misc = miscellaneous; Inf = information; serv = server; prim_metal = primary metal; fab_metal = fabricated metal; trans_truck = transportation truck.

Fig. 14. 
Fig. 14. 

Principal component (PC2) scores (“barriers to entry”) versus import shares. PC2 scores as of 2012 versus industry-level import shares, defined as the ratio of industry-level imports to gross output plus imports. Imports from Peter Schott’s website; gross output from the Bureau of Economic Analysis gross domestic product by industry accounts. Dur = durable; Nondur = nondurable; Misc = miscellaneous; transp = transport.

Nondurable chemical manufacturing, information (data), and information (publishing) present a mix of intangible-driven concentration and barriers to entry. These industries include Pfizer and Dow DuPont; Google and Facebook; and Microsoft, respectively. They are good examples of industries with large amounts of intangible assets—including patents—where leaders have become more efficient but also more entrenched over time.

Information (telecom), banking, and air transportation score near the top according to PC2. As discussed in Gutiérrez and Philippon (2018), these industries exhibit higher concentration, prices, and profitability in the US than in Europe, despite using similar technologies. Accommodation/food (i.e., restaurants) scores near the bottom according to both measures. This is an industry with limited use of intangible assets that remains largely competitive. The fact that education is the only real outlier is also comforting.

The PCA shows that both the κ theory and a combination of σ and γ are important for explaining the evolution of US industries over the past 20 years. But are they equally important at each point in time? Figure 15 plots the average PC1 and PC2 scores over time. The conclusions are striking. The average PC1 score, reflecting “good” concentration, was substantially higher and increased faster from 1997 to 2002. But PC2 (i.e., barriers to entry) caught up afterward. By 2012, most industries weighted heavily on PC2 while the average PC1 score remained close to zero (with wide dispersion, of course, as shown in fig. 13).

Fig. 15. 
Fig. 15. 

Evolution of the average scores for principal components PC1 and PC2

V.  Conclusion

A.  Internal Consistency of Macro-market Power Literature

We have used a wide range of measures of competition throughout this paper, sometimes independently and sometimes jointly, albeit nonparametrically. But all of these measures are connected by economic theory. Let us conclude by bringing together estimates from the macro-market power literature to validate the internal consistency of our conclusions. A decomposition first made by Susanto Basu in his discussion of De Loecker et al. (2019) is useful. We describe the decomposition briefly and refer the reader to Syverson (2019) for a discussion of the underlying assumptions.

Consider a standard profit-maximizing economy, and rewrite the markup by multiplying and dividing by average costs:

μ=PMC=PACACMC=ACMCRevenueCost.
The ratio of average to marginal costs, AC/MC, equals the returns to scale for a cost-minimizing firm taking factor prices as given while Revenue/Cost can be written as 1/(1sπ) using the profit share in revenues sπ. Therefore,
(5)μ=γ1sπ.
Using equation (5) for two time periods, we obtain
μ2016μ1980=(1sπ,19801sπ,2016)γ2016γ1980,
which can be used to assess the internal consistency of the macro-market power literature.

Let us begin by reiterating the discrepancy raised by Syverson (2019) and Basu (2019). De Loecker et al. (2019) report an increase in markups from 1.21 to 1.61 between 1980 and 2016 and an increase in returns to scale from 1.03 to 1.08. Barkai (2017) estimates rising profit shares from 3% to 16% of value added over the same period, which (roughly) equate to 1.5% and 8% of sales. Plugging in these values, we obtain

1.611.21=(10.01510.08)1.081.03,1.33=1.12.
The relationship appears widely inconsistent but there is an issue with this comparison. The markup estimates of De Loecker et al. (2019) are based on public firms, which likely have higher intangible (and SG&A) intensity than private firms—certainly more than small and medium enterprises. For the reasons discussed in Section II, this leads to an overestimation of the rise in markups for the full economy. As a rough approximation, let us assume that markups of private firms remained stable, in line with the median Compustat firm as reported in figure 8a of De Loecker et al. (2019). This is valid if the distribution of high intangible firms, and therefore markup increases, is concentrated at the top. We can then obtain a rough estimate of the change in economy-wide markups as the product of the Compustat markup increase (33%) times the Compustat share of sales in the total economy (40% as reported by Grullon et al. 2019). The resulting markup increase is then 13.2%, which seems consistent with the estimates above. Using our return to scale estimates, the last term would be 0.8/0.78—again broadly in agreement.25

B.  Explaining the Rise in κ

Estimates from the macro-market power literature appear reasonably consistent with each other. They include a sharp increase in profits unique to the US, concentrated in the post-2000 period and explained mostly by rising barriers to entry. The next question is, of course: What might explain the rise in κ in the US? Gutiérrez and Philippon (2018) argue that this is partly explained by weakening competition policy (i.e., antitrust and regulation) compared to Europe. Gutiérrez and Philippon (2019) show that the decline in the elasticity of entry to Q is partly explained by lobbying and increasing federal and state-level regulations.26 Last, Jones et al. (2019) combine a rich structural dynamic stochastic general equilibrium model with cross-sectional identification from firm and industry data. They use the model to structurally estimate entry cost shocks and show that model-implied entry shocks correlate with independently constructed measures of entry regulation and merger and acquisition activities.

Endnotes. 

Author email addresses: Covarrubias (), Gutiérrez (), Philippon (). This paper was prepared for the NBER Macroeconomics Annual 2019. Some of the results presented in the text were first published in Gutiérrez and Philippon (2017). We are grateful to the Smith Richardson Foundation for a research grant, to Janice Eberly and Chad Syverson for their discussion, and to Erik Hurst and participants at the NBER Macro Annual conference for helpful comments and suggestions. For acknowledgments, sources of research support, and disclosure of the authors’ material financial relationships, if any, please see https://www.nber.org/chapters/c14237.ack.

1. See Autor et al. (2017a) for a longer time-series of US Census–based concentration measures under a consistent segmentation. The series in Autor et al. (2017a) exhibit similar trends: concentration begins to increase between 1992 and 1997 for retail trade and services and between 1997 and 2002 for the remaining sectors.

2. One could entertain other hypotheses, such as weak demand or credit constraints, but previous research has shown that they do not fit the facts. See Gutiérrez and Philippon (2016) for detailed discussions and references.

3. Furman (2015) also emphasizes the weakness of corporate fixed investment and points out that low investment has coincided with high private returns to capital, implying an increase in the payout rate (dividends and share buybacks).

4. For figure 2, we measure concentration as the ratio of sales by the eight largest firms in Compustat that belong to a given capital (K), labor (L), energy (E), materials (M), and service (S; KLEMS) industry × region to total Gross Output reported in OECD STAN. Corporate consolidation is therefore accounted for, as dictated by accounting rules. Appendix A (see https://www.nber.org/data-appendix/c14237/appendix.pdf) provides additional details on the calculation, while Gutiérrez and Philippon (2018) provide a detailed comparison across a wide range of concentration measures for the US and Europe. Bajgar et al. (2019) use Orbis data to include private firms and take into account that some firms are part of larger business groups. When they measure concentration at the business group level within two-digit industries, they find a moderate increase in concentration in Europe, with the unweighted average CR8 increasing from 21.5% to 25.1%. In North America, CR8 increases from 30.3% to 38.4%.

5. These comparisons aggregate across industry categories and may therefore be affected by changes in industry mix. However, Gutiérrez and Philippon (2018) reach similar conclusions using industry-level data. Moreover, in Gutiérrez and Philippon (2017), we compare the evolution of the five industries that concentrate the most in the US against Europe. We find that concentration, profits, and Q increased in the US, while investment decreased. By contrast, concentration and investment remained (relatively) stable in Europe, despite lower profits and lower Q. This is true even though these industries use the same technology and are exposed to the same foreign competition. For more details on the labor share, see Gutiérrez and Piton (2019) and Cette, Koehl, and Philippon (2019).

6. We derive most but not all of these in Section I. For predictions on leader investment, see Gutiérrez and Philippon (2017).

7. In this model, an increase in returns to scale corresponds to a shift toward a high productivity, high fixed cost technology.

8. Gutiérrez and Philippon (2017) report similar results using Herfindahls and the data of Feenstra and Weinstein (2017).

9. We follow Autor et al. (2016) and define import penetration for industry j at time t as ΔIPjt=ΔMjtUC/(Yj,91+Mj,91Ej,91), where ΔMjtUC denotes the change in US imports from China from 1991 to t; and Yj,91+Mj,91Ej,91 denotes the initial absorption (defined as output, Yj,91, plus imports, Mj,91, minus exports, Ej,91). Yj,91 is sourced from the NBER-CES database and Mj,91 and Ej,91 are based on Peter Schott’s data. Only NAICS level 6 industries where data are available across all sources are included in the analyses.

10. Until 2001, China was considered a nonmarket economy. It was subject to relatively high tariff rates (known as “non-normal trade relations” tariffs or “non-NTR rates”) as prescribed in the Smoot-Hawley Tariff Act of 1930. From 1980 onward, US presidents began temporarily granting NTR tariff rates to China but required annual reapproval by Congress. The reapproval process introduced substantial uncertainty around future tariff rates and limited investment by both US and Chinese firms (see Pierce and Schott 2016 for a wide range of anecdotal and news-based evidence). This ended in 2001, when China entered the WTO and the US granted PNTR. The granting of PNTR removed uncertainty around tariffs, leading to an increase in competition.

11. SALE/COGS is related to the benchmark measure of De Loecker et al. (2019) up to a measurement error correction and a (time-varying) industry-level scaling factor, which measures the elasticity of SALES to COGS. Both the measurement error correction and the elasticity of output remain largely stable even in the long run so that SALE/COGS dominates the evolution of markups.

12. In unreported tests, we find similar conclusions using the firm-level user-cost markups first reported in the appendix of Gutiérrez and Philippon (2017), and studying regulatory shocks (the entry of Free Mobile in France and the implementation of large product market regulations, as compiled by Duval et al. 2018).

13. See De Loecker and Eeckhout (2018) for actual markup estimates globally. As expected, their results closely follow the SALE/COGS series.

14. This is not to say that profits are a perfect measure. Accounting rules often deviate from economic concepts, and estimates of economic profits are prone to errors given the difficulty in measuring the capital stock and the user cost of capital. We can gain some comfort, however, by comparing a wide range of measures from alternate sources. In Gutiérrez and Philippon (2018), for example, we show that accounting profits from Compustat and national accounts, economic profits in the style of Barkai (2017) and firm-level user-cost implied profits are consistent with each other in both the US and Europe.

15. See Rossi-Hansberg, Sarte, and Trachter (2018), among others, for related evidence, but note that their conclusions are controversial (Ganapati 2018).

16. We use a constant number of leaders because they account for a roughly stable share of sales. In unreported tests, we consider the top 10% of firms and obtain similar results, though this broader group accounts for a rising share of sales.

17. Figure A20 (see https://www.nber.org/data-appendix/c14237/appendix.pdf) presents an additional test, based on the correlation of firm rankings over time. It yields consistent results.

18. For manufacturing, the NBER-CES database includes nominal output, prices, wages, and employment. For nonmanufacturing, the concentration accounts include nominal output, payroll, and employment, while the BEA’s “detailed” GDP by industry accounts include prices. The detailed GDP by industry accounts include about 400 industries so that our nonmanufacturing data set is more aggregated than that of Ganapati (2018). We use the more aggregated data set given the concerns with skewness described in the text and because, even at that level of aggregation, the BEA cautions of potential measurement error. That said, our results are largely consistent.

19. Ganapati (2018, 9) estimates the following relationship:

Δ5log(Pjt)=0.00992×Δ5log(CR4)0.0520×Δ5log(LP)+γs,t+εjt,
which implies that “a one standard deviation increase in monopoly power offsets 1/5 of the price decrease from a one standard deviation increase in productivity.” He argues that “the most pessimistic reading is that after controlling for productivity, monopolies do increase prices. But this argument assumes that all other conditions including productivity remain constant. In the light of the close linkage of productivity and concentration, this seems untenable.”

20. Operating income before depreciation shares are stable, which is consistent with stable shares of market value and stable relative discount factors. Because firms are discrete, the actual share of market value in each grouping varies from year to year. To improve comparability, we scale measured shares as if they each contained 33% of market value.

21. Gutiérrez and Philippon (2017) present results excluding ΔIPj,t¯ to mirror the specification of Pierce and Schott (2016) as well as following the approach of Autor et al. (2016)—which instruments ΔIPj,tUS with the import penetration of eight other advanced economies (ΔIPj,tOC).

22. These industry characteristics are sourced from the NBER-CES database. We include the (i) percent of production workers, (ii) log-ratio of capital to employment, (iii) log-ratio of capital to value added, (iv) log-average wage, and (v) log-average production wage.

23. In unreported robustness tests, we confirm that our results are robust to including only balance sheet intangibles or excluding goodwill in the Peters and Taylor (2016) measure.

24. De Loecker et al. (2019) perform the same estimation in levels and find an increase in returns to scale from 0.97 to 1.08. However, levels regressions are likely affected by the inability to control for differences in firm-level prices or to accurately measure intangible capital. For example, an increase in the markups of large relative to small firms would appear as an increase in quantities and result in an overestimation of the increase in returns to scale. The estimation based on changes better controls for this, hence is likely more robust.

25. We can perform a similar exercise since 2000, using the results of Diez et al. (2019), which are based on Orbis data and therefore include private firms. According to their estimates, US markups increased by 12% since 2000 while returns to scale increased from 0.91 to 0.93. Over the same period, Barkai (2017) reports profit shares of value added rising from 4.5% to 16%. We then have

1.12=(10.02310.08)0.930.91,1.12=1.09.
We may also want to consider total economy profit shares, instead of nonfinancial corporation profit shares. Gutiérrez (2017) uses BEA data for the nonfinancial private economy. He finds an increase in the profit share from 11% to 21% from 1988 to 2015, which closely aligns with Barkai (2017) over the same period. Last, performing the same exercise for Europe with markup and returns to scale estimates from Diez et al. (2019) and profit share estimates from the appendix of Gutiérrez and Philippon (2018; accounting only for the cost of debt to mirror Barkai [2017]), we obtain
1.06=(10.03610.038)0.930.91,1.06=1.03.
Again, broadly in agreement.

26. In unreported tests, we confirm there is a positive relationship between PC2 and industry-level lobbying intensity.

References