You are on page 1of 24

JID: JEBO

ARTICLE IN PRESS [m3Gsc;November 30, 2020;16:23]


Journal of Economic Behavior and Organization xxx (xxxx) xxx

Contents lists available at ScienceDirect

Journal of Economic Behavior and Organization


journal homepage: www.elsevier.com/locate/jebo

Labor costs and the adoption of robots in China


Haichao Fan a, Yichuan Hu b, Lixin Tang c,∗
a
Institute of World Economy, School of Economics, Fudan University, Shanghai, China
b
Faculty of Economics and Management, East China Normal University, Shanghai, China
c
Institute for Economic and Social Research, Jinan University, Guangzhou, China

a r t i c l e i n f o a b s t r a c t

Article history: We study the effects of higher labor costs on the adoption of industrial robots by Chinese
Received 11 May 2020 firms. We use the number of robots imported by firms from the customs data to mea-
Revised 19 October 2020
sure the adoption of robots. To address endogeneity concerns, we focus on the effects of
Accepted 18 November 2020
higher minimum wages from contiguous city pairs that straddle provincial borders. We
Available online xxx
find negligible effects of minimum wages on firms’ robot adoption over 20 04–20 07, but
JEL classification: much larger and statistically significant effects over 2008–2012. For a 10% increase in the
J3 minimum wage over 2008–2012, the probability of a firm adopting robots increases by 0.11
O33 percentage points, compared to an average adoption probability of 0.19% over the same pe-
riod. The effects of higher minimum wages on robot adoption are stronger for firms with
Keywords: higher productivity, those in the coastal region, those with private ownership, and those
Robot
from skilled-labor-intensive industries.
Labor costs
Minimum wage © 2020 Elsevier B.V. All rights reserved.
China

1. Introduction

The use of industrial robots has become a central feature of modern manufacturing. Although advanced industrial coun-
tries, such as the United States, Japan and Germany, have been traditionally dominant players in developing and adopting
robotics technology, China plays an increasingly important role in the adoption of industrial robots. Estimated operation
stock of industrial robots in China increased from around 1.6 thousand in 2001 to 97 thousand in 2012, according to data
from the International Federation of Robotics (henceforth “IFR”). Since 2016, China has been the largest user of industrial
robots in the world (Cheng et al., 2019).
The rapid adoption of industrial robots in China is commonly attributed to rising labor costs.1 However, empirical evi-
dence on this issue is scarce. In this paper, we study the causal effects of higher labor costs on Chinese firms’ adoption of
robots.
An empirical study on the effects of labor costs on firms’ adoption of robots has to confront two major challenges. First,
measuring robot adoption at the firm level is difficult. Data from the International Federation of Robotics (IFR), which is the
most influential source of data for robot adoption, are available only at the country-industry level. To make progress on this
front, we use information about imported robots from the customs data to measure the adoption of robots in China. This
is a sensible approach, because according to the 2014 IFR report, the vast majority of robots used in China over our sample


Corresponding author.
E-mail addresses: fan_haichao@fudan.edu.cn (H. Fan), ychu@jjx.ecnu.edu.cn (Y. Hu), lixintang@jnu.edu.cn (L. Tang).
1
See Jacob and Mishkin (2012) and Bradsher (2017) for examples of newspaper articles that cite rising labor costs as a probable cause for the adoption
of robots in China.

https://doi.org/10.1016/j.jebo.2020.11.024
0167-2681/© 2020 Elsevier B.V. All rights reserved.

Please cite this article as: H. Fan, Y. Hu and L. Tang, Labor costs and the adoption of robots in China, Journal of Economic
Behavior and Organization, https://doi.org/10.1016/j.jebo.2020.11.024
JID: JEBO
ARTICLE IN PRESS [m3Gsc;November 30, 2020;16:23]

H. Fan, Y. Hu and L. Tang Journal of Economic Behavior and Organization xxx (xxxx) xxx

period (2004–2012) were imported from countries such as the United States, Japan, and Germany. We merge the information
on imported robots with a dataset on manufacturing firms from the Annual Surveys of Industrial Firms (henceforth “ASIF”).
Doing so allows us to study adoption behaviors in a panel of manufacturing firms.
The second challenge is the potential endogeneity of labor costs with respect to firms’ adoption of robots. To isolate
exogenous shocks to labor costs, we focus on the minimum wages in China, which vary across cities. Minimum wages in
China have increased substantially in recent years and have been cited as an important factor in rising labor costs (Gan et al.,
2016). The large variation in minimum wages across cities and over time is particularly helpful for studying the effects of
rising labor costs. To address endogeneity concerns, we follow the methodology in Dube et al. (2010), who use contiguous
county pairs across state borders in the United States to identify the employment effects of minimum wages. We construct
a dataset of all city pairs in China that cross provincial borders, as in Fan et al. (2018b). We then include city-pair-year fixed
effects in our panel regressions to purge out spatially correlated unobserved factors.
In Section 3, we develop a model in which workers and robots are substitutes in production. Firms decide whether to
adopt robots by comparing the marginal profits from adoption versus the fixed cost of adoption. In the model, only more
productive firms choose to adopt robots. We derive two testable hypotheses from this simple model. First, a rise in labor
costs increases a firm’s probability of adopting robots. Second, the effect of higher labor costs on the probability of adoption
is stronger for firms with higher productivity and lower fixed cost of adoption.
We study the effects of higher minimum wages on robot adoption using city-pair firm-level data over 2004–2012. We
find that an increase in the minimum wage has only a small effect on the probability of adopting robots in the full sample
of firms. However, this finding masks large variation in the effects over time and across firms. When we examine the effects
of labor costs on robot adoption separately for 20 04–20 07 and 2008–2012, we find negligible effects of minimum wages on
robot adoption for the earlier period, but a much larger and statistically significant effect over the later period.2 For a 10%
increase in the minimum wage over 2008–2012, the probability of adoption increases by 0.11 percentage points, relative to
firms in the other cities. Since the average probability of robot adoption is only 0.19% over the later period, the magnitude
of this estimated effect is large.
We study the heterogeneous effects of higher minimum wages on firms’ adoption of robot along several dimensions. For
each of the dimensions studied, we find the effects to be more dispersed for 2008–2012 than for 20 04–20 07, in line with
the finding that minimum wages have a larger and more significant effect on firms’ adoption of robots over the later half of
our sample period. Consistent with our model, we find the effects to be stronger for firms with higher productivity. We find
the effects of minimum wages on robot adoption to apply to firms from the coastal region, but not to those from the inland
region. In terms of firms’ ownership type, although minimum wages have statistically significant effects on the adoption by
both private firms and state-owned enterprises (SOEs) over 2008–2012, the magnitude is much larger for private firms than
for SOEs. The effects of minimum wages on robot adoption hold for a large number of industries, but the magnitudes of the
effects substantially vary across industries. Finally, relating the heterogeneous effects to industry characteristics, we find the
effects of minimum wages on robot adoption to be stronger for industries that employ skilled labor more intensively. These
results suggest that the relationship between labor costs and robot adoption is influenced by a myriad of factors, such as
firm productivity, geographic location, and industry-specific robotics technology.3
We conduct additional analysis to examine the robustness of our baseline results. To account for other potential con-
founding factors, we add additional control variables, including measures for trade exposure, fixed capital investment and
demographic characteristics, to the regressions. Since our main coefficients of interest are not substantively changed from
our baseline regressions, our results are not likely to be driven by these factors. To address the issue that our baseline
firm-level analysis omits smaller firms outside the ASIF dataset, we construct a city panel of robot imports based solely on
the customs database, which include robot imports by firms both inside and outside the ASIF. We then study the effects of
minimum wages on total quantity of adopted robots in a panel of cities. Consistent with our firm-level results, we find large
effects of higher minimum wages on robot adoption over 2008–2012, but we do not find any effects for 20 04–20 07.
To address the concern that our estimates may be driven by the omission of domestically produced robots, we make
use of the fact that the main customers of Chinese robot suppliers are concentrated in four industries (2014 IFR). We show
that higher minimum wages have large and positive effects on robot adoption also for the industries to which Chinese
robot manufacturers do not supply many robots. To address the issue that robots observed in the custom database may be
imported by robot manufacturers in China either as intermediate input or for R&D purpose, we use firm names to identify
potential manufacturer of robots in the ASIF. We then repeat our baseline analysis after dropping these firms as well as
those in the same 4-digit industry. Our baseline results are robust to the exclusion of these potential robot manufacturers
from the sample.
Our paper contributes to understanding the intricate relationship between industrial robots and the labor markets.
Perhaps because of the widespread concerns in many countries that robots may make workers redundant (Ford, 2015),

2
We divide our sample period into 20 04–20 07 and 20 08–2012 because there are several important changes around 2008, including crucial advances
in robotic technology, the global financial crisis and the subsequent policy response by the government, and the introduction of Labor Contract Law. We
discuss these changes in greater details in Section 5.2.
3
As noted by Adachi et al. (2020), if the price of the welding robot falls relative to the assembling robot, the incentive to adopt robots would be
increased in the automobile industry, which intensively uses welding processes, relative to the electric machine industry which intensively uses assembling
processes.

2
JID: JEBO
ARTICLE IN PRESS [m3Gsc;November 30, 2020;16:23]

H. Fan, Y. Hu and L. Tang Journal of Economic Behavior and Organization xxx (xxxx) xxx

the predominant focus of the literature is on the employment-reducing effects of robot adoption. Related papers include
Acemoglu et al. (2020), Bessen et al., 2019, Carbonero et al., 2018, Dauth et al., 2018, Dixon et al., 2020, Dinlersoz and
Wolf (2018), Giuntella et al., 2019 and Koch et al., 2019, among others. Nevertheless, a smaller number of papers, including
those by Acemoglu and Restrepo, 2018 and Cheng et al. (2019), study factors affecting robot adoptions. Our paper belongs
to the latter category. Much of the empirical literature on robot adoption is based on the data provided by the IFR (2014
IFR). Examples of studies using the IFR data include Graetz and Michaels (2018), Acemoglu and Restrepo, 2018, de Backer
et al., 2018, and Krenz et al. (2018). The industry-level variation in the IFR data can combine with variation in local indus-
try composition to measure local exposure to robot adoption, as in Acemoglu and Restrepo (2020) for the United States,
Chiacchio et al., 2018 for the EU. In contrast, our procedure of relying on customs data to infer robot adoption at the firm
level is similar to that in Humlum (2019).4 Different from Humlum (2019) which studies the interaction between robot
adoption and labor market dynamics, we focus on the causal link between rising labor costs and robot adoption in China.
We see our paper, and other papers using the customs transaction to measure robot adoption such as Humlum (2019), as
complements to the papers based on aggregate IFR data.
The literature on minimum wages primarily focuses on the employment effects in developed countries, particularly in the
United States (Brown et al., 1982; Card and Krueger, 1994; Neumark and Wascher, 20 0 0; Dube et al., 2010; Autor et al., 2016,
among others). Recent studies, including those by Aaronson and Phelan (2017) and Lordan and Neumark (2018), emphasize
the differences between jobs in terms of their susceptibility to automation, for example as measured by the routine task
intensity (RTI) index from Autor and Dorn (2013). These studies find that an increase in the minimum wage leads to a
reduction in employment in the more automatable occupations. An assumption underlying these studies is that the adoption
of automation technology, such as industrial robots, is responsible for these effects. However, because of data availability,
few studies consider the effects of higher minimum wages on the adoption of industrial robots. Our paper fills this gap.
Our paper is more closely related to the strand of papers on the minimum wages in China, including Wang and Gun-
derson (2011), Huang et al., 2014, Fang and Lin (2015), Long and Yang (2016), Hau et al., 2016, Gan et al. (2016) and
Mayneris et al. (2018). The literature consistently finds that minimum wages in China have large effects on firms’ per-
formance and operating decisions. The closest paper to ours is Fan et al. (2018b), who study the effects of minimum wages
on firms’ decisions to conduct foreign direct investment (FDI) in other countries, thereby substituting domestic labor in-
puts with foreign labor inputs in production. This paper complements Fan et al. (2018b) by studying the effects of higher
minimum wages on the adoption of robots, which is another way to substitute away from labor use.5
The rest of the paper is organized as follows: Section 2 discusses the relevant institutional background; Section 3 presents
a simple theoretical model and discusses the testable implications; Section 4 presents our empirical specification and the
data; Section 5 presents the main results; Section 6 discusses heterogeneous effects; Section 7 discusses the robustness of
the main results; and Section 8 concludes.

2. Background

2.1. Minimum wage in China

In this section, we briefly discuss the institutional background on minimum wages in China. Related studies, such as
those by Fan et al. (2018b) and Fang and Lin (2015), provide further details. China’s first national minimum wage regulations
were issued in 1993 and written into the Labor Law in 1994. Provincial governments are granted substantial flexibility in
setting their own minimum wages (Wang and Gunderson, 2011). Local minimum wages can vary by city within a province
according to local economic conditions.
The Chinese Ministry of Labor and Social Security introduced the Minimum Wage Regulations in 2004. Provincial gov-
ernments from then on are required to renew the standard of the minimum wage at least once every 2 years. Adjustments
to the local minimum wages have become more frequent and are accompanied by rapid increases. The 2004 regulations
significantly strengthened the enforcement of the minimum wage system by increasing the penalties for violations. Com-
pliance with the minimum wage regulation has substantially improved over time (Du et al., 2016). Recent studies, such as
those by Gan et al. (2016) and Fan et al. (2018b), find larger effects of minimum wages after 2004 on firms’ export and FDI
decisions, respectively. Therefore, we focus our analysis on the years after 2004.
In 2008, the government introduced a new Labor Contract Law (Du et al., 2016; Gallagher et al., 2015). The new law
is considered be highly protective of workers, with extensive regulations on labor contracts and severance conditions. Im-
portantly, the new law reinforced the importance of minimum wage standards for the labor market in China (Huang et al.,
2014). Du et al., 2016 find the 2008 law significantly contributed to the uniform enforcement of labor regulations across
Chinese cities.

4
Both Humlum (2019) and our paper take care to deal with potential issues with distributors. In our case, we exclude trade intermediaries in the dataset.
Both papers also make use of the IFR data to validate the measurement strategy. However, while Humlum (2019) focuses on single-establishment firm, we
consider adoption at all establishments, which might potentially belong to the same firm. We do not have information on the headquarter-subsidiary
relationship for these establishments.
5
Faber (2020) studies how robot adoption in the US affects employment by US multinational firms in Mexico.

3
JID: JEBO
ARTICLE IN PRESS [m3Gsc;November 30, 2020;16:23]

H. Fan, Y. Hu and L. Tang Journal of Economic Behavior and Organization xxx (xxxx) xxx

2.2. Robots in China

Following studies like Graetz and Michaels (2018) and Acemoglu and Restrepo, 2018, we focus on the adoption of in-
dustrial robots. According to ISO 8373:2012, an industrial robot refers to “automatically controlled, reprogrammable, mul-
tipurpose manipulator, programmable in three or more axes, which can be either fixed in place or mobile for use in in-
dustrial automation applications.”6 Industrial robots have played an increasingly important role in modern manufacturing.
Graetz and Michaels (2018) calculates that robot density, defined as the stock of robots per million hours worked, increased
by more than 150%, from 0.58 in 1993 to 1.48 in 2007, for the seventeen developed countries in their study. The operational
stock of industrial robots worldwide increased from 1.04 million units in 2008 to 1.33 million units in 2013, an increase of
7.8% per year (2014 IFR). Industries with the heaviest use of industrial robots are the automotive industry and the electri-
cal/electronics industries, accounting for 38% and 18% of the total stock at the end of 2013.
Growth in the adoption of robots in China is even more striking. The estimated operation stock of industrial robots
increased from around 1.6 thousand in 2001 to 97 thousand in 2012.7 By 2013, China ranked 5th among countries with
highest operational stocks of industrial robots (Cheng et al., 2019). Most of the robots in China were imported from Japan,
Europe or North America before 2013. According to the 2014 IFR report, of the 23 thousand new robot shipments to China
in 2012, only about 3 thousand were produced by Chinese suppliers, and this domestic share of new robots was even lower
before 2012.8
The adoption of industrial robots in China has continued at a rapid pace since 2012. In 2016, China overtook Japan as the
country with the largest operational stock, accounting for 19% of the total worldwide stock of industrial robots (Cheng et al.,
2019). Rapid adoption has been accompanied by growing robot production in China by both domestic and foreign firms.
Between 2012 and 2017, the number of robots produced in China grew by as much as 86% per year. In 2017, about 30% of
the robots produced in China were made by domestic firms.
Government policy plays a prominent role in the development of robotics industry in China (Cheng et al., 2019; de Backer
et al., 2018). In 2013, Ministry of Industry and Information Technology (MIIT) announced its “Guidance on the Promotion and
Development of the Robot Industry,” including a specific goal of developing three to five world-leading robot companies. In
2015, the well-known “Made in China 2025” plan was released, and robotics was identified as one of the ten key sector areas
in order to achieve digitization and intelligent manufacturing. Finally, in the 13th Five-Year Plan (2016–20), the government
introduced the “New Robotic Technology Roadmap” with large budgets for research and development on robotics.
To implement the policy initiatives above, many local governments set up some investment capital and allocated funds
to subsidize robot usage and innovation. For example, the government of Guangdong Province set up a fund of $150 billion
US dollars in 2015 to encourage firms to invest in automation technology and promote robotics innovation, though it is not
clear how much of the fund is spent on promoting robot production versus robot adoption (Yang, 2017). Lin (2018) reports
that government subsidies accounted for up to 40% of the net profits of the four public robotics firms in 2016. These new
developments made the analysis of the causal link between labor costs and robot adoption after 2012 more complicated.
For this reason, we focus our analysis on the years until 2012.

3. Model

In this section, we present a simple model to study firms’ decisions to adopt robots.9 Consider a country with population
L. Each worker supplies one unit of labor inelastically. There are two sectors, one sector with a numeraire good and the
other with differentiated goods. There is perfect competition in the numeraire sector and the output from that sector is
freely traded with the rest of the world. The productivity in that sector, denoted by w, pins down the wage of the home
country.

3.1. Preferences and endowments

We assume that there is a representative consumer in the economy, with the following preference:
  σσ−1 1−α
σ −1
U = yα0 y (ω ) σ dω ,
ω ∈

6
https://www.iso.org/standard/55890.html. For comparison, service robots are robots that perform useful tasks for humans or equipment, excluding
industrial automation applications.
7
Values reported by the IFR report do not include Foxbot robots manufactured by Foxconn Electronics, a Taiwanese-owned enterprise. The estimated
units of Foxbot robots installed in mainland China by 2013 ranges from 10,0 0 0 to 30,0 0 0.
8
Based on their reading of various government reports, Cheng et al. (2019) report that about 5,800 robots were produced in China in 2012. Robot
production by foreign suppliers in China likely accounts for the discrepancy in the number between the 2014 IFR report and Cheng et al. (2019).
9
Our model shares a number of features with the framework in Helpman et al. (2004). First, both models feature firms with heterogeneous productivity
in a parsimonious setup. Second, in both models, firms face fixed costs, either in conducting FDI as in Helpman et al. (2004), or in adopting robots as in
our model. These fixed costs introduce an advantage in favor of larger firms in conducting FDI or adoption robots. Third, the trade-offs faced by firms are
similar across the two models. Specifically, firms pay a fixed cost to reduce variable costs.

4
JID: JEBO
ARTICLE IN PRESS [m3Gsc;November 30, 2020;16:23]

H. Fan, Y. Hu and L. Tang Journal of Economic Behavior and Organization xxx (xxxx) xxx

where y0 is the numeraire good; y(ω ) is a variety of the differentiated good; α ∈ (0, 1 ) governs the relative share of expen-
diture on the numeraire good; ω indexes varieties in product set  in the sector with differentiated goods; and σ > 1 is
the elasticity of substitution between varieties.
Consumer optimization yields the following demand function for variety ω:
y(ω ) = p(ω )−σ P σ −1 E,
where p(ω ) is the price of variety ω; P is an aggregate price index; and E is the aggregate expenditure on differentiated
goods.

3.2. Firms’ behavior

Each firm in the differentiated goods sector produces a unique variety. These firms are heterogeneous in productivity. A
firm with productivity ϕ produces according to
y(ϕ ) = ϕ (l + bm ), (1)
where l and m denote the numbers of workers and robots used in production, respectively, and b is a parameter governing
the relative productivity between a worker and a robot. Eq. (1) implies that workers and robots are perfect substitutes in
production. If workers are employed in production, the marginal cost of one unit of the differentiated variety is w
ϕ , where w
Rrobot
is the wage rate. If robots are used instead, the marginal cost is ϕb , where Rrobot is the rental rate for robots. We assume
Rrobot < w · b, so that the marginal cost of producing with robots is always lower than that of producing without robots. The
assumption that marginal cost is lower when producing with robots follows Acemoglu and Restrepo (2020).
To use robots in production, a firm needs to pay a fixed cost to install the robots. We assume a fixed cost of f · ε , where
ε ∈ (0, ∞ ) is a random variable given by the cumulative distribution function (CDF) G(ε ). The CDF G(ε ) is independent of
the distribution for ϕ and has a mean of one.
To simplify the model, we consider the partial equilibrium where the wage for labor w and the rental rates for robots
Rrobot are assumed to be fixed. Conditional on the demand of consumers, each firm decides whether to adopt robots in
production by comparing the profits from producing without and with using robots. Given the wages for labor and the
rental rates for robots, the firm’s profit without using robots is given by
1
 σ w 1 − σ
πo ( ϕ ) = P σ −1 E.
σ σ −1 ϕ
In comparison, the profit of the same firm when using robots is given by

1
 σ R 1 − σ
πr ( ϕ ) = robot
P σ −1 E − f · ε .
σ σ − 1 bϕ
Therefore, a firm adopts robots if and only if

1
 σ R 1 − σ 1
 σ w 1 − σ
robot
P σ −1 E − P σ −1 E > f · ε . (2)
σ σ −1 ϕ σ σ −1 ϕ
Intuitively, because of the fixed cost associated with robot adoption, only more productive firms find it profitable to
adopt them. Rearranging Eq. (2), we have
 
 Rrobot 1−σ  w 1−σ
ε < ε∗ = − ,
f bϕ ϕ
P σ −1 E
σ
1−σ
where = σ σ −1 . Therefore, the probability of a firm adopting robots is given by Pr(ε < ε ∗ (ϕ )) = G(ε ∗ (ϕ ) ). Fur-
thermore, the impact of an increase in the wage on the probability of adopting robots is
∂ Pr(ε < ε ∗ (ϕ ))
= (σ − 1 )g(ε ∗ (ϕ ) )w−σ ϕ σ −1 > 0,
∂w f
dG (ε )
where g(ε ) = dε
is the probability density function (PDF) of ε . The following proposition summarizes this result:

Proposition 1. Conditional on firm productivity, a firm’s probability of adopting robots increases with an increase in the wage.

To study the heterogeneous effects, we take the second-order derivatives with respect to ϕ and f . These derivatives are
given by:
 
∂ 2 Pr(ε < ε ∗ ) ∂ g( ε ∗ ) ∗
= g( ε ) +

ε (σ − 1 )2 w−σ ,
∂ w∂ ϕ ∂ε ϕ 2 −σ f
 
∂ 2 Pr(ε < ε ∗ ) ∂ g( ε ∗ ) ∗
= − g( ε ) +

ε (σ − 1 )w−σ .
∂ w∂ f ∂ε ϕ 1 −σ f 2
5
JID: JEBO
ARTICLE IN PRESS [m3Gsc;November 30, 2020;16:23]

H. Fan, Y. Hu and L. Tang Journal of Economic Behavior and Organization xxx (xxxx) xxx

∂ g( ε ∗ ) ∂ 2 Pr(ε <ε ∗ )
For convenience, we assume that ε follows a uniform distribution, so that 10
∂ϕ = 0. By inspection, ∂ w∂ ϕ > 0 and
∂ 2 Pr(ε <ε ∗ )
∂ w∂ f < 0. These results give rise to our second proposition:

Proposition 2. The effects of higher wages on the probability of adopting robot are larger for firms with higher productivity and
for firms with lower fixed cost of installing robot.

Taking the model to data requires a source of exogenous variation in labor costs. We rely on the changes in the minimum
wage across cities, which is closely linked to the rising labor costs in China discussed in Li et al. (2012). Theoretically, both
Burdett and Mortensen (1998) and Van Den Berg (2003) show that an increase in the minimum wage will raise the wage
paid by firms even when firms’ wages are already above the minimum wage.
Our theoretical model admits only a single task. Once robots are adopted, firms do not use any labor inputs. To address
the concern that labors and robots are not perfectly substitutable in reality, we follow Acemoglu and Restrepo (2020) and
extend our model into a framework with a task-based production function.11 In Appendix A, we show that our previous
results regarding the effects of labor costs on the extensive margin (the probability of adopting robot) continue to hold in the
model extension. On the intensive margin (defined as the number of robot quantity for robot-adopting firms), an increase
in labor costs has two distinct effects. Besides the substitution effect highlighted above, there is a scale effect acting in the
opposite direction. Specifically, an increase in labor costs may raise the output price and reduce output quantity, thereby
decreasing the demand for robots. Therefore, the overall impact of higher labor costs on the number of robots for robot-
adopting firms is ambiguous. This additional theoretical prediction is tested in Table A.4 and discussed in Section 5.2.

4. Empirical strategy

4.1. Specification

Our empirical strategy builds on Dube et al. (2010), who use contiguous county pairs across state borders in the United
States to identify the employment effects of minimum wages. We start by constructing a dataset of all city pairs in China
that straddle provincial borders. We classify two cities as a pair if they (1) share a border and (2) belong to two different
provinces. A city may belong to multiple city pairs and the resultant city-pair dataset may contain multiple replicates of
the same city.12 We merge this city-pair dataset with the merged firm-level dataset (the ASIF dataset, plus the customs
database). A firm-year observation will appear multiple times in the new dataset if the city of this firm belongs to multiple
city pairs. To resolve the issue of duplicates in the regressions, we weight the observations with the inverse of the number
of duplicates.
This city-pair-firm dataset offers two advantages over standard approaches. First, contiguous regions have relatively sim-
ilar underlying economic conditions (Dube et al., 2010). In this dataset, we can use city-pair-year fixed effects to account
for time-varying and spatially correlated unobserved factors. Second, cities on provincial borders are on average more dis-
tant from the provincial government that sets the minimum wage. Minimum wages in the border cities are thus arguably
more exogenous to local economic conditions. Because minimum wages in the municipalities Beijing, Tianjing, Shanghai and
Chongqing are set by the municipal government, we drop pairs containing these cities from our city-pair dataset. Finally, we
limit our analysis to 2004–2012. We exclude the years before 2004, a time period during which the minimum wage system
was less effective and robot adoption was rare. We exclude the years after 2012 because of the changes in the Chinese
robotics industry discussed in Section 2. Our final sample contains 2191421 observations from 279 city pairs and 210 unique
cities. Fig. 1 illustrates the sample of cities covered by our city-pair dataset.
We consider the following specification for our empirical investigation

Yi,t = β0 + β1 ln(mwagec,t−1 ) + β2 Zc,t−1 + ϕi + ϕ p,t + εi,t (3)


where Yi,t is a measure of robot adoption for firm i in year t; mwagec,t−1 is the minimum wage in city c in year t − 1, and
ϕ p,t are the fixed effect for city pair p in year t. We lag the local minimum wage and the city control Zc,t−1 by one year, on
the grounds that firms may take time to respond to changes in the minimum wage and to reduce endogeneity concerns. The
city-pair-year fixed effects ϕ p,t in Eq. (3) account for the city-pair-specific time-varying shocks. The vector of city controls
Zc,t−1 include the city population, the city gross domestic product (GDP) per capita, the price level and the unemployment
rate. By including ϕ p,t , we are effectively comparing firms’ robot adoption behaviors to those in the bordering city within
the same city pair. We cluster the standard errors at the city level.13
While the research design based on city pairs helps to address endogeneity concerns, a caveat on the interpretation of
the estimation results should be noted. The identification strategy above hinges on comparisons of the outcome variable

10
Proposition 2 can be derived with a weaker assumption, namely, that ∂ g(∂ϕε ∗ |ϕ )
is not too negative.
11
We thank an anonymous referee for suggesting this extension.
12
As an example, Meizhou of the Guangdong Province borders Zhangzhou and Longyan of the Fujian Province to the east, and Ganzhou of the Jiangxi
Province to the north. Meizhou appears in three city pairs as a result.
13
We also cluster the standard errors at both the city level and the city-pair-year level following Cameron et al. (2011). As Table A.2 shows, our results
are robust to this alternative method of clustering,

6
JID: JEBO
ARTICLE IN PRESS [m3Gsc;November 30, 2020;16:23]

H. Fan, Y. Hu and L. Tang Journal of Economic Behavior and Organization xxx (xxxx) xxx

Fig. 1. Neighboring city pairs across provincial borders. Notes: Cities included in our empirical analysis are colored in dark. Provincial borders are colored
in white. We consider two cities to be a pair if they share a border and belong to two different provinces. The municipalities Beijing, Tianjing, Shanghai
and Chongqing are not included (see the text for an explanation).

within city pairs. However, an increase in the minimum wage in a city could have spillover effects on the paired city, for
example through the migration of workers. Our estimation strategy relates relative changes in the minimum wage within a
city pair to the overall differences in robot adoption between the two cities. To the extent spillover effects within city pairs
exist, they are included in the estimated effects. However, recent research shows that frictions in inter-province migration
in China are much larger than those in intra-province migrations (Fan, 2019; Tombe and Zhu, 2019). Since each city pair in
our sample straddles a provincial border, the scope for spillovers through migration is more limited in our setting.

4.2. Data

Our empirical analysis relies on a number of data sources containing rich information, both at the firm level and at the
prefecture-city level. To obtain the firm-level variables, such as total factor productivity (TFP) and robot adoption behavior,
we use the Chinese Annual Survey of Industrial Firms (ASIF) and the transaction-level custom database provided by the
Chinese General Administration of Customs (CGAS). These two datasets have been widely used in combination in various
studies of the Chinese economy such as Yu (2015) and Fan et al. (2018a). The ASIF surveys are conducted by the Chinese
National Bureau of Statistics (NBS) and contain firm-level information, including firm identification (e.g., company name and
location), ownership structure, employment, capital stock, gross output, value added and various financial variables. These
surveys cover all large and medium-sized non-SOEs with annual sales above 5 million RMB before 2011 (approximately
770 thousand USD in 2007) and above 20 million RMB after 2011 (approximately 3.1 million USD in 2011). Until 2007, the
surveys also covered SOEs with annual sales below the threshold of 5 million RMB.14 Although the ASIF data are one of
the most reliable sources for studying Chinese firms, some concerns regarding data quality should be noted (Brandt et al.,
2014). For 2010 in particular, the number of firms in the ASIF data is substantially smaller than those in adjacent years and
no explanation has been offered by the Chinese NBS. For this reason we exclude 2010 from our firm-level analysis.

14
We include all observations in our baseline analysis. Our results are robust to sample adjustment where we exclude the smaller SOEs or apply a sales
threshold of 20 million RMB for all years. Results are available on request.

7
JID: JEBO
ARTICLE IN PRESS [m3Gsc;November 30, 2020;16:23]

H. Fan, Y. Hu and L. Tang Journal of Economic Behavior and Organization xxx (xxxx) xxx

Fig. 2. Adoption of Robots in China.

The customs database covers the universe of all Chinese exports and imports from 2001 to 2012 at the HS 8-digit level.
For each trade transaction, the customs database records detailed information including import and export values, quantities,
product information, the source and the destination countries, firms’ contact information, and the customs regime.15 We
identify imports of industrial robots by firms in the customs database using the HS6 code for industrial robots.16 We then
match the ASIF data with the customs database using the procedure in Fan et al. (2015).
We supplement the firm-level information with city-level variables. Data on minimum wages are manually collected
from local government websites, statistical bulletins, and other public statistics sources on the Internet and have been used
in Fan et al. (2018b). Finally, data on other city-level socioeconomic characteristics come from the commercial data provider
CEIC.

4.3. Measuring robot adoption

A crucial ingredient of our analysis is a disaggregate measure of robot adoption at the firm level. To construct this
measure, we first identify the imports of robots in the customs database (see footnote 16). Fig. 2 plots the total number
of imported robots based on the customs database from 2001 to 2012. The quantity of robot imports increased at an annual
rate of 23%, from 3.7 thousand units in 2001 to 36.5 thousand units in 2012. Fig. 2 also plots the total number of robot sales
from the IFR, which calculates annual robot shipments by country based on statistics provided by national robot associations
and robot suppliers worldwide. Two patterns emerge from this figure. First, the quantity of robot sales in the customs
database is higher than that in the IFR data throughout the period. One potential explanation is that the IFR data might be
based on a narrower definition of industrial robots.17 Another potential explanation is that the robot suppliers and national
robot associations under-report their sales data to the IFR for strategic reasons. The second pattern from Fig. 2 is that the
numbers from the customs database closely track their IFR counterparts. This helps to validate our use of the customs-based
measure for the empirical analysis.
After merging the customs database and the ASIF dataset, we compute the operational stock of robots for each firm
(denoted by Qrobots ) in the ASIF as the cumulative number of imported industrial robots.18 This allows us to estimate the

15
An important feature of the Chinese economy is the prominent role of processing trade, in which a domestic firm imports raw materials or intermediate
inputs and exports the value-added final goods after processing (Feenstra and Hanson, 2005). The customs database classifies each trade transaction into
several specific types of processing trade, such as “processing with assembly” and “processing with inputs.”
16
Industrial robots are counted under the HS6 code 847950, which comprises two HS8 categories: 84795010 (“industrial robots for multiple uses”) and
84795090 (robots, end maneuvers, and other industrial robots”). Acemoglu and Restrepo, 2018 use the same HS6 code to calculate the imports and exports
of robots for their cross-country analysis.
17
In our paper, to identify industrial robots, we use the HS6 code 847950, which comprises two HS8 categories: 84795010 (ǣindustrial robots for mul-
tiple usesǥ) and 84795090 (robots, end maneuvers, and other industrial robotsǥ). By contrast, the IFR report states that the robot statistics are based on
consolidated world data reported by robot suppliers and on the statistics of various national robot associations but is not very clear about the definition
of industrial robots used. Consequently we do not know for certain whether the definition of robots in the IFR statistics coincides with those in the HS
system.
18
The average service life of a robot is estimated to be 12–15 years (2014 IFR).

8
JID: JEBO
ARTICLE IN PRESS [m3Gsc;November 30, 2020;16:23]

H. Fan, Y. Hu and L. Tang Journal of Economic Behavior and Organization xxx (xxxx) xxx

Table 1
Summary statistics.

Variable No. obs. Mean Std. dev. Median

City-level variables
Monthly minimum wage (yuan) 1625 497.2 182.2 460
City population (thousands) 1625 3921.7 2329.4 3460
City GDP per capita (yuan) 1625 17861.9 16273.7 12786
City price level 1625 2.206 0.907 1.962
Unemployment rate 1625 0.073 0.053 0.067
Firm-level variables
Having adopted robots 2,191,421 0.001 0.035 0
Robot quantity 2,191,421 0.017 2.829 0
SOE firm 2,191,421 0.115 0.319 0
Log firm TFP 1,341,886 3.698 0.998 3.702
Skill intensity 1,748,296 0.105 0.049 0.105

effects of minimum wages on the overall quantity of robots. We exclude processing-trade imports of robots in the calculation
for our baseline analysis. However, our results are robust if we include them in the calculation. Finally, we define a firm as
having adopted robots if its operational stock is greater than zero. The resulting dummy variable allows us to examine the
effects of minimum wages on whether a firm adopts any robots.
Our measure of robot adoption offers two distinct advantages over the alternatives. First, The customs database is ad-
ministrative in nature, in the sense that presumably a customs official has to verify the machinery in question. In contrast,
firm-level surveys (that are not the same as the ASIF surveys) such as those used in Cheng et al. (2019), Koch et al., 2019 and
Bessen et al., 2019 rely to some extent on the judgement of the questionaire-taker on what counts as a robot. Second, our
dataset covers a longer time period with a wider geographic coverage compared to other datasets with information on firms’
robot adoption behavior.
Nevertheless, there are some limitations of using the customs database to measure robot adoption. First, the merging
of custom database and the ASIF dataset is not perfect. In particular, smaller firms below a revenue threshold are not
included in the ASIF and the matching procedure are able to link about only 43% of the nonprocessing robot-importing
firms (52% if we exclude trade intermediaries) in the customs data to ASIF firm data.19 Those smaller firms are not included
in our firm-level analysis. Second, some firms in the ASIF data may be importing robots indirectly with the help of trade
intermediaries. This would result in underestimating robot adoption behavior for these firms. Finally, while industrial robots
produced within China were relatively uncommon before 2013, to the extent they do exist, they are not included in our
robot measure. We discuss the robustness of our results with respect to the robot adoption measure in Section 7.
Table 1 presents the summary statistics of the key variables used in our analysis. As Table 1 shows, robot adoption is
relatively rare in China over the sample period, with only 0.12% of firms having adopted any robots. Table A.1 reports the
average robot adoption probability and robot quantity by SOE status, by firm TFP, by region and by industry skill intensity.

5. Empirical results

5.1. Baseline results

Table 2 presents the estimation results based on Eq. (3). Panel A reports results when the dependent variable is whether
a firm has adopted robots. In column (1), besides the variable log minimum wage, we include only firm fixed effects and
year fixed effects. The coefficient on the log minimum wage is -0.0 0 054 with a standard error of 0.0 0 044. We introduce
city-pair-year fixed effects and a vector of city controls in columns (2) and (3), respectively. According to column (3), a
10% increase in the local minimum wage raises a firm’s probability of adopting robots by 0.0074 percentage point, a small
number even after accounting for the fact that only 0.12% of firms adopt any robots in the full sample. Moreover, the point
estimate is not statistically significant.
Panel B of Table 2 presents results when the dependent variable is log robot quantity (more precisely, log(1 + Qrobots ))
where the quantity Qrobots is the cumulative number of robots imported by the firm. According to our preferred estimates
in column (3), an increase in the minimum wage has no significant effect on the quantity of robots adopted by firms.

5.2. Effects over time

According to the results in Table 2, although the effects of minimum wages on robot adoption are positive, the coef-
ficients are quite small and statistically insignificant. However, because of many changes in the Chinese labor market and
in the evolution of robotics technology over the sample period, one might expect the effects of minimum wages on robot

19
These numbers are comparable to those obtained for the matching procedure performed by other researchers. For example, Yu (2015) matches 27% of
firms in the customs database to the ASIF dataset for 20 0 0–20 06.

9
JID: JEBO
ARTICLE IN PRESS [m3Gsc;November 30, 2020;16:23]

H. Fan, Y. Hu and L. Tang Journal of Economic Behavior and Organization xxx (xxxx) xxx

Table 2
Effects of minimum wages on robot adoption.

(1) (2) (3)

A. Adoption probability
Log min wage 0.00054 0.00092 0.00074
(0.00044) (0.00079) (0.00065)
R-squared 0.850 0.851 0.851
B. Log robot quantity
Log min wage 0.00084 0.00201 0.00161
(0.00077) (0.00154) (0.00127)
R-squared 0.874 0.875 0.875
Firm FE Yes Yes Yes
Year FE Yes Yes Yes
City-pair-year FE No Yes Yes
City controls No No Yes
Observations 2,191,421 2,191,421 2,191,421
∗∗∗ ∗∗
Notes: Standard errors, which are in parentheses, are clustered by city. p < 0.01; p < 0.05; ∗ p < 0.1.

Table 3
Effects of minimum wages over time.

(1) (2) (3)

A. Adoption probability
Log min wage×2004–2007 0.00056 0.00026 0.00011
(0.00054) (0.00063) (0.00056)
Log min wage×2008–2012 0.00449 0.01156∗ ∗ 0.01130∗ ∗
(0.00379) (0.00553) (0.00535)
R-squared 0.850 0.851 0.851
B. Log robot quantity
Log min wage×2004–2007 0.00086 0.00077 0.00043
(0.00095) (0.00120) (0.00107)
Log min wage×2008–2012 0.00842 0.02179∗ ∗ 0.02129∗ ∗
(0.00723) (0.01057) (0.01017)
R-squared 0.874 0.875 0.875
Firm FE Yes Yes Yes
Year FE Yes Yes Yes
City-pair-year FE No Yes Yes
City controls No No Yes
Observations 2,191,421 2,191,421 2,191,421
∗∗∗ ∗∗
Notes: Standard errors, which are in parentheses, are clustered by city. p < 0.01; p < 0.05; ∗ p < 0.1.

adoption to vary over time. First, as described by Ford (2015), crucial breakthroughs in related technology have made robots
much more affordable. Consistent with this view, data from the 2014 IFR show that the total sales of robots worldwide
increase at a faster rate starting in 2009 compared with previous years. Second, the global financial crisis in 2008 and the
subsequent policy response by the government, including a large fiscal stimulus package, could have altered firms’ incentive
to invest in robots (Fardoust et al., 2012). Through the lens of the model, these two changes can be thought of as a reduc-
tion in the fixed cost of robot adoption. According to Proposition 2 in Section 3, we would expect the effects of labor costs
on robot adoption to be stronger for the later period. Third and finally, as discussed in Section 2.1, the Labor Contract Law
introduced in 2008 significantly increased firms’ burdens of complying with labor regulations. In view of these factors, we
analyze the effects separately for 20 04–20 07 and 2008–2012 by including relevant interaction terms.
Table 3 present results when we include interactions between log minimum wage, and the dummy variables for 2004–
20 07 and 20 08–2012. Panel A presents results when the dependent variable is whether a firm has adopted robots. As in
the previous table, we include only firm fixed effects and year fixed effects in column (1) and introduce city-pair-year fixed
effects and city controls in columns (2) and (3), respectively. Across the different specifications, we do not find any effects
for 20 04–20 07. For 20 08–2012, the coefficient on log minimum wage is 0.0 0449 in column (1) and 0.01156 in column (2).
The differences between columns (1) and (2) demonstrate the importance of limiting the comparison of robot adoption to
firms from neighboring city pairs with city-pair-year fixed effects. Column (3) presents our preferred estimates with city
controls. The coefficient on log minimum wage for the years for 2008–2012 is 0.011, with a standard error of 0.0054. This is
statistically significant at 5%. According to this point estimate, a 10% increase in the minimum wage raise a firm’s probability
of adopting robots by 0.11 percentage points, relative to firms in the other cities. Because the share of firms that adopted
robots is only 0.12% over the full sample period and 0.19% for the years after 2008, the magnitude of this effect is large.20

20
In our sample, the standard deviation of log minimum wage is 0.3498. Based on our coefficient estimate of 0.01130 for 2008–2012 in column (3) of
Panel A, for a one-standard-deviation increase in log minimum wage, the average adoption probability increases by 0.39 percentage points, relative to firms
in other cities.

10
JID: JEBO
ARTICLE IN PRESS [m3Gsc;November 30, 2020;16:23]

H. Fan, Y. Hu and L. Tang Journal of Economic Behavior and Organization xxx (xxxx) xxx

Panel B of Table 3 presents results on robot quantity. According to the estimates in Column (3), a 10% increase in the
minimum wage leads to an 0.213% increase in the number of robots for the average firm. A caveat is that, because of the
large number of zero values in the variable Qrobots , comparisons between zero-value firms and non-zero-value firms play an
important role in identifying the coefficients in Panel B.
Given the rapid pace of changes in China in recent decades, any partition of the sample period into two sub-periods is
not perfect. In Table A.3, we include a full set of interaction terms between log minimum wage and year dummies to study
the full time dynamics of the effects of minimum wages on robot adoption. Consistent with Table 3, we find that the effects
to larger and statistically more significant for the years after 2007. Additionally, the coefficients corresponding to the years
2008 and 2009 are larger than those for later years. Kopytov et al. (2018) construct a model, in which the adoption of new
technologies by firms is concentrated in downturns due to low opportunity costs, to account for the acceleration of skill-
biased technological change during the Great Recession. While the Great Recession is relatively mild in China compared to
other countries, GDP growth slowed down from an annual rate of 14.2% in 2007 to 9.7% in 2008 and 9.4% in 2009. A greater
number of firms may decide to upgrade their technology by adopting robots during the downturns, because the opportunity
cost associated with adjustment costs due to production disruption would be lower.
By analyzing the effects on whether a firm chooses to adopt any robots, Panel A of both Tables 2 and 3 speaks to the
extensive margin of robot adoption. In contrast, by analyzing the effects on robot quantity for all firms, Panel B of both
tables speaks to the overall effects, which also include the intensive margin where firms that already adopted robots choose
to increase their quantity of robots. The large number of zero values in the variable Qrobots suggests that the patterns of robot
adoption are primarily driven by the first-time robot adopters. In other words, the extensive margin plays a more important
role than the intensive margin. To examine the intensive margin more carefully, we limit the sample to only firms that have
ever adopted robots in Table A.4. As Table A.4 shows, we do not find any significant effects of minimum wages on robot
adoption on the intensive margin. This is consistent with the model extension in Appendix A which predicts that higher
labor costs have ambiguous effects on the intensive margin due to the existence of scale effects.

6. Heterogeneous effects

We analyze the heterogeneous effects of minimum wages on robot adoption in this section. We start by testing
Proposition 2, which states that the effects of minimum wages on robot adoption should be larger for firms with higher
TFP. We then study the effects of minimum wages on robot adoption along three additional dimensions: location, SOE sta-
tus, and industry skill intensity.

6.1. Effects of minimum wage by initial TFP

According to Proposition 2, the effects of higher minimum wages on robot adoption should be larger for firms with higher
TFP. To test the proposition, we introduce an interaction term between log minimum wage and TFP to Eq. (3). Because the
effects on robot adoption are concentrated on the years after 2008, as shown by our results in Table 3, it is interesting to
examine the interaction between log minimum wage and TFP for 20 04–20 07 and 2008–2012 separately. To this end, we
introduce triple interaction terms of log minimum wage, TFP, and sub-period dummies.21
Because firms’ contemporary TFP might be influenced by its robot adoption behavior, we use the initial TFP (i.e., the
TFP when a firm appears in our dataset for the first time) instead. We compute TFP following Olley and Pakes (1996) and
Ackerberg et al. (2015). Because the ASIF data do not have information on either intermediate goods or value added for the
years after 2007, we cannot compute initial TFP for firms entering the dataset after 2007. Consequently, the sample size for
the analysis in this subsection is reduced relative to the baseline.
Table 4 presents the results from our analysis of the heterogeneous effects by initial TFP. As column (1) shows, for the
full sample period, the coefficient on the interaction term of interest is 0.0 0 052 and significant at the 5% level. According to
the estimates in column (1), a 10% increase in the minimum wage raises the probability of a firm adopting robots by 0.018
percentage points for a firm at the 90th percentile of the TFP distribution, compared to 0.005 percentage point for a firm
at the 10th percentile. Column (2) of Table 4 examines the interaction between log minimum wage and TFP for 20 04–20 07
and 2008–2012 separately. Consistent with the finding in Table 3 that minimum wages have no effects on robot adoption for
20 04–20 07, the coefficient on the corresponding triple interaction term in column (2) is very close to zero. Put differently,
before 2007, the effects of minimum wages on robot adoption are negligible, regardless of whether firms have high or low
TFP. Through the lens of the model, the period before 2007 corresponds to the case in which the fixed cost of robot adoption
is so high such that an increase in labor costs have little effect on robot adoption regardless of the level of TFP.
By comparison, the coefficient on the corresponding triple interaction term for 2008–2012 is 0.00076, with a p-value
of 0.11.22 According to the estimates, for 2008–2012, a 10% increase in the minimum wage raises the probability of robot
adoption by as much as 0.16 percentage points for a firm at the 90th percentile of the TFP distribution, compared to 0.14
percentage points for a firm at the 10th percentile.

21
We also include an interaction term between TFP and a 2008–2012 dummy to fully saturate the specification.
22
A likely reason for the lower statistical significance is that when we include triple-interaction terms in Column (2) instead of double-interaction terms
as in Column (1), we are effectively requiring the same data to estimate a greater number of parameters. Consequently there is a loss of statistical power.

11
JID: JEBO
ARTICLE IN PRESS [m3Gsc;November 30, 2020;16:23]

H. Fan, Y. Hu and L. Tang Journal of Economic Behavior and Organization xxx (xxxx) xxx

Table 4
Effects of minimum wages by initial TFP.

(1) (2) (3) (4)


Adoption probability Log robot quantity

Log min wage 0.00082 0.00251


(0.00082) (0.00167)
Log min wage×TFP 0.00052∗ ∗ 0.00133∗ ∗ ∗
(0.00024) (0.00050)
Log min wage×2004–2007 0.00112 0.00022
(0.00155) (0.00243)
Log min wage×2008–2012 0.01240∗ ∗ 0.02398∗ ∗
(0.00603) (0.01109)
TFP×2008-2012 0.00457 0.00373
(0.00334) (0.00521)
Log min wage×TFP×2004–2007 0.00001 0.00073
(0.00030) (0.00046)
Log min wage×TFP×2008–2012 0.00076 0.00135∗
(0.00047) (0.00081)
R-squared 0.720 0.720 0.746 0.747
Firm FE Yes Yes Yes Yes
City-pair-year FE Yes Yes Yes Yes
City controls Yes Yes Yes Yes
Observations 1,341,886 1,341,886 1,341,886 1,341,886
∗∗∗ ∗∗
Notes: Standard errors, which are in parentheses, are clustered by city. p < 0.01; p < 0.05; ∗ p < 0.1.

The results on robot quantity, reported in columns (3) and (4), are similar to those in columns (1) and (2), the main
difference being that the coefficient on the triple interaction term for 2008–2012 is statistically significant at the 10% level.
In summary, in line with Proposition 2, we find the positive effects of minimum wages on robot adoption to be larger for
more productive firms.
Because of missing values for TFP, the sample size in Table 4 is smaller than in the baseline regressions in Table 2.
To mitigate concerns about sample attrition, we use firm size, as measured by total employment, as a proxy for firm TFP,
because there is a one-to-one mapping between TFP and employment size through the lens of the model. We examine the
effects by firm size in Table A.5. We find the effects of minimum wages to be stronger for larger firms. Therefore, the results
in Table A.5 are consistent with those in Table 4. The difference is that employment size is directly observed while firm TFP
has to be estimated from the data using procedures such as Olley and Pakes (1996).

6.2. Effects by region

The coastal region accounts for about 52% of the firms in the ASIF data, whereas the inland region, which covers a much
greater geographic area, accounts for the rest.23 The effects of minimum wages on robot adoption may be stronger for
firms from the coastal region for two reasons. First, the theoretical model suggests the effects of rising labor costs on robot
adoption to be larger for more productive firms. The data show that the more productive firms are indeed clustered in the
coastal region. Second, holding other things constant, due to the high levels, minimum wages are more likely to be binding
on firms in the coastal region and to have an impact on robot adoption decisions.
Similar to our analysis on the effects by firm productivity in Section 6.1, we introduce an interaction term between log
minimum wage and an dummy variable for the coastal region to Eq. (3). To analyze the effects by sub-period, we further
introduce triple interaction terms of log minimum wage, the coastal region dummy, and period dummies. Table 5 presents
the results. According to Columns (1) and (3), an increase in labor costs has no effects at all for firms located in the inland
region, over the full sample period. In contrast, the coefficients on the interaction terms with the coastal region dummy are
positive and statistically significant, indicating that an increase in labor costs has a much larger effect on firms from the
coastal region and the differences are statistically significant. Columns (2) and (4) show that differences between coastal
and inland regions mainly driven by the 2008–2012. In summary, we find the positive effects of minimum wages on robot
adoption to apply to the coastal region, but not to the inland region.

6.3. Effects by SOE status

As documented by Berkowitz et al. (2017), SOE in China are often required to pursue noneconomic objective by hiring
excess labor. Consequently, SOEs are systematically different from private firms. Furthermore, minimum wages are less likely
to be binding for SOEs as the wages at these firms are higher than those at non-SOEs on average. Therefore, an increase in

23
The National Bureau of Statistics of China (NBSC) classifies Chinese provinces into four regions: coastal (east), central, west, and northeast. The coastal
region comprises ten provinces: Beijing, Tianjin, Hebei, Shanghai, Jiangsu, Zhejiang, Fujian, Shandong, Guangdong, and Hainan. We combine the latter three
NBSC regions into one region (“the inland region”) for simplicity.

12
JID: JEBO
ARTICLE IN PRESS [m3Gsc;November 30, 2020;16:23]

H. Fan, Y. Hu and L. Tang Journal of Economic Behavior and Organization xxx (xxxx) xxx

Table 5
Effects of minimum wages by region.

(1) (2) (3) (4)


Adoption probability Log robot quantity

Log min wage 0.0000 0.0002


(0.0005) (0.0009)
Log min wage × Coastal 0.0020∗ ∗ 0.0037∗ ∗
(0.0009) (0.0014)
Log min wage × 2004–2007 0.0005 0.0010
(0.0005) (0.0009)
Log min wage × 2008–2012 0.0038∗ 0.0076∗ ∗
(0.0020) (0.0038)
Coastal × 2008–2012 0.0853∗ ∗ 0.1569∗ ∗
(0.0329) (0.0627)
Log min wage × Coastal × 2004–2007 0.0011 0.0023
(0.0008) (0.0015)
Log min wage × Coastal × 2008–2012 0.0144∗ ∗ ∗ 0.0269∗ ∗
(0.0054) (0.0104)
R-squared 0.851 0.851 0.875 0.875
Firm FE Yes Yes Yes Yes
City-pair-year FE Yes Yes Yes Yes
City controls Yes Yes Yes Yes
Observations 2,191,421 2,191,421 2,191,421 2,191,421
∗∗∗ ∗∗
Notes: Standard errors, which are in parentheses, are clustered by city. p < 0.01; p < 0.05; ∗ p < 0.1.

Table 6
Effects of minimum wages by SOE status.

(1) (2) (3) (4)


Adoption probability Log robot quantity

log minimum wage 0.00093 0.00199


(0.00070) (0.00137)
log min wage×SOE 0.00113∗ ∗ 0.00226∗ ∗
(0.00057) (0.00106)
log min wage×2004-2007 0.00020 0.00065
(0.00059) (0.00113)
log min wage×2008–2012 0.01167∗ ∗ 0.02199∗ ∗
(0.00547) (0.01042)
SOE×2008-2012 0.02311∗ 0.03952
(0.01294) (0.02438)
log min wage×SOE×2004–2007 0.00039 0.00108
(0.00041) (0.00076)
log min wage×SOE×2008–2012 0.00411∗ 0.00742∗
(0.00220) (0.00416)
R-squared 0.851 0.851 0.875 0.875
Firm FE Yes Yes Yes Yes
City-pair-year FE Yes Yes Yes Yes
City controls Yes Yes Yes Yes
Observations 2,191,421 2,191,421 2,191,421 2,191,421
∗∗∗ ∗∗
Notes: Standard errors, which are in parentheses, are clustered by city. p < 0.01; p < 0.05; ∗ p < 0.1.

the minimum wage may have smaller effects on the adoption of robots by SOEs. We examine the heterogeneous effects by
firms’ SOE status in Table 6. Columns (1) and (3) show that the effects of minimum wages on robot adoption, in terms of
both probability and quantity, are much smaller for SOE firms than for private firms. The difference, as indicated by the co-
efficient on the interaction term, is statistically significant at the 5% level. Columns (2) and (4) examine these heterogeneous
effects separately for 20 04–20 07 and 2008–2012. According to the estimates, for 20 04–20 07, an increase in the minimum
wage has no effects on robot adoption, in terms of probability or quantity, for either type of firms. For 2008–2012, a 10%
increase in the minimum wage raises the probability of robot adoption of 0.117 percentage and the quantity of robots by
0.220% for non-SOEs. These effects are significant at the 5% level. The corresponding numbers for SOEs are smaller in mag-
nitude at 0.076 percentage points and 0.146%, respectively. Therefore, robot adoption by SOEs is less sensitive to increases
in the minimum wage. This is consistent with the finding in Fan et al. (2018b) that the outward FDI decisions of SOEs are
unresponsive to changes in the minimum wage.

6.4. Effects by industry

The adoption of robots is not uniform across industries. To examine this heterogeneity in details, we conduct our analysis
separately for selected 2-digit industries based on the China Industry Classification (CIC). In our data, the seven 2-digit

13
JID: JEBO
ARTICLE IN PRESS [m3Gsc;November 30, 2020;16:23]

H. Fan, Y. Hu and L. Tang Journal of Economic Behavior and Organization xxx (xxxx) xxx

Table 7
Effects of minimum wages for selected industries.

(1) (2) (3) (4) (5) (6) (7) (8)


Communication, Instruments, Transportation Special Plastic Rubber Electric others
Industry description computers & meters, cultural & equipment purpose products products machinery &
electronic equip. clerical machinery equipment equipment

A. Adoption probability
Log min wage×2004–2007 0.00181 0.00881 0.00182 0.00181 0.00495 0.00100 0.00202 0.00008
(0.01815) (0.00554) (0.00346) (0.00146) (0.00395) (0.00090) (0.00248) (0.00008)
Log min wage×2008–2012 0.06997∗ ∗ 0.06852∗ ∗ ∗ 0.02037 0.02061∗ ∗ 0.05540∗ ∗ 0.01985∗ 0.01738∗ ∗ ∗ 0.00120
(0.03253) (0.01934) (0.01447) (0.00806) (0.02469) (0.01118) (0.00507) (0.00085)
R-squared 0.874 0.898 0.861 0.872 0.833 0.923 0.830 0.865
B. Log robot quantity
Log min wage×2004–2007 0.01604 0.01304∗ 0.00008 0.00189 0.01134 0.00162 0.00389 0.00008
(0.03249) (0.00706) (0.00693) (0.00274) (0.00832) (0.00146) (0.00493) (0.00012)
Log min wage×2008–2012 0.16723∗ ∗ ∗ 0.07440∗ ∗ ∗ 0.03686 0.04172∗ ∗ 0.10606∗ ∗ 0.03195∗ 0.02729∗ ∗ 0.00173
(0.06220) (0.02105) (0.02552) (0.01654) (0.04734) (0.01799) (0.01061) (0.00112)
R-squared 0.879 0.959 0.880 0.893 0.884 0.937 0.816 0.908
Firm FE Yes Yes Yes Yes Yes Yes Yes Yes
City-pair-year FE Yes Yes Yes Yes Yes Yes Yes Yes
City controls Yes Yes Yes Yes Yes Yes Yes Yes
Observations 46,262 20,674 74,263 72,896 69,269 19,503 92,747 1,792,193

Notes: The 2-digit industries are arranged in descending order of average robot adoption probability. The industry description corresponding to the 2-digit
industry codes is as follows: 40, communication equipment, computers, and other electronic equipment; 41, instruments, meters, cultural, and clerical
machinery; 37, transportation equipment; 36, special purpose equipment; 30, plastic products; 29, rubber products; 39, electric machinery and equipment.
Standard errors, which are in parentheses, are clustered by city. ∗ ∗ ∗ p < 0.01; ∗ ∗ p < 0.05; ∗ p < 0.1.

industries with the highest probability of robot adoption accounts for only 18.3% of the firms in our data and 84.3% of the
robot-adopting firms over 2004–2012. We repeat our analysis in Section 5 for each of the seven industries and a subsample
combining all other industries.
Table 7 presents the estimation results. Throughout the table, we do not find any effects over 20 04–20 07. Consider
Panel A, column (1), which corresponds to the ”communication equipment, computers, and other electronic equipment”
industry. This industry has the highest probability of robot adoption in our data, with an average probability of 2.0%. The
coefficient on log minimum wage over 2008–2012 is 0.0700, implying a 0.700 percentage point increase in the probability
of robot adoption for a 10% increase in the minimum wage. The corresponding coefficients in columns (2)–(7) in Panel A
of Table 7 range from 0.0174 in column (7) to 0.0685 in column (2). These coefficients in columns (2)–(7) are statistically
significant at the 10% level or higher, with the exception of column (3). Finally, column (8) reports results for firms from
the other industries. For these firms, the average probability of robot adoption is much lower at 0.0230%. Consequently,
the coefficient on log minimum wage over 2008–2012 for these firms is also smaller at 0.00120, with a standard error of
0.0 0 085. In summary, the positive effects of minimum wages on robot adoption continue to hold when we study selected
industries in isolation. Furthermore, the magnitudes of these effects vary considerably across industries.
We explore the heterogeneous effects across industry by relating them to industry characteristics. We focus on the skill
intensity of an industry, as measured by the share of workers with at least a four-year college education.24 We conduct
the analysis at the 4-digit industry level. The data on skill intensity are calculated from the 2004 Firm Census in China
by Che and Zhang (2018). According to the data, the most skill-intensive 4-digit CIC industry is petrochemical refineries,
whereas the least skill-intensive is leathers manufacturing.
Columns (1) and (3) of Table 8 present the overall effects on adoption probability and quantity over the sample period.
We find the coefficient on the interaction between log minimum wage and the skill intensity measure to be positive and
statistically significant. A potential explanation is that recent advances in robotics technology are especially applicable to
several skill-intensive industries, such as the manufacturing of precision instruments and electronics. In columns (2) and (4)
we introduce triple interactions—of log minimum wage, skill intensity, and time dummies—to analyze the heterogeneous
effects for 20 04–20 07 and 20 08–2012 separately. The results show that the effects of minimum wages on firms’ robot
adoption are stronger for firms in more skill-intensive sectors for 2008–2012.
One concern about the results in Table 8 is that the skill-labor intensity of an industry could be influenced by robot
adoption. In Table A.6, we use data on skill intensity from the United States in 1980 (available from Ciccone and Papaioan-
nou, 2009) to replace the measure from China to address the endogeneity concerns. The correlation in industry-level skill
intensity between this measure and that for China in 2004 is 0.7471. As Table A.6 show, our results are robust to the alter-
native measure of industry-level skill intensity.

24
One could use a production function with workers with different skill levels to consider the role of skill intensity for robot adoption in the model.
However, this would substantially complicate our model. We refrain from pursuing this route.

14
JID: JEBO
ARTICLE IN PRESS [m3Gsc;November 30, 2020;16:23]

H. Fan, Y. Hu and L. Tang Journal of Economic Behavior and Organization xxx (xxxx) xxx

Table 8
Effects of minimum wages by skill intensity.

(1) (2) (3) (4)


Adoption probability Log robot quantity

Log min wage 0.0002 0.0006


(0.0003) (0.0006)
Log min wage × Skill 0.0014∗ ∗ 0.0018∗ ∗
(0.0007) (0.0009)
Log min wage × 2004–2007 0.0001 0.0003
(0.0003) (0.0006)
Log min wage × 2008–2012 0.0031∗ 0.0052∗
(0.0016) (0.0030)
Skill × 2008–2012 0.1658 0.3091
(0.1031) (0.2098)
Log min wage × Skill × 2004–2007 0.0001 0.0005
(0.0005) (0.0011)
Log min wage × Skill × 2008–2012 0.0273∗ 0.0501
(0.0165) (0.0332)
R-sq 0.859 0.860 0.884 0.884
Firm FE Yes Yes Yes Yes
City-pair-year FE Yes Yes Yes Yes
City controls Yes Yes Yes Yes
Observations 1,748,296 1,748,296 1,748,296 1,748,296
∗∗∗ ∗∗
Notes: Standard errors, which are in parentheses, are clustered by city. p < 0.01; p < 0.05; ∗ p < 0.1.

To summarize this section, we find the effects of minimum wages on robot adoption to be stronger for more productive
firms, those located in the coastal region, those with private ownership, and those in more skill-intensive industries.

7. Robustness

We discuss several potential concerns regarding our empirical approach in this section. In our baseline results, we include
a number of control variables including city population, city GDP per capita, city unemployment and city price level. To
address concerns regarding other potential confounding factors, we add additional control variables to the regressions. We
include a measure of city-level trade exposure, defined as the sum of import and export values as a ratio of city GDP.
To account for changes in stock of other capitals, we control for city-level fixed capital investment (in log). The preceding
two variables are both obtained from the CEIC database. To account for demographic characteristics, we include as control
variables the female share of the population aged between 15 and 60 and the share of workers with university degree or
higher, both measured at the provincial level and available from provincial statistical yearbooks. As Table A.7 in the appendix
shows, the main coefficients of interest are not substantively changed from our baseline regressions. Therefore, our results
are not likely to be driven by these factors.
Another concern is that some firms in the ASIF data may be indirectly importing robots with the help of trade interme-
diaries. This would result in underestimating robot adoption behavior for these firms. To address this issue, we first identify
trade intermediaries in the custom database. Following Ahn et al. (2011) and Yu (2015), we identify intermediary firms
based on Chinese characters with the meaning of “importer”, “exporter”, or “trading” in the firm’s name. Fig. A1 shows
that the share of robots imported by trade intermediaries is less than 25% throughout 2004–2012. Further, importing robots
through trade intermediaries should be less of a concern for the medium and large firms in the ASIF dataset, because firms
importing through trade intermediaries tend to be smaller in size (Ahn et al., 2011).
A further concern is that, even putting aside the role of trade intermediaries, our baseline results are based on only a
subset of the remaining robot imports in the customs database, as a result of the imperfect matching procedure between
the customs data and the ASIF data. To address concerns about the role of firms outside the ASIF, we construct a city panel
of robot imports using only the customs database. For each year, we aggregate the robot imports to the city level, based on
the location of the importing firm. We exclude imports by trade intermediaries but include all other firms, both inside and
outside the ASIF. We then study the effects of minimum wages on the total quantity of adopted robots in a panel of cities.
We conduct the analysis on the sample of city pairs, with the following specification:
ln(robotc,t ) = γ0 + γ1 ln(mwagec,t−1 ) + γ2 Zc,t−1 + ωc + ω p,t + εc,t (4)
where robotc,t is the cumulative number of robots in city c in year t, and ωc and ω p,t are city and city-pair-year fixed effects,
respectively.
Table 9 presents the results. Columns (1) and (2) present the overall effects on robot adoption over 2004–2012 from
regressions without and with city controls, respectively. For both columns, the coefficient on log minimum wage is about
0.22, but is statistically insignificant. Columns (3) and (4) report the effects by period. According to our preferred estimates
in column (4), for 2008–2012, a 10% increase in the minimum wage in a city raises the quantity of robots by about 8.5%

15
JID: JEBO
ARTICLE IN PRESS [m3Gsc;November 30, 2020;16:23]

H. Fan, Y. Hu and L. Tang Journal of Economic Behavior and Organization xxx (xxxx) xxx

Table 9
Results from city-level regressions.

Log city robot quantity

(1) (2) (3) (4)

Log min wage 0.223 0.220


(0.257) (0.259)
Log min wage×2004–2007 0.216 0.194
(0.215) (0.207)
Log min wage×2008–2012 0.901∗ ∗ 0.851∗ ∗
(0.419) (0.426)
R-squared 0.940 0.942 0.942 0.943
City-pair-year FE Yes Yes Yes Yes
City controls No Yes No Yes
Observations 3,578 3,578 3,578 3,578
∗∗∗ ∗∗
Notes: Standard errors, which are in parentheses, are clustered by city. p < 0.01; p < 0.05; ∗ p < 0.1.

relative to other cities. This large effect on robot quantity from city-level regressions contrasts with that from our firm-level
regressions. Because many firms in the data have not adopted any robots, the coefficients from our firm-level regressions in
Tables 2–3 are dominated by comparisons between zero-value firms and non-zero-value firms. Table 9, however, shows that
increases in the minimum wage have large and positive effects on the adoption of robots in a city, especially for 2008–2012.
Finally, one might be concerned that while robots produced in China do not appear in the customs database, they might
still be used by firms in the ASIF. As discussed earlier, only about 3,0 0 0 robots were produced by Chinese suppliers in 2012,
according to data by the IFR (2014 IFR). However, one still may be concerned that the estimates may be driven by the
omission of these domestically produced robots. The 2014 IFR report notes that the main customers of Chinese robot sup-
pliers are concentrated in four industries corresponding to manufacturing electrical, electronic, rubber, and plastic products,
respectively. In our industry analysis in Section 6.4, we find that an increase in the minimum wage has positive effects on
robot adoption not only for the four industries above (columns (1), (5), (6), and (7) of Table 7) but also for those in which
Chinese manufacturers do not supply many robots. Furthermore, the results for the former industries do not systematically
differ from those for the latter ones. These observations mitigates concerns about the role of domestically produced robots.
A related concern is that the robots observed in the custom database may be imported by robot manufacturers in China,
either as intermediate input or for R&D purpose. To address this concern, we make use of the ASIF variable that has the
names of firms. We identify a firm as a manufacturer of robots if its name contains the Chinese characters for “robot.” This
step results in a list of 43 robot-manufacturing firms. We then repeat our analysis after dropping these firms and find the
results to be unchanged. Next, we drop not only these 43 firms, but also all firms sharing the same 4-digit CIC industry,
based on the assumptions that robot-producing firms should share the same disaggregate industrial classification. This step
eliminates seven 4-digit CIC subindustries from our baseline sample.25 Table A.8 in the appendix presents the regression
results based on the resultant sample. As Table A.8 shows, the coefficients on log minimum wage over 2008–2012 are
slightly reduced, but they remain significant at the 5% level.
One also may be concerned that our results could be driven by the construction of our city-pair-firm dataset. To address
this concern, we show that our results hold in a panel of firms from all cities (instead of just those in a city pair straddling
a provincial border) in Table A.9. A drawback to this alternative approach is that we can no longer include city-pair-year
fixed effects to reduce endogeneity concerns.
Finally, one may be concerned by our choice of the sample period. We focus our analysis on 2004–2012, because the
minimum wage system was less effective and robot adoption was very rare before 2004. In Table A.10, we include data
from 2001 to 2012 in our analysis as a robustness check. As Table A.10 shows, an increase in the minimum wage continues
to have positive effects on firms’ robot adoption for 2008–2012, but not for the earlier years.

8. Conclusion

With rising labor costs in China, robots are widely expected to play a greater role in the Chinese economy. Our analysis
of the effects of minimum wages on firms’ robot adoption supports this view to some extent. However, our heterogeneity
analysis shows that the effects of higher minimum wages on robot adoption are far from uniform over time or across firms.
This suggests that the relationship between labor costs and robot adoption is influenced by a myriad of factors such as
institutional setting, firm productivity, and industry-specific robotics technology. Given the diversity of the Chinese economy
along these dimensions, the patterns of robot adoption should continue to vary at the firm level. These diverse patterns of
robot adoption could have important implications for the distribution of firm productivity and labor income in the economy.

25
The CIC industry classification was introduced in 2002, when robots were less common. This may explain why robot-producing firms are not assigned
the same 4-digit CIC industry code in the ASIF data.

16
JID: JEBO
ARTICLE IN PRESS [m3Gsc;November 30, 2020;16:23]

H. Fan, Y. Hu and L. Tang Journal of Economic Behavior and Organization xxx (xxxx) xxx

Declaration of Competing Interest

The authors declare that they have no conflict of interest.

Acknowledgments

We thank the editor Scott Adams and two anonymous referees for helpful comments and suggestions. Fan acknowledges
financial support from the Shanghai Institute of International Finance and Economics, the Ten Thousand Talents Program
(Young Talents) of China, and the Youth Innovative Team on Humanities and Social Sciences of Fudan University. Tang ac-
knowledges the support of the National Natural Science Foundation of China (No. 71903071) and the 111 Project of China
(No. B18026).

Appendix A. Model with task-based production function

Following Acemoglu and Restrepo (2020), we extend our model into a multi-task framework. Under this framework, the
production function satisfies

y(ϕ ) = ϕ min (l (s ) + b(s )m(s ) ), (5)


s ∈ [0 , 1 ]

where l (s ) and m(s ) denote the numbers of workers and robots used in task s, respectively, and b(s ) is a parameter gov-
erning the relative productivity between a worker and a robot. The comparative advantage of robots differ across tasks, as
captured by the term b(s ). Following Acemoglu and Autor (2011), we incorporate a Ricardian relation (Dornbusch et al.,
1977) into our theoretical model by assuming b(s ) to be continuously differentiable and strictly decreasing. Therefore, there
exists a cutoff s∗ for robot-adopting firms such that robots are used in tasks s < s∗ and labors are used in the tasks s > s∗ ,
R
where b(s∗ ) = robot
w .
On the extensive margin, the firm compares its profits from producing without and with using robots. Given the wage
for labor and the rental rate for robots, the firm’s profit without using robots is given by
 σ w 1 − σ
1
πo ( ϕ ) = P σ −1 E.
σ σ −1 ϕ
In comparison, the profit of the same firm when using robots is given by
  1−σ
1 σ s∗ Rrobot (1 − s∗ )w
πr ( ϕ ) = ds + P σ −1 E − f · ε .
σ σ −1 0
b(s )ϕ ϕ
Therefore, a firm adopts robots if and only if
  1−σ  σ w 1 − σ
1 σ s∗ Rrobot (1 − s∗ )w 1
ds + P σ −1 E − P σ −1 E > f · ε . (6)
σ σ −1 0
b(s )ϕ ϕ σ σ −1 ϕ
Intuitively, because of the fixed cost associated with robot adoption, only more productive firms find it profitable to
adopt them. Rearranging Eq. (6), we have
 1−σ
s∗ Rrobot 1 −σ
ε<ε = ∗
ds + (1 − s∗ )w −w ,
ϕ 1 −σ f 0
b( s )

σ −1
1−σ
where = P σ E σσ−1 . The probability of a firm adopting robots is given by Pr(ε < ε ∗ (ϕ ) ) = G(ε ∗ (ϕ ) ). Furthermore,
the impact of an increase in the wage on the probability of adopting robots is
  −σ 
∂ Pr(ε < ε ∗ ) (σ − 1 ) ∗ −σ
s∗ Rrobot
= g ( ε ) w − ds + ( 1 − s ∗
)w ( 1 − s ∗
) > 0,
∂w ϕ 1 −σ f 0
b( s )
 −σ
s∗ Rrobot
where the inequality stems from 0 b( s ) ds + (1 − s∗ )w (1 − s∗ ), which equals to w−σ for s∗ = 0 and equals to 0 for
s∗ = 1, is decreasing in s∗ . Hence, Proposition 1 in Section 3 continues to hold:

Proposition A1. Conditional on firm productivity, a firm’s probability of adopting robots increases with an increase in the wage.

17
JID: JEBO
ARTICLE IN PRESS [m3Gsc;November 30, 2020;16:23]

H. Fan, Y. Hu and L. Tang Journal of Economic Behavior and Organization xxx (xxxx) xxx

To study the heterogeneous effects, we take the second-order derivatives with respect to ϕ and f . These derivatives are
given by:
    −σ 
∂ 2 Pr(ε < ε ∗ ) ∂ g(ε ∗ ) ∗ (σ − 1 )2 −σ
s∗ Rrobot
= g( ε ) +

ε w − ∗
ds + (1 − s )w ∗
(1 − s )
∂ w∂ ϕ ∂ε ϕ 2 −σ f 0
b( s )
    −σ 
∂ 2 Pr(ε < ε ∗ ) ∂ g(ε ∗ ) ∗ (σ − 1 ) −σ
s∗ Rrobot
= − g( ε ∗ ) + ε w − ds + ( 1 − s ∗
)w ( 1 − s ∗
)
∂ w∂ f ∂ε ϕ 1 −σ f 2 0
b( s )

ε ∗ |ϕ )
If ε follows a uniform distribution, we have ∂ Pr∂(wε <
2 (ε <ε ∗ )
> 0 and ∂ Pr
2
∂ϕ ∂ w∂ f < 0. Therefore, Proposition 2 in Section 3 con-
tinues to hold as well.

Proposition A2. The effects of higher wages on the probability of adopting robot are larger for firms with higher productivity
and for firms with lower fixed cost of installing robot.

On the intensive margin, the impact of the increase in labor costs is different. The number of adopted robots for robot-
adopting firms equals to
  −σ
s∗ 1 σ s∗ Rrobot (1 − s∗ )w
Q= ds ds + P σ −1 E
0
b(s )ϕ σ −1 0
b(s )ϕ ϕ
Different from the extenstive margin, there are two opposite effects of an increase in labor costs on the intensive margin,
which satisfies
1
∂ ln Q b( s∗ ) ∂ s∗ 1 − s∗
= s∗ 1 −σ s∗
∂w 0 b(s ) ds
∂w 0
Rrobot
b( s )
ds + (1 − s∗ )w
     
Substitution Effects>0 Scale Effects<0

On the right-hand side of the above equation, the first term corresponds to the substitution effects. The substitution effects
1
∗) Rrobot
could be written as − s∗b(s ∂ b( s∗ ) 2 , which is larger than zero since b(s ) is decreasing in s. The second term corresponds
1 ds
0 b( s ) ∂ s∗ w
to the scale effects and is negative. These results give rise to the following proposition:

Proposition A3. There are two opposite effects of an increase in labor costs on the intensive margin. The overall impact on the
number of adopted robots for robot-adopting firms is ambiguous.

Appendix B. Additional Figures and Tables

Fig. A1. Share of robot imports by trade intermediaries.

18
JID: JEBO
ARTICLE IN PRESS [m3Gsc;November 30, 2020;16:23]

H. Fan, Y. Hu and L. Tang Journal of Economic Behavior and Organization xxx (xxxx) xxx

Table A.1
Summary Statistics by Subsamples.

Variable No. obs. Mean Std. Dev. Median

by ownership status
SOE
Adoption probability 255,577 0.000 0.012 0
Robot quantity 255,577 0.000 0.043 0
Non-SOE
Adoption probability 1,971,224 0.001 0.036 0
Robot quantity 1,971,224 0.019 2.983 0
by firm TFP
High TFP Firms
Adoption probability 677,442 0.001 0.035 0
Robot quantity 677,442 0.014 0.967 0
Low TFP Firms
Adoption probability 677,447 0.001 0.025 0
Robot quantity 677,447 0.005 0.432 0
by region
Coastal Region
Adoption probability 1,164,830 0.002 0.045 0
Robot quantity 1,164,830 0.024 1.282 0
Inland Region
Adoption probability 1,061,971 0.000 0.017 0
Robot quantity 1,061,971 0.010 3.837 0
by skill intensity
High Skill Intensity Firms
Adoption probability 783,329 0.001 0.038 0
Robot quantity 783,329 0.021 4.455 0
Low Skill Intensity Firms
Adoption probability 965,878 0.000 0.025 0
Robot quantity 965,878 0.002 0.272 0

Table A.2
Results using two-way clustered standard errors.

(1) (2) (3) (4)


Adoption probability Log robot quantity

Log min wage 0.00074 0.00161


(0.00094) (0.00183)
Log min wage×2004-2007 0.00011 0.00043
(0.00082) (0.00160)
Log min wage×2008-2012 0.01130∗ 0.02129∗
(0.00578) (0.01100)
R-squared 0.851 0.851 0.875 0.875
Firm FE Yes Yes Yes Yes
City-pair-year FE Yes Yes Yes Yes
City controls Yes Yes Yes Yes
Observations 2,191,421 2,191,421 2,191,421 2,191,421

Notes: Standard errors, which are in parentheses, are twoway-clustered by city and
city-pair-year. ∗ ∗ ∗ p < 0.01; ∗ ∗ p < 0.05; ∗ p < 0.1.

19
JID: JEBO
ARTICLE IN PRESS [m3Gsc;November 30, 2020;16:23]

H. Fan, Y. Hu and L. Tang Journal of Economic Behavior and Organization xxx (xxxx) xxx

Table A.3
Effects on robot adoption by year.

(1) (2)
Adoption probability Log robot quantity

Log min wage 0.00183 0.00363


(0.00134) (0.00265)
Log min wage×2005 0.00017 0.00042
(0.00071) (0.00141)
Log min wage×2006 0.00057 0.00139
(0.00095) (0.00193)
Log min wage×2007 0.00485∗ ∗ 0.00981∗ ∗
(0.00246) (0.00493)
Log min wage×2008 0.01436∗ ∗ 0.02676∗ ∗
(0.00706) (0.01337)
Log min wage×2009 0.01353∗ ∗ 0.02611∗ ∗
(0.00596) (0.01159)
Log min wage×2011 0.00203 0.00344
(0.00188) (0.00305)
Log min wage×2012 0.00333∗ 0.00576∗
(0.00199) (0.00344)
R-squared 0.851 0.875
Firm FE Yes Yes
City-pair-year FE Yes Yes
City controls Yes Yes
Observations 2,191,421 2,191,421

Notes: Data for 2010 are included in the regression, because the number of firms in the
2010 ASIF data is substantially smaller than those in adjacent years and no explanation
has been offered by the Chinese NBS. Standard errors, which are in parentheses, are
clustered by city. ∗ ∗ ∗ p < 0.01; ∗ ∗ p < 0.05; ∗ p < 0.1.

Table A.4
Effects of minimum wages on robot quantity of robot-adopting firms.

(1) (2) (3) (4)

Log min wage 0.1070 0.1246


(0.3723) (1.4385)
Log min wage × 2004–2007 0.1015 0.1160
(0.3715) (1.4207)
Log min wage × 2008–2012 0.3826 0.0552
(0.7166) (1.1818)
R-squared 0.831 0.850 0.832 0.850
Firm FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
City-pair-year FE No Yes No Yes
City controls No Yes No Yes
Observations 3,700 3,456 3,700 3,456
∗∗∗
Notes: Standard errors, which are in parentheses, are clustered by city. p < 0.01;
∗∗
p < 0.05; ∗ p < 0.1.

20
JID: JEBO
ARTICLE IN PRESS [m3Gsc;November 30, 2020;16:23]

H. Fan, Y. Hu and L. Tang Journal of Economic Behavior and Organization xxx (xxxx) xxx

Table A.5
Effects of minimum wages by firm size.

(1) (2) (3) (4)


Adoption probability Log robot quantity

Log min wage 0.00022 0.00033


(0.00057) (0.00117)
Log min wage×Size 0.00017∗ 0.00035
(0.00010) (0.00024)
Log min wage×2004-2007 0.00851∗ ∗ 0.01841∗
(0.00399) (0.00971)
Log min wage×2008–2012 0.02565 0.05742
(0.01980) (0.04492)
Size×2004–2007 0.01050∗ ∗ 0.02301∗
(0.00502) (0.01212)
Size×2008–2012 0.04589 0.09811
(0.02949) (0.06518)
Log min wage×Size×2004–2007 0.00179∗ ∗ 0.00391∗
(0.00085) (0.00207)
Log min wage×Size×2008–2012 0.00748 0.01596
(0.00477) (0.01055)
R-squared 0.851 0.852 0.875 0.875
Firm FE Yes Yes Yes Yes
City-pair-year FE Yes Yes Yes Yes
City controls Yes Yes Yes Yes
Observations 2,191,421 2,191,421 2,191,421 2,191,421

Notes: Firm size is measured by log firm employment. Standard errors, which are in parentheses, are
clustered by city. ∗ ∗ ∗ p < 0.01; ∗ ∗ p < 0.05; ∗ p < 0.1.

Table A.6
Effects by skilled intensity from the US in 1980.

(1) (2) (3) (4)


Adoption probability Log robot quantity

Log min wage 0.00036 0.00069


(0.00057) (0.00105)
Log min wage×Skill 0.00196∗ 0.00236∗ ∗
(0.00104) (0.00104)
Log min wage×2004–2007 0.00005 0.00002
(0.00053) (0.00097)
Log min wage×2008–2012 0.00712∗ ∗ 0.01264∗
(0.00358) (0.00653)
Skill×2008–2012 0.14360∗ 0.23162
(0.08359) (0.14599)
Log min wage×Skill×2004–2007 0.00099∗ 0.00069
(0.00052) (0.00060)
Log min wage×Skill×2008-2012 0.02441∗ 0.03858
(0.01397) (0.02361)
R-squared 0.855 0.856 0.886 0.886
Firm FE Yes Yes Yes Yes
City-pair-year FE Yes Yes Yes Yes
City controls Yes Yes Yes Yes
Observations 1,817,352 1,817,352 1,817,352 1,817,352
∗∗∗ ∗∗
Notes: Standard errors, which are in parentheses, are clustered by city. p < 0.01; p < 0.05; ∗ p < 0.1.

21
JID: JEBO
ARTICLE IN PRESS [m3Gsc;November 30, 2020;16:23]

H. Fan, Y. Hu and L. Tang Journal of Economic Behavior and Organization xxx (xxxx) xxx

Table A.7
Results with additional city-level control variables.

(1) (2) (3) (4)


Adoption probability Log robot quantity

Log min wage 0.0010 0.0021


(0.0009) (0.0017)
Log min wage × 2004–2007 0.0001 0.0003
(0.0006) (0.0010)
Log min wage × 2008–2012 0.0115∗ ∗ 0.0218∗ ∗
(0.0052) (0.0098)
Log population 0.0011 0.0008 0.0036 0.0030
(0.0009) (0.0007) (0.0023) (0.0020)
Log GDP per capita 0.0004 0.0000 0.0008 0.0016
(0.0009) (0.0010) (0.0021) (0.0022)
Price level 0.0015∗ 0.0011∗ 0.0027 0.0018
(0.0008) (0.0006) (0.0016) (0.0012)
Unemployment rate 0.0037 0.0054 0.0066 0.0098
(0.0031) (0.0034) (0.0058) (0.0065)
Trade exposure 0.0012∗ 0.0022∗ ∗ 0.0031∗ ∗ ∗ 0.0051∗ ∗ ∗
(0.0006) (0.0009) (0.0010) (0.0014)
Log fixed capital investment 0.0010∗ ∗ 0.0004 0.0018∗ ∗ 0.0007
(0.0005) (0.0003) (0.0009) (0.0006)
Female share 0.0277∗ 0.0339∗ ∗ 0.0504∗ 0.0621∗ ∗
(0.0150) (0.0147) (0.0280) (0.0299)
Share of university graduates 0.0329 0.0280∗ ∗ 0.0642 0.0551∗ ∗
(0.0217) (0.0140) (0.0406) (0.0262)
R-squared 0.851 0.851 0.875 0.875
Firm FE Yes Yes Yes Yes
City-pair-year FE Yes Yes Yes Yes
Observations 2,191,421 2,191,421 2,191,421 2,191,421
∗∗∗ ∗∗
Notes: Standard errors, which are in parentheses, are clustered by city. p < 0.01; p < 0.05; ∗ p < 0.1.

Table A.8
Regression results excluding robot manufacturers.

(1) (2) (3) (4)


Adoption probability Log robot quantity

Log min wage 0.00069 0.00144


(0.00063) (0.00121)
Log min wage×2004–2007 0.00010 0.00034
(0.00054) (0.00101)
Log min wage×2008–2012 0.01060∗ ∗ 0.01982∗ ∗
(0.00527) (0.00983)
R-squared 0.848 0.848 0.872 0.872
Firm FE Yes Yes Yes Yes
City-pair-year FE Yes Yes Yes Yes
City controls Yes Yes Yes Yes
Observations 2,173,231 2,173,231 2,173,231 2,173,231

Notes: We identify all firms in the data that have the Chinese characters for ”robot” in the firm name. We
drop these firms, as well as those sharing the same 4-digit CIC industry code. Standard errors, which are in
parentheses, are clustered by city. ∗ ∗ ∗ p < 0.01; ∗ ∗ p < 0.05; ∗ p < 0.1.

Table A.9
Results from a standard firm panel covering all cities.

(1) (2) (3) (4)


Adoption probability Log robot quantity

Log min wage 0.0001 0.0003


(0.0003) (0.0006)
Log min wage×2004–2007 0.0002 0.0002
(0.0003) (0.0006)
Log min wage×2008–2012 0.0021∗ 0.0046∗ ∗
(0.0011) (0.0020)
R-squared 0.870 0.870 0.898 0.898
Firm FE Yes Yes Yes Yes
Year FE Yes Yes Yes Yes
City controls Yes Yes Yes Yes
Observations 2,138,279 2,138,279 2,138,279 2,138,279
∗∗∗ ∗∗
Notes: Standard errors, which are in parentheses, are clustered by city. p < 0.01; p < 0.05; ∗ p < 0.1.

22
JID: JEBO
ARTICLE IN PRESS [m3Gsc;November 30, 2020;16:23]

H. Fan, Y. Hu and L. Tang Journal of Economic Behavior and Organization xxx (xxxx) xxx

Table A.10
Robustness check for 2001–2012.

(1) (2) (3) (4)


Adoption probability Log robot quantity

Log min wage×2001–2003 0.00031 0.00095 0.00073 0.00250∗ ∗


(0.00062) (0.00058) (0.00111) (0.00113)
Log min wage×2004–2007 0.00011 0.00006 0.00047 0.00031
(0.00054) (0.00051) (0.00103) (0.00098)
Log min wage×2008–2012 0.01136∗ ∗ 0.01115∗ ∗ 0.02132∗ ∗ 0.02095∗ ∗
(0.00537) (0.00526) (0.01014) (0.00996)
R-squared 0.821 0.835 0.850 0.862
Firm FE Yes Yes Yes Yes
City-pair-year FE Yes Yes Yes Yes
City controls No Yes No Yes
Observations 2,744,966 2,525,825 2,744,966 2,525,825
∗∗∗ ∗∗
Notes: Standard errors, which are in parentheses, are clustered by city. p < 0.01; p
< 0.05; ∗ p < 0.1.

References

Aaronson, D., Phelan, B.J., 2017. Wage shocks and the technological substitution of low-wage jobs. Econ. J. 129 (617), 1–34.
Acemoglu, D., Autor, D., 2011. Skills, tasks and technologies: implications for employment and earnings. In: Handbook of Labor Economics, 4. Elsevier,
pp. 1043–1171.
Acemoglu, D., LeLarge, C., Restrepo, P., 2020. Competing with robots: firm-level evidence from france.
Acemoglu, D., Restrepo, P., 2018. Demographics and automation. NBER Working Paper.
Acemoglu, D., Restrepo, P., 2020. Robots and jobs: evidence from us labor markets. J. Polit. Econ. 128 (6), 2188–2244.
Ackerberg, D.A., Caves, K., Frazer, G., 2015. Identification properties of recent production function estimators. Econometrica 83 (6), 2411–2451.
Adachi, D., Kawaguchi, D., Saito, Y. U., 2020. Robots and employment: evidence from japan, 1978–2017.
Ahn, J., Khandelwal, A.K., Wei, S.-J., 2011. The role of intermediaries in facilitating trade. J. Int. Econ. 84 (1), 73–85.
Autor, D., Dorn, D., 2013. The growth of low-skill service jobs and the polarization of the us labor market. American Economic Review 103 (5). 1553–97
Autor, D.H., Manning, A., Smith, C.L., 2016. The contribution of the minimum wage to us wage inequality over three decades: a reassessment. Am. Econ. J.
8 (1), 58–99.
Berkowitz, D., Ma, H., Nishioka, S., 2017. Recasting the iron rice bowl: the reform of China’s state-owned enterprises. Rev. Econ. Stat. 99 (4), 735–747.
Bessen, J.E., Goos, M., Salomons, A., Van den Berge, W., 2019. Automatic reaction-what happens to workers at firms that automate?. Boston Univ. School of
Law, Law and Economics Research Paper.
Bradsher, K., 2017. A robot revolution, this time in china. New York Times. https://www.nytimes.com/2017/05/12/business/
a- robot- revolution- this- time- in- china.html, accessed July 9th,2019]
Brandt, L., Van Biesebroeck, J., Zhang, Y., 2014. Challenges of working with the chinese NBS firm-level data. China Econ. Rev. 30, 339–352.
Brown, C., Gilroy, C., Kohen, A., 1982. The effect of the minimum wage on employment and unemployment. J. Econ. Lit. 20 (2), 487–528.
Burdett, K., Mortensen, D.T., 1998. Wage differentials, employer size, and unemployment. Int. Econ. Rev. 257–273.
Cameron, A.C., Gelbach, J.B., Miller, D.L., 2011. Robust inference with multiway clustering. J. Bus. Econ. Stat. 29 (2), 238–249.
Carbonero, F., Ernst, E., Weber, E., 2018. Robots worldwide: the impact of automation on employment and trade. Working Paper.
Card, D., Krueger, A.B., 1994. Minimum wages and employment: a case study of the fast-food industry in new jersey and pennsylvania. Am. Econ. Rev. 84
(4), 772–793.
Che, Y., Zhang, L., 2018. Human capital, technology adoption and firm performance: impacts of China’s higher education expansion in the late 1990s. Econ.
J. 128 (614), 2282–2320.
Cheng, H., Jia, R., Li, D., Li, H., 2019. The rise of robots in China. J. Econ. Perspect. 33 (2), 71–88.
Chiacchio, F., Petropoulos, G., Pichler, D., 2018. The impact of industrial robots on EU Employment and wages: a local labour market approach. Working
Paper.
Ciccone, A., Papaioannou, E., 2009. Human capital, the structure of production, and growth. Rev. Econ. Stat. 91 (1), 66–82.
Dauth, W., Findeisen, S., Suedekum, J., Woessner, N., 2018. Adjusting to robots: worker-level evidence. Working Paper.
Dinlersoz, E., Wolf, Z., 2018. Automation, labor share, and productivity: plant-level evidence from us manufacturing.
de Backer, K., DeStefano, T., Menon, C., Suh, J.R., 2018. Industrial robotics and the global organisation of production. Working Paper.
Dixon, J., Hong, B., Wu, L., 2020. Employment consequences of robots: firm-level evidence. NYU Stern School of Business.
Dornbusch, R., Fischer, S., Samuelson, P.A., 1977. Comparative advantage, trade, and payments in a Ricardian model with a continuum of goods. Am. Econ.
Rev. 67 (5), 823–839.
Du, Y., Park, A., Giles, J., 2016. Labor regulation and manufacturing employment in China. Working Paper.
Dube, A., Lester, T.W., Reich, M., 2010. Minimum wage effects across state borders: estimates using contiguous counties. Rev. Econ. Stat. 92 (4), 945–964.
Faber, M., 2020. Robots and reshoring: evidence from mexican labor markets. J. Int. Econ..
Fan, H., Lai, E.L.-C., Li, Y.A., 2015. Credit constraints, quality, and export prices: theory and evidence from China. J. Comp. Econ. 43 (2), 390–416.
Fan, H., Li, Y.A., Yeaple, S.R., 2018. On the relationship between quality and productivity: evidence from China’s accession to the WTO. J. Int. Econ. 110 (1),
28–49.
Fan, H., Lin, F., Tang, L., 2018. Minimum wage and outward FDI from china. J. Dev. Econ. 135, 1–19.
Fan, J., 2019. Internal geography, labor mobility, and the distributional impacts of trade. Am.Econ. J. 11 (3), 252–288.
Fang, T., Lin, C., 2015. Minimum wages and employment in china. IZA J. Lab. Policy 4 (1).
Fardoust, S., Lin, J.Y., Luo, X., 2012. Demystifying China’s fiscal stimulus. World Bank Policy Research Working Paper.
Ford, M., 2015. Rise of the Robots: Technology and the Threat of a Jobless Future. Basic Books.
Gallagher, M., Giles, J., Park, A., Wang, M., 2015. Chinas 2008 labor contract law: implementation and implications for Chinas workers. Hum. Relat. 68 (2),
197–235.
Gan, L., Hernandez, M.A., Ma, S., 2016. The higher costs of doing business in China: minimum wages and firms’ export behavior. J. Int. Econ. 100, 81–94.
Giuntella, O., Lu, Y., Wang, T., 2019. Is an army of robots marching on Chinese jobs? Working Paper.
Graetz, G., Michaels, G., 2018. Robots at work. Rev. Econ. Stat. 100 (5), 753–768.
Hau, H., Huang, Y., Wang, G., 2016. Firm response to competitive shocks: evidence from China’s minimum wage policy. Working Paper.
Helpman, E., Melitz, M.J., Yeaple, S.R., 2004. Export versus FDI with heterogeneous firms. Am. Econ. Rev. 94 (1), 300–316.

23
JID: JEBO
ARTICLE IN PRESS [m3Gsc;November 30, 2020;16:23]

H. Fan, Y. Hu and L. Tang Journal of Economic Behavior and Organization xxx (xxxx) xxx

Humlum, A., 2019. Robot adoption and labor market dynamics.


International Federation of Robotics, 2014. World robotics: Industrial robots (2014).
Huang, Y., Loungani, P., Wang, G., 2014. Minimum wages and firm employment: evidence from China. Working Paper.
Jacob, R., Mishkin, S., 2012. China turns to robots as labour costs rise. Financial Times. https://www.ft.com/content/
2a804e04- 0c95- 11e2- a776- 00144feabdc0, accessed July 9th,2019]
Koch, M., Manuylov, I., Smolka, M., 2019. Robots and firms. Working Paper.
Kopytov, A., Roussanov, N., Taschereau-Dumouchel, M., 2018. Short-run pain, long-run gain? Recessions and technological transformation. J. Monet. Econ.
97, 29–44.
Krenz, A., Prettner, K., Strulik, H., 2018. Robots, Reshoring, and the Lot of Low-skilled Workers. Working Paper.
Li, H., Li, L., Wu, B., Xiong, Y., 2012. The end of cheap chinese labor. J. Econ. Perspect. 26 (4), 57–74.
Lin, O., 2018. Chinese robotic firms thrive, thanks to subsidies. Asia Times. May 8th
Long, C., Yang, J., 2016. How do firms respond to minimum wage regulation in china? Evidence from chinese private firms. China Econ. Rev. 38, 267–284.
Lordan, G., Neumark, D., 2018. People versus machines: the impact of minimum wages on automatable jobs. Lab. Econ. 52, 40–53.
Mayneris, F., Poncet, S., Zhang, T., 2018. Improving or disappearing: firm-level adjustments to minimum wages in china. J. Dev. Econ. 135, 20–42.
Neumark, D., Wascher, W., 20 0 0. Minimum wages and employment: a case study of the fast-food industry in New Jersey and pennsylvania: comment. Am.
Econ. Rev. 90 (5), 1362–1396.
Olley, G.S., Pakes, A., 1996. The dynamics of productivity in the telecommunications equipment industry. Econometrica 64 (6), 1263–1297.
Tombe, T., Zhu, X., 2019. Trade, migration, and productivity: A quantitative analysis of china. American Economic Review 109 (5). 1843–72
Van Den Berg, G.J., 2003. Multiple equilibria and minimum wages in labor markets with informational frictions and heterogeneous production technologies.
Int. Econ. Rev. 44 (4), 1337–1357.
Wang, J., Gunderson, M., 2011. Minimum wage impacts in China: estimates from a prespecified research design, 20 0 0–20 07. Contemp. Econ. Policy 29 (3),
392–406.
Yang, Z., 2017. Who will satisfy chinas thirst for industrial robots? Diplomat. May 19th
Yu, M., 2015. Processing trade, tariff reductions and firm productivity: evidence from chinese firms. Econ. J. 125 (585), 943–988.

24

You might also like