Rising Markups or Changing Technology?

by

Lucia Foster
U.S. Census Bureau

John Haltiwanger
University of Maryland and NBER

Cody Tuttle
University of Texas at Austin

CES 22-38 September 2022

The research program of the Center for Economic Studies (CES) produces a wide range of
economic analyses to improve the statistical programs of the U.S. Census Bureau. Many of these
analyses take the form of CES research papers. The papers have not undergone the review
accorded Census Bureau publications and no endorsement should be inferred. Any opinions and
conclusions expressed herein are those of the author(s) and do not represent the views of the
U.S. Census Bureau. All results have been reviewed to ensure that no confidential information is
disclosed. Republication in whole or part must be cleared with the authors.

To obtain information about the series, see www.census.gov/ces or contact Christopher Goetz,
Editor, Discussion Papers, U.S. Census Bureau, Center for Economic Studies, 4600 Silver Hill Road,
Washington, DC 20233, [email protected]. To subscribe to the series, please click
here.
 Abstract

Recent evidence suggests the U.S. business environment is changing, with rising market
concentration and markups. The most prominent and extensive evidence backs out firm-level
markups from the first-order conditions for variable factors. The markup is identified as the ratio
of the variable factor’s output elasticity to its cost share of revenue. Our analysis starts from this
indirect approach, but we exploit a long panel of manufacturing establishments to permit output
elasticities to vary to a much greater extent - relative to the existing literature - across
establishments within the same industry over time. With our more detailed estimates of output
elasticities, the measured increase in markups is substantially dampened, if not eliminated, for
U.S. manufacturing. As supporting evidence, we relate differences in the markups’ patterns to
observable changes in technology (e.g., computer investment per worker, capital intensity,
diversification to non-manufacturing), and we find patterns in support of changing technology as
the driver of those differences.

*
Contact information: Foster: Center for Economic Studies, Bureau of the Census, Washington, DC 20233;
Haltiwanger: Department of Economics, University of Maryland, College Park, MD 20742. Tuttle: Department of
Economics, University of Texas at Austin, Austin, TX 78712. John Haltiwanger and Cody Tuttle were part-time
Schedule A employees at Census at the time of the writing of this paper. Any opinions and conclusions expressed
herein are those of the authors and do not represent the views of the U.S. Census Bureau. The Census Bureau's
Disclosure Review Board and Disclosure Avoidance Officers have reviewed this data product for unauthorized
disclosure of confidential information and have approved the disclosure avoidance practices applied to this release
(DRB Approval Number: CBDRB-FY20-CED006-0022, CBDRB-FY22-CED006-0016). We thank Daron Acemoglu, Jan
De Loecker, Mert Demirer, Jan Eeckhout, John Eltinge, Cheryl Grim, Devesh Raval, Pascual Restrepo, and
participants at the NBER 2021 Summer Institute, the 2022 Montreal Conference on Markets with Frictions, and
seminar participants at Dartmouth and the Erasmus School of Economics for their helpful comments on earlier
drafts.
I. Introduction

Increasing evidence suggests that markups of prices relative to marginal costs have been

rising in the U.S. as well as other countries. The most definitive evidence for the U.S. is the

recent research of De Loecker, Eeckhout and Unger (2020) (hereafter DEU). DEU present

evidence from public firms for the entire private sector and supporting evidence from

manufacturing, retail trade and wholesale trade establishments. Using establishment-level data,

we examine whether the DEU methodology attributes underlying cross-sectional and/or time-

series changes in technology to rising markups. We examine results from three different

estimation methodologies allowing for greater flexibility and granularity in the estimation of

output elasticities. We find that with more granular estimates of output elasticities, the increase

in markups is substantially dampened if not eliminated. Finally, we provide indirect evidence of

the connection to changes in technology to our dampened increase in markups.

The methodology for detecting the rise in the average and dispersion of markups builds

on Hall (1988) and De Loecker and Warzynski (2012) and is clever and simple in principle.

Using the first-order condition for a variable factor, the markup at the firm level is the ratio of

the output elasticity of the variable factor to the cost share of revenue of that factor. This

“production approach” (or “ratio approach” as denoted by Bond et al. (2020)) requires an

estimate of the output elasticity of the variable factor. The simplest implementation assumes a

constant output elasticity at the industry level so rising markups and dispersion relate to changes

in the empirical cost share of revenue of the variable factor. DEU recognized this simple

approach is potentially misleading since there might be variation over time and across firms in

output elasticities. They show their results are robust to considering output elasticities that vary

across time and firms. They permit elasticities to vary at the 4-digit level by year when using

2
cost shares of total costs to estimate output elasticities using the Economic Census data. When

estimating a Cobb-Douglas specification with COMPUSTAT data, they permit estimates to vary

at the 2-digit level by year and when using a translog specification (in DEU’s 2018 draft, see

Figure 18), they allow estimates to vary at the 2-digit industry level. For their estimation, they

use the control function methodology but innovate on the standard approach in the literature as

they recognize they are estimating revenue functions that depend on both markups and output

elasticities. 1

We explore specifications that push the potential for changing technology to a much

greater extent. For this purpose, we use establishment-level data from the Annual Survey of

Manufactures for 1972-2014. This yields a data infrastructure with approximately 2.2 million

establishment-year observations at the annual frequency. Specifically, we compare results across

three estimation methodologies with varying degrees of flexibility and allow for varying levels

of granularity in industry and time. For the cost share approach (CS), we compare results using

output elasticities that vary at the 4-digit by year level (similar to DEU) with those that vary at

the establishment-level every year. For the control function approach, we consider both Cobb-

Douglas (CD) and translog (TL) specifications. For the CD specification, we compare a

1
As we discuss below, the DEU approach recognizes that there is likely a bias from estimating output elasticities
from the revenue function. They propose an adjustment to their control function approach to address this issue, and
we adopt their approach in our analysis. Specifically, following DEU, we include market share in the estimation of
the revenue function. De Ridder, Grassi, Morzenti (2022) (hereafter DGM) raise a variety of questions about the
estimation of output elasticities from the revenue function in their analysis comparing estimates that emerge from
estimating the output function when firm-level prices are available. Their main findings are that using the revenue
function rather than the output function biases the level but the correlation is high between the two approaches
(mitigating concerns about the implications for changes in markups). Our analysis is distinct from DGM as we
focus on allowing for more granularity (detailed industry output elasticities that vary over time) in the estimation of
output elasticities. If the DGM concerns apply to both DEU and our analysis even with the addition of the market
share as a covariate in the estimation of the revenue function, we find it reassuring that they find the primary bias is
in the level of markups and not in the variation. Our main finding is that using a more granular, time varying
technology has a large impact on the implied change in markups.

3
specification at the 2-digit by year level (similar to DEU) with a 4-digit by year level. 2 For the

TL specification, we consider a 2-digit specification with time-invariant parameters (similar to

DEU, 2018) with a 4-digit specification with time-varying (every five year) parameters.

Importantly, the large annual panel of establishment-level data permits us to use the

control function approach to estimate Cobb-Douglas and translog specifications of the revenue

function at a more disaggregated level with time-varying coefficients. When using Census data,

DEU restrict themselves to using the cost share of total costs method for output elasticities. This

reflects their use of Economic Census data that is available at a five-year frequency. The control

function approach relies on the innovation to unobserved revenue shocks being uncorrelated with

predetermined variables (e.g., lagged inputs) and thus this approach is not well suited to

Economic Census data. Their control function estimation described above only applies to their

use of the COMPUSTAT data at the firm level.

We find that the increase in the average sales-weighted markup declines systematically

when allowing for output elasticities that vary more by detailed industry, by establishment and

by time. For our cost-share (CS) specification at the 4-digit level with annual data, we find the

sales-weighted markup increases by 47% from 1977 to 2007 and 29% from 1977 to 2012. These

patterns are broadly similar to those found by DEU using the Economic Census data for

manufacturing. The analogous changes using output elasticities that vary at the plant-by-year

level yields dampened sales-weighted markup increases of 24% from 1977 to 2007 and 16%

from 1977 to 2012.

For our Cobb-Douglas (CD) specification at the 2-digit level with annual estimates we

find the sales-weighted markup increases by 24% from 1977 to 2007 and 7% from 1977 to 2012.

2
For the control function by year estimates we follow DEU using samples that are rolling 5-year windows around
the focal year.

4
The analogous changes using a 4-digit specification with annual estimates yield only an 8%

increase from 1977 to 2007 and a decline of -5% from 1977 to 2012.

Even more dramatic differences occur when using the translog (TL) specification. Using

the translog specification at the 2-digit level with time-invariant parameters, we find that the

average sales-weighted markup increases by 41% from 1977 to 2007 and by 32% from 1977 to

2012. The analogous changes using a translog specification at the 4-digit level with annual

estimates are -3% and -6%, respectively. Throughout, we refer to those output elasticities

estimated with more granular measures of industry and time as “more detailed” estimates.

Our analysis does not just explore the robustness of the “production approach” to

estimating markups, it also opens a more extended investigation into differences in production

technologies across establishments and firms. It has long been known there are large differences

in revenue productivity measures across establishments within the same measured industry (see,

e.g., Baily, Hulten and Campbell (1992), Foster, Haltiwanger and Krizan (2001), and Syverson

(2004)). Such differences are present in revenue per composite input taking into account

multiple inputs (a TFPR type measure as defined by Foster, Haltiwanger and Syverson (2008))

and in revenue per unit input such as labor.

Hsieh and Klenow (2009) highlight that such dispersion potentially reflects wedges

relative to a frictionless and distortionless allocation of activity. Such wedges include markups.

The production method advocated by DEU is closely related theoretically and empirically to the

Hsieh and Klenow (2009) approach as the production method uses the dispersion in the cost

shares of revenue of variable inputs (e.g., materials and/or labor). Since firm and plant-level

deflators are not typically available, the measured cost share of revenue is closely related to

5
revenue per unit of nominal expenditures of the inputs. It may be that markups are the primary

factor driving such measured dispersion.

Alternatively, differences in production technologies (as well as differences in input

prices) may be driving the observed dispersion. We regard this as the natural flip side of the

production identification approach. The “production approach” to estimating markups identifies

dispersion in cost shares of revenue across firms and establishments as stemming from

differences on the demand side without imposing much structure on the demand side. We

investigate the alternative hypothesis that the variation is mostly coming from the supply

(cost/production) side.

Differences in our results using more detailed output elasticities raise a variety of

questions. Specifically, if the differences are consistent with greater variation in the production

technologies over time, then presumably, we should be able to find some direct evidence of such

changes. To investigate the potential link to changes in the way establishments do business, we

use observed variation in indicators of changing technology at the establishment and detailed (4-

digit) industry level. At the establishment-level, we explore measures of capital per worker,

computer investment per worker, a diversification measure based on the ratio of non-

manufacturing to manufacturing activity of the parent firm, and a relative size measure based on

the share of sales accounted for by the parent firm in the industry. We find that all four of these

indicators of how establishments change the way they do business exhibit increases in the mean

and dispersion across establishments over time.

For each of our three estimation approaches (CS, CD, TL), we compare markups and

output elasticities estimated at “less detailed” and “more detailed” levels, where these levels

differ in terms of variation over time and industry detail. As noted above, the “less detailed”

6
levels target DEU specifications. All four indicators of changing technology or business

structure are positively associated with the difference in the “less detailed” and “more detailed”

markup estimates at the establishment-level. Similarly, all four indicators are positively

associated with the difference in the “less detailed” and “more detailed” output elasticity

estimates. We also find that the industries with above median changes in these indicators of

changing technology exhibit increasing differences over time between the “less detailed” and

“more detailed” markup estimates.

Our findings that output elasticities of variable factors and in turn markups are

increasingly upward biased are consistent with recent findings of Hubmer and Restrepo (2021)

and Demirer (2020). Hubmer and Restrepo (2021) use COMPUSTAT data to estimate a Cobb-

Douglas specification with output elasticities of the variable factor of production permitted to

vary across time, industry and firm size classes. Demirer (2020) also uses COMPUSTAT data

for the U.S. and develops a novel methodology for estimating production functions. Both of

these papers present evidence that output elasticities of the variable factor of production are

lower and falling for larger firms. Moreover, both present evidence this translates into smaller

increases in variable markups.

We interpret our results as complementary with this recent literature finding that output

elasticities for variable factors are lower and declining for larger firms. Our contribution is to

use a large, representative panel of manufacturing establishments that permits flexible

specification of technologies (e.g., translog) with fewer restrictions than the existing literature.

We do not impose any structure that inherently yields differences in elasticities across firm size,

but with our more flexible specifications, we find that output elasticities for materials and

7
markups are smaller for larger firms. In turn, this implies the shift towards larger firms yields

less of an increase in measured markups taking into account the more flexible specifications.

An important differentiating feature of our analysis is that our rich data permits us to

focus on materials input as the variable factor while the analysis with COMPUSTAT requires

using a composite measure of the variable input or constructing a decomposition of the

components indirectly. As we argue below, the firm adjustment costs literature suggests that

labor should not be treated as a variable factor even at an annual frequency. 3

Another differentiating feature of our results is that we show that this pattern of

differences in output elasticities extends to indicators of adopting more advanced and capital-

intensive technologies. As with firm size, we find that these indicators are associated with

smaller estimated output elasticities and smaller estimated markups. We also find that

establishments with greater non-manufacturing activity of parent firms (diversification) have

smaller estimated output elasticities and smaller estimated markups. This finding is consistent

with the arguments in Fort, Pierce, and Schott (2018) that some firms are shifting away from the

production of physical goods and more towards the design and marketing of goods. Critical here

is that these firms retain some manufacturing activity but are leaner in terms of variable inputs.

The paper proceeds as follows. Section II sets out the conceptual framework and

estimation methodology. Data and measurement are discussed in section III. Output elasticity

estimates and implied markups are presented in section IV. Section V presents analysis of the

3
Raval (2020) presents insightful analysis that markup patterns estimated from using materials and labor inputs as
alternative variable factors yield inconsistent patterns. Our findings also yield inconsistent patterns across markups
estimated from materials and labor. Raval (2020) investigates the hypothesis that this inconsistency can be
reconciled by considering labor-augmenting technical change. From our perspective, we think the adjustment costs
for labor imply that labor is not a variable factor even at an annual frequency. It may be that the appropriate
frequency for labor adjustment costs is at a monthly or quarterly frequency. However, as shown in Cooper,
Haltiwanger and Willis (2020) adjustment costs for labor at this higher frequency have important implications for
annual moments of firm-level employment adjustment (that differ from the frictionless model where labor is a
variable factor).

8
factors driving the differences in markups across less and more detailed output elasticity

estimation. Concluding remarks are provided in section VI.

II. Conceptual Framework and Estimation

The DEU approach (along with earlier and subsequent papers by the authors) to

estimating markups relies on the following equation derived from a cost-minimizing

establishment’s objective function.

𝜃𝜃𝑖𝑖𝑖𝑖 𝑉𝑉
𝜇𝜇𝑖𝑖𝑖𝑖 = (1)
𝛼𝛼𝑖𝑖𝑖𝑖 𝑉𝑉

where 𝜇𝜇𝑖𝑖𝑖𝑖 is the markup for establishment i in year t, 𝜃𝜃𝑖𝑖𝑖𝑖 𝑉𝑉 is the output elasticity for input v for

establishment i in year t, and 𝛼𝛼𝑖𝑖𝑖𝑖 𝑉𝑉 is input v’s share of total revenue for establishment i in year t.

In other words, the markup is the ‘wedge’ between the establishment’s output elasticity for any

variable input v and that input’s share of the establishment’s revenue. 4

The input’s share of revenue, 𝛼𝛼𝑖𝑖𝑖𝑖 𝑉𝑉 , can be measured directly in firm or establishment-

level data. It is the establishment’s total expenditure on the input divided by the total revenue in

the establishment (the cost share of revenue). This leaves equation (1) with two unknown

quantities, the markup (μ) and the output elasticity (θ). To recover the markup, the output

elasticity must be estimated, and typically, it is estimated at relatively coarse levels of industry

and time.

Our primary question is whether the relatively coarse variation in estimated output

elasticities attributes to markups cross-sectional differences in technology and/or time-series

changes in technology occurring at more disaggregated levels. We use a large, annual dataset on

4
As noted in equation 1, the markup is defined for any variable input (v). While in theory, the markup is defined to
be the same over any variable input, in practice the measured markup may differ. We discuss our preference for
measuring markups using materials as opposed to labor as the variable input in footnote 2 and later in the paper.

9
U.S. manufacturing establishments to estimate production technologies flexibly and demonstrate

how estimated markups change when using this flexible approach. We do this in two ways.

First, we estimate output elasticities using a cost-share approach, which, under certain

assumptions, allows technology to be estimated at the establishment-by-year level. Second, we

estimate the production function using proxy methods at finer levels of industry and time.

It is common to estimate output elasticities using averages of cost shares of total costs at

the industry level. The motivation for averaging to the industry level (and often over time) is

that the first-order conditions for cost minimization underlying this approach are unlikely to hold

for all factors at each instant of time at the micro level (see, e.g., discussion in Syverson (2011)).

Still, this leaves open questions as to the level of industry detail that should be used and whether

time averaging is needed. We push as far as we can on these dimensions by using cost shares of

total costs of variable factors at the establishment-by-year level. We also compare this to a range

of alternative less detailed approaches (e.g., 2-digit, 4-digit, 6-digit industry-based estimates that

are constant over time or vary by year). We acknowledge using establishment-by-year estimates

requires very strong assumptions but think it useful as an attempt to permit as much

establishment-level variation in technology as possible. 5

In our second approach, we follow DEU in estimating output elasticities by directly

estimating the revenue function at varying levels of industry-by-time. Like DEU, we use a

control function approach to estimate the output elasticities for the Cobb-Douglas and translog

specifications. 6 Moreover, we take advantage of their contribution to this literature that

5
While it may seem like an extreme, using the establishment-level cost share of total costs yields a common
approach for measuring markups given by Rit / TCit where Rit is establishment-level revenue and TCit is
establishment-level total costs. Autor et al. (2020) denote this the accounting measure of markups.
6
We discuss details in Appendix C but note that we follow the specification of DEU closely. As is clear from our
results, when we use the level of aggregation in terms of time and industry, we obtain results very similar to theirs
both qualitatively and quantitatively.

10
recognized that since the dependent variable is firm or establishment-level revenue, accounting

for the wedge between unobserved output and input prices is potentially needed. Specifically,

following DEU, we include as a covariate the establishment’s market share in their 4-digit

industry (instrumented using three lags of the establishment’s market share).

We are pushing the data hard in our analysis; the control function estimation is often used

at a more aggregate industry level given that the polynomial approximations are sensitive to

smaller samples. As a robustness check, we conduct our analysis for the 50 largest industries (in

terms of number of establishments) since these are the industries where sample size restrictions

are less binding. In addition, we also explore the relationship between observable indicators of

changing technology and the growing gap in markup estimates we find when using “less

detailed” and “more detailed” output elasticity estimates.

In sum, we have three different estimation methods for output elasticities: cost-share

(CS), production function using Cobb-Douglas (CD), and production function using translog

(TL). We estimate these over two samples (full and top 50) and with varying degrees of

flexibility in industry (2-digit, 3-digit, 4-digit, 6-digit, plant-level) and time (constant, annual). In

the penultimate section of the paper, we explore how differences in these markup estimates relate

to changing technology and business structure.

Before proceeding to the analysis, it is instructive to consider how to interpret potential

differences in estimated output elasticities. For this purpose, we find it useful to consider

conceptually the difference between the estimated output elasticity and the true elasticity. We

specify this difference as: θˆit − θit =

ε it . Plugging this into the expression for the markup, the

difference between the actual and estimated markup is equal to: ε it / α it . Several inferences can

be drawn from this simple expression.

11
 First, at the establishment-level, the average bias depends on the mean of ε it / α it . This is

given by: cov(ε it ,1/ α it ) + E (ε it ) E (1/ α it ) . Thus, the average bias in the markup depends not only

on the bias in the output elasticity but on the covariance between the error and the (inverse) of

cost share of revenue. Second, at the establishment-level the bias may vary systematically with

the technology adopted by the establishment. Such systematic relationships can help account for

the dispersion in errors in estimated markups across establishments. The error in the revenue-

weighted average markup depends on the mean of ωit ε it / α it where ωit is the revenue share. This

expression reminds us that the average bias in the revenue-weighted markup will depend further

on covariances of the error and the cost share of variable inputs of revenue with the revenue

share.

This discussion highlights that, on the one hand, it is instructive to examine differences in

output elasticities across estimation methods directly. Other things equal, a rising bias in the

output elasticity itself will yield an increase in average (weighted or unweighted) markups. On

the other hand, examining the sign or change in magnitude of the bias in output elasticities is

insufficient. We build on this insight in the analysis that follows.

III. Data and Measurement

This paper takes advantage of a dataset that has been created in the Collaborative Micro

Productivity (CMP) project at Census that tracks large (roughly 55,000 establishments per year)

representative samples of U.S. manufacturing establishments from the Annual Survey of

Manufactures (ASM) from 1972 to 2014. The ASM is a series of five-year panels (starting in

years ending in “4” and “9”) with probability of panel selection being a function of industry and

size. We use ASM sample weights in all our analyses. We provide an overview of our

measurement methodology in the main text but provide more details in the data appendix.

12
A. Nominal Measures

We require nominal measures of revenue and input expenditures to compute the two

types of cost share measures (cost shares of revenue and cost shares of total costs). Nominal

revenue is measured as the total value of shipments adjusted for changes in final and

intermediate inventories. Nominal materials are measured as the sum of the cost of materials and

parts, the cost of resales and the cost of contract work done for the establishments by others on

the establishment’s materials. Nominal labor costs are measured as salary and wages for all

workers. Nominal energy expenditures are the sum of the cost of purchased electricity and the

cost of purchased fuels consumed for heat, power, or electricity generation. Nominal

expenditures for capital (calculated separately for structures and equipment) are the product of

the user cost of capital we obtain from the Bureau of Labor Statistics (BLS) at the 3-digit

industry level times the real capital stock. Real capital stocks are constructed using a perpetual

inventory method. Nominal expenditures are deflated with industry-level investment deflators.

We use 3-digit industry-level deflators from BLS for both investment expenditures and the

depreciation rate.

These nominal measures permit us to construct cost shares of revenue for materials and

labor. We focus on the cost share of revenue for materials since materials is more plausibly a

variable input. While we show results for labor in the appendix, the firm-level adjustment costs

literature provides evidence that labor is not a variable factor of production even at an annual

frequency (see Cooper, Haltiwanger, and Willis (2020) and Decker et al. (2020)). We also use

these data to construct cost shares of total costs in our cost-share based estimation of output

13
elasticities at the establishment-by-year level. For our output elasticities measured from cost

shares at the industry-level, we use appropriately weighted establishment-level cost shares.

B. Real Measures

For our production/revenue function estimation we follow standard practice of converting

the nominal revenue and input expenditure measures into real measures using industry-level

deflators. For nominal revenue, materials, and energy we use 6-digit NAICS deflators from the

NBER-CES database (extended to 2014). 7 For the labor input measure for estimating output

elasticities, we use the measure of total hours constructed as the production worker hours times

the ratio of salary and wages for all workers to those for production workers. This method

includes an adjustment for difference in labor quality for production and non-production

workers.

IV. Estimates of Output Elasticities and Markups

We start by providing the results of the estimations in three tables (corresponding to our

three methodologies: CS, CD, and TL). Panels A and B of Table 1 show the distribution of

estimated output elasticities from cost shares of materials of total costs (CS) at different levels of

aggregation. Panel A shows results for the entire manufacturing sector while Panel B shows the

results for the top 50 industries. Panels A and B of Table 2 show the distribution of estimated

output elasticities for materials from control function estimates of the revenue function using the

Cobb-Douglas (CD) specification. Panels A and B of Table 3 show the distribution of estimated

output elasticities for materials from control function estimates of the revenue function using the

translog specification (TL).

7
See https://www.nber.org/research/data/nber-ces-manufacturing-industry-database.

14
 As we consider specifications with more industry detail and greater time variation, the

estimated output elasticities for materials exhibit substantially more dispersion. For example, the

standard deviation for the cost share (CS) approach rises from 0.0344 for the least detailed

estimation (2-digit, constant) to 0.2051 for the most detailed estimation (plant-level, yearly). The

Cobb-Douglas (CD) specification has an increase in similar magnitude (0.01838 to 0.1093), but

the translog (TL) specification has a less dramatic increase (0.1765 to 0.1951). The patterns for

the top 50 industries are broadly similar to the full sample of industries. We focus on the full

sample for the remainder of the analysis (but show results for the top 50 industries in the

appendix). Results for estimates of output elasticities for labor show similar patterns and are

reported in Tables A.1-A.3. 8

We now turn to the estimated markups. 9 Figures 1 to 3 show the implied pattern of

changing markups on a sales-weighted basis (for cost-shares, Figure 1; Cobb-Douglas, Figure 2;

and translog estimations, Figure 3). Panel (a) in each figure shows long differences from 1980 to

2014 for alternative cases. The color of the bars denotes differences in time variation (black is

more restrictive and is denoted by “Constant”, red striped is less restrictive and is denoted by

“1yr”) and the bars are grouped by industry level. Panel (b) in each figure shows annual markups

for two key benchmark cases: (1) dotted black lines shows “less detailed” that corresponds

closely to the level of aggregation used by DEU and (2) the red solid line shows “more detailed.”

Focusing first on panel (a), as we consider specifications with more industry detail and

greater time variation, the increase in markups is substantially dampened. In some cases, the time

8
The output elasticities in Tables A.1-A.3 are reported for the primary “less” and “more” detailed specifications.
9
All the markup estimates are winsorized in each year at the 1st and 99th percentiles. Our reading of DEU is that they
trim the 1% tails rather than winsorize. Given that we consider a wide range of alternative markup estimates,
winsorized markups facilitate avoiding disclosure issues from trimming each of the alternative estimates. Figure
A.1 shows that the long differences for our benchmark “less detailed” and “more detailed” cases are very similar for
the results based on winsorized versus trimmed markup distributions.

15
variation is driving this decrease, in other cases it appears that the industry variation is driving

the decrease. For example, industry differences appear to dominate for cost shares (CS)

approach, but time variation appears to dominate for both proxy methods. These patterns of

implied markups are robust to limiting to the top 50 industries (see Figure A.2). 10

Turning to the time series pattern of markups in panel (b) of Figures 1-3 shows further

interesting patterns. In all three cases, the “more detailed” cases (red lines) are everywhere

below the “less detailed” cases (black dotted lines) but the gap between the two series widens

starting in the late 1990s. For the “less detailed” specifications (black dotted lines), there is still

an overall increase in markups from 1980 to 2014. However, with the “more detailed”

specifications (red solid lines), we find only a moderate increase in markups using the cost-share

approach (CS), little change using the Cobb-Douglas specification (CD), and a decline using the

translog specification (TL).

Comparing our results with those of DEU, we note that for these results we start, as they

do for their analysis of Economic Census data, at the establishment level. They aggregate to the

firm level within manufacturing, then to the industry level, and finally to the total manufacturing

level. The findings in Figures 1 to 3 focus on the total manufacturing level patterns although we

explore results at a more disaggregated level below. Our results at the total manufacturing level

are comparable conceptually to the estimates in DEU. While appropriate caution is required in

direct comparisons given their focus on the cost share approach with the Economic Census, a

comparison of Figure 1 using the 4-digit by year benchmark to their results from the Census of

10
Long differences from 1980 to 2014 for implied change in markups using labor as the variable factor are in
Figures A.3 (all industries) and A.4 (top 50 industries) for the less detailed and more detailed specifications. For the
cost share approach to estimating elasticities, we obtain similar results to those for materials. Results are less
systematic using less detailed versus more detailed for Cobb-Douglas and translog. Even so, we find markups
decline overall from 1980 to 2014 using the translog specification for labor as the variable input whether using less
or more detailed specifications for the top 50 industries. For all industries, the less detailed translog yields a sharp
decline in markups while the more detailed yields little change.

16
Manufactures also using 4-digit by year cost shares shows broadly similar patterns. Also, while

it is an apples-to-oranges comparison, our results using the control function approach with the

manufacturing establishment data are similar to those they report using the control function

approach for COMPUSTAT using manufacturing firms (see Appendix C for more discussion).

Notably, the increase in markups from 1972 to 2014 peaks in the mid-2000s, and from

2006 to 2014, markups decline substantially. This peak in markups around 2005 occurs in all

three “less detailed” cases and in the “more detailed” cost share and Cobb-Douglas cases. 11 The

analysis of Economic Census data in DEU offers a glimpse at this fall in markups. In their work,

the average markup for manufacturing decreases from 2007 to 2012, falling below the level of

markups from 1992-2002. Our analyses with annual data confirm that this decrease is not simply

a one-year dip, but rather a persistent decline from 2005 through 2014. Averaging across the

three less detailed specifications, markups decrease by about 20% from 2005 to 2014, returning

to the levels estimated for the mid-to-late 90s. Although we find a smaller rise in the more

detailed cases using cost share and Cobb-Douglas approaches, we likewise find a smaller

decrease of around 14% from 2005 to 2014, with markups again returning to 1990s levels. This

decrease in markups is robust to estimation strategy and is not present in COMPUSTAT data

(see DEU 2019 draft, Appendix 12.1). This further highlights the value of using ASM/CM data

and abstracting from the greater measurement issues raised in this paper, suggests that estimated

markups for manufacturing have fallen dramatically in recent years.

We believe the time series patterns in Figures 1-3 provide reassurance that our findings

are not being driven by a greater impact of measurement or specification error with our more

detailed output elasticities. The patterns in Figures 1-3 show that the sales-weighted markup

11
The more detailed translog case does not exhibit a rise in markups, and thus there is no corresponding decrease.

17
estimates from the “less detailed” and “more detailed” specifications are quite similar for about

the first ten years of our sample (e.g., 1972 to the mid-1980s). In the middle part of our sample

there is a growing gap between the sales-weighted markup from the “less detailed” and “more

detailed” output elasticity specifications. Finally, in the last ten years of our sample, this gap

either stays about the same or even falls. Also, it is notable that markups from both “less

detailed” and “more detailed” output elasticity specifications decline in the last ten years of our

sample. These time series patterns would require a time series evolution of

measurement/specification error that was minimal in the first part of our sample, increased

substantially in the middle part of our sample and then stabilized or declined in the last part of

our sample.

V. Factors Driving Differences in Results

What drives differences in markups between the “less detailed” and “more detailed”

specifications? We explore this question with several exercises examining three potential factors

driving differences in results. These are measurement issues (aggregation and weights), shifting

shares as evidenced through decompositions, and fundamental changes in production technology

as captured by capital intensity, computer investment per worker, diversification, and relative

firm size.

A. Measurement Issues: Output Elasticities, Revenue Shares, and Total Cost Weighting

First, we highlight some measurement issues related to aggregation. We show that the

results cannot simply be interpreted through the lens of separately examining the patterns of

output elasticities ( θ ) and cost shares of revenue (α). The sales-weighted mean of the estimated

markup at any level of aggregation is:

18
 𝜃𝜃𝑖𝑖𝑖𝑖 𝑉𝑉
� 𝜔𝜔𝑖𝑖𝑖𝑖 𝜇𝜇𝑖𝑖𝑖𝑖 = � 𝜔𝜔𝑖𝑖𝑖𝑖 (2)
𝛼𝛼𝑖𝑖𝑖𝑖 𝑉𝑉
𝑖𝑖 𝑖𝑖

Where the sales weight of plant i is given by 𝜔𝜔𝑖𝑖𝑖𝑖 . It is apparent that the sales-weighted average

of markups is not equal, in general, to the ratio of the sales-weighted output elasticities to the

sales-weighted cost shares of revenue. We refer to the latter as the naïve markup given by: 12

∑𝑖𝑖 𝜔𝜔𝑖𝑖𝑖𝑖 𝜃𝜃𝑖𝑖𝑖𝑖 𝑉𝑉

𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁 𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀 = (3)
∑𝑖𝑖 𝜔𝜔𝑖𝑖𝑖𝑖 𝛼𝛼𝑖𝑖𝑖𝑖 𝑉𝑉

Figure 4 shows the long differences of the naïve markups for the selected benchmark cases. It is

evident that the patterns in Figure 4 are distinct from those in Figures 1-3. Under the less

detailed specifications, the naïve markup exhibits little change for the cost share (CS) approach,

declines under Cobb-Douglas (CD) and increases under the translog (TL) but much less than

implied by Figure 3. For the more detailed specification, the naïve markup declines for the

Cobb-Douglas (CD) and translog (TL) specifications.

While the naïve markup is not directly informative about the actual markup, it is still

interesting to consider the numerator (sales-weighted output elasticities) and denominator (sales-

weighted revenue cost shares of inputs) of the naïve markup. Recall from the discussion earlier

that the bias in the estimated aggregate markup will depend on revenue-weighted average output

elasticity and the revenue-weighted average cost share of variable inputs. The point of our

earlier discussion is that examining these moments independently is insufficient given

underlying covariances, but they are still informative.

12
The naïve markup is not exactly what one would compute from aggregate data (see e.g., equation (11) from DEU
when output elasticities are constant) since we use sales weighting for both the output elasticity and the cost share of
revenue. We use this formulation to highlight that caution needs to be used in drawing inferences from the
“aggregate” patterns of output elasticities and cost shares of revenue regardless of the weighting used in the
aggregation.

19
 We analyze these two moments in Figures 5 and 6. Figure 5 shows the long difference in

output elasticities for materials. 13 Figure 6 shows the sales-weighted revenue cost shares for all

inputs (materials as well as labor, energy, and capital) as well as the ratio of sales-weighted total

costs to sales-weighted revenue. In Figure 5, we find that sales-weighted output elasticities

exhibit different patterns across the estimation approaches and using less versus more detailed

specifications. For both Cobb-Douglas (CD) and translog (TL) the more detailed specification

yields a decline in the sales-weighted output elasticity for materials. Turning now to the cost

share of revenue for inputs (Figure 6), we find that the (sales-weighted) materials share rises

slightly, the labor and energy shares decline, the capital share rises and the overall ratio of total

costs to revenue declines. We note that the capital costs in this case are based on perpetual-

inventory- based capital stocks and detailed industry-specific user-costs of capital from the BLS.

Figure 7 depicts the long differences in the sales-weighted returns to scale. For the “less

detailed” Cobb-Douglas (CD) specification and the “more detailed” translog (TL) there is some

mild evidence of rising (sales-weighted) returns to scale. For the “more detailed” Cobb-Douglas

(CD) and “less detailed” translog (TL), there is, if anything, evidence of an even more modest

decline in (sales-weighted) returns to scale.

As a further cross-check on the basic patterns, we follow DEU and Edmond, Midrigan and

Xu (2018) by computing total-cost-share weighted markups. We show in Figure 8 the long

differences of the changes of this alternate measure of markups (again using materials as the

variable input). Broadly consistent with these papers, we find smaller increases in total-cost-share

weighted markups even using the “less detailed” specifications (and a decline with Cobb-Douglas).

13
Figure A.5 shows the analogous plot for labor.

20
Consistent with Figures 1-3, we find that “more detailed” specifications yield a smaller increase or

larger decline in markups.

B. Shifting Shares: Within vs. Reallocation Components of Changing Markups

Underlying the finding of rising sales-weighted measured markups by DEU and the

related literature is a rising dispersion across businesses in markups -- especially with an increase

in the upper tail of the distribution. Accompanying this change in dispersion and skewness is a

shift in sales to high markup businesses. DEU use a decomposition developed by Haltiwanger

(1997) to decompose aggregate changes in sales-weighted markups into within, between, cross

and net entry terms. They find that the reallocation components dominate the increase in sales-

weighted markups. We use this same methodology to compare these composition effects

between the more and less detailed cases. 14 We are interested in whether the differences we

observe are driven by specific components. The decomposition is given by:

∆𝜇𝜇𝑡𝑡 = ∑𝑖𝑖𝑖𝑖𝑖𝑖 𝜔𝜔𝑖𝑖𝑖𝑖−1 ∆𝜇𝜇𝑖𝑖𝑖𝑖 + ∑𝑖𝑖𝑖𝑖𝑖𝑖(𝜇𝜇𝑖𝑖𝑖𝑖−1 − 𝜇𝜇 𝑡𝑡−1 )∆𝜔𝜔𝑖𝑖𝑖𝑖 + ∑𝑖𝑖𝑖𝑖𝑖𝑖 ∆𝜇𝜇𝑖𝑖𝑖𝑖 ∆𝜔𝜔𝑖𝑖𝑖𝑖 + ∑𝑖𝑖𝑖𝑖𝑖𝑖(𝜇𝜇𝑖𝑖𝑖𝑖 − ������)𝜔𝜔
������ 𝜇𝜇𝑡𝑡−1 𝑖𝑖𝑖𝑖 −

∑𝑖𝑖𝑖𝑖𝑖𝑖(𝜇𝜇𝑖𝑖𝑖𝑖−1 − ������)𝜔𝜔
𝜇𝜇𝑡𝑡−1 𝑖𝑖𝑖𝑖−1 (4)

The first term in equation (4) is the within term. The second (between effect) and third (cross

effect) terms together capture reallocation across continuing establishments. The last two terms

combined reflect net entry as the penultimate term captures entry and the final term captures exit.

Bars over terms denote weighted means.

Before showing the results of the decomposition, we first examine the dispersion in our

measures of markups. We focus on two measures of dispersion: an overall measure (standard

14
We apply the decomposition at the establishment rather than the firm-level. Our objective is to quantify the
relative contribution of the different components for less and more detailed output elasticity specifications.

21
deviation) and one that focuses on the right tail (the 90th-75th percentiles differential). Figure 9

illustrates that we also find rising dispersion (panel (a)) and a rising right tail (measured by the 90-

75 differential in panel (b)) in markups across establishments for both “less detailed” and “more

detailed” specifications. The rising dispersion and skewness are mitigated by the “more detailed”

specifications (except for the cost share approach for skewness). This pattern is intuitive since the

“more detailed” specifications absorb more of rising dispersion with dispersion in output

elasticities (see Tables 1-3).

The decomposition of the changing markups for both the “less detailed” and “more

detailed” specifications is reported in Table 4. We compute the terms in Table 4 first for the five-

year intervals between Economic Census years from 1977 to 2012. We then cumulate the

components over the entire time period. Recall the specification that is closest to DEU’s analysis

of the Economic Census is in the first row of the table (CS, 4-digit industry, 1 year). For that

specification, we find results broadly consistent with theirs, showing a positive contribution of the

within, net entry and reallocation terms, with the latter component dominating. More generally, for

the less detailed specifications, we find that the reallocation from continuing establishments

dominates the increase in markups although net entry also contributes substantially.

For the more detailed specifications, the much smaller increase in markups is associated

with a general tendency of all components to fall in magnitude. Especially noticeable is the

substantial negative within contribution for all of the more detailed specifications. That is, the

markup is declining substantially on a sales-weighted basis within businesses. The reallocation

terms are all positive for the more detailed specifications offsetting the declining within terms. In

that respect, reallocation continues to play a critical role. Our findings suggest that if there had not

been this shift towards high markup businesses then there would have been a decline in aggregate

22
markups in manufacturing. While reallocation plays a critical role with the more detailed

specifications, the magnitude of the reallocation terms is smaller than it is in the less detailed (with

the exception of the cost share, plant, 1 year approach). The findings in Figure 9 help explain this

declining contribution of reallocation. There is a shift in activity towards higher markup

businesses, but since dispersion in markups rises by a smaller amount in more detailed

specifications, this shift yields less of an increase in the sales-weighted markup.

C. Changing Technology?

Our findings imply that permitting greater variation in the estimation of output elasticities

across time and firms substantially dampens the measured increase in markups in U.S.

manufacturing. This inference depends on the robustness of estimating output elasticities at this

level of disaggregation. As discussed above, there are multiple factors that provide support for

this robustness. In this section, we take an additional step by exploring the relationship between

differences in the “less detailed” and “more detailed” markup patterns and observable measures

of changing technology and firm structure. This analysis also provides insights into why the

more detailed specifications yield smaller increases in estimated markups.

We exploit observable plant-level indicators of capital intensity (capital per worker),

computer intensity (computer investment per worker), diversification (ratio of non-

manufacturing to manufacturing activity in the parent firm) and relative firm size (share of the

parent firm’s sales in industry sales). Capital intensity is measurable for all establishments from

1972 to 2014. Computer investment is available in the Economic Census for 1977, 1982, 1987,

2002, and 2007 and in the ASM in 2000. U.S. firms with activity in manufacturing often have

activity in non-manufacturing. Fort, Pierce, and Schott (2018) document there has been a positive

trend in this direction with some firms with only modest levels of manufacturing being described

23
as a form of factory-less production. Based on this work, we use the Longitudinal Business

Database (LBD) to construct a measure of the extent of this activity using the parent firm for each

establishment. 15 The share of the parent firm’s sales in industry sales is measurable in Economic

Census years when all establishments are covered.

Figure 10 shows that all four indicators exhibit an increase in mean and three of the four

indicators exhibit an increase in dispersion over time. 16 These findings are important indicators

that establishments are changing the way they are doing business with increased differentiation

across establishments. The log firm share is related to changing market structure with the shift

towards superstar firms. As discussed in both DEU and Autor et al. (2020), the shift towards

superstar firms is connected to rising measured markups as larger firms have higher measured

markups. However, it may be that this reflects differences in output elasticities between smaller

and larger firms as well as differences in the covariance between output elasticities and cost

shares across firm size. We investigate that question as we explore the connection between the

“less detailed” and “more detailed” measured markups and estimated output elasticities and these

indicators of changing technology and changing structure of the economy.

Before presenting regression results that investigate this question, we provide summary

statistics for the dependent and explanatory variables in Table 5. A highlight is the enormous

variation across establishments in these variables. Turning to the regression results, the top panel

of Table 6 presents bivariate establishment-level regressions that relate the difference in

establishment-level “less detailed” minus “more detailed” estimated markups using the translog

specification to these technology/business structure measures. All specifications control for

15
Specifically, we measure the ratio of non-manufacturing to manufacturing employment for the parent firm.
16
It is not surprising that as the log firm share rises rapidly in the post-2000 period that dispersion falls as large
firms increasingly dominate.

24
detailed industry (6-digit) by year effects. All four of the measures are positively related to the

less minus more detailed estimated markups.

The bottom panel of Table 6 presents the analogous bivariate establishment-level

regressions with the dependent variable as the “less detailed” minus “more detailed” estimate of

the output elasticity. Again, we find that all four of the measures are positively related to the less

minus more detailed estimated output elasticities at the establishment-level. 17

These findings are consistent with the hypothesis that establishments that have adopted

different ways of doing business within industries have systematically different estimated

markups and output elasticities. The results on log firm size imply that larger firms have smaller

estimated output elasticities of variable factors (and smaller measured markups) when using

specifications that permit greater differences in output elasticities across establishments within

and across industries as well as time. These findings on firm size are consistent with those in

Hubmer and Restrepo (2021) who present evidence that output elasticities of variable factors of

larger firms are smaller using COMPUSTAT data. 18 There is large literature on technology

adoption that provides theory and evidence that larger and growing businesses are more likely to

adopt advanced technologies (see, e.g., Dunne, Haltiwanger and Troske (1997) and Dunne,

Foster, Haltiwanger and Troske (2004) for evidence early in our sample; Acemoglu et al. (2022)

for more recent evidence). 19 The logic is that there are fixed costs associated with changing

17
In Appendix Table A.4, we explore whether the relationships in Table 5 have changed over time. The basic
answer is no.
18
Hubmer and Restrepo (2021) is more theoretically focused and much of the attention in their analysis is on the
declining labor share. However, in an extension of their framework they consider variable markups estimated in a
manner similar to DEU using COMPUSTAT data. Rather than estimate flexible functional forms (as we do with,
for example, translog) they estimate a Cobb-Douglas specification with output elasticities of the variable factor of
production permitted to vary across time, industry and firm size classes. Their imposition of constant returns to
scale implies capital output elasticities must be higher and rising for large firms.
19
As noted, there is a large theoretical literature as well. Early papers include Cooper, Haltiwanger and Power
(1999) while more recent papers include Acemoglu and Restrepo (2018).

25
technology. The results on capital intensity, computer investment per worker and diversification

combined with those on firm size are consistent with this interpretation. Within industries,

establishments with higher indicators of these variables have lower estimated output elasticities

of the variable factor and in turn lower estimated markups.

It is instructive to compare the magnitude of the coefficients in the upper and lower

panels of Table 6. The estimated coefficients are uniformly higher in the upper panel (markups)

as compared to the lower panel (output elasticities of materials). The difference in these

magnitudes depend on the difference in the covariance between the “less detailed” minus “more

detailed” markup with the explanatory variable and the covariance of the “less detailed” minus

“more detailed” output elasticity with the explanatory variable. If, for the purpose of discussion,

we treat the differences between “less detailed” and “more detailed” using the notation from

section II, then these differences reflect the differences in cov(ε it / α it , X it ) (where X is the

explanatory variable) and cov(ε it , X it ) . The findings imply both these covariances are positive,

but the former is larger than the latter. Put differently, this is a reminder that it is insufficient to

only examine the impact of differences across businesses in output elasticities, one needs to take

into account covariances including the cost share of the variable input.

To provide additional perspective, we exploit industry-level variation in the “less detailed”

minus “more detailed” markups and related industry-level changes in indicators of technology and

business structure. For the latter, we classify industries based upon the long difference from 1977-

2007 for capital intensity, computer intensity, diversification and a measure of concentration. We

use this window of time since this corresponds to the time interval (using Census years) of the

largest increases in markups using the “less detailed” specifications in Figures 1-3. As discussed

above, markups decline from the mid-2000s to 2014. For computer intensity and capital intensity,

26
we use the value of each industry’s change and classify industries as above/below the median

change for each variable (using the revenue-weighted median for the industry). For the

diversification measure, we use the absolute value of the change since industries with either

increases or decreases are changing business structure. For concentration, we use the 20-firm

concentration ratio at the 4-digit level for this purpose (this is closely related to the superstar firm

measures used by Autor et al. (2020)).

Figures 11-14 plot the mean difference between the “less detailed” and “more detailed”

markup estimates in each year for industries with above median industry-level technology/business

structure changes versus below median industry-level technology/business structure changes. We

find that the industries with above median changes in capital intensity (Figure 11), computer

intensity (Figure 12), and diversification (Figure 13) exhibit an increasing difference between the

less detailed and more detailed based markups. Industries with larger changes in concentration

ratios (Figure 14) have about the same increase in the difference between industry differences in

“less detailed” minus “more detailed” markups.

The industry-level findings provide further support for the interpretation that the increase

in “less detailed” minus “more detailed” markups reflects changes in technology and business

structure. In other words, if the rise in markups from the “less detailed” estimates is attributable

to a change in technology, then markups under the “less detailed” estimates should increase

particularly so (beyond the “more detailed” estimates) in industries with greater indicators of

technological change and change in business structure. 20

20
We provide further evidence on the industry long differences in Table A.5. These specifications are broadly
similar to analogous to those reported in Tables 5 and 6. The RHS variable is a dummy variable equal to one if the
industry has a long difference change from 1977-2007 above the sales-weighted median for the technology change
(or concentration ratio) interacted with sub-period dummy variables. The omitted subperiod is 1972-80 with
subperiod dummies for 1981-89, 1990-2005 and 2006-14. The estimated coefficients are positive for all the
technology change measures under all output elasticity estimation approaches for all periods after 1990 and for

27
 Putting the pieces together, we interpret the findings of this section along with those in the

earlier sections as consistent with the following narrative. The way that manufacturing businesses

are producing output has changed substantially over time (the mean increases in Figure 10) with an

uneven pattern across establishments (the standard deviation increases in Figure 10). These

indicators of uneven changing patterns of production are significantly related to the differences

between the “less detailed” and “more detailed” estimates of markups and output elasticities. 21

The main finding from the “more detailed” estimates of production technologies is that they yield

less of an increase in markup. The findings in this section provide supporting evidence that these

“more detailed” estimates of production technologies are related to observable changes in

technology at the establishment and industry-level.

Our findings do not provide causal evidence about why some establishments, their parent

firms, and industries are changing their technology and ways of doing business in ways that

differ from others. Instead, we show that indicators of such within- and between-industry

heterogeneity are closely related to estimates of differences in the estimates of the output

elasticities of the production technology. Our findings highlight that exploring the causes and

consequences of such heterogeneity should be a high priority for future research.

VI. Conclusions and Future Research

virtually all approaches after 1980. They are statistically significant for computer intensity for the 1990-2005
subperiod for all output elasticity estimation approaches and for selected other subperiods for specific estimation
approaches. For capital intensity, the estimates are statistically significant for translog for both the 1990-2005 and
2006-14 subperiods. For the absolute change in diversification, the estimates are statistically significant for both the
1990-05 and 2006-14 subperiods for Cobb-Douglas and translog approaches (and for the cost share approach in the
2006-14 subperiod). In contrast, the estimates for the concentration measure are small in magnitude and never
statistically significant.
21
The uneven nature of technology adoption is a core feature of the empirical evidence (see, e.g., Acemoglu et al.
(2022)) with accompanying evidence that large firms are more likely to adopt capital intensive, advanced
technologies.

28
 Measuring markups from firm or establishment-level data using the “production

approach” on U.S. data yields a striking pattern of rising (sales-weighted) first and second

moments of markups. The rising first and second moments are related since a substantial fraction

of the rising sales-weighted mean is accounted for by the reallocation of sales activity away from

low to high measured markup businesses. The “production approach” depends critically on

accurate estimates of the output elasticities of the variable factors of production. There is a large

literature estimating output elasticities either from cost shares of total costs or from estimates of

the production/revenue function. Much of this literature imposes the same time-invariant output

elasticities across businesses within the same industry.

In the recent pathbreaking work of DEU, output elasticities are permitted to vary across

businesses within industries and over time. They find that permitting output elasticities to

exhibit variation across time and businesses mitigates the measured increase in sales weighted

markups but the residual increase in markups is still substantial. DEU use annual firm-level data

for publicly traded firms and the quinquennial Economic Census data for manufacturing, retail,

and wholesale trade establishments. This limits the degree to which output elasticities can be

permitted to vary across businesses and time. We can use a more flexible approach by relying on

the dataset developed by the Collaborative Micro Productivity (CMP) project at the Census

Bureau that tracks large (roughly 55,000 establishments per year) representative samples of U.S.

manufacturing establishments from the ASM from 1972 to 2014. 22 These data permit much

greater flexibility in estimating output elasticities across establishments.

Using either cost share or estimation methods, we find greater flexibility in output

elasticities (over time and industry) substantially mitigates the measured increase in sales-

22
The CMP data underlie the public domain DiSP data on within industry productivity dispersion (see Cunningham
et al. (2021)).

29
weighted markups. Using the 2-digit translog specification with time-invariant parameters as in

DEU, we find the sales-weighted markup in U.S. manufacturing increases by about 30 log points

from 1980-2014. Using the 4-digit translog specification with parameters that vary by year, we

find the sales-weighted markups declines by about 5 log points from 1980-2014. Similar

substantial differences are evident using either cost share or Cobb-Douglas revenue estimation

approaches.

We find that the substantially mitigated increases in markups with more flexible and

changing production technologies are associated with declines in the sales-weighted markups

within businesses, smaller increases in the dispersion of markups, and smaller roles for

reallocation in accounting for the changing mean. A key finding in the literature is that there has

been a shift towards businesses with higher markups within industries. We also find this pattern,

but the differences in markups across business within industries are less pronounced. Moreover,

the reallocation component is offsetting a substantial within business decline in markups when

using the more flexible production function specifications.

Our results are consistent with the hypothesis that much of measured increases in

markups instead reflect changing production technology. We present supporting evidence for

this hypothesis using observable indicators of changing technology and business structure. We

find that the mean and dispersion across establishments of capital intensity, computer intensity,

diversification into non-manufacturing and relative industry size are increasing over time.

Moreover, all of these indicators are positively related with establishment-level differences in the

“less detailed” minus “more detailed” markups and “less detailed” minus “more detailed” output

elasticity estimates. We also show there is an important between industry component of these

relationships. Our findings are consistent with related findings in the recent literature that part of

30
the explanation for estimated rising markups is lower and declining output elasticities of variable

factors at larger firms.

More research is needed on several dimensions. First of these is whether our results

extend beyond manufacturing. 23 Unfortunately, the CMP database developed for U.S.

manufacturing establishments is not easily replicated for other sectors. A second more

fundamental question is how we should characterize the production technology at the

establishment and firm level. Our findings suggest that the common practice of imposing the

same technology across all establishments in the same (even detailed) industry is likely

problematic. This practice has had a large influence on the literature on misallocation and now

this more recent related literature on markups. Our results suggest we need to open the black

box of different production technologies across businesses in the same industry. In many

respects, we regard this inference as more important than the inference that markups may not be

rising as much as recent work suggests. We think the task approach developed in a series of

recent papers (e.g., Acemoglu and Restrepo (2019) and Acemoglu et al. (2022)) may be helpful

for this important research agenda of characterizing differences across businesses in how they

conduct business.

23
Hubmer and Restrepo (2021) is an important step in this direction for publicly traded firms.

31
References

Acemoglu, Daron and Pascual Restrepo (2018) “The Race between Man and Machine:
Implications of Technology for Growth, Factor Shares, and Employment,” American Economic
Review, 108(6), 1488-1542.

Acemoglu, Daron, and Pascual Restrepo, (2019), “The Task Content of Production,” Working
Paper.

Acemoglu, Daron, Gary Anderson, David Beede, Catherine Buffington, Eric Childress, Emin
Dinlersoz, Lucia Foster, Nathan Goldschlag, John Haltiwanger, Zachary Kroff, Pascual
Restrepo, and Nikolas Zolas (2022) “Automation and the Workforce: A Firm-Level View from
the 2019 Annual Business Survey,” Paper Presented at NBER/CRIW Conference in March 2022.

Autor David, David Dorn, Lawrence F. Katz, Christina C. Patterson, and John Van Reenen
(2020) “The Fall of the Labor Share and the Rise of Superstar Firms,” Quarterly Journal of
Economics, Vol. 135, No. 2, pp. 645-709.

Baily, Martin Neil, Charles Hulten, and David Campbell (1992) “Productivity Dynamics in
Manufacturing Establishments” (pp. 187-249), in M. Baily and C. Winston (Eds.), Brookings
Papers on Economic Activity: Microeconomics (Washington: Brooking's Institution).

Bond, Steve, Arshia Hashemi, Greg Kaplan, Piotr Zoch (2020) “Some Unpleasant Markup
Arithmetic: Production Function Elasticities and Their Estimation from Production Data,” NBER
Working Paper No. 27002.

Cooper, Russell, John Haltiwanger, Jonathan Willis (2020) “Declining Dynamism at the
Establishment Level: Sources and Productivity Implications,” July, Working Paper.

Cooper, Russell, John Haltiwanger, and Laura Power (1999) “Machine Replacement and the
Business Cycle: Lumps and Bumps,” American Economic Review, 89 (4): 921-946.

Cunningham, Cindy, Lucia Foster, Cheryl Grim, John Haltiwanger, Sabrina Wulff Pabilonia, Jay
Stewart, and Zoltan Wolf (2021) “Dispersion in Dispersion: Measuring Establishment-Level
Differences in Productivity,” CES Working Paper 18-25RRR, Center for Economic Studies, U.S.
Census Bureau.

De Loecker, Jan, Jan Eeckhout, and Gabriel Unger (2020) “The Rise of Market Power and the
Macroeconomic Implications,” Quarterly Journal of Economics, May, 135(2), 561–644

De Loecker, Jan, Jan Eeckhout, and Gabriel Unger (2019) “The Rise of Market Power and the
Macroeconomic Implications,” Working Paper, November.

De Loecker, Jan, Jan Eeckhout, and Gabriel Unger (2018) “The Rise of Market Power and the
Macroeconomic Implications,” Working Paper, November.

32
De Loecker, Jan, and Frederick M. P. Warzynski (2012) “Markups and Firm-level Export Status,”
American Economic Review, 102(6), 2437–2471.

De Ridder, Maarten, Basile Grassi, and Giovanni Morzenti (2022) “A Hitchhikers Guide to
Markup Estimation,” mimeo.

Decker, Ryan, John Haltiwanger, Ron Jarmin and Javier Miranda (2020) “Changing Business
Dynamism and Productivity: Shocks vs. Responsiveness,” American Economic Review,
December.

Demirer, Mert (2020) “Production Function Estimation with Factor-Augmenting Technology: An

Application to Markups,” mimeo.

Dunne, Timothy, John Haltiwanger, and Kenneth R. Troske (1997) “Technology and Jobs: Secular
Change and Cyclical Dynamics.” Carnegie-Rochester Public Policy Conference Series 46 (June):
107–78.

Dunne, Timothy, Lucia Foster, John Haltiwanger, and Kenneth R. Troske (2004) “Wages and
Productivity Dispersion in U.S. Manufacturing: The Role of Computer Investment,” Journal of
Labor Economics, April.

Edmond, Chris, Virgiliu Midrigan, and Daniel Yi Xu (2018) “How Costly are Markups,” NBER
Working Paper No. 24800.

Fort, Teresa C. and Shawn D. Klimek (2018) “The Effects of Industry Classification Changes on
US Employment Composition,” CES Working Papers 18-28, Center for Economic Studies, U.S.
Census Bureau.

Fort, Teresa, Justin Pierce and Peter Schott (2018) “New Perspectives on the Decline of US
Manufacturing Employment,” Journal of Economic Perspectives, 32(2), 47-72.

Foster, Lucia, Haltiwanger, John, and C. J. Krizan (2001) “Aggregate Productivity Growth:
Lessons from Microeconomic Evidence,” in Edward Dean, Michael Harper, and Charles Hulten
(Eds.), New Developments in Productivity Analysis (Chicago: University of Chicago Press).

Hall, Robert (1988) “The Relation between Price and Marginal Cost in U.S. Industry,” Journal of
Political Economy, 96(5), 921–947.

Haltiwanger, John (1997) “Measuring and Analyzing Aggregate Fluctuations: The Importance of
Building from Microeconomic Evidence,” Federal Reserve Bank St. Louis Review, 79(3),
55–77.

Hsieh, Change-Tai and Peter J. Klenow (2009) “Misallocation and Manufacturing TFP in China
and India,” The Quarterly Journal of Economics, 124 (4), 1403-48.

33
Hubmer, Joachim and Pascual Restrepo (2021) “Not a Typical Firm: The Joint Dynamics of
Firms, Labor-Shares and Capital Substitution,” mimeo.

Raval, Devesh (2020) “Testing the Production Approach to Markup Estimation” (October 21,
2020), SSRN: https://ssrn.com/abstract=3324849.

Syverson, Chad, (2004) “Product Substitutability and Productivity Dispersion,” Review of

Economics and Statistics, 86:2, 534-550.

Syverson, Chad (2011) “What Determines Productivity?” Journal of Economic Literature,

49(2):326-365.

Wooldridge, Jeffrey (2009) “On Estimating Firm-Level Production Functions using Proxy
Variables to Control for Unobservables,” Economic Letters, 104 (112-114).

34
Table 1. Output Elasticities for Materials, Cost Share (CS) Approach

Panel A. All industries

Level of aggregation Mean SD
2-digit, constant over time 0.5856 0.0344
3-digit, constant over time 0.07435
4-digit, constant over time 0.1026
6-digit, constant over time 0.1199
Plant-level, constant over time 0.1813
2-digit, yearly 0.03707
3-digit, yearly 0.0797
4-digit, yearly 0.107
6-digit, yearly 0.1259
Plant-level, yearly 0.2051
Panel B. Top 50 industries
Level of aggregation Mean SD
2-digit, constant over time 0.5761 0.0631
3-digit, constant over time 0.09867
4-digit, constant over time 0.1253
6-digit, constant over time 0.1325
Plant-level, constant over time 0.1911
2-digit, yearly 0.06506
3-digit, yearly 0.1024
4-digit, yearly 0.1286
6-digit, yearly 0.1359
Plant-level, yearly 0.2122
Notes: Simple means and standard deviations reported for the pooled
full sample. The mean statistics in the first row of each panel applies
to all following rows in the panel. Panel A has about 2.16 million
establishment-year observations. Panel B has about 750 thousand
establishment-year observations.

35
Table 2. Output Elasticities for Materials, Cobb-Douglas Proxy Method (CD) Approach

Panel A. All industries

Level of aggregation Mean SD
2-digit, constant over time 0.5381 0.01838
3-digit, constant over time 0.5314 0.07713
4-digit, constant over time 0.5193 0.09744
6-digit, constant over time 0.5058 0.1249
2-digit, yearly 0.5295 0.03857
3-digit, yearly 0.5185 0.08687
4-digit, yearly 0.4953 0.1093
Panel B. Top 50 industries
Level of aggregation Mean SD
2-digit, constant over time 0.4802 0.02689
3-digit, constant over time 0.4993 0.06252
4-digit, constant over time 0.4814 0.09022
6-digit, constant over time 0.4801 0.1057
2-digit, yearly 0.4695 0.04653
3-digit, yearly 0.4797 0.0912
4-digit, yearly 0.4579 0.1323
6-digit, yearly 0.4577 0.1416
Notes: Simple means and standard deviations reported. See notes to
Table 1. Yearly estimates are from rolling 5-year windows around
focal year.

36
Table 3. Output Elasticities for Materials, Translog Proxy Method (TL) Approach

Panel A. All industries

Level of aggregation Mean SD
2-digit, constant over time 0.5653 0.1765
3-digit, constant over time 0.5498 0.1857
4-digit, constant over time 0.527 0.1887
6-digit, constant over time 0.5138 0.206
2-digit, yearly 0.5617 0.1813
3-digit, yearly 0.5424 0.1897
4-digit, yearly 0.5019 0.1951
Panel B. Top 50 industries
Level of aggregation Mean SD
2-digit, constant over time 0.5344 0.1965
3-digit, constant over time 0.5237 0.2019
4-digit, constant over time 0.5074 0.1977
6-digit, constant over time 0.503 0.2019
2-digit, yearly 0.5268 0.1963
3-digit, yearly 0.5018 0.2031
4-digit, yearly 0.4713 0.2049
6-digit, yearly 0.468 0.214
Notes: Simple means and standard deviations reported. See notes to
Table 1. Yearly estimates are from rolling 5-year windows around focal
year.

37
 Table 4. Decomposition of the Change in Markups 1982-2012
Total % of Diff., % of Diff., % of Diff.,
Reallocation Within Net Entry Change Realloc. Within Net Entry
CS, Ind4, 1yr 0.1855 0.04112 0.08917 0.3158
CS, Plant, 1yr 0.4041 -0.2469 0.02755 0.1847 -1.667 2.197 0.47
CD, Ind2, 1yr 0.1537 -0.1307 0.08452 0.1075
CD, Ind4, 1yr 0.1393 -0.2341 0.04467 -0.05014 0.09163 0.6556 0.2528
TL, Ind2, Constant 0.3485 -0.09166 0.05682 0.3137
TL, Ind4, 1yr 0.1401 -0.2544 0.06164 -0.05268 0.5688 0.4443 -0.01314
Notes: The markups in the above table are estimated using materials as the variable input. The decomposition above uses revenue
weights. 1yr for CD and TL are from five year rolling windows around focal year.

38
Table 5: Summary Statistics for Plant-Level Regressions Relating Less minus More Detailed Markups
and Less minus More Detailed Output Elasticities
Variable Mean SD

Less minus more detailed markup 0.6097 1.003

Less minus more detailed output elasticity 0.2238 0.2005

Capital Intensity (log (K/L)) 4.908 1.325

Computer Intensity (IHS Comp Inv Per Worker) 0.3706 0.6534

Diversification Index 0.5746 0.7485

Log(firm share) -3.494 1.954

Notes: Summary statistics for the dependent and explanatory variables in Tables
6. All variables are measured at the establishment level. The less minus detailed
markups and output elasticities are from the translog specification. The
diversification index is the IHS of the ratio of nonmanufacturing to manufacturing
employment for the parent firm. The firm share is the share of sales of the parent
firm in the industry of the establishment.

39
 Table 6: Relationship Between Less minus More Markups and Output Elasticities and Indicators of
Technology and Firm Structure
Less minus More Detailed Markup
log(Capital IHS(Computer Inv Diversification Log(Firm
Intensity) Per Worker) Index Share)
Slope Coefficient 0.1441*** 0.0853*** 0.0799*** 0.1139***
(0.0279) (0.0281) (0.0200) (0.0150)
Constant -0.0974 0.5748*** 0.5750*** 0.9749***
(0.1372) (0.0104) (0.0115) (0.0523)
R-squared 0.371 0.3956 0.364 0.421

Less minus More Detailed Output Elasticity, Materials

Slope Coefficient 0.0377*** 0.0175*** 0.0239*** 0.0371***

(0.0056) (0.0063) (0.0059) (0.0033)
Constant 0.0390 0.2220*** 0.2134*** 0.3514***
(0.0273) (0.0023) (0.0034) (0.0114)
R-squared 0.461 0.464 0.456 0.522

Observations 2164000 394000 1924000 472000

Notes: All specifications control for 6-digit industry by year effects using establishment-level observations. Less minus more
detailed markup and output elasticity from translog specification. See notes to Table 5.

40
 Figure 1. Markups Estimated Using Cost Shares (CS)

(a) Long difference in markups 1980-2014

0.4

0.3

0.2

0.1

0
Ind2 Ind3 Ind4 Ind6 Plant
-0.1

-0.2
Constant 1yr

(b) Markups from 1972-2014, benchmark cases

2.5

1.5

0.5

0
1972
1974
1976
1978
1980
1982
1984
1986
1988
1990
1992
1994
1996
1998
2000
2002
2004
2006
2008
2010
2012
2014

Ind4, 1yr (Less Detailed) Plant, 1yr (More Detailed)

Notes: The markups above are estimated using materials as the variable input. Aggregate markups are revenue-weighted
means. Long differences are log differences.

41
 Figure 2. Markups Estimated Using Cobb-Douglas (CD)

(a) Long difference in markups 1980-2014

0.4

0.3

0.2

0.1

0
Ind2 Ind3 Ind4 Ind6
-0.1

-0.2

Constant 1yr

(b) Markups from 1972-2014, benchmark cases

2.5

1.5

0.5

0
1972
1974
1976
1978
1980
1982
1984
1986
1988
1990
1992
1994
1996
1998
2000
2002
2004
2006
2008
2010
2012
2014

Ind2, 1yr (Less Detailed) Ind4, 1yr (More Detailed)

Notes: The markups above are estimated using materials as the variable input. Aggregate markups
are revenue-weighted means. Long differences are log differences.

42
 Figure 3. Markups Estimated Using Translog (TL)

(a) Long difference in markups 1980-2014

0.4

0.3

0.2

0.1

0
Ind2 Ind3 Ind4 Ind6
-0.1

-0.2

Constant 1yr

(b) Markups from 1972-2014, benchmark cases

2.5

1.5

0.5

0
1972
1974
1976
1978
1980
1982
1984
1986
1988
1990
1992
1994
1996
1998
2000
2002
2004
2006
2008
2010
2012
2014

Ind2, Constant (Less Detailed) Ind4, 1yr (More Detailed)

Notes: The markups above are estimated using materials as the variable input. Aggregate markups
are revenue-weighted means. Long differences are log differences.

43
 Figure 4. Long Difference in Naïve Markups 1980-2014

0.3

0.2

0.1

0
CS CD TL
-0.1

-0.2

-0.3
Less detailed More detailed

Notes: The markups above are estimated using materials as the variable input. See equation (3) for
definition of naïve markup. Long differences are log differences.

44
 Figure 5. Long Difference in Materials Output Elasticities 1980-2014

0.3

0.2

0.1

0
CS CD TL
-0.1

-0.2

-0.3
Less detailed More detailed

Notes: The output elasticities above are estimated for materials. Output elasticities are revenue-
weighted means. Long differences are log differences.

Figure 6. Long Difference in Input Shares of Revenue 1980-2014

0.2
0.1
0
M L E K TC
-0.1
-0.2
-0.3
-0.4
-0.5
-0.6

Notes: Input shares of revenue are revenue-weighted means. Long differences are log differences.

45
 Figure 7. Long Differences of Returns to Scale 1980-2014
0.1
0.08
0.06
0.04
0.02
0
CD TL
-0.02
-0.04
Less detailed More detailed

Notes: Returns to scale measured as the sum of estimated output elasticities. Aggregate returns to
scale are revenue-weighted means. Long differences are log differences.

Figure 8. Long Difference of Markups from 1980-2014. Robustness to Total Cost Weighting

0.2
0.15
0.1
0.05
0
CS CD TL
-0.05
-0.1
-0.15
-0.2
Less detailed More detailed

Notes: The markups above are estimated using materials as the variable input. Aggregate markups
are total cost-weighted means. Long differences are log differences.

46
 Figure 9. Dispersion in Markups over Time
(a) Long difference in standard deviation 1980-2014

1.4
1.2
1
0.8
0.6
0.4
0.2
0
CS CD TL
Less detailed More detailed

(b) Long difference in p90-p75 1980-2014

1.4

1.2

0.8

0.6

0.4

0.2

0
CS CD TL
Less detailed More detailed

Notes: The markups above are estimated using materials as the variable input. The markup moments
are computed from revenue-weighted distribution. Long differences are log differences.

47
 Figure 10. Changes in Indicators of Plant-Level Technology
(a) Capital Intensity (log(K/L)) (b) Computer Investment Per Worker

6 1.6 0.8
1.4 0.7
5
1.2 0.6
4
1 0.5
3 0.8 0.4
0.6
2 0.3
0.4
0.2
1
0.2
0.1
0 0
0
1972
1974
1976
1978
1980
1982
1984
1986
1988
1990
1992
1994
1996
1998
2000
2002
2004
2006
2008
2010
2012
2014
1977 1982 1987 1992 2000 2002 2007 2012

Mean Std Dev (Right axis) Mean Std Dev

(c) Diversification Index (IHS Ratio of Non-Mfg/Mfg Emp for Parent Firm) (d) Log firm share

0.9 -3 2.2
0.8 1972 1977 1982 1987 1992 2000 2002 2007 2012
-3.1
0.7 2.1
0.6 -3.2
2
0.5 -3.3
0.4 -3.4 1.9
0.3
-3.5
0.2 1.8
0.1 -3.6
0 1.7
-3.7
1976
1978
1980
1982
1984
1986
1988
1990
1992
1994
1996
1998
2000
2002
2004
2006
2008
2010
2012
2014

-3.8 1.6

Mean Std Dev Mean Std Dev (Right Axis)

Notes: Tabulations from the ASM, CM and LBD. Computer Investment Per Worker uses the inverse hyperbolic sine (IHS). The log firm share is the share of sales of the parent
firm in total industry sales. These are moments not weighted by activity.

48
 Figure 11. Markups and Changes in Capital per Worker

1.3

Less-More Detailed Markup

1.1
0.9
0.7
0.5
0.3
0.1
-0.1 1972 1979 1986 1993 2000 2007 2014
Below median Above median

Notes: The markups above are estimated using materials as the variable input. Less detailed is translog
with parameters that vary at the 2-digit level and are constant over time. More detailed is translog
with parameters that vary at the 4-digit level and vary in rolling annual intervals. Reported are
differences between the less detailed and more detailed markups by year for the two groups defined
by whether establishment is an industry with above or below median long differences in computer
intensity (from 1977 to 2007). Aggregate markups are revenue-weighted means.

Figure 12. Markups and Changes in Computer Intensity

1.2
Less-More Detailed Markup

1
0.8
0.6
0.4
0.2
0
1972 1979 1986 1993 2000 2007 2014
Below median Above median

Notes: The markups above are estimated using materials as the variable input. Less detailed is translog
with parameters that vary at the 2-digit level and are constant over time. More detailed is translog
with parameters that vary at the 4-digit level and vary in rolling year intervals. Reported are
differences between the less detailed and more detailed markups by year for the two groups defined
by whether establishment is an industry with above or below median long differences in computer
intensity (from 1977 to 2007). Aggregate markups are revenue-weighted means.

49
 Figure 13. Markups and Absolute Changes in Diversification

1.3

Less-More Detailed Markup

1.1
0.9
0.7
0.5
0.3
0.1
-0.1 1972 1979 1986 1993 2000 2007 2014
Below median Above median

Notes: The markups above are estimated using materials as the variable input. Less detailed is translog
with parameters that vary at the 2-digit level and are constant over time. More detailed is translog
with parameters that vary at the 4-digit level and vary in rolling year intervals. Reported are
differences between the less detailed and more detailed markups by year for the two groups defined
by whether establishment is an industry with above or below median long differences in diversification
(from 1977 to 2007). Aggregate markups are revenue-weighted means.

50
 Figure 14. Markups and Changes in Concentration

1.3

Less-More Detailed Markup

1.1
0.9
0.7
0.5
0.3
0.1
-0.1 1972 1979 1986 1993 2000 2007 2014
Below median Above median

Notes: The markups above are estimated using materials as the variable input. Less detailed is translog
with parameters that vary at the 2-digit level and are constant over time. More detailed is translog
with parameters that vary at the 4-digit level and vary in rolling annual intervals. Reported are
differences between the less detailed and more detailed markups by year for the two groups defined
by whether establishment is an industry with above or below median long differences in concentration
(from 1977 to 2007). Aggregate markups are revenue-weighted means.

51
Appendix A: Additional Tables and Figures
Table A.1 Output Elasticities for Labor, Cost Share (CS) Approach

Panel A. All industries

Level of aggregation Mean SD
4-digit, yearly 0.2926 0.1015
Plant-level, yearly 0.1706
Panel B. Top 50 industries
Level of aggregation Mean SD
4-digit, yearly 0.2968 0.1183
Plant-level, yearly 0.1773
Notes: Simple means and standard deviations for the full sample are
reported. The mean statistics in the first row of each panel applies to
all following rows in the panel.

52
Table A.2 Output Elasticities for Labor, Cobb-Douglas Proxy Method (CD) Approach

Panel A. All industries

Level of aggregation Mean SD
2-digit, yearly 0.2382 0.05379
4-digit, yearly 0.2362 0.08864
Panel B. Top 50 industries
Level of aggregation Mean SD
2-digit, yearly 0.2345 0.06633
4-digit, yearly 0.22 0.1031
Notes: Simple means and standard deviations for the full sample are
reported.

53
Table A.3 Output Elasticities for Labor, Translog Proxy Method (TL) Approach

Panel A. All industries

Level of aggregation Mean SD
2-digit, constant over time 0.2533 0.1175
4-digit, yearly 0.2575 0.1807
Panel B. Top 50 industries
Level of aggregation Mean SD
2-digit, constant over time 0.2682 0.1446
4-digit, yearly 0.2439 0.1888
Notes: Simple means and standard deviations for the full sample are
reported.

54
 Table A.4: Relationship Between Less minus More Markups and Output Elasticities and Indicators of
Technology and Firm Structure, Time Varying Coefficients

Dependent Variable: Less minus More Detailed Markup

log(Capital IHS(Computer Inv Diversification log(firm
Intensity) Per Worker) Index share)
Slope Coefficient 0.1092** -0.2881 0.0928 0.0923***
(0.0427) (0.1835) (0.0673) (0.0097)
Slope X 81-89 0.0095 0.3124** -0.0256 0.0228***
(0.0212) (0.1499) (0.0285) (0.0052)
Slope X 90-05 -0.0001 0.3276* -0.0026 0.0139
(0.0452) (0.1818) (0.0759) (0.0102)
Slope X 06-14 -0.0215 0.3578* -0.0360 0.0182
(0.0697) (0.1887) (0.0989) (0.0182)
Constant -0.1358 0.3373*** 0.2992*** 0.6758***
(0.1686) (0.0519) (0.0280) (0.0451)
R-squared 0.381 0.407 0.378 0.434
P-value
81-89 = 90-05 0.726 0.8214 0.6501 0.3434
81-89 = 06-14 0.5549 0.539 0.888 0.7812
90-05 = 06-14 0.5179 0.6056 0.3517 0.7906

Dependent Variable: Less minus More Detailed Output Elasticity Materials

Slope Coefficient 0.0383** -0.1170 0.0534** 0.0337***
(0.0182) (0.0775) (0.0263) (0.0032)
Slope X 81-89 0.0062 0.1267** -0.0161 0.0096***
(0.0108) (0.0585) (0.0118) (0.0026)
Slope X 90-05 -0.0102 0.1329* -0.0265 0.0007
(0.0210) (0.0776) (0.0295) (0.0038)
Slope X 06-14 -0.0221 0.1213 -0.0444 0.0001
(0.0250) (0.0781) (0.0345) (0.0049)
Constant -0.0249 0.1517*** 0.1212*** 0.2681***
(0.0679) (0.0259) (0.0104) (0.0216)
R-squared 0.481 0.484 0.482 0.543
P-value
81-89 = 90-05 0.1363 0.8232 0.5931 0.0267
81-89 = 06-14 0.0581 0.8508 0.2452 0.0078
90-05 = 06-14 0.0531 0.1276 0.0047 0.9275

Observations 2164000 394000 1924000 472000

Notes: All specifications control for 6-digit industry by year effects using establishment-level observations. Less minus more
detailed markup and output elasticity from translog specification.

55
 Table A.5. Difference in Markups and Changes in Industry-Level Measures
Dependent Variable: Less detailed markup – more detailed markup
Computer Capital
Change in… intensity intensity Diversification Concentration
(1) (2) (3) (4)
Panel A. Cost share
Above med. X 1981-1989 0.0546*** 0.0019 0.0013 -0.0251
(0.0160) (0.0181) (0.0169) (0.0167)
Above med. X 1990-2005 0.2684* 0.1026 0.1488 -0.0569
(0.1423) (0.1548) (0.1427) (0.1421)
Above med. X 2006-2014 0.1507 0.2654 0.3777* -0.0952
(0.1876) (0.2235) (0.1953) (0.2012)
Above med. -0.0454 0.0691 0.0746* 0.0203
(0.0515) (0.0439) (0.0449) (0.0483)
Constant 0.0936*** 0.0467 0.0387 0.0643***
(0.0220) (0.0332) (0.0340) (0.0229)
Panel B. Cobb-Douglas
Above med. X 1981-1989 -0.0075 0.0269 0.0783*** 0.0072
(0.0207) (0.0242) (0.0195) (0.0215)
Above med. X 1990-2005 0.1609* 0.0783 0.2321*** 0.1025
(0.0887) (0.0961) (0.0881) (0.0882)
Above med. X 2006-2014 0.1433 0.2167 0.4034*** 0.1185
(0.1354) (0.1642) (0.1399) (0.1483)
Above med. -0.0201 0.0926 -0.0370 -0.1355***
(0.0616) (0.0569) (0.0592) (0.0504)
Constant 0.0607* 0.0146 0.0704 0.1239***
(0.0321) (0.0362) (0.0456) (0.0327)
Panel C. Translog
Above med. X 1981-1989 0.0428 0.0624 0.0722* -0.0092
(0.0473) (0.0418) (0.0416) (0.0441)
Above med. X 1990-2005 0.2888** 0.2577** 0.3092*** 0.0939
(0.1282) (0.1287) (0.1187) (0.1345)
Above med. X 2006-2014 0.2073 0.5193** 0.5624*** 0.0518
(0.2334) (0.2237) (0.2014) (0.2381)
Above med. 0.1195** 0.1196** 0.0165 -0.1224**
(0.0495) (0.0573) (0.0517) (0.0526)
Constant 0.1987*** 0.1989*** 0.2398*** 0.3123***
(0.0318) (0.0241) (0.0236) (0.0358)
Observations 2,123,000 2,123,000 2,123,000 2,123,000
Notes: The markups above are estimated using materials as the variable input. All specifications use revenue weights. Standard errors
are clustered at the 6-digit FK-NAICS industry. “Above med.” is a dummy variable equal to one if the change in the industry from 1977-
2007 is above the revenue-weighted median change for all industries. The “change in…” row indicates the relevant measure for
calculating “above med.” in each column. “1981-1989”, “1990-2005”, and “2006-2014” are dummy variables equal to one when the
year is in that year range. The reference years for these specifications are 1972-1980.

56
Figure A.1 Long Differences in Markups 1980-2014 Comparing Trimming versus Winsorizing

0.4

0.3

0.2

0.1

0
CS CD TL
-0.1
Less detailed, trimming More detailed, trimming
Less detailed, winsorizing More detailed, winsorizing

Notes: The markups above are estimated using materials as the variable input. Aggregate markups
are revenue-weighted means. Long differences are log differences.

Figure A.2 Long Difference in Markups 1980-2014, Top 50 Industries

0.5
0.4
0.3
0.2
0.1
0
CS CD TL
-0.1
-0.2
-0.3
Less detailed More detailed

Notes: The markups above are estimated using materials as the variable input. Aggregate markups
are revenue-weighted means. Long differences are log differences.

57
 Figure A.3 Long Difference in Markups 1980-2014

0.5

0
CS CD TL
-0.5

-1

-1.5

-2
Less detailed More detailed

Notes: The markups above are estimated using labor as the variable input. Aggregate markups are
revenue-weighted means. Long differences are log differences.

Figure A.4 Long Difference in Markups 1980-2014, Top 50 Industries

0.8
0.6
0.4
0.2
0
-0.2 CS CD TL

-0.4
-0.6
-0.8
-1
Less detailed More detailed

Notes: The markups above are estimated using labor as the variable input. Aggregate markups are
revenue-weighted means. Long differences are log differences.

58
 Figure A.5 Long Difference in Labor Output Elasticities 1980-2014

0.1
0
CS CD TL
-0.1
-0.2
-0.3
-0.4
-0.5
-0.6
-0.7
Less detailed More detailed

Notes: The output elasticities above are for labor. Output elasticities are revenue-weighted means.
Long differences are log differences.

59
Appendix B. Data Appendix

Our analysis uses the Annual Survey of Manufactures (ASM) from 1972 to 2014. The
ASM surveys roughly 50,000-70,000 establishments. The ASM is a series of five-year panels
(starting in years ending in “4” and “9”) with probability of panel selection being a function of
industry and size. We use the ASM sample weights to adjust for the probability of selection.
A. Output and production factors
We calculate real establishment-level real revenue as 𝑄𝑄𝑗𝑗𝑗𝑗 =
�𝑇𝑇𝑇𝑇𝑆𝑆𝑗𝑗𝑗𝑗 + 𝐷𝐷𝐹𝐹𝑗𝑗𝑗𝑗 + 𝐷𝐷𝑊𝑊𝑗𝑗𝑗𝑗 �⁄𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑡𝑡 , where 𝑇𝑇𝑇𝑇𝑆𝑆𝑗𝑗𝑗𝑗 is total value of shipments, 𝐷𝐷𝐹𝐹𝑗𝑗𝑗𝑗 is the change in
(the value of) finished goods inventories, 𝐷𝐷𝑊𝑊𝑗𝑗𝑗𝑗 is the change in (the value of) work-in-progress
inventories, and 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑡𝑡 is the industry-level shipments deflator, which varies by detailed
industry (4-digit SIC prior to 1997 and 6-digit NAICS thereafter) and is taken from the NBER-
CES Manufacturing Productivity Database and updated as part of the Collaborative Micro
Productivity Project (CMP) (see Cunningham et al. (2021). If the resulting 𝑄𝑄𝑗𝑗𝑗𝑗 is not greater
than zero, then we simply set 𝑄𝑄𝑗𝑗𝑗𝑗 = 𝑇𝑇𝑇𝑇𝑆𝑆𝑗𝑗𝑗𝑗 ⁄𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑡𝑡 . Nominal revenue just uses the numerators
of these measures.
We construct labor from the ASM in terms of total hours (𝑇𝑇𝐻𝐻𝑗𝑗𝑗𝑗 ) as follows:
𝑆𝑆𝑊𝑊𝑗𝑗𝑗𝑗 (B1)
𝑃𝑃𝐻𝐻𝑗𝑗𝑗𝑗 if 𝑆𝑆𝑊𝑊𝑗𝑗𝑗𝑗 > 0 and 𝑊𝑊𝑊𝑊𝑗𝑗𝑗𝑗 > 0
𝑇𝑇𝐻𝐻𝑗𝑗𝑗𝑗 = � 𝑊𝑊𝑊𝑊𝑗𝑗𝑗𝑗
𝑃𝑃𝐻𝐻𝑗𝑗𝑗𝑗 otherwise
where 𝑃𝑃𝐻𝐻𝑗𝑗𝑗𝑗 is production worker hours, 𝑆𝑆𝑊𝑊𝑗𝑗𝑗𝑗 is total payroll, and 𝑊𝑊𝑊𝑊𝑗𝑗𝑗𝑗 is the payroll of
production workers. Nominal labor costs are measured as 𝑆𝑆𝑊𝑊𝑗𝑗𝑗𝑗
We measure capital separately for structures and equipment using the perpetual inventory
method: 𝐾𝐾𝑗𝑗𝑗𝑗+1 = (1 − 𝛿𝛿𝑡𝑡+1 )𝐾𝐾𝑗𝑗𝑗𝑗 + 𝐼𝐼𝑗𝑗𝑗𝑗+1 where 𝐾𝐾 is the capital stock, 𝛿𝛿 is a year- (and industry-)
specific depreciation rate, and 𝐼𝐼 is investment. At the earliest year possible for a given
establishment, we initialize the capital stock by multiplying the establishment’s reported book
value by a ratio of real capital to book value of capital derived from BEA data (where the ratio
varies by 2-digit SIC or 3-digit NAICS). Thereafter, we observe annual capital expenditures and
update the capital stock accordingly, where we deflate capital expenditures using BLS deflators. 1

1
See Cunningham et al. (2020) for more detail.

60
 We calculate real materials as 𝑀𝑀𝑗𝑗𝑗𝑗 = �𝐶𝐶𝑃𝑃𝑗𝑗𝑗𝑗 + 𝐶𝐶𝑅𝑅𝑗𝑗𝑗𝑗 + 𝐶𝐶𝑊𝑊𝑗𝑗𝑗𝑗 �/𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑇𝑇𝑡𝑡 , where 𝐶𝐶𝐶𝐶 is the
cost of materials and parts, 𝐶𝐶𝐶𝐶 is the cost of resales, 𝐶𝐶𝐶𝐶 is the cost of work done for the
establishment (by others) on the establishment’s materials, and 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 is the industry materials
deflator. We calculate energy costs as 𝑁𝑁𝑗𝑗𝑗𝑗 = �𝐸𝐸𝐸𝐸𝑗𝑗𝑗𝑗 + 𝐶𝐶𝐹𝐹𝑗𝑗𝑗𝑗 �/𝑃𝑃𝑃𝑃𝑃𝑃𝑁𝑁𝑡𝑡 , where 𝐸𝐸𝐸𝐸 is the cost of
purchased electricity, 𝐶𝐶𝐶𝐶 is the cost of purchased fuels consumed for heat, power, or electricity
generation, and 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃 is the industry energy deflator. The nominal materials and energy just use
the numerators for these measures.
We use the production factor and output measures described above for our estimation of
the control function approach for estimation of output elasticities. For this estimation, we
combine structures and equipment into a total capital stock. We use the nominal values for cost
shares of revenue and cost shares of total costs. For the latter we use user cost of capital
measures from BLS following Cunningham et al. (2020).
We use the Fort and Klimek (2018) (FK) NAICS consistent industry codes back to 1976.
In turn, we build on that methodology to assign NAICS consistent codes to establishments in the
ASM from 1972 to 1975. The first step of that methodology is that any establishment in the
1972-75 ASM that has an FK NAICS code from the 1976 on period is assigned that code. The
second step is to use SIC-NAICS concordances to assign codes with probabilistic assignment
based on revenue shares when there is a one-to-many or many-to-many concordance.

61
Appendix C. Estimation issues
We follow the approach of DEU in recognizing that in estimating output elasticities that
we don’t observe establishment (firm) level output or input prices. We illustrate the implied
estimation issues with a Cobb-Douglas specification but the same issues apply for the more
general translog specification. Consider a production function for a given industry and time
period (all variables logged):

(A1) yit = θtk kit + θtl lit + θtm mit + θte eit + ωit + ε it

Where yit is output, kit is capital, lit is labor, mit is materials, eit is energy, ωit is a serially

correlated productivity shock and ε it is i.i.d noise. Beyond the well-known issues of endogeneity

of inputs that the control function approach addresses, the additional challenge is prices of both
outputs and inputs at the micro level are not observed. Thus, the relevant revenue equation
(building on equation (29) of Appendix A of DEU) is given by:
(A2) yit + pit= θtk kit + θtl lit + θtm mit + θte eit + ωit + ε it + pit − ∑ θt j jit
j

where j indexes inputs. The error term thus includes the wedge between output and input prices
(the latter weighted by technology parameters). We follow DEU by assuming that this wedge is
related to market share.
In practice, we implement the control function method using Wooldridge (2009) GMM
estimation method. We use the conditional input demand for energy as the control as a nonlinear
function of productivity and capital. However, following Wooldridge we also allow for a
nonlinear relationship between current and lagged productivity. We also include market share of
the establishment at the 4-digit level as in DEU to account for variation in input and factor
markets. The Wooldridge transformation of the revenue function yields revenue as a function of
inputs, market share and a nonlinear function of lagged capital and the control (see equation 2.11
in Wooldridge). Lagged inputs and market shares are instruments. For the translog we include
additional interactions of lagged inputs.
As noted in the main text, DEU implement their control function estimation only for the
COMPUSTAT data which are quite distinct from the establishment-level data we use. When
DEU use the Economic Census data they focus on cost shares. Thus, the most comparable
results with ours are the cost share based output elasticities. Moreover, the details of their

62
implementation of the control function estimation differ from ours especially given the
differences in the firm vs establishment level data. However, we note that when we implement a
control function estimation with the establishment-level data for manufacturing, we obtain
results when using Cobb-Douglass or Translog (see our Figures 2 and 3) that are similar to those
in Figure 12.1 (in Appendix 12 of the 2019 Draft of DEU) when we use similar levels of
aggregation in terms of industry and time. 2 We interpret these patterns as implying that using
our data and methods that we can largely replicate their findings using what we denote as the
“less detailed” estimates.

2
An exception is that they do not find a decline in markups post 2007 in their control function based results for
COMPUSTAT for manufacturing. However, we note that they do find such a decline using the cost share based
results for manufacturing using Economic Census data.

63