Documentamos la evolución del poder de mercado a partir de datos de empresas de la economía estadounidense desde 1955. Medimos tanto los márgenes comerciales como la rentabilidad. En 1980, los márgenes agregados empiezan a aumentar desde el 21% por encima del coste marginal hasta el 61% actual. El aumento se debe principalmente a la cola superior de la distribución de los márgenes: los percentiles superiores han aumentado considerablemente. Sorprendentemente, la mediana no cambia. Además del aumento de la cola superior de la distribución de los márgenes, hay una reasignación de la cuota de mercado de las empresas con márgenes bajos a las que tienen márgenes altos. Este aumento se produce sobre todo dentro de un mismo sector. También se observa un aumento de la tasa media de beneficios del 1% al 8%. Aunque también se produce un aumento de los gastos generales, el aumento del margen de beneficio es superior al de los gastos generales. Analizamos las implicaciones macroeconómicas de un aumento del poder de mercado medio, que puede explicar varias tendencias seculares de las últimas cuatro décadas, sobre todo el descenso de las cuotas de trabajo y capital, así como la disminución del dinamismo del mercado laboral.
II. Empirical Framework and Data
We present the empirical framework that allows
us to derive a markup measure for each firm covering the entire economy,
over more than six decades. The framework uses the cost minimization
approach, where firms choose the optimal bundle of variable inputs of
production. This reasonable assumption on firm behavior only relies on
firm-level revenue and input expenditure data for firms across the U.S.
economy. As such we do not impose restrictions on product market
competition and consumer demand.
In this
section, we present the model and then discuss the particular
implementation in the data sets we use. Our focus is to provide a robust
description and analysis of markups across producers using different
methods and approaches.
II.A. Obtaining Markups from Producer Behavior
The markup is commonly defined as the output
price divided by the marginal cost. Measuring markups is notoriously
hard as marginal cost data is not readily available, let alone prices.
There exist three distinct approaches to measure markups. First, the
accounting approach relies on directly observable gross (or net) margins
of profits. Although this approach is straightforward to implement, it
suffers from well-known problems, chief among them the inability to
directly measure the marginal cost of production. A straightforward way
to circumvent this problem is to equate average to marginal costs, but
this imposes strong and unrealistic restrictions on firm-level cost
structures.8
The second approach was developed in the modern industrial organization literature (see Berry, Levinsohn, and Pakes 1995; Bresnahan 1989)
and relies on the specification of a demand system that delivers price
elasticities of demand. Combined with assumptions on how firms compete,
the demand approach delivers measures of markups through the first-order
condition associated with optimal pricing. This approach, while
powerful in other settings, is not applicable here for two distinct
reasons. First, we do not want to impose a specific model of how firms
compete across a large data set of firms active in very different
industries, or commit to a particular demand system for all the products
under consideration. Second, even if we wanted to make all these
assumptions, there is simply no information on prices and quantities at
the product level for a large set of sectors of the economy over a long
period of time. This is necessary to successfully estimate price
elasticities of demand, and specify particular models of price
competition for all sectors.
Instead, we rely on a third way: the production approach. This approach is based on the insight of Hall (1988) to estimate markups from the firm’s cost minimization decision. Hall (1988) used industry aggregates; De Loecker and Warzynski (2012)
recently proposed to estimate firm-level markups. The method uses
information from the firm’s financial statements and does not require
any assumptions on demand and how firms compete. Instead, markups are
obtained by exploiting cost minimization of a variable input of
production. This approach requires an explicit treatment of the
production function to obtain the output elasticity of at least one
variable input of production.
Before we discuss the production approach, on which we
rely to measure markups, it is instructive to go back to the underlying
assumptions of the accounting and so-called demand approaches.
Throughout we define markups as the price-to-marginal cost ratio:
In essence, the simplicity of the accounting approach is to simply multiply through by total output (
Q) and obtain:
The entire approach rests on the assumption that the object cQ
is directly observable in the data. There are three main assumptions
and therefore complications. First, this approach relies crucially on
the equality of marginal and average cost of production. This requires
constant returns to scale (CRS) in production and the absence of
economies of scale, that is, there are no fixed costs. Second, it
implicitly relies on the assumption that all relevant factors of
production are perfect substitutes in production. Third, and related,
the measure of cost (cQ) is not equal to marginal cost if it
includes cost items that do not vary with output. Note that in the
accounting approach the markup equals the profit rate when all cost
items (including fixed factors like capital, and investment activities
such as R&D and advertising) are included in the measure cQ.
The
demand approach relies on an estimated demand curve (having data
separately on prices and quantities for all products in a prespecified
market) and a particular model of competition to back out c
from a first-order condition resulting from profit maximization. The
production approach frees up all these restrictions on conduct and
demand by computing the marginal cost of production directly from the
cost minimization condition for a single variable input of production.
II.B. The Production Approach
Consider an economy with N firms, indexed by i = 1, …
,
N. Firms are heterogeneous in terms of their productivity Ω
it and production technology
Qit(.).
9 In each period
t, firm
i minimizes the contemporaneous cost of production given the production function:
where V=(V1,...,VJ)
is the vector of variable inputs of production (including labor, intermediate inputs, materials,
…),
Kit is the capital stock and Ω
it
is productivity. The key assumption is that within one period (a year
in our data), variable inputs frictionlessly adjust, whereas capital is
subject to adjustment costs and other frictions. Because in the
implementation we use information on a bundle of variable inputs and not
the individual inputs, in the exposition we treat the vector
V as a scalar
V.
10 We consider the Lagrangian objective function associated with the firm’s (conditional) cost minimization:
11where
PV is the price of the variable input,
r is the user cost of capital,
12Fit is the fixed cost,
Q(·) is the technology specified in
equation (3),
Q¯¯¯¯ is a scalar and λ is the Lagrange multiplier. We assume that variable input prices are given to the firm.
13 We consider the first-order condition with respect to the variable input
V, and this is given by:
Multiplying all terms by
VitQit and rearranging terms yields an expression for the output elasticity of input
V:
The Lagrange multiplier λ is a direct measure of marginal
cost (tracing out the value of the objective function as we relax the
output constraint), and we define the markup as the price–marginal cost
ratio μ=Pλ
, where
P is the output price. Substituting marginal cost for the markup to price ratio, we obtain a simple expression for the markup:
The
expression of the markup is derived without specifying conduct or a
particular demand system. Note that with this approach to markup
estimation there are in principle multiple first-order conditions (of
each variable input in production) that yield an expression for the
markup. Regardless of which variable input of production is used, two
key ingredients are needed to measure the markup: the revenue share of
the variable input, PVitVitPitQit
, and the output elasticity of the variable input,
θvit.
The markup formula (7)
derived under the production approach highlights that the marginal cost
of production is derived from a single variable input in production,
without imposing any particular substitution elasticity with respect to
other inputs (variable or fixed) in production or returns to scale. It
is instructive to contrast it to the accounting approach introduced
above: only in the case of a CRS single variable input (V) production function without fixed costs will the correct markup be measured by the sales to the variable input expenditure.
An important component of the markup formula under the production approach is therefore the output elasticity θvit
. In Appendix A,
we discuss in detail the different approaches we take to measure this,
and we appraise the merits and shortcomings of each approach. We
distinguish between obtaining output elasticities from estimating the
production function and from cost shares.
II.C. Data
To cover the longest possible period of time and
to have a wide coverage of economic activity, we use data on publicly
traded firms. To our knowledge, Compustat is the only data source that
provides substantial coverage of firms in the private sector over a long
period of time, spanning the period 1950 to 2016. While publicly traded
firms are few relative to the total number of firms, because the public
firms tend to be the largest firms in the economy, they account for 29%
of private U.S. employment (Davis et al. 2007).
There
is a serious concern that the sample of publicly traded firms is not
representative of the distribution of the universe of firms. Listed
firms are bigger, older, more capital- intensive, and more
skill-intensive. They also involve a bigger role for multinationals. The
industry mix of Compustat firms differs from that of the private sector
as a whole.14 We deal with the selection bias from studying the publicly traded firms in two ways. In Section III.D
we repeat our analysis on the U.S. Censuses. For a number of sectors,
we have the universe of firms. Second, we use the population weights of
each sector to adjust the weights in the Compustat sample. Although we
still only use publicly traded firms to calculate the markups, we
account for any bias because of the sectoral composition.
The
Compustat data contains information about firm-level financial
statements, which allows us to rely on the so-called production approach
for measuring markups. In particular, we observe measures of sales,
input expenditure, capital stock information, and detailed industry
activity classifications.15
The item from the financial statement of the firm that we will use to
measure the variable input is cost of goods sold (COGS). It bundles all
expenses directly attributable to the production of the goods sold by
the firm and includes materials and intermediate inputs, labor cost,
energy, and so on.16
In addition, we observe relevant and direct accounting information of
profitability and stock market performance. The latter information is
useful to verify whether our measures of markups, as discussed below,
also relate to the overall evaluation of the market. Appendix Table B.1 provides basic summary statistics of the firm-level panel data used throughout the empirical analysis.
From our data, we construct a measure of the user cost of capital. We follow the standard procedure in the literature and use rt = (It − Πt) + Δ, where It, Πt,
and Δ are the nominal interest rate, the inflation rate, and a
depreciation rate. We use gross capital (PPEGT) that we adjust for the
industry-level input price deflator (PIRIC from FRED), for the federal
funds rate and for an exogenous depreciation rate and risk premium
jointly that we set at 12%.17
Our
data also have a measure of overhead, booked under selling, general and
administrative expenses (SG&A). This item includes selling expenses
(salaries of sales personnel, advertising, rent), general operating
expenses, and administration (executive salaries, general support
related to the overall administration). We use SG&A to calculate
total costs—not just the cost of factors of production—to measure the
profits of the firms. In addition, we will consider a production
technology, different from the conventional technology, where we treat
overhead as a factor of production.
II.D. Censuses
As a robustness exercise and to verify the
extent of selection bias in our sample of publicly traded firms, we
repeat this exercise for the Economic Census. The Economic Census is
administered every five years. It is composed of censuses of different
sectors: a Census of Manufacturing, a Census of Retail Trade, a Census
of Wholesale Trade, and so on. Within each sector, it covers the
universe of employer establishments (establishments that hire workers
and are not just one-person sole proprietorships); compliance is legally
required.
The Census of Manufacturing
contains establishment-level data on sales, in addition to very
comprehensive data on inputs (the total labor wage bill, capital,
materials, etc.). However, most of the other sector censuses (retail,
wholesale, etc.) only contain data on establishment-level sales and wage
bills, and not other nonlabor inputs. The census does not include
information on overhead directly.18 In Section III.D we analyze markups for manufacturing, retail, and wholesale. A detailed description of the census data is in Appendix B.
III. The Evolution of Markups in the U.S. economy
The bulk of our analysis is for the Compustat
data where we observe firms across a wide range of sectors and time.
Because we have firm-level markups, the main focus of attention is on
the evolution of the distribution of markups. We first report the
average markup, then detailed properties of the distribution, and
finally we decompose the average markup to single out the reallocation
of economic activity toward high-markup firms.
III.A. Aggregate Markups
The measure of markups in equation (7) is the product of the output elasticity θ and the inverse of the variable input’s revenue share PQPVV
.
The latter is directly measured in the firm’s income statement, and we
estimate the former. Our estimated output elasticities are sector- and
time-specific and thus capture technological differences across sectors
and time.
We calculate the average markup as follows:
where
mit is the weight of each firm. In our main specification, we use the share of sales in the sample as the weight.
Figure I
reports the evolution of our baseline measure of average markups across
the economy over time. In the beginning of the sample period, markups
were relatively stable, initially slightly increasing to 1.34 in the
1960s and then decreasing to 1.21 in 1980. Since 1980 there has been a
steady increase to 1.61. In 2016, the average markup charged is 61% over
marginal cost, compared with 21% in 1980. In
Online Appendix 5 we report a few examples of individual firms’ markups.
Figure I
Average Markups
Output elasticities θst
from the estimated production function are time-varying and
sector-specific (two-digit). The average is revenue weighted. The figure
illustrates the evolution of the average markup from 1955 to 2016.
In
broad terms, three sources can account for this rise in aggregate
markups: (i) the inverse ratio of the cost share of sales, (ii) the
output elasticity, (iii) the weight. To show the sensitivity of the
average markups to each of these determinants, in Figure II we plot the average markup with input weights and the average markup with a fixed, time-invariant output elasticity.
When
we fix the output elasticity to be time-invariant (calibrated to 0.85,
the average cost share), we find that the pattern of markups (Figure II,
Panel A) is similar to that in the benchmark with estimated output
elasticities. This tells us that the rise in markups is not due to the
change in the estimated output elasticity, which captures technological
change under our production function specification. Consistent with this
evidence, we find that the output elasticities vary very little over
time (see also Figure XII, Panel B, later).
Next we investigate the role of the input weight, the importance of which has first been flagged by Grassi (2017) and Edmond, Midrigan, and Xu (2019).
When firms have market power, they charge higher prices and, as a
result, dampen demand. With lower demand, the quantity sold and the
inputs used to produce are lower. Nonetheless, revenue (price times
quantity) is higher. As a result, firms with higher markups tend to have
higher revenue weights relative to their input weights.
This is exactly what we see in Figure II,
Panel B. The level of markups is lower throughout, and the rise is less
pronounced, which indicates that the gap between inputs and sales has
grown. The widening gap indicates that there is a change in the
equilibrium outcome and the market structure. Moreover, as we will see
in the next two sections, the widening of the markup distribution and
the reallocation of sales toward high-markup firms can explain why the
gap has widened.
Here we use the total cost (computed as the sum of COGS, SG&A, and rK)
as the input weight and for different weighting measures the gap
between the sales and the input-weighted aggregate markup is larger.19
Because we are interested in the properties of the entire distribution
of markups, we believe it is instructive to show as many different
moments as possible. In particular, the gap between the input-weighted
and the revenue-weighted aggregate markup informs us about the
underlying mechanism, the underlying distribution, and the reallocation.
We
use as our benchmark the revenue-weighted markup for the following
reasons. First, a substantial portion of what is going on in the output
market is reallocation (see below) of revenues toward high-markup firms.
We cannot capture this crucial phenomenon with input-weighted markups.
The revenue-weighted markup therefore informs us about the economic
mechanism and we show in a companion paper (De Loecker, Eeckhout, and Mongey 2018)
that this is an important determinant in explaining the rise of market
power. Second, to study market power, we link markups with profit rates (Section IV).
Profit rates are traditionally aggregated with revenue weights, and
consistency then calls for revenue weighting of the markup as well.
Finally, revenue weighting is a common benchmark that is commonly used,
most notably for widely used economic indicators, such as GDP and, in
the context of market power, HHI.
The bottom line is that our finding for the
benchmark measure of aggregate markups is robust. This implies that the
bulk of the action comes from the increase in the wedge of sales to
COGS. The rise is not driven by technological change (changing output
elasticities) and the weighting scheme informs us about the underlying
mechanism where the increasing gap between the revenue-weighted and
input-weighted aggregate markup tells us that firms spend less on
variable inputs.
III.B. The Distribution of Markups
Although average markups make for a good
headline, they do not fully capture the underlying distributional change
in markups. The advantage of our method to calculate markups is that we
obtain one for each firm, so we have a distribution of markups. A key
finding is that the increase in markups is driven by a few firms,
without any increase for most.
To get an idea
of the evolution of the entire distribution of markups, we plot the
kernel density of the unweighted markups for 1980 and 2016 (Figure III,
Panel A). We find that the variance has increased and that, in
particular, the upper tail has considerably fattened and become longer.
It is the upper tail that drives the increase in the average markup.
Figure III
The Distribution of Markups μit
Because
the kernel density does not take into account the weights, we next plot
the different moments of the distribution of sales-weighted markups
over time (Figure III,
Panel B). We rank the firms by markup; to obtain the percentiles we
weight each firm by its market share in the entire sample. This makes
the percentiles directly comparable to our share-weighted average. The
ranking is updated each year, so the firms at the top may be different
each year (later we investigate the persistence in the markup process).
The
increase in the average markup comes entirely from the firms with
markups in the top half of the markup distribution. The median (P50) and
the percentiles below the median are invariant over time. Most firms
see no increase in markups.20
For the higher percentiles, markups increase. For the 90th percentile
in particular, the increase is sharpest. Between 1980 and 2016, it
increases from 1.5 to 2.5. This indicates that the change in average
markup is largely driven by a few firms that currently have much higher
markups than decades ago.21
III.C. Reallocation of Economic Activity
The rise in the average markup is driven by a
few firms at the top of the distribution. Most firms see no increase in
markups, while a few firms see a large increase. We can further
decompose the increase in the weighted average markup into the component
that is attributable to the increase in the markup itself, and the
component that is attributable to the reallocation of economic activity
towards high-markup firms.
Inspection of Figure III,
Panel A already shows that there is a change in the distribution of
unweighted markups. The fatter tail is evidence that more firms have
higher markups. Even if the distribution of unweighted markups had
remained unchanged, the weighted aggregate markup could have gone up if
the firms with higher markups now obtain a higher share of the market.
This reallocation of economic activity toward higher-markup firms is
important to understand the implication that market power has on the
concentration of economic activity in the hands of a few dominant firms.
Though not in all, in most theories of market power, firms that have
higher market power also increase their market share (in the Cournot
model in particular, the market share is a sufficient statistic of
market power).
Because the change in aggregate markups is a combination
of the rise in unweighted markups and a reallocation of economic
activity, we decompose the average markup at the firm level as follows:
where
μ~it=μit−μt−1 and
μ~it−1=μit−1−μt−1We
apply the insights from the productivity-decomposition literature, and
while this decomposition appears very similar to that in equation (10),
it is different, first because it has one additional term, and second
because its interpretation is very different. There is an additional
term here because there is entry and exit of firms, whereas in the
sectoral decomposition the number of sectors is fixed.23 The interpretation also differs. Following Haltiwanger (1997),
we consider a theoretical counterfactual where the Δwithin term
measures the average change that is merely due to a change in the
markup, while keeping the market shares unchanged from last period.
Instead, the Δmarket share term measures the change due to an increase
in market share while keeping the markup fixed. If this term is
increasing, it captures the fact that firms with higher markups now have
a higher market share, and hence there is an increase in the weight of
the high-markup firms. This in turn raises the average markup without
raising the markup itself. The Δcross term measures the joint change in
markups and market share. We denote by Δreallocation the joint effect of
Δmarket share + Δcross term.24
Finally, the new last term measures the effect of entry and exit on
markups. This captures the change in the composition of firms in the
market. If the entering firms have higher markups than the exiting
firms, for example, this term will be positive.
We perform this decomposition across firms in the entire economy. To best present this decomposition, Figure IV
plots the average markup (solid line in print; solid red in color
version online; color version online), as well as three counterfactual
experiments based on the decomposition starting in 1980. We set the
initial level to 1980 and then cumulatively add the changes of each
component term in equation (9).25
Figure IV
Decomposition of Markup Growth at the Firm Level
The
first experiment (long dashed line in print; solid blue in color
version online) shows the evolution of the average markup as if there
was only component Δwithin and all other components were 0. This shows
that the rise in average markups in the 1980s and 1990s from 1.21 to 1.3
in 2000 is about one-third of the total increase from 1.21 to 1.47.
From 2000 onward, this term decreases and picks up again after the Great
Recession. The change in the average markup is also evident from Figure III, Panel A, where we see an increase in the upper right tail.
The
second experiment (short dashed line in print; solid black in color
version online) shows the path of the markup if the only change had been
due to Δreallocation. All markups remain unchanged from the previous
period, and we apply only the change in the market shares. The plot
shows that accumulated over the whole time period, reallocation accounts
for about two-thirds of the change in the weighted markup. The main
takeaway here is that there are two forces at work. On the one hand, the
markup (the within term) increases, which is an indication of the
change in pricing power of firms. In De Loecker, Eeckhout, and Mongey (2018),
we show that this can be due to a change in the market structure (less
competition) or due to technological change (bigger spread in firm
productivity). On the other hand, there is also a reallocation of sales
activity away from low-markup firms toward high-markup firms (the
reallocation term). This is entirely consistent with a model of
imperfect competition where firms with higher markups also attract a
higher market share. This reallocation effect is in accordance with the
findings in Autor et al. (2020) and Hartman-Glaser, Lustig, and Zhang (2016),
who establish that large firms have grown in size relative to small
firms, and those firms tend to operate in more concentrated markets.
While we find that the reallocation term is important, it is not the
only force at work. Unweighted markups have gone up (measured by the
Δwithin term and visualized by the density of markups in Figure III,
Panel A, especially in the upper tail), which is an important force
behind the rise in market power. In a general equilibrium model with
input-output linkages, Baqaee and Farhi (2020) find a similar decomposition of the within and the reallocation component.
The
third experiment (dash and dot line in print; solid green in color
version online) shows the evolution of markups if the only change was
net entry of firms. The net entry component rises early on and is more
or less constant afterward, indicating that the rise in markup is not
exclusively driven by the changing composition of firms in the sample.
The net entry component can simply be driven by the fact that the panel
of firms is not balanced and more firms enter than exit. In part, it can
also be driven by mergers and acquisitions. Consider two firms that
merge. If their joint market share is unchanged but they now charge
higher markups, then the net entry term will be positive. Or it could be
driven by the fact that the net entry accounts for a higher market
share than the sum of the individual premerger shares.
In
summary, the rise in aggregate markups is driven in part by a change in
the markup distribution itself, by a reallocation from low-markup firms
to high-markup firms, and by some net entry. In the first decade of the
sample, the 1980s, all three forces are equally at work. But by the end
of the period, reallocation dominates. Cumulatively over the whole
period, reallocation accounts for two-thirds of the rise in markups.
The decomposition exercise implies that the reallocation component captures movements of firms across all sectors. In Online Appendix 4
we perform the same decomposition for each of the broad sectors of the
economy, where reallocation of economic activity is measured within
sector.
In contrast to the firm-level decomposition (economy-wide
and within sector), we also analyze the decomposition of the rise of
markups by firm size at the sectoral level, that is, within and between
sectors. Is the increase in markup over time due to a change of markup
at the industry level (Δwithin), due to a change in the composition of
the firms—there are more firms with a high markup—(Δbetween), or due to
the joint change in markup and the firm composition (Δcross term)? This
can be expressed in the following formula:
We consider the change over 10-year periods starting in 1956 in Table I.26
The decomposition shows that the change in markup is mainly driven by
the change within industry. Most of the Δmarkup is driven by Δwithin.
There is some change in the composition between industries, but that is
relatively minor compared to the within industry change. The change due
to reallocation, the joint effect, is mostly small.
Table ISectoral Decomposition of 10-Year Change in Markup
| Markup | Δmarkup | Δwithin | Δbetween | Δcross |
|---|
| 1966 | 1.337 | 0.083 | 0.057 | −0.017 | 0.041 |
| 1976 | 1.270 | −0.067 | −0.055 | 0.002 | −0.014 |
| 1986 | 1.312 | 0.042 | 0.035 | 0.010 | −0.003 |
| 1996 | 1.406 | 0.094 | 0.098 | 0.004 | −0.008 |
| 2006 | 1.455 | 0.049 | 0.046 | 0.007 | −0.005 |
| 2016 | 1.610 | 0.154 | 0.133 | 0.014 | 0.007 |
In sharp contrast
with the firm-level decomposition where most of the increase is due to
reallocation between firms, the sectoral decomposition shows that most
of the increase in markups occurs within all sectors, not between
sectors. This is an important and unexpected discovery. Intuitively, we
would expect that certain sectors, such as technology, would see a much
bigger increase in the markup. But as the sector-specific markups in Online Appendix Figure 12.1
illustrate, there are no sectors that systematically have higher market
power. This confirms that the increase in market power occurs in all
sectors and industries.
Further evidence that
most of the rise in market power occurs within industry comes from
comparison of our results with those based on aggregate data
(industry-level or economy-wide). Using national accounts data by
sector, Hall (2018), extending his original work (Hall 1988), finds a rise in market power but only by about 20 points, half of the increase we find with firm-level data.
To
investigate where the discrepancy when using aggregate data comes from,
we use our firm-level data and aggregate them at the industry level. In
Figure V,
we plot our benchmark aggregate markup together with three series of
industry averages, based on our firm-level data, summed up to industry
averages: one where we treat the entire economy as one industry
(dashed-dotted line in print; solid blue in color version online), one
where we aggregate at the industry level with constant elasticities
(short-dashed line in print; long-dashed black line in color version
online), and one where we aggregate at the industry level and use the
estimated, time-varying and industry-specific elasticities (long-dashed
line in print; solid black line in color version online).
Figure V
Using Industry and Economy-Wide Averages versus Aggregating Microdata
The three series with averages look similar. The average
markups are below our benchmark, and it grows at half the rate. The
increase between 1980 and 2016 is from 1.15 to 1.35 approximately, by
about 20 points, as in
Hall (2018).
This clearly establishes that a substantial part of the increase occurs
within industry and that some of that change is lost when taking
averages. To see how that can occur, consider the following comparison.
To make the comparison as transparent as possible, we abstract from any
technological change or sectoral heterogeneity in output elasticities
and simply keep θ constant throughout. We compare our aggregate markup
with the one obtained using aggregate data:
The reason the two objects are not equal to each other is because of the heterogeneity in markups across firms.27
Aggregation of a nonlinear function (Jensen’s inequality) leads to
different outcomes. This is the case for any cross-section, but
importantly with the reported increasing skewness in the underlying
markup distribution, this difference becomes larger over time. The
widening gap between the micro and the macro ratios is simple economics:
if market share is reallocating toward the higher-markup firms, this
reinforces the process of increased skewness, due to the increased
correlation of markups and market share (in a given industry or in the
entire economy depending on the focus).
It is
clear from the figure that the aggregate-based series trend up but to a
much lesser extent, and this is to be expected given the increased
dispersion. This tells us that the dispersion and skewness of the
distribution have increased over time. Much of the rise that we observe
in the average markup disappears once we use industry or economy-wide
averages. This tells us that most of the heterogeneity in markups is
within industry and that the reallocation of market shares (see also
below) occurs mainly within industries.
III.D. Results from the U.S. Censuses
The data on publicly traded firms suffers from
selection. So far, what we have analyzed cannot be generalized to the
entire U.S. economy. The publicly traded firms tend to be large, and the
number of firms (fewer than 10,000) is small relative to the
approximately 6 million firms in the economy. Moreover, entry and exit
in the sample of publicly traded firms is nonrandom. Even though the
shares of GDP and of employment are large (because the firms are large),
we want to find out whether our results are representative for the
entire economy.
To that end, we repeat the
exercise for the censuses in different industries. The advantage of the
censuses is that they represent the universe of firms within a sector
and are therefore representative of the whole economy in that sector. We
focus on three censuses: manufacturing (NAICS codes 31-32-33),
wholesale (NAICS code 42), and retail (NAICS codes 44-45). We provide
more detail on the sample construction and measurement of the key
variables in Appendix B.
The measurement of markups in the census data relies on the framework outlined in Section II.
The implementation, however, differs because we do not observe the same
detailed information as in Compustat regarding a firm’s balance sheet
and income and loss statement, with the exception of the Census of
Manufacturing, for which we do observe most of the traditional
production and cost variables. The analysis of the manufacturing sector
will therefore closely track the analysis applied to the universe of
Compustat firms. There remains one big difference: to our knowledge
there is no analogue to the reporting of SG&A (or overhead cost) in
the census data.28
With
the exception of the Census of Manufacturing data, we only observe the
wage bill and sales consistently across plants and time. This implies
that output elasticities cannot be measured or estimated due to the
limited information on costs. For manufacturing, where there is more
detailed reporting of costs, we use the industry-time specific cost
shares as measures for output elasticities. For retail and wholesale, we
cannot impute the cost shares. Instead, we use the sector- and
time-specific output elasticities that we estimated from the publicly
traded firms.
In the Census of Manufacturing, we use the cost shares to
construct the output elasticity of any variable input (labor and
materials) at the four-digit NAICS industry level (denoted by
n) by census year.
29 This leads to the standard recovery of the output elasticity for the variable input:
where
j denotes a plant active in industry
n, in this case a unique four-digit NAICS code.
30 For manufacturing we can use information on materials and on the wage bill for the variable input
V.
This allows us to check the robustness of our findings. For the other
censuses, we only observe the wage bill. In the absence of information
on cost shares, we infer the output elasticities of labor using the
cost-share approach in Compustat. In particular for each two-digit NAICS
sector (
s), we compute the median labor cost share, by year, for the sample of active firms, as in
Section VI.A.
Finally,
we aggregate the plant-level markups to obtain firm-level markups, the
ultimate object of interest in this analysis. This also makes our
results consistent with the analysis performed for the Compustat sample.31
More specifically, we compute markups at the plant level and aggregate
to the firm level using plant-level revenue shares. The sector-specific
aggregate markup is computed as before, using a firm’s share in total
sectoral sales.
Figure VI
reports the weighted average (left panels) for each of the three
censuses, as well as the percentiles of the markup distribution (right
panels), weighted by sales (the equivalent of Figure III, Panel B). With data only in five-year intervals, the patterns are obviously less detailed.
Figure VI
Markups in the U.S. Censuses: Manufacturing, Retail, and Wholesale
The
variable input is employment. Averages and percentiles are revenue
weighted. Manufacturing firm-level markups rely on the industry-specific
cost shares. Retail trade relies on the output elasticities computed in
the Compustat sample. Wholesale relies on a calibrated output
elasticity.
Starting with Manufacturing (Figure VI,
Panels A and B), we see average markups that start to increase from
1977 onward, from around 1.55 up to around 1.8. This pattern mirrors
what we find in the whole sample of publicly traded firms and in the
publicly traded firms in manufacturing.
We
also calculate the markup using materials as the variable input, instead
of employment, and we find a very similar pattern. In the Compustat
sample, we cannot separate the labor and material expenditures, instead
we have to rely on the bundle COGS. The results indicate that all three
series (Compustat COGS-based, census labor-based, and census
materials-based) indicate the same pattern of rising aggregate markups.32
Like for the publicly traded firms, the pattern in retail (Figure VI,
Panels C and D) until 2002 is flat or only slightly increasing. This is
the case also for the percentiles. There is instead a sharp increase of
the weighted average in 2012 that we do not observe in the publicly
traded firms.
The figures for wholesale are
again in line with the series obtained from our analysis in the
Compustat sample. We observe a continuous decline in the aggregate
markup until 2002, after which we see an increase of about 15 percentage
points in the markup over the course of 10 years. The percentiles
highlight again that the rise is concentrated at the top of the
(weighted) markup distribution. In contrast to the results for the
manufacturing and retail census, we could not rely on reliable labor
cost shares to approximate the time-specific output elasticity. We
describe the procedure and compare the results to reported (aggregate)
profit margins in Online Appendix
Section 18, but the same message holds: the time-series markup pattern
is dominated by the dynamics in the sales-to-expenditure (here the wage
bill) ratio, and the output elasticity mostly affects the level.
IV. Market Power and Profitability
The documented rise in markups does not
necessarily imply that firms have more market power and therefore higher
economic profits. In fact, increasing markups can come from a variety
of reasons that are not associated with a decline in aggregate welfare.33
For example, a decrease in marginal costs, an increase in fixed costs
or innovation, an increase in demand or in its elasticity, a change in
the market structure, or new product varieties all lead to increasing
markups without necessarily implying higher profits.
Although
the textbook definition of market power is the case whereby a firm can
command a price above the marginal cost of production (markup), any
conclusions regarding whether market power increased will greatly depend
on the pattern of overhead costs, or any other factor affecting the
cost structure of firms (like innovation activities such as R&D).
Therefore, before we can conclude whether the higher markups are
associated with market power, we need to analyze profits. In the absence
of detailed data, the mapping from markups to market power (and
therefore welfare) can only be done through a particular model of the
economy.
With the accounting data available,
we assume that we can observe profits as the wedge between sales and all
variable and fixed costs (including innovation, advertising, and
others). In what follows, we consider higher market power a situation
whereby a firm can generate higher profits.34
Key
here is the evolution of overhead and capital as a share of
expenditure. If those have increased and markups have increased at the
same rate, then the higher markups are charged only to recover the
higher overhead costs and capital investment. In Figure VII we plot the evolution of overhead and capital as a share of total costs.35 In our data, the capital share has been fairly constant, in line with the findings by Barkai (2017).
Instead, overhead as a share of total expenditure has seen an increase.
The rise in overhead costs thus requires us to analyze profits to
conclude whether the rise in markups is associated with a rise in market
power.
Figure VII
Aggregate Overhead and Capital Cost Shares of Total Cost
We
proceed in two steps. First, we relate markups to recorded profits at
the firm level and contrast the observed markups to counterfactual
markups generated by a zero-profit condition. Second, we consider
aggregate profits and ask whether these are consistent with our
estimates of firm-specific markups and recorded fixed costs.
IV.A. Markups and Profits at the Firm Level
To calculate profits, we use the markup measure
and properly account for all costs, including the overhead (or fixed)
costs and the expenditure on capital. We then interpret this profit rate
as a measure of market power.
Let Πi=Sit−PVtVit−rtKit−PXtXit
denote net profits, where
PXtXit =
Fit denotes expenditure on overhead as measured by SG&A and is equal to the fixed cost.
36 Then the net profit rate
πit=ΠitSit can be written as:
where
we have substituted the expenditure on variable inputs as a share of
sales with the output elasticity over the markup, from
equation (7).
This measure of the profit share is different from the accounting
profits because it uses a measure of capital obtained from the balance
sheet, not the income statement. With adjustment frictions, the
accounting measure does not adequately reflect the expenditure on
capital. Note also that our measure of profits incorporates the output
elasticity of the production technology, which takes into account that
the variable factors of production
V adjust while the fixed factors do not.
Figure VIII,
Panel A plots the average revenue-weighted profit rate for the data in
our sample. We find that profits have gone up by about 7 percentage
points between 1980 and 2016.37 Underlying the rise in profits is the increase in the upper tail of the profit distribution. In Figure VIII,
Panel B we plot the kernel density of the unweighted profit rate
distribution in 1980 and 2016. The rise in average profit rate is nearly
exclusively driven by the increase in the upper percentiles of the
profit distribution. More firms have extremely high profit rates of 15%
and higher. Consistent with the results on markups, the average profit
rate increase is driven in part by the reallocation of economic activity
toward high-profit, dominant firms.
Figure VIII
Average Profit Rate and Profit Rate Distribution
Our
measure of the profit rate is the firm profits as a share of sales,
which effectively scales those profits by the firm size as measured by
its revenue. From an investment viewpoint, we may want to measure the
return on assets. The return on assets is calculated as the firm profits
divided by its assets. We define profits by sales minus all costs,
COGS, SG&A, and the expenditure on capital. Because the expenditure
on capital is included, our measure of return on assets is the return
over and above r, which includes the inflation-adjusted
risk-free rate, as well as an adjustment for depreciation and risk.
Therefore, it is the excess return on assets. We plot this in Online Appendix
Figure 9.1a together with our baseline profit rate. The return on
assets is remarkably similar to the profit rate, with an increase
starting in 1980 and rising from around 1% to around 8% in 2016. This
average return on assets is weighted by the capital of each firm. When
we weight it by the sales of each firm (Online Appendix
Figure 9.1b), then the average return on assets is higher and also
rising faster. Firms with high sales have higher returns on assets, and
the large firms have seen a bigger rise in their returns.
All
this seems to suggest that at least based on the flows reported in the
accounting data, starting in 1980 there has been an increase in the
profitability of firms, and therefore an increase in market power. Note
that the profit rate we have reported accounts for the increase in
contemporaneous overhead costs as measured by SG&A. Of course, some
costs may have been incurred earlier. Still, it is not clear what those
startup costs may be as they are not booked in the firms’ accounts, and
firms have incentives to book as many costs as possible to reduce
corporate taxes on profits. The only possibility is that those startup
costs were incurred before the firms were observed in our data. As a
result, profits based on contemporaneous costs may therefore be
overstated. What the data are indicating, however, is that if such costs
are incurred earlier, there must be an increase in those startup costs
as a share of the sales of a firm since 1980. With free entry and hence
zero ex ante expected profits, what we expect is that over the past four
decades, the unmeasured startup cost as a share of future sales has
gone up from 1% of sales to 8% of sales (roughly from 2% of value added
to 16%). Some of those costs could be R&D costs that were incurred
before the firms were observed in our data. We turn to the impact of
recorded R&D costs below.
The flow of
profits may not be the best measure of profitability of the firm,
because it mixes up the firm’s result with investment decisions. To that
effect, we consider a measure of profitability based on what firms
generate as a return to their shareholders. For that we have two
measures: (i) the market value (or market capitalization), and (ii)
dividends. Our second measure, dividends, is the return an investor
receives on holding equity in the firm. Of course, dividends may vary
for reasons that have nothing to do with the actual flow of profits. In
particular, they will be closely related to the investment opportunities
the firm has. Still, over a long enough horizon and averaging out over a
large number of firms, we would expect that dividends are a good
indicator of profits. Our first measure, market value, is essentially
the discounted sum of dividends, since a shareholder who sells shares in
a firm gives up the opportunity value of receiving the indefinite
stream of dividend payments. In contrast to the actual dividends, the
market price is more a measure of future expected profits, not just
contemporaneous profits, since it takes into account the flow of all
expected future dividends.
Figure IX, Panel A shows the evolution of the market value as a share of sales, averaged by the sales share in the entire economy: ∑iSi∑iSiMktValiSi=∑iMktVali∑iSi
.
Unlike standard composite indices of stock market values like the
S&P 500, this measure is a “rate” that can be interpreted in
conjunction with the profit rate π (profits as a share of sales) from
our model. As such, first, it is not affected by inflation38
and second, this measure is independent of the size of firms or the
composition of firms because it is normalized by sales. For example,
even if there are 500 firms in the index, the index will artificially
grow when firms become larger, for example, due to mergers.39
Figure IX
Market Value and Dividends
If
the flow of profits and dividends as a share of sales were constant,
then the market value that reflects the discounted stream of dividends
would be constant as a share of sales. This is clearly not the case.
Market value as a share of sales rises from less than 50% in 1980 to
over 150% in 2016 (Figure IX,
Panel A, right scale). A similar pattern arises for dividends, where
dividends as a share of sales increases from 1.7% in 1980 to over 3.5%
in 2016 (Figure IX, Panel B).40
This
is not just an artifact of the aggregate data. At the individual firm
level, firms with higher markups also have higher market values and
dividends. In Table II we report the regression results.41
Not surprisingly, contemporaneous firm-level markups are correlated
with both market value and dividends. For all specifications, the
coefficient is highly significant (even in the presence of firm fixed
effects, see columns (4) and (8)). At the firm level, this is consistent
with the fact that higher markups reflect higher profits and therefore
higher dividends and market values.
Table IIFirm-Level Regressions: Market Values and Dividends on Markups
| (1) | (2) | (3) | (4) |
| (5) | (6) | (7) | (8) |
|---|
| | ln(marketvaluesales) |
| | | ln(market value) |
| ln(markup) | 0.71 | 0.64 | 0.56 | 0.17 | | 0.71 | 0.65 | 0.58 | 0.27 |
| | (0.03) | (0.02) | (0.02) | (0.03) | | (0.02) | (0.02) | (0.02) | (0.02) |
| ln(sales) | | | | | | 0.81 | 0.81 | 0.83 | 0.68 |
| | | | | | | (0.00) | (0.00) | (0.00) | (0.01) |
| Year fixed effects | | Y | Y | Y | | | Y | Y | Y |
| Sector fixed effects | | | Y | | | | | Y | |
| Firm fixed effects | | | | Y | | | | | Y |
| R2 | 0.05 | 0.13 | 0.21 | 0.68 | | 0.68 | 0.71 | 0.73 | 0.89 |
| |
| | ln(dividendssales) |
| | | ln(dividends) |
| ln(markup) | 1.05 | 0.97 | 0.80 | 0.26 | | 1.03 | 0.93 | 0.78 | 0.26 |
| | (0.04) | (0.03) | (0.04) | (0.05) | | (0.04) | (0.04) | (0.04) | (0.05) |
| ln(sales) | | | | | | 0.94 | 0.92 | 0.93 | 0.76 |
| | | | | | | (0.01) | (0.01) | (0.01) | (0.02) |
| Year fixed effects | | Y | Y | Y | | | Y | Y | Y |
| Sector fixed effects | | | Y | | | | | Y | |
| Firm fixed effects | | | | Y | | | | | Y |
| R2 | 0.06 | 0.11 | 0.17 | 0.70 | | 0.66 | 0.68 | 0.70 | 0.89 |
Table IIIRegressions:
Effect of SG&A, R&D Expenditure, and Advertising Expenditure on
Markups and Profit Rate; Extensive Margin Effect of R&D and
Advertising
|
| Markup (log) | Profit rate (log) |
|---|
|
| (1) | (2) | (3) | (4) | (5) |
|---|
| SG&A (log) | | 0.56 | | | 0.15 | |
| | | (0.01) | | | (0.03) | |
| R&D exp. (log) | | | 0.16 | | | 0.10 |
| | | | (0.01) | | | (0.01) |
| Advertising exp. (log) | | | 0.05 | | | 0.03 |
| | | | (0.00) | | | (0.01) |
| R&D dummy | | | | 0.06 | | |
| | | | | (0.01) | | |
| Advertising dummy | | | | −0.00 | | |
| | | | | (0.03) | | |
| R2 | | 0.61 | 0.07 | 0.43 | 0.04 | 0.05 |
| N | | 26,743 | 247,615 | 26,743 |
Based on the
evidence from the firm’s fixed overhead as measured by SG&A and the
resulting profits and by market value and dividends, we find evidence
that the rise in markups is associated with the rise in market power.
To complete this section, we investigate the relation between profits, markups, and overhead costs (SG&A). In Figure X,
Panel A we plot the relation between the share of sales of SG&A and
the markup for different percentiles in the (unweighted) markup
distribution. This shows that the firms with a higher SG&A share of
sales have higher markups. For a given year, the higher percentiles in
the distribution of markups have higher overhead shares. This is as
expected in a competitive economy: higher prices relative to marginal
cost are required to offset the overhead and avoid making losses. In
addition, over time, the overhead share is increasing which
automatically implies that the markup increases, even in a competitive
economy. Note that if we plot the markup against the share of COGS in
sales, then by construction this relation is downward sloping,
indicating that unlike SG&A, COGS is a variable input.
.
Figure X
Markup, Excess Markup, and SG&A Share (Markup PF2)
Figure XI
Decomposition of Equation (15) due to Overhead Costs and Aggregation
Now we want to evaluate whether the increase in markups
that we observe is merely to offset the rise in overhead. To that
effect, we calculate a fictitious markup, denoted by μ
⋆, that corresponds to zero profits. We obtain that markup from setting profits π
it to zero in
equation (15) and solving for μ:
This
zero-profit markup is a weak upper bound, however, and the true
zero-profit markup is weakly lower (provided there are no costs in
addition to COGS, SG&A, and capital). This is because we do not know
what sales
Sit would be under competition.
To predict sales under perfect competition, we need to know the
properties of demand. Only in the case of unit elasticity demand will
sales be invariant for different markups. In all other cases, however,
sales under perfect competition will be lower than when there is market
power. This is due to the fact that firms are charging higher prices
only if the marginal revenue is positive, which by definition
necessarily implies higher sales for higher markups. Therefore, sales
under perfect competition (
S⋆it) will be weakly lower than under market power. Under our assumption that in the short run
Kit and
Xit are not variable, the expression in
equation (14) where we use
Sit instead of
S⋆it is weakly higher than the true zero-profit markup.
In Figure X, Panel B we also plot μit−μ⋆it
for different percentiles in the markup distribution. Because
μ⋆it is the upper bound of the zero-profit markup, the gap between the actual markup and
μ⋆it
indicates the extent of the excess markup, over and above the markup
that arises under perfect competition. We see that the excess markup is
highest for the highest percentiles of the markup distribution, where
incidentally the SG&A share is the highest as well. High-overhead
firms have high markups but also high excess markups; and this became
stronger over time (the excess markup rose from about 0.2 in 1980 to
about 0.6 in 2016).
When we analyze the
relation between markups (and profits) and overhead at the individual
firm level, we find a strong positive relation, as expected. As we have
pointed out all along, one of the reasons for raising prices and markups
is that overhead has increased. The elasticity is 0.56 (see Table III):
only just over half of the SG&A increases are passed on to markups.
In a competitive economy this should be 1. Interestingly, firms with
higher SG&A also have higher profits. In a competitive market, this
coefficient should be 0. We can decompose the change in SG&A into
R&D expenditure and advertising expenditure. These are often
signaled as the components of SG&A that are important for intangible
capital. Indeed, R&D expenditure has risen from 5% in 1980 to 20%
of SG&A, and advertising from 4% to 10%. Even in 2016, these remain
relatively minor shares of SG&A. The majority is still sales related
and administrative expenditure. We find that the elasticity of R&D
expenditure on markups is 16% and 5% for advertising expenditure.
Interestingly, most of that effect remains when the dependent variable
is the profit rate. This elasticity should be 0 under competition. Most
of R&D and advertising expenditures translate into profits as much
as they do into higher markups. These are all at the intensive margin.
When we evaluate the extensive margin—whether a firm does or does not
have expenditures on R&D or advertising—we find an elasticity of 6%
from R&D and no significant effect from advertising (since nearly
all firms have advertising expenditure, there is not enough variation;
only about 10% of the firms report R&D expenditure).
In
sum, at the firm level, we find consistent evidence that profits and
the market valuation of firms have gone up together with markups.
Markups are not higher only to compensate for higher fixed costs, they
are also higher because firms exert market power.
IV.B. Aggregate Profits and Markups
Even though markups and profit rates are different
concepts—most notably because of the inclusion in profits of total
costs, including overhead costs—they are related. In particular, there
is an identity that links profit rates and markups and that holds for
any technology
C(
Q), as has been pointed out by
Syverson (2019) and
De Loecker and Eeckhout (2018b):
where
ACitMCit is the ratio of average cost to marginal cost and because
ACit=C(Qit)Qit and
μit=PitMCit.
Now
there is a puzzle. The aggregate markup of 1.61 that we calculate in
2016 cannot be reconciled with the profit rate of 8%. In particular, Basu (2019)
has pointed out that something must be wrong with our markup measure,
because the implied profit is too high. If we plug in the aggregate
markup in 2016 and assume that the ratio of average cost to marginal
cost is equal to 1, then the implied profit rate is 38%.42
There
are two problems with this argument. The first is that in this thought
experiment, we have assumed that the average to marginal cost ratio is
constant and equal to 1. We know from Figure VII
that the fixed cost is sizable and has gone up. Therefore the average
to marginal cost ratio is neither constant nor equal to 1.
The second problem with this argument is that it erroneously relies on a representative-firm framework.
Equation (15) strictly holds at the firm level. In the aggregate, this translates into:
where μ
t = ∑
imitμ
it,
ACt = ∑
imitACit,
MCt = ∑
imitMCit. Therefore, the premise of a representative-firm framework is counterfactual.
Once
we correct for these counterfactual assumptions—that the average cost
to marginal cost ratio has increased and that we properly aggregate
without assuming a representative-agent framework—the implied average
profit rate of 8% in 2016 and the markup of 1.61 are indeed consistent.
In Figure XI, we decompose equation (15). We also report the actual values in Table IV.43
When we assume both a representative firm and a constant average to
marginal cost ratio equal to 1 (no fixed cost), we see profits rise from
18% in 1980 to 38% in 2016 (solid line). When we adjust for the
observed average to marginal cost ratio but keep the representative firm
assumption (long dashes in print; solid green in the color version
online), the profits drop by more than half over the entire period. When
we adjust for proper aggregation (drop the representative firm) and
keep a constant average to marginal cost ratio (very short dashes in
print; solid black in the color version online), profits drop by about
one-third. Note that the gap is larger towards 2016 than in 1980, which
is consistent with the fact that the distribution of firm sizes and
markups has become more dispersed, resulting in a bigger gap between the
aggregate and the average (due to Jensen’s inequality). Finally, when
we adjust for both proper aggregation and the observed average to
marginal cost ratio (dash-dot in print; solid purple in the color
version online), profits are close to the observed profits in the data
(short dashes in print; solid red in the color version online).
|
| Profit rate |
|---|
|
| 1980 | 2016 |
|---|
| Use average | No fixed cost | 0.17 | 0.38 |
| Use average | Fixed cost | 0.04 | 0.20 |
| Aggregation | No fixed cost | 0.14 | 0.24 |
| Aggregation | Fixed cost | 0.00 | 0.05 |
Overall, the
relation that predicts average profit rates as a function of markups
fits the data once we properly account for returns to scale (fixed
costs) and once we properly aggregate. This indicates that our measure
of markups does not predict an outlandish profit rate. What it does
confirm is that markups and profit rates are different objects and that
we should be careful comparing them. Too often, they are used
interchangeably.
Finally,
Traina (2018)
proposes a different measure of market power that includes both COGS
and SG&A. His measure is therefore closely related to the profit
rate. Denote by τ:
where
pVV is the expenditure on the variable input,
pXX is the expenditure on overhead as measured by SG&A and where θ
V + X ≈ 0.95 (though he estimates a separate elasticity for each sector).
This ratio is directly related to the operating profit rate, the definition of which is
We can therefore write the measure τ
it as
Given this identity, this measure is closely related to the aggregate operating profit rate πOPXt=∑imitπOPXit
(see Figure F.1 in Appendix F).
We find an increase in the operating profit rate between 1980 and 2016
of about 7–8 percentage points, and for the measure τ, which we
interpret as an alternative measure of the profit rate, we see an
increase of about 10 points, from 1.08 to 1.18.
In
sum, aggregate markups and profitability are both increasing. Therefore
the rise in markups is not exclusively due to the rise in overhead
costs. This is evidence of the rise in market power.
V. The Macroeconomic Implications
The focus of our analysis so far has been on
documenting in detail the time-series and cross-sectional evolution of
markups and profitability. We now turn to discussing the macroeconomic
implications of the rise in market power in the past decades.
V.A. The Secular Decline in the Labor Share
In the national accounts, the labor share of
income measures the expenditure on labor (the wage bill) divided by the
total income generated (value added). Although there are business cycle
fluctuations, the labor share has been remarkably constant since World
War II up to the 1980s, at around 62%. Since 1980, there has been a
secular decline all the way down to 56% (Bureau of Labor Statistics
Headline measure).44 The decline since the 1980s occurs in the large majority of industries and across countries (see Gollin 2002; Karabarbounis and Neiman 2013).
Economists
have struggled to understand the mechanism behind the decline in the
labor share. One obvious hypothesis, ex ante, would be a within-firm
substitution of labor for capital. This hypothesis is explored most
prominently in Karabarbounis and Neiman (2013),
which argues that a secular decrease in the relative price of
investments goods led firms to substitute away from labor toward capital
and can explain half of the decline in labor’s share of income. The
basic problem with this mechanism is that it rests crucially on a high
elasticity of substitution between capital and labor (higher than 1).
While Karabarbounis and Neiman (2013) claim that this elasticity is 1.25, the overwhelming majority of several decades of empirical studies (Antràs (2004),
among many others) find that this elasticity is much lower than 1. The
combination of a low elasticity of substitution between capital and
labor, with the fact of a declining labor share of income, has been
especially puzzling.
Koh, Santaeulalia-Llopis, and Zheng (2017)
offer yet another explanation, which is based on the increasing
importance of intangible capital and its incomplete measurement as part
of capital in aggregate data. Firms now invest substantially more in
intellectual property products, and this leads to a lower expenditure on
labor.45
However, in their world with perfect competition, this measurement
issue should not lead to an increase in the total profit share. As we
have documented, there is a substantial increase in the profit rate. If
intangibles play a role, it must allow firms to exert more market power,
which is the central thesis of our article. We do find evidence that
expenditure on overhead has increased (see below), which could certainly
include intangibles, but we also find that economic profits increase
even if we interpret overhead (and hence intangibles) as a factor of
production (see Figure VIII, Panel A). Finally, Elsby, Hobijn, and Şahin (2013)
find little support for capital-labor substitution, nor for the role of
a decline in unionization. They do find some support for offshoring
labor-intensive work as a potential explanation.
In the context of our setup, the change in the markup has
an immediate implication for the labor share. Although we have
calculated the markup from all variable inputs, we could do so as well
for labor alone. Then rewriting the first-order condition (
7) where
V =
L,
PV =
w, and θ
V = θ
L, the output elasticity of labor, we obtain that at the firm level the labor share satisfies
Observe
that if there are multiple inputs that are fully variable, the
estimated markup should be the same. So even if the markup is calculated
for the bundle V, it should also hold for L as long as both V and L
are variable. Profit maximization by individual firms thus implies that
the labor share is inversely proportional to the markup. As the markup
increases, we expect to see a decrease in the labor share.
Unfortunately
Compustat does not have good data for the wage bill. Because reporting
compensation to the SEC is not compulsory, the variable XLR for total
compensation is heavily underreported.46
Because of selection in the sample of those firms that do report total
compensation, we need to be cautious interpreting the aggregate labor
share outcomes.
Despite the shortcomings of our data, we can nonetheless verify the firm’s optimization condition (20) at the firm level. In Table V
we report the regression coefficients of the log of the labor share on
the log of the firm’s markup. The first four specifications only differ
in the fixed effects that are included. We consistently find a negative
coefficient of around −0.20 to 0.24. As a firm’s markup increases by,
say, 10%, its labor share decreases by 2–2.4%.
Table VRegressions: Log (Labor Share) on Log (Markup)
|
| Labor share (log) |
|---|
|
| (1) | (2) | (3) | (4) | (5) | (6) |
|---|
| Markup (log) | | −0.24 | −0.23 | −0.20 | −0.24 | −0.68 | −0.73 |
| | | (0.03) | (0.03) | (0.03) | (0.03) | (0.02) | (0.02) |
| Cost share (log) | | | | | | 0.91 | 0.96 |
| | | | | | | (0.01) | (0.01) |
| Year FE | | | X | X | X | X | X |
| Industry FE | | | | X | | X | |
| Firm FE | | | | | X | | X |
| R2 | | 0.02 | 0.08 | 0.21 | 0.88 | 0.93 | 0.99 |
| N | | 24,838 |
To extrapolate
these firm-level results to the aggregate economy, we need to keep in
mind that there is no such thing as a representative firm in this
context. The rise of average markups is distributed unequally, and
increasingly so. Most important, since two-thirds of the rise in market
power is due to reallocation of economic activity toward high-markup
firms, the effect of markups on the labor share in the aggregate is
predominantly driven by a few large firms with high markups and a low
labor share. Our findings for the firm-level markups are thus consistent
with those in Autor et al. (2020) and Kehrig and Vincent (2017)
for the Census of Manufacturing. In sum, we find firm-level evidence of
the direct inverse relation between markups and the labor share that we
obtain from the first-order condition (20).
In
the table, we also analyze whether we can reject any evidence that
there is perfect competition. The fifth and sixth columns report the
same regression where we now include the log of the cost share (labor
over total cost) as a covariate. Under perfect competition, the
coefficient is 1. Here we find a coefficient significantly smaller than
1, indicating that there is a wedge between sales and costs. Equally
important, any other covariate (in this case markup) should be
insignificant. We find instead that the coefficient on the markup is
highly significant and negative. This indicates that there is evidence
of noncompetitive price setting.
V.B. The Secular Decline in the Capital Share
The same logic for the decline in the labor share also applies to materials M,
that is, variable inputs that are used in production. Those are
included in our variable cost measure COGS. Now if we consider the
evolution of capital expenses, which is not included in our measure of
variable cost and which adjusts at a lower and more long-run frequency,
then the increase in markup has implications for the capital share.47
In the long run and once the adjustment frictions are taken into
account, higher output prices and lower output quantities eventually
will lead to a decrease in the capital share. While the decline in the
labor share is widely discussed, the decline in the capital share has
received much less attention.48
Assuming a static environment, the following equality has to hold:
The
labor share and the capital share sum up to 1 minus the profit share
minus the overhead share. We have established that the profit share and
the overhead share increase, so the right-hand side decreases. With
complementary capital and variable inputs, and over a long enough time
horizon for capital to adjust, the expenditure on capital rK as
a share of output will be decreasing over time. In fact, if capital
were fully flexible, it would adjust according to the equivalent of
first-order condition (7) rKPQ=θKμ
which relates the capital share to the inverse of the markup.
In Online Appendix
Figure 14.1b we document the evolution of the capital share for the
firms in our data. Not surprisingly this measure is quite volatile
because it is a long-term measure that adjusts at a lower frequency and
is more subject to aggregate fluctuations. Also, before the 1980s,
capital investment was particularly low because of tumultuous financial
times: inflation was high and financial frictions were considered
higher. What we learn from the figure is that there was a decrease in
the capital share from around 12% in 1980 to 8–10% toward the end of the
sample. In the aggregate, the capital share is correlated with the
inverse of our markup measure. With a long enough horizon, capital
investment adjusts and hence there will be a reduction in capital
investment as markups increase.49
As with the labor share, we can also investigate the firm-level relation between the capital share and markups. In Table VI
we report the regression coefficients for different specifications. We
find that without firm fixed effects, there is no significant relation
between markups and the capital share. This may be indicative of the
adjustment costs that firms face when investing in capital. Instead,
with firm fixed effects, there is a significant negative effect, with an
elasticity of −0.14. When we include the cost share, the coefficient on
the cost share is larger than 1. Under variable adjustment of capital,
perfect competition would require this to be equal to 1, and less than 1
with market power (see, for example, Table V
for the labor share). The fact that the coefficients on the cost share
here are larger than 1 indicates that capital does not adjust
frictionlessly.
Table VIRegressions: Log(Capital Share) on Log(Markup)
|
| Capital share (log) |
|---|
|
| (1) | (2) | (3) | (4) | (5) | (6) |
|---|
| Markup (log) | | 0.03 | 0.03 | −0.02 | −0.14 | −0.90 | −0.86 |
| | | (0.02) | (0.02) | (0.01) | (0.02) | (0.00) | (0.00) |
| Cost Share (log) | | | | | | 1.13 | 1.11 |
| | | | | | | (0.00) | (0.00) |
| Year FE | | | X | X | X | X | X |
| Industry FE | | | | X | | X | |
| Firm FE | | | | | X | | X |
| R2 | | 0.00 | 0.02 | 0.31 | 0.83 | 0.98 | 1.00 |
| N | | 242,692 |
V.C. The Secular Decline in Low-Skill Wages and Labor Force Participation
An increase in markups implies a decrease in
aggregate output produced, whenever demand is not perfectly inelastic.
Lower output produced then implies lower demand for labor. This results
in both lower labor force participation and lower wages. Even if supply
is perfectly elastic, real wages decrease with market power because the
price of the output goods has increased.
There
is ample evidence of the stagnation of wages in the lower half of the
distribution. The median weekly wage in constant prices has changed
barely since 1980, from $330 to $345 (1982 prices, source Current
Population Survey). But there has been technological progress, and the
share of median wages out of GDP has nearly halved, because over three
and a half decades GDP has nearly doubled. In the past few decades,
labor force participation has also been decreasing from 67% in the 1990s
to 63% now. Most strikingly, while the gender gap has continued to
close, in the past two decades female labor force participation is also
decreasing.
The quantitative investigation of
the effect of market power on low-skill wages and labor force
participation is beyond the scope of the current article. In De Loecker, Eeckhout, and Mongey (2018)
we construct an oligopolistic framework for firm dynamics that
quantitatively accounts for these general equilibrium implications of
the rise in market power. We find that market power indeed has an effect
on equilibrium wages, and that quantitatively, that effect is large.
Our quantitative model predicts that real wages as a share of GDP drop
by over 26%, consistent with what we see in the data.
V.D. The Secular Decline in Labor Reallocation and Migration Rates
It is well known that in an environment with
market power, shocks to productivity and costs are not translated one
for one into prices. In a competitive market, firms face a perfectly
elastic demand and any decrease in costs is passed on to the consumer,
where prices decrease by the same amount as the decrease in costs. With
market power however, the pass-through of cost shocks to prices is
generally incomplete.50 Crucial for our finding is that the higher the degree of market power by firms, the lower the pass-through.
Now
consider an environment where firms have market power and face shocks
to their productivity. With positive shocks, firms face lower costs and
adjust their inputs (say, labor) upward. With negative shocks, they
adjust inputs downward. Because pass-through is lower in the presence of
higher market power, the rise in market power will give rise to lower
degree of adjustment of the variable inputs, including labor, for the
same shock process.
This is precisely what Decker et al. (2014)
find for the U.S. economy over the past three decades. The volatility
of shocks has not decreased, but rather the responsiveness of firm’s
output and labor force decisions to the existing shocks has declined.51
The rise in market power can thus rationalize the decrease in labor
reallocation across firms, even if the observed shocks to firm
productivity have remained constant.
The
decrease in labor market dynamism is evident in the decrease of labor
reallocation as well as in the decrease of job-to-job transitions,
nonemployment to employment transitions, and employment to nonemployment
transitions.52
The decrease in market power and the resulting decrease in labor
reallocation can also rationalize the fact that migration rates across
U.S. states and metropolitan areas have decreased by nearly half from
around 3% in 1980 to 1.5% in 2016.53
If firms are based in different local labor markets and a fraction of
all job relocation decisions are between local labor markets, then lower
job flow rates will automatically give rise to lower migration rates.
We assess the quantitative significance of the impact of the rise of
market power on labor reallocation and migration in a companion work (De Loecker, Eeckhout, and Mongey 2018).54
VI. Discussion and Robustness
Here we discuss the features of our model and report a number of robustness exercises.
VI.A. Cost Shares
We repeat the analysis where we obtain the
output elasticity from cost shares. For each firm, we have an
observation for the cost share αVit=PVtVitPVtVit+rtKit
. Within an industry, we use the median of the distribution as the measure for the output elasticity:
θst=mediani∈s{αVit}.
Figure XII,
Panel A reports the sales-weighted average of the markups with the
output elasticity derived from the cost share for the traditional
production technology where overhead is a fixed cost and denoted by CS.
The pattern is very similar to that in Figure I.
There is a moderate decrease from the 1960s and then an increase from
1980 up to 2016. The level is slightly higher, and the increase by 50
points is somewhat more pronounced.
Figure XII
Cost-Share Based Aggregate Markups and Technology
Panel
A reports the aggregate markups using cost shares (median cost shares
for each year and sector). Panel B compares the sector-weighted cost
share to the estimated output elasticities; weights are sectoral total
sales.
From inspection of the definition of the markup in equation (7),
the rise in the markup could potentially be attributed to two sources:
(i) an increase in the ratio of sales to expenditure on variable inputs;
or (ii) technological change, an increase in the output elasticity θV over time. In Figure XII, Panel B we plot the average cost share of the factors of production V and K
as well as the average output elasticity estimated from the benchmark
technology. There is some volatility in the cost shares, but they are in
line with the estimated output elasticity. This indicates that the
steep increase in markups is driven by the increase of sales over
expenditure on inputs. Firms are selling their goods at higher margins.
This is also evident from inspection of Figure II,
Panel B, confirming again that the evolution of the share-weighted
average markup is mainly driven by the ratio of sales to expenditure on
variable inputs and not by changes in the output elasticity.
VI.B. Production Function with Overhead as a Factor of Production
The conventional production function uses as factors of production the variable input V and capital K.
All other expenditures accounted for as not directly related to the
production of the goods sold are overhead. They are considered fixed
costs, a cost incurred that is independent of the output produced. This
is the standard approach in the industrial organization literature on
markup estimation.
In contrast to the
conventional interpretation of the production technology, we propose an
alternative interpretation where a portion of the overhead is a factor
of production. Higher expenditure on getting more and better logistics
managers will lead to an increase in the units produced. More sales
people increases the units sold. To interpret overhead as a factor of
production, we denote its expenditure by pXX, where the quantity that enters the production technology is X and the unit price is pX.
We
now take this nonconventional interpretation of overhead as a factor of
production seriously and assume that all of it is a factor. The
production function can then be written as Q(V, K, X) and firm profits are PQ(V, K, X) − PVV − rK − PXX. We can apply the same cost-based method for the derivation of markups as laid out in Section II. We treat X as a factor of production that enters the production function, but it is nonvariable, just like capital K. The treatment of the variable input V
remains as before. The difference relevant for the measurement of
markups therefore stems from the production function estimation and the
resulting estimate for θV. To differentiate, we
denote the estimates from this production function by PF2. When we
calculate elasticities based on cost shares that take into account
overhead as a factor of production, we refer to it as CS2. To further
differentiate the graphical representation, PF1 and CS1 are plotted in
red, and PF2 and CS2 are in blue (color versions of all figures are
available online).
Figure XIII,
Panel B plots the cost share of variable factors in the total cost
(consisting of variable factors, capital, and overhead), as well as the
cost share of overhead. We see that there is a slight decrease in the
cost share of the variable factor of production from 80% in the
beginning of the sample to 70% in 2016. The share of the fixed cost has
increased from 18% at the beginning to 24% toward the end. This is
indicative of the fact that the overhead cost, and thus the technology,
has changed. The estimated output elasticities confirm this pattern,
although importantly they do not necessarily have to sum to one
(including rK of course).
Figure XIII
Average Markups, Elasticities, and Cost Shares for Production Function with Overhead as a Factor
Output elasticities from estimated PF2 and from CS2: time-varying, sector-specific (two-digit) output elasticity θst (revenue-weighted average).
In Figure XIII,
Panel A, we report the evolution of the average markup with this new
production technology. Qualitatively, we see a similar pattern for the
increase in the average markup starting in 1980. Initially around 1, the
average markup increases by about 30 points by 2016. The increase for
this technology is 10 percentage points lower than for the traditional
production function (Figure I). This difference is driven by the fact that the cost share of overhead (and the estimated output elasticity θX) is increasing over time (see Figure XIII, Panel B). What matters for the markup estimate however is the elasticity θV. We know that it is roughly constant for the conventional production function (Figure XII, Panel B). For the production technology with overhead as a factor of production, θV is slightly decreasing (Figure XIII,
Panel B). Therefore the estimated markup shows a more moderate increase
(30 points) than under the conventional production technology (40
points).
VI.C. Returns to Scale
VI.C. Returns to Scale
With the estimated technologies, we can evaluate
any technological change that affects the returns to scale. Because the
technology is Cobb-Douglas, the returns to scale are measured by the
sum of the output elasticities: θV + θK for PF1 and θV + θK + θX for PF2. We find that the estimated technology shows a rise in the degree of increasing returns over time. In Figure XIV,
Panel A we report the sum of the output elasticities for both
technologies PF1 and PF2. For the conventional technology (PF1), from
the start of the sample, the estimated returns to scale go from around
1.02 in 1980 to 1.08 in 2016. We estimate an increase in the returns to
scale of the technology with overhead as a factor of production, from
1.07 up to 1.13, reaching 1.22 in 2010. The fact that the production
function with overhead as an input has higher returns to scale confirms
that overhead X is in part a fixed cost that generates
increasing returns. Moreover, those returns to scale are increasing more
over time as overhead increases, which establishes that the role of
overhead as a source of returns to scale is growing.
Figure XV
Average Markups for Production Function with Overhead as a Factor
Output elasticities from estimated PF2 and from CS2: time-varying, sector-specific (two-digit) output elasticity θst (revenue-weighted average).
An alternative way to measure returns to scale is with a method first used in
Syverson (2004). While using cost shares implicitly assumes that the technology is constant returns,
Syverson (2004)
adjusts the technology based on cost shares and derives the returns to
scale. He assumes the following functional form for the technology based
on cost shares but without constant returns:
with all variables in logs, where
αV=PVVPVV+rK+PXX is the cost share of the variable input, and likewise for αK and αX.
While
each cost share determines the output elasticity, the technology need
not be constant returns and the curvature is captured by γ. In Figure XIV,
Panel B we plot two measures of the estimated γ, one for the average
firm-level γ and one where we impose a common γ at the two-digit
industry level for the technology with overhead as a factor of
production.55
These graphs reveal that also with this method, returns to scale have
increased throughout the sample. There were decreasing returns to scale
before 1980 and since 1980 returns to scale have been increasing, up to
1.05 at the end of the sample.
The increase
in the returns to scale also explains why the markup estimate based on
cost shares only shows an increase of 20 percentage points (Figure XV,
Panel A), whereas under the elasticity estimated from the production
function the increase since 1980 is 30 percentage points (Figure XIII,
Panel A). By construction, cost shares add up to 1, and therefore the
implied elasticities are derived under the assumption of constant
returns. As a result, the increase in the elasticity θX due to an increase in the expenditure share of overhead must necessarily lead to a decrease in θV.56 With θV decreasing, from equation (7),
the increase in the markup must necessarily be dampened. This
illustrates that directly using the cost shares can by construction not
account for any change in the returns to scale in the technology.
The evolution of returns to scale helps us understand the difference between Figures XIII, Panel A and XV,
Panel A. In the latter, we ignore the change in the returns to scale
because cost shares are implicitly assuming CRS. If instead we use the
elasticities obtained for the Syverson (2004) technology in equation (22), which is equal to γαv
, we obtain an average markup (see Figure XV,
Panel B) that is very similar to the one using the elasticity estimated
with the production function (PF2). The increase in γ in Figure XIV, Panel B implies that the elasticity αV used in Figure XV, Panel A is multiplied by γ.
Finally, in Online Appendix 13 we also analyze the returns to scale using the data from the censuses.
VI.D. Input Weights and Joint Distributions
We have shown in Section III
that to calculate aggregate markups, the choice of the weighting
measure matters. Because we are interested in the entire joint
distribution of markups and firm characteristics (revenue, costs,
inputs, etc.), having information on as many moments as possible
provides more detailed insights about the evolution of markups.57
Here
we report the average markup using different weights. To get a better
idea of the definition of each of the weights, we report them in Table VII.
Table VIIAggregate Markups: Variation by Technology (θ) and Weighting (mit)
| Output elasticity | Revenue weight |
| Input weight |
|---|
|
|
|
| 1 Input | All inputs |
|---|
|
| mit=RitRt |
|---|
|
|---|
| Economy-wide θ | | θ∑imitRitCOGSit |
| |
| Sector-specific θst | | ∑iθstmitRitCOGSit |
Table B.1Summary Statistics (1955–2016)
|
| Sample A |
|---|
| Acronym, var. | Mean | Median | No. obs |
|---|
| Sales | SALE, PQ | 1,922,074 | 147,806 | 247,644 |
| Cost of goods sold | COGS, V | 1,016,550 | 55,384 | 247,644 |
| Capital stock | PPEGT, K | 1,454,210 | 57,532 | 247,644 |
| SG&A | XSG&A, X | 342,805 | 29,682 | 247,644 |
| Wage bill | XLR, WL | 1,093,406 | 130,486 | 28,116 |
| Employment | EMP, L | 8,363 | 863 | 221,121 |
|
| Sample B |
|---|
| Acronym, var. | Mean | Median | No. obs |
|---|
| Sales | SALE, PQ | 5,894,779 | 578,912 | 28,116 |
| Cost of goods sold | COGS, V | 2,970,693 | 195,087 | 28,116 |
| Capital stock | PPEGT, K | 5,193,319 | 345,592 | 28,116 |
| SG&A | XSG&A, X | 926,542 | 78,487 | 28,116 |
| Wage bill | XLR, WL | 1,093,406 | 130,486 | 28,116 |
| Employment | EMP, L | 24,861 | 4,522 | 25,527 |
For our data, in Section III we have already reported the revenue-weighted aggregate markups in Figures I and II, Panel A, as well as the input-weighted markup with total cost as the input weight (Figure II, Panel A).
Here
we plot the aggregate markups with one input weight and for a
sector-specific output elasticity (the results are very similar with
economy-wide, constant output elasticities). We present two versions of
the production technology (PF1 and PF2). In row 1 (Figure XVI, Panels A and B) we plot the aggregate markup measure proposed by Edmond, Midrigan, and Xu (2015) and Grassi (2017)
on the grounds of its representativeness of welfare measures in a
setting with CES preferences in models such as monopolistic competition
and Atkeson and Burstein (2008). Here we do not take a stance on welfare, but we do in De Loecker, Eeckhout, and Mongey (2018).
Figure XVI
Markups with Input Weights: COGS and Employment for the Benchmark Technology and PF2
We
find a rise in the COGS-weighted aggregate markups that is only about
half of the rise in revenue-weighted markups and is substantially lower
than the total cost–weighted aggregate markup (Figure XVI, Panels A and B). We can see from Table VII
why. With one input equal to the variable input used for calculating
the markup, the aggregate markup is simply the ratio of revenue over
COGS (multiplied by the output elasticity). As a result, the markup is a
function of aggregates only. This implies that the aggregate is not
sensitive to within-sector (or within-economy) variation. In fact, the
aggregate markup measure is identical to that obtained by Hall (1988) and that we report in Figure V.
Rather
than using expenditure, we can also use quantities. Although we do not
have quantities of COGS (because we do not know the unit prices), we
have quantities of labor, as measured by the number of employees. In Figure XVI,
Panels C and D for the two technologies we use employment weights.
Unlike aggregate markups with weights from expenditure shares of inputs,
those with employment weights track the benchmark aggregates. This
seems to indicate that at least for employment, the quantities do not
adjust as much as the input prices (in this case wages) do. When we
quantify the economy in De Loecker, Eeckhout, and Mongey (2018), we find large general equilibrium effects on wages and smaller effects on labor supply.
Finally, we also do the same decomposition exercise on our input-weighted measures in Figure XVII.
Not surprisingly, for the COGS-weighted aggregate, there is no role for
reallocation (the reallocation term is even negative). Because by
construction the COGS-weighted measure is based on averages only, there
is no impact of within-industry reallocation, and we have shown earlier
that the reallocation occurs predominantly within industry. What the
decomposition shows is that virtually all of the change in the
COGS-weighted markups is driven by the within term, the rise in markups
themselves.
Figure XVII
Decomposition of Input-Weighted Average Markups
Figure C.1
Distribution of Markups μit: Kernel Density Plots (Unweighted)
Figure D.1
Industry-Specific Average Markups Using Labor Cost
Instead,
for the employment-weighted measure, the picture looks much more
similar to that of the benchmark revenue-weighted aggregate markup.
About two-thirds of the rise in the employment-weighted average markup
can be attributed to reallocation.
An
important conclusion to take away from these alternative measures for
average markups is that they are different moments of a much richer
distribution of markups. We have documented that the distribution has a
fairly constant median, that the upper tail has become a lot fatter, and
within a market, larger firms tend to have higher markups. The
different weights give us further insights into the joint distribution
of markups, revenue, and all inputs.
VI.E. Comparison of Our Estimates with Those in the Literature
n Online Appendix Section 7, we compare our estimates with those obtained in the literature using the demand approach (Berry, Levinsohn, and Pakes 1995)
for seven industries for which there are data: beer, breakfast cereal,
steel, autos, airlines, department stores, and electronic shopping and
mail order. For the companies in our data set that fall in the same
industry classification, we construct an average markup and plot them
jointly with the markups obtained in the literature (Online Appendix Figure 7.1).
Whenever
there is overlap, the patterns of markups obtained with the demand
approach closely follow those obtained with our cost-based approach.
This is remarkable because not only are the methods different, they rely
on different data. This is testament to the fact that the estimates we
obtain are robust across different methods and data sources.
We perform further robustness exercises in the Appendix and the Online Appendix.
VII. Concluding Remarks
Using firm-level data on the accounts of all
publicly traded firms and of the census of private firms (in
manufacturing, retail and wholesale trade) in the United States, we
study the evolution of market power. For each firm, we estimate both
markups and profitability, and we document the properties of their
distribution. We find that from 1980 onward, markups have risen from 21%
to nearly 61% in 2014, an increase of 40 points. For the same period,
average profit rates have increased from 1% of sales to 8%.
We
attribute this rise in market power nearly exclusively to the increase
for the firms with the highest markups already. The distribution of
markups has become more skewed with a fat upper tail while the median of
the distribution remains unchanged. Because of this increasingly skewed
distribution, we must be cautious not to use the average markup as that
of a representative firm to draw any conclusion about the aggregate
economy. When markets are noncompetitive, aggregation is generally
nonlinear. In particular, the rise in revenue-weighted markups is due in
part to the rise of the markups themselves and in part to the
reallocation of sales shares from low- to high-markup firms. We find
that reallocation accounts for two-thirds of the rise.
We
further establish that the rise in markups is not merely to offset a
rise in overhead costs. Although overhead costs have risen, the rise in
markups exceeds that of overhead. We thus find that there are excess
markups, and that the excess markups are highest for those firms with
high overhead costs. This is consistent with the increase in our measure
of profits. We also find substantial increases in the market value as a
share of sales. All this indicates that the rise in markups is evidence
of a rise in market power.
We use our
evidence to investigate the macroeconomic implications of the rise of
markups. We focus our attention on the decrease in the labor share. From
the first-order condition of the firm’s optimization problem, there is a
negative relation between the labor share and the markup. We establish
that this negative relation exists at the firm level. This provides a
compelling justification for the secular decline in the labor share that
the aggregate U.S. economy has experienced. We further discuss the
impact of the rise in market power on the decrease in the capital share,
on the decrease in low-skill wages and labor force participation, and
on the decrease in labor market dynamism and migration rates.
Markups
of some firms are reaching heights multiple times higher than ever
seen, at least since World War II, when our data start. It is open to
speculation whether this trend will continue, but for now there are no
signs that markups will decrease substantially any time soon.
Appendix A: Estimating Output Elasticities
A crucial component to measure markups is to obtain an estimate of the output elasticity of a variable input of production (θV). Although the production approach to markup estimation described in De Loecker and Warzynski (2012)
does not restrict the output elasticity, when implementing this
procedure, it depends on a specific production function and assumptions
of underlying producer behavior to identify and estimate the elasticity
in the data. We use two distinct methods to estimate the output
elasticity of the production function. First, we estimate a parametric
production function for each sector-year using recent techniques that
take into account the well-known potential biases discussed in the
literature. Second, we nonparametrically estimate the output elasticity
using (constructed) cost shares. Both approaches have their advantages
and disadvantages, which we discuss.
A.A. Production Function Estimation
We follow standard practice and rely on a panel of firms,
for which we estimate production functions for each (two-digit)
industry. For the benchmark specification, we consider a
sector-year-specific Cobb-Douglas production function, with a variable
input bundle and capital as inputs. For each industry
s we consider the production function (PF1 in the main text):
where lowercase letters denote logs and ω
it = ln Ω
it, where
yit is a measure of realized firm’s output, and ε
it captures measurement error in output—that is,
yit = ln (
Qitexp(ε
it)).
We
depart from the standard specification in the literature by considering
time-varying production function parameters. In particular, in the
baseline model we estimate production functions with both time-varying
and sector-specific coefficients, for each of the 22 sectors (i.e.,
two-digit NAICS).58
There are good reasons to believe that technology varies across sectors
of the economy, from retail with giants like Walmart and Amazon, to
highly specialized medical devices companies. Equally or more important
for the evolution of markups is that the technology is time varying.
Over a period of seven decades, technology is likely to change. This is
important for the estimation of markups because systematic technological
change will imply a time-varying output elasticity θVit
. From inspection of equation (7), imposing a constant technology and hence a constant θV will therefore yield an overestimate of the markup if θV is decreasing and an underestimate if θV is increasing.59
Allowing the production function coefficients to vary over time is also
a parsimonious way to account for factor-biased technological change.
When we consider the production function with overhead (PF2 in the main text), the specification is given by:
with
x =
ln(
X), and
X captures (deflated) SG&A.
The
challenges in estimating production functions, using any data set, be
it the Compustat data or plant-level manufacturing census data, can be
grouped into two main categories: dealing with unobserved productivity
shocks (ωit); and extracting units of output and
inputs from revenue and expenditure data (i.e., the omitted price
variable bias). Both issues are of course not independent, and we rely
on methods that aim to deliver consistently estimated output
elasticities, dealing with them adequately.
We follow the literature and control for the simultaneity and selection bias, inherently present in the estimation of equation (23),
and rely on a control function approach, paired with a law of motion
for productivity, to estimate the output elasticity of the variable
input.60 This method accounts for the fact that the variable factor of production V adjusts in response to a productivity shock, whereas the fixed factor K
does not react to contemporaneous shocks to productivity, but it is
correlated with the persistent productivity term. This requires us to
restrict the production function to a particular class to guarantee that
the coefficients of interest—which determines the output elasticity—are
identified.
1. Control Function
We build on the insight from
Olley and Pakes (1996) that (unobserved) productivity ω
it
can be expressed as an (unknown) function of the firm’s state variables
and observables. This is obtained by considering input (or investment)
demand, and inverting out for productivity to yield:
where
dit is the control variable. We consider two cases: a variable input in production (in our case COGS,
v), and investment (
i), and
Ackerberg et al. (2007) provide an excellent treatment of the two types of control variables.
zit captures output and input market factors that generate variation in factor demand (for input
d)
across firms, conditional on the level of productivity and capital. The
latter is critical to allow for imperfectly competitive product markets
when estimating production functions. Standard approaches in the
literature on production function estimation are restricted to either
perfect competition or models of common markups (monopolistic
competition paired with CES demand). Instead, we rely on
De Loecker and Warzynski (2012) and
De Loecker et al. (2016)
to allow for imperfect competition in product markets and thus markup
heterogeneity across firms. In practice this amounts to allowing for
input demand shifters that move around the optimal amount of a variable
input, conditional on a firm’s productivity and capital stock.
Regardless
of which control variable is used, this method relies on a so-called
two-stage approach. In the first stage, the measurement error and
unanticipated shocks to output are purged using a nonparametric
projection of output on the inputs and the control variable.
In the case of a static control, dit = v
it, this is given by:
The output elasticity is obtained by constructing moments
of the productivity shock, which is obtained by considering a
productivity process given by ω
it =
g(ω
it − 1) + ξ
it. It gives rise to the following moment condition to obtain the industry-year-specific output elasticity:
where ξ
it(θ
t) is obtained by projecting productivity ω
it(θ
t) on its lag ω
it−1(θ
t), with
θt={θVt,θKt}, where productivity is in turn obtained from
ϕit−θVtvit−θKtkit, using the estimate φ
it
from the first-stage regression. This approach identifies the output
elasticity of a variable input under the assumption that the variable
input use responds to productivity shocks but the lagged values do not,
and that lagged variable input use is correlated with current variable
input use, through serially correlated input and output market
conditions, captured in
zit. In the case of PF2, an additional moment identifies the output elasticity of the overhead (SG&A) input,
E(ξit(θt)xit)=0.
In the case of the Olley-Pakes approach, we can identify
and estimate the output elasticity using a simple nonlinear regression:
and rely on the identification arguments made in
Ackerberg, Caves, and Frazer (2015)—that is, if the variable input bundle
v (COGS) is nondynamic and chosen at
t−b(0<b<1), while the investment decision is made at
t,
allowing for productivity shocks to hit the firm between these two
subperiods. This approach has the advantages that it is simple to
implement and does not require us to consider the subsequent second
stage. Compared with the static control, discussed above, the investment
policy function needs to be increasing in productivity (conditional on
capital and variables captured by
z), and the specification adopted here limits the scope of strategic interaction among firms.
We
consider both controls (COGS and investment) and find very similar
results for the estimated output elasticities. Below we plot the two
series, using the static and dynamic control variable, and we aggregate
the industry-year-specific output elasticities of COGS using industry
sales.
A.B. Units
As pointed out in
De Loecker and Goldberg (2014)
standard production data, whether it is Compustat or census data,
records revenue and expenditures, rather than physical production and
input use (with the exception of a few manufacturing industries). In the
presence of product differentiation (be it through physical attributes
or location) an additional source of endogeneity presents itself through
unobserved output and input prices. This has been the topic of recent
research, for a recent treatment see
De Loecker et al. (2016). A first observation is that the error term, ε
it, will generally contain output and input prices (scaled by the relevant technology parameters).
De Loecker et al. (2016)
show that the correlation of input expenditures with this error yields
biased estimates of the output elasticity. However, in their setting
physical output quantities are observed, and the unobserved input
prices, reflecting differentiation, are the only source of the price
error. We do not observe output price variation, and we are therefore
left with the following structural error term:
where
we let the user cost of capital be industry-time specific, but input
prices potentially vary across firms reflecting variation in quality,
location, and other exogenous factors. We follow
De Loecker et al. (2016)
and let the wedge between the output and input price (scaled by the
output elasticity) be a function of the demand shifters and productivity
difference.
61 In the case of Olley and Pakes, the inclusion of the variable
z should
therefore capture the relevant output and input market forces that
generate differences in output and input price. Of course, productivity
differences that influence the wedge between output and input prices are
automatically captured by the inclusion of the control function. Note
that not observing output prices has the perhaps unexpected benefit that
output price variation absorbs input price variation, thus eliminating
part of the variation in the error term. In the extreme case, we are
left with just the productivity unobservable, and this puts us back in
the standard framework introduced above.
Under
the alternative DGP, where the static control (COGS) is used and where
the output elasticity is identified in the second stage, we follow De Loecker et al. (2016).
The main difference lies in the fact that we cannot rely on observed
output prices, and we therefore have to rely on constructed measures of
market share (at various levels of aggregation) to eliminate the
variation in the price error wedge.
In
practice, we consider market share, measured at various levels of
aggregation (two, three, and four digit), to take into account
additional variation in output and input markets. As discussed in De Loecker et al. (2016)
this is an exact control when output prices, conditional on
productivity, reflect input price variation, and when demand is of the
(nested) logit form. We have subjected our analysis to a host of
different specifications of the production function (such as the
translog production function), and we find similar results for the
estimated output elasticities. As discussed in the main text, the main
findings on aggregate markups are furthermore not sensitive to the use
of a common time-invariant calibrated output elasticity of 0.85. We also
considered an alternative specification, including SG&A as a factor
of production, and document a comparable rise in aggregate markups.
Finally, there are a host of possible measurement error
and endogeneity concerns with any single specification we could
consider. We do not attempt to provide the one final set of output
elasticities for all sectors of the U.S. economy, using the Compustat
data. Rather we consider a variety of specifications, and show that the
main facts we are interested in (under the maintained year-sector
specific Cobb-Douglas production function) are not sensitive to these.
The aggregate markup can be expressed in terms of the potential bias ψst=θ^st−θst
in the production function coefficient:
We
have no prior belief that there is a particular correlation between the
weight of an industry in the economy, or in the sample, and the bias
introduced by either the simultaneity, selection, or omitted price
variable bias.
Appendix B: Data: Summary Statistics
B.A. Compustat
We obtain firm-level financial variables of all
U.S.-incorporated publicly listed companies active at any point during
the period 1950–2016. We access the Compustat North America Fundamentals
Annual (through WRDS) and download the annual accounts for all
companies. The results in this article are obtained with a download on
March 25, 2018. We keep unique records for each firm and assign a firm
to a unique two-digit industry, as reported. We exclude firms that do
not report an industry code. All financial variables are deflated with
the appropriate deflators. The main results, unless reported otherwise,
rely on the sample of firms over the period 1950–2016, where we
eliminate firms with reported cost-of-goods to sales and SG&A to
sales ratios in the top and bottom 1%, where the percentiles are
computed for each year separately. Our results are invariant to trimming
up to 5% (bottom and top).62
As such, a firm-year observation requires information on both sales and
COGS, two essential ingredients to measure markups. Appendix Table B.1
below presents a few basic summary statistics for a few leading
variables used in our analysis (sales, COGS, capital, wage bill,
employment, and SG&A), for two samples. Sample A, observations with
information on sales, COGS, and SG&A; and Sample B, observations
with information on the wage bill.
B.B. Economic Censuses
The focus of our analysis of the census data is
on manufacturing (NAICS codes 31-32-33), retail (NAICS codes 44-45), and
wholesale (NAICS code 42). In 2012, manufacturing consists of about
297,000 establishments, retail of about 1,060,000 establishments, and
wholesale of about 420,000 establishments. These establishments
aggregate into about 650,000 retail firms, about 314,000 wholesale
firms, and about 250,000 manufacturing firms. Together these three
sectors make up a little over 20% of U.S. GDP. In principle, each
economic census spans the universe of every single employer
establishment in its sector, across the size distribution; only
nonemployer establishments (sole proprietorships with no employees) are
omitted.
The other censuses that we do not
use are the Census of Services, the Census of Construction Services; the
Census of Mining; the Census of Transportation, Communications, and
Utilities; the Census of Finance, Insurance, and Real Estate; and the
Census of Auxiliary Establishments.63
The
data are organized around the most discrete unit of production in the
microdata, an “establishment,” which is a single physical plant.
Establishments can be aggregated to the EIN level (Employer
Identification Number, the most discrete legal unit of production; an
EIN is a unique tax ID associated with a distinct legal entity), and
higher up to the firm level (major corporations are usually collections
of EINs, which in turn are collections of multiple establishments). The
microdata associate each establishment with an EIN and a firm ID: the
EIN is considered part of the firm if the firm has complete or majority
ownership of the EIN.
Perhaps the most common
way of defining “firm” in the recent firm heterogeneity literature is
to say that all of a firm’s establishments in a given four-digit SIC
industry (roughly equivalent to a six-digit NAICS industry) are a
distinct firm (this is the approach taken by Hsieh and Klenow 2009; Autor et al. 2020;
and others). Under this definition, Walmart’s establishments listed as,
for example, SIC 5411 (Retail - Grocery Stores) are one firm, and
Walmart’s establishments listed in SIC 5412 (Retail - Convenience
Stores) are a separate firm. Our preferred default approach is to define
“firm” as all of the firm’s establishments in a single sector census
(e.g., all of Walmart’s firms in all of retail, NAICS codes 44-45, are a
single firm).
Notice that even though we do
not have information on overhead directly, one can obtain multiple
sources of information about overhead costs in the census data: (i)
Census flags the auxiliary establishments of multiunit firms and links
them to other establishments of the same firm. Auxiliary establishments
include headquarters, other facilities mainly engaged in general
management functions, and facilities that mainly engage in R&D.
Census has taken a systematic approach to identifying and flagging
auxiliary establishments across most sectors of the economy since 1997.
(ii) Census conducts various business surveys that elicit information
about various types of overhead. For example, the 2012 Annual Survey of
Manufactures includes questions about software expenses, the cost of
purchased communication services, advertising and promotional expenses,
and the cost of purchased professional and technical services. As a
second example, the Survey of Industrial R&D collects data on
research and development expenses for all large firms and a sample of
smaller ones. Although the data are scattered across a variety of
sources and databases, there is great potential in constructing
firm-level measures of overhead costs using the census data, and
combining it with the markup analysis. This is left for future work and
lies beyond the scope of this paper.64
https://academic.oup.com/qje/article/135/2/561/5714769?login=false
