High-tech Clusters, Labor Demand, and Inequality

Post on 30-Mar-2023

0 views 0 download

transcript

High-tech Clusters, Labor Demand, and Inequality:

Evidence from Online Job Vacancies in China

Qin Chen Klint Mane Geunyong Park Ande Shen*

February 24, 2022

AbstractThis paper evaluates the impact of high-tech clusters on inequality by focusing on a re-

cent Chinese placed-based industrial policy called “Made in China 2025.” Exploiting the

staggered roll-out of the policy and the representative online job posting data in China,

we conduct an event-study analysis to investigate the heterogeneous effects of high-tech

clusters on labor demand, wages, and living costs across occupations and regions. We

find that targeted cities experience a significant increase in labor demand measured by

online job vacancies relative to non-targeted cities, partially due to an increase in new

firm entry. The gap in the offered wages between non-routine and routine jobs grows at

the same time. This policy also causes labor demand and wages to drop in neighboring

areas of the targeted cities. Combining the labor market effects with increasing living

costs in the pilot cities, the welfare of non-routine job workers in the targeted cities dis-

proportionately declines by construction of high-tech clusters. Our results suggest that

policymakers should be cautious about occupational and regional inequalities when con-

structing high-tech clusters.

Keywords: Made in China 2025, Inequality, Labor Demand, Industrial Policy, Vacancy

Postings

*Qin Chen: Business Big Data Company; Klint Mane: University of Rochester (email:kmane@ur.rochester.edu); Geunyong Park: University of Rochester (email: gpark17@ur.rochester.edu);Ande Shen: University of Rochester (email: ande.shen@rochester.edu). We are grateful to Travis Baseler,Lisa Kahn, Ronni Pavan, John Singleton, and Kegon Tan for their guidance and support. We thank GeorgeAlessandria, Victor Hernandez, Cameron LaPoint, Arvind Sharma, and seminar participants at RochesterApplied Reading Group, Rochester Student Seminar, 15th North America Meeting of Urban EconomicsAssociation, 92nd International Atlantic Economic Conference, and 2021 Jinan-SMU-ABFER Conference onUrban and Regional Economics for their valuable comments. All errors are ours.

1

1 Introduction

Many countries recently have set up high-tech clusters to foster the adoption of modern

technologies into their production.1 The implications of high-tech clusters, however, are am-

biguous and not well understood, especially on inequalities. When it comes to inequalities

across workers, advanced technologies such as robots and AI newly adopted by high-tech

clusters might replace some types of workers, while raising labor demand for high-skilled

workers (Acemoglu et al., 2020; Acemoglu and Restrepo, 2020). On the contrary, high-tech

clusters may improve the overall productivity of the workers in the treated areas as tradi-

tional place-based industrial policies (Busso, Gregory, and Kline, 2013; Kline and Moretti,

2014a; Ehrlich and Seidel, 2018; Criscuolo et al., 2019). When it comes to spatial inequal-

ities, high-tech clusters might have negative spillover effects on nearby areas by draining

their resources (Hanson and Rohlin, 2013). On the contrary, the technology spillover from

high-tech clusters may lead to increased productivity in neighboring areas.

The implications are highly important for the governments of developing countries to boost

inclusive economic growth by technology upgrading. However, it is far from conclusive on

the impacts across different economic players due to three challenges. First, construction of

high-tech clusters and technological adoption usually happen in a gradual and cumulative

way, which leads to the difficulty of estimating their causal effects. Second, the locations

of high-tech clusters are endogenous in most of cases. High-tech firms intentionally choose

their locations to maximize their profit and a high-tech cluster is naturally built as more firms

decide to locate in the same advantageous place. For example, high-tech companies and IT

professionals choose to locate in San Francisco Bay Area to take the exclusive advantages

of the location. Lastly, the empirical evidence on the impacts of technological adoption

and constructing high-tech clusters usually comes from developed countries due to data

restrictions. Unlike developed countries, it is enormously hard to get detailed information

across individuals and regions for developing countries.

In this paper, we tackle these challenges by evaluating a modern industrial place-based pol-

icy called “Made in China 2025” (henceforth MIC25). The goal of MIC25 is to make China

a global leader in the manufacturing of high-technology products. From 2016 to 2017, 30

1For example, South Africa’s Inter-ministerial Committee on Industry 4.0, The Republic of Korea’s I-Koreastrategy in 2017, Indonesia’s Making Indonesia 4.0, and Uganda’s National 4IR Strategy in 2020.

2

pilot cities of different levels were chosen to be the front runners for this technological ad-

vancement. MIC25 targeted ten high-tech fields, including next-generation IT, robots, and

electric vehicles. With this policy, the government provided firms in the pilot cities’ targeted

industries with financial instruments such as tax incentives and direct subsidy. We empir-

ically explore the policy’s impact on labor demand, wages, and living costs across regions

by utilizing representative online job postings data and housing rent records. Combining

the reduced-form estimates, we examine the overall welfare implications of such industrial

policies which induced high-tech clusters.

In our setting, there are three key advantages that help elevate the aforementioned issues.

First, in MIC25, the decisions to upgrade technology and build high-tech clusters in specific

regions are made by the central government, leading to a plausibly exogenous shock to firms

and individuals around the targeted areas. Second, 30 cities were selected as pilots out of

more than 300 cities in China, which leaves a way to identify the causal impact by comparing

each pilot city with a non-pilot but similar city and to evaluate the inequality issues across

cities induced by the policy. The implementation date of the policy also varies across the

pilot cities. The regional and time variations created by this policy are useful in solving the

identification challenges. Finally, the representative online job-posting data collected since

2015 provide detailed information on labor demand and wages at a high frequency, which

is not the case for most of Chinese official data. This enables us to cleanly identify the labor

market dynamics induced by MIC25 across occupations and regions both before and after

the policy.

By utilizing the place-based implementation and staggered roll-out of MIC25, we conduct

an event-study analysis to identify the causal impacts of high-tech clusters on labor de-

mand, employment, firm entry, and housing costs. To deal with the potential selection bias

of governments’ choice on pilot cities, we compare the 30 pilot cities to the 30 control cities

matched based on observable characteristics before the implementation of MIC25. Our re-

sults are robust to using different control groups. We further compare the neighborhood

cities of pilot cities to the neighborhood cities of non-pilot cities to examine the spillover ef-

fects and inequalities across cities. By comparing wage and rent changes across occupations

and places, we identify who are the winners and losers of the high-tech clusters. We also

explore the mechanisms behind the welfare implications by analyzing the location choices

of both firms and workers.

3

We find that the number of job openings in pilot cities has significantly increased by 24%

compared to control cities for 3 years after the implementation of MIC25. During the same

period, the policy does not have a statistically significant effect on offered wages in the job

postings, but we could observe an slightly increasing trend of offered wages in the pilot cities

toward the latter periods. These results are supported by city-level yearly employment and

wage data. On the other hand, the neighboring cities of pilot cities suffer 25% fewer job

openings for 3 years after MIC25 compared neighboring cities of control cities. Neighbor-

ing cities also suffer wage decreases of 3%. The negative effects on the neighboring areas

are larger than the positive impact of MIC25 on the pilot cities. In pilot cities, wages are

increasing for non-routine occupations and targeted industries. At the same time, housing

rent also increases by 7% in the pilot cities during the same period, reducing the welfare of

residents in the pilot cities. The main mechanisms behind the results are inflow of workers

and firms to the pilot cities. Using China’s firm registration data, we provide evidence that

10% more firms enter the pilot cities and that 32% less firms enter the neighboring cities for

3 years after MIC25. We also observe that the population of the pilot cities increases by 2%

even with migration restrictions in place.

To quantify the welfare impact of MIC25 and disentangle the channels behind it, we con-

struct a theoretical model inspired by Kline and Moretti (2014b). Workers’ utilities depend

on wages, local amenities, and rents and vary across locations and occupations. Given that

wage increases for any type of workers do not catch up with the hike in rent prices in the

pilot cities, the welfare of workers is worse off there. In the neighborhood cities, wages

significantly decline for any type of workers due to the reduction in labor demand though

housing costs do not change. So, the welfare of workers in the neighborhood cities is also

worse off. The back-of-the-envelope calculation weighing both wage reductions and rent

hikes suggests that average workers would lose around $260 per year due to the implemen-

tation of MIC25 for 3 years after the policy, which accounts for around 3% loss in yearly

wages, and the total welfare loss of the workers accounts for around 0.15% ($ 19 billion) of

China’s annual GDP in 2017. Moreover, the heterogeneity in the response of labor demands

and wages indicates that the welfare loss is more significant for routine job workers than

non-routine-job workers in the pilot cities. Readers, however, should be cautious to inter-

pret our results because we only focus on the first 3 years of the policy. Long-run effects

could be different from our estimates.

4

This paper is related to several strands of the literature. First, our study is related to the

growing literature on the local effects of high-tech facilities. The majority of these stud-

ies focus on identifying productivity spillovers on knowledge production or R&D (Bloom,

Schankerman, and Van Reenen, 2013; Carlino and Kerr, 2015; Moretti, 2021), but some of

the studies present the effects on local economies. For instance, Kantor and Whalley (2014)

report the long-term positive effects of research university activity on their local economies.

They find that the spillover effects grow over time as the composition of local industries ad-

justs to the comparative advantage of the location. Qian and Tan (2021) identify the mixed

welfare effects of the entry of a large high-skilled firm on local residents. Our idea is closely

related to this study, but we focus on national-level high-tech clusters rather than a single

high-tech firm.

This paper is also part of a much larger literature on place-based industrial policies. Busso,

Gregory, and Kline (2013) and Criscuolo et al. (2019) identify strong positive employment

and wage effects due to subsidies in disadvantaged areas in the US and UK, respectively.

Both studies find that there is not much negative spillover effect of the traditional place-

based industrial policies. In China, Alder, Shao, and Zilibotti (2016) find a positive spillover

effect of Special Economic Zones(SEZs) along with a 20% increase in GDP in SEZs. On the

contrary, Kline and Moretti (2014a) focus on the Tennessee Valley Authority (TVA) imple-

mented in the 1930s of the US, and find that a positive effect of subsidies for disadvantaged

areas is offset by the negative effect on nearby areas. We contribute the literature by in-

vestigating a modern version of place-based industrial policies, building high-tech clusters,

which focus on adopting and distributing advanced technology as a national strategy rather

than developing disadvantaged areas.

Lastly, this study follows the ideas in the literature on the adoption of recent technologies

and inequality. Acemoglu and Autor (2011) provide a theoretical framework where recent

technological progress could have the opposite effects on workers based on replaceability

and complementarity of their tasks. Graetz and Michaels (2018) is the first paper to report

that industrial robots are reducing the labor share of low-skilled workers at national level.

Acemoglu and Restrepo (2020) also find that robots’ impact is distinct from traditional capi-

tal and technologies, having negative effects on wages and employment in local economies.

Using establishment level data on online job vacancies in the US, Acemoglu et al. (2020)

report a rapid growth in AI-related job postings from 2010 and find that AI-exposed estab-

5

lishments are reducing hiring in non-AI positions while they expand AI-related hiring. Our

study complements this literature by investigating such a composition change in labor force

induced by national-level high-tech clusters.

The remainder of this paper unfolds as follows. We first review the background of China’s

place-based industrial policy MIC25 in Section 2. In Section 3, we introduce the data used

in our analysis. Section 4 demonstrates our matching and estimation strategies. Section 5

presents the results and analyze the mechanisms. In Section 6 we build a conceptual frame-

work and investigate the welfare impact of building a high-tech cluster by combining the

estimation results. Section 7 concludes.

2 Background

This section provides a detailed description on the policy “Made in China 2025” and rele-

vant institutional background. After about 40 years of development since the opening and

reform in 1978, China has gained tremendous development in the manufacturing sector.

China’s industrial production, however, is still backward compared to industrial countries.

Unlike technology-leading countries, such as the US, Germany and Japan, where there is

intensive use of production lines and management processes based on modern information

technology and highly automated machines, China’s manufacturing production is still very

labor-intensive and in a rudimentary level of automation compared to advanced countries

(Hanson, 2020). The data from the International Federation of Robotics (IFR) demonstrate

that on average Chinese enterprises utilize just 19 industrial robots per 10,000 industry em-

ployees in 2015, which is much lower than 531 in South Korea, 301 in Germany, or 176 in the

US.

In May 2015, the Chinese government issued the national strategic plan “Made in China

2025” — as part of the Thirteen and Fourteen Five-year Plans — aimed to further develop

the manufacturing sector in China and transform China from a labor-intensive workshop

to a more technology-intensive powerhouse by the first half of the century. China used

to be a producer of cheap low-technology goods by utilizing the advantage of low labor

costs. This plan is intended to develop the production of high-technology goods, with a

goal of increasing the Chinese-domestic content of core materials to 40% by 2020 and 70%

by 2025.

6

To concentrate resources and fulfill the goal of MIC25 more efficiently, the Chinese gov-

ernment selected ten key industries to focus on: Information Technology, Robotics, Green

energy and green vehicles, Aerospace equipment, Ocean engineering and high tech ships,

Railway equipment, Power equipment, New materials, Medicine and medical devices, and

Agriculture machinery. All of these industries are regarded as crucial technology sectors,

which can lead China to a strong position in the Fourth Industrial Revolution.

The Chinese government implemented MIC25 in a “pilot-and-expansion” way which took

a step-by-step procedure of implementation following the experience of the other reforms

such as the Special Economic Zone (Lu, Wang, and Zhu, 2019).2 In August 2016, Ningbo, a

city in Eastern coast of China, was selected as the first pilot city of MIC25. In the following

couple of months, the other 29 cities were selected, which makes a total of 30 pilot cities

among 335 cities. In 2018, the pilot cities were upgraded to National Demonstration Zones

(NDZs) to lead the MIC25 initiative under centralized guidance. These cities are widely

located across regions in China, as shown in Figure 1.

Figure 1: The Spatial Distribution of Pilot Cities

Notes: The map shows the boundaries of all cities (both province-level and prefecture-level) in China. The highlighted regions are 30 pilot cities of MIC25.

2Xin Guobin, the vice-minister of the Ministry of Industry and Information Technology (MIIT) states thatthese pilot cities were chosen to prevent the trend of local governments rushing in implementation and there-fore causing inefficiencies. See https://www.chinadaily.com.cn/business/2016-08/19/content 26536807.htm.

7

After the Ministry of Industry and Information Technology (MIIT) of the Chinese govern-

ment approved a city’s application to be a pilot city, then the city announced its own de-

velopment plan for MIC25 later. Table 1 shows the quarter when each pilot city published

its implementation plan for MIC25. Before the announcements, firms and workers have

no way of knowing which cities are going to be selected or how the policy will be imple-

mented, which is consistent with our results showing no anticipation effects. All these pilot

cities focus on their own advantageous industries, rather than being involved in all targeted

industries of MIC25.

Table 1: The Policy Starting Dates of Pilot Cities in MIC25

Quarter Number of cities City name

2016Q2 1 Ningbo

2017Q1 8 Shenyang, Changchun, Nanjing, Wuxi, Changzhou, Suzhou, Zhenjiang,Quanzhou

2017Q2 15 Huzhou, Zhengzhou, Luoyang, Xinxiang, Changsha, Zhuzhou, Xiangtan,Hengyang, Zhuhai, Foshan, Jiangmen, Zhaoqing, Yangjiang, Zhongshan,Chengdu

2017Q3 3 Hefei, Wuhan, Wuzhong

2017Q4 3 Ganzhou, Qingdao, Guangzhou

Notes: The starting date is the time when the city published its own development plan for MIC25 after being se-lected as a pilot city. Usually the starting date is a few month after the approved date. All the date informationis collected from government announcements, reports, official newspapers, and policy documents database.

Though MIC25 was inspired by the other modern industrial plans in other countires such as

“Industry 4.0” of Germany, its scale is incomparable with the other policies in many aspects

(Li, 2018; Malkin, 2018; Zenglein and Holzmann, 2019). First, the president Xi Jinping made

MIC25 his signature project, reflecting how crucial it is in the central government side. Sec-

ond, both central and local governments provide a wide range of policy benefits including

protecting intellectual property, allocating more lands, and simplifying government review

and approval procedures. Third, the governments offer huge amount of financial support.

Firms with MIC25-related projects can get higher subsidies, lower loan interest and higher

tax benefits. For example, MIIT and China Development Bank promised $45 billion in di-

rect loans, bond sales and leasing to support major MIC25 projects. The Special Construc-

tive Fund provided an estimated $270 billion funding for numerous MIC25-related projects.

Shanxi MIC25 Fund offers $117 billion of financial support for approximately 100 MIC25-

related projects. This huge scale makes MIC25 the best example of “big push” development

strategy in developing countries and may generate huge increases in welfare (Rosenstein-

Rodan, 1943; Murphy, Shleifer, and Vishny, 1989; Kline and Moretti, 2014a).

8

3 Data

Our empirical analysis draws on four main data sources: (1) online job posting data; (2)

publicly-available city-level statistical yearbooks; (3) housing rent database; and (5) firm

registration database. This section describes each of these data sources and their represen-

tativeness.

We begin with the online job posting data from a research company called Business Big

Data. The company collects job postings from six major job hunting platforms in China.3

To avoid duplicate job postings, the company only keeps the first appearance of a specific

job from a specific company. The data span from 2015 to 2020, covering periods before

and after the adoption of MIC25. The job postings from the six online platforms can cover

more than 80% of all online job postings in China.4 For each job posting, the company

scraps detailed information about the job position including job title, job description, offered

wages, required education and specific skills, etc. It also collects the information about the

employer such as name, industry, location, size, founding year, etc.

Online job posting data is necessary in our research for two reasons. First, it is a unique data

source for investigating labor demand and wages across detailed occupations and locations

in China. For instance, the China City Statistical Yearbooks only report city-level wages and

China Census does not report wages at all. So, Chinese job postings data have been used

in literature to infer detailed aspects in the labor market of China (Kuhn and Shen, 2013;

Fang et al., 2020; Helleseter, Kuhn, and Shen, 2020; He, Mau, and Xu, 2021). Second, it also

provides high-frequency data of job vacancies. This feature is helpful for the identification

of policy effects because we utilize the staggered adoption across the pilot cities in differ-

ent quarters. Considering the short time window (2015 - 2019) of our research, it is also

important to expand the number of periods by using quarterly data.

Our online job posting data is most close to the US job posting data set assembled by Burning

Glass Technologies (BG). It has been widely applied in literature (Hershbein and Kahn, 2018;

Acemoglu and Restrepo, 2020; Forsythe et al., 2020). The Chinese job posting data set, how-

ever, provides two additional advantages. First, we can observe the job posting itself and

351job, zhaopin.com, liepin.com, lagou.com, 58.com, and ganji.com451job and liepin.com have more than half of the market share in 2019 and 51job, 58.com

and zhaopin.com account for over 80% share of the online-hiring market in 2018. Seehttp://report.iresearch.cn/report/202004/3572.shtml.

9

Figure 2: Representativeness of Job Postings Data

(a) Industry Composition (b) Employment-Job Postings Share Deviations

Notes: In Panel (a), industry composition of employment are from the City StatisticalYearbooks and job ads are from the main data of online job postings for the periods from2015 to 2018. In Panel (b), we plot the deviation of the online job ads share from theyearbook employment share in 2015 on x-axis and that in 2018 on y-axis. Each pointrepresents a city-industry pair. If a point (city-industry) is on the 45-degree line, then therelative representativeness of that city-industry pair in the online job postings does notchange from 2015 to 2018.

job openings, which refers to how many workers the firm intends to hire in each job post.

The US data does not report how many workers an employer would like to hire for a job

vacancy. So, we can better capture the changes of labor demand in response to the policy by

using the number of job openings rather than the number of job postings. More interestingly

and importantly, roughly 80% of the job postings have salary information in contrast with

the US data where salary information is usually unavailable. This unique wage information

constitutes a major component in our empirical study on inequality across occupations, and

regions.

The main concern of using online job vacancy data is that online job ads may be biased

towards certain types of occupations among all job vacancies. Following Hershbein and

Kahn (2018),5 we investigate the representativeness of online job postings by comparing our

data with employment from the China City Statistical Yearbooks for the periods from 2015

to 2018. Figure 2a compares industry composition of employment from the City Statistical

Yearbooks (blue bars) and that of job ads from the online job posting data (red bars). The

5Hershbein and Kahn (2018) compare the online job vacancies in the US with public data sets including JobOpenings and Labor Turnover Survey (JOLTS), Current Population Survey (CPS), and Occupational Employ-ment Statistics (OES) and find that the changes of the job distributions across time fit well with other publicnationally representative data, though the online postings overrepresent high-skill occupations and industriesoverall.

10

online job ads over-represent several industry groups such as business, retailing, and IT,

while under-represent manufacturing, construction, and education. These results are not

surprising considering that firms in business, retailing, and IT sectors are more likely to hire

their employees via internet than other sectors.

What matters more for our analysis is whether the representativeness changes over time. If it

is the case, our results could be deteriorated by the composition changes of the collected job

postings across industries or locations. Figure 2b shows each city-industry pair’s deviation

of the online job ads share from the employment share in 2015 (on the x-axis) and that in

2018 (on the y-axis), as well as the 45-degree line. If a point (an industry in a city) is on the

45-degree line, then the representativeness of that city-industry pair does not change from

2015 to 2018. The figure indicates that the representativeness of the online job ads for each

city-industry is very stable over the periods of interest because most of the dots lie around

the 45-degree line. Thus, we believe that our results are not driven by a growing importance

of certain cities or industries in the job vacancy data.

Due to the fluctuation and seasonality issue, we aggregate the firm-specific data to a city-

quarter-industry level, starting from the first quarter of 2015 and covering all cities (pre-

fectures) in China. We use the 2-digit industry code (18 in total) due to data availability.

This paper examines two main variables related to job postings: the number of job openings

in each city-quarter-industry cell and the weighted average wage where the weight is the

number job openings in each posting.

Second, we assemble publicly-available data including the China City Statistical Yearbooks

(2010-2018) and the China City Development Yearbooks (2010-2018) to construct annual eco-

nomic outcomes at the city level. The integrated data include city-level characteristics such

as GDP, employment, wage, land area, population, government expenditure, and others.

These data provides the measure on the similarity of cities, which serves to construct our

matching sample. We also investigate the effect of MIC25 on city-level employment from

this data set to make sure that the changes of job vacancies generate corresponding changes

of employment in local labor market.

To account for the impact of MIC25 on housing market, we employ monthly housing rent

from the Wind Financial Terminal.6 The data set includes average housing rent of 100 cities

6See https://www.wind.com.cn/en/wft.html.

11

out of 335 cities in China, but 26 out of 30 pilot cities are included. The change in rent also re-

veals the migration pattern of the labor force. Pilot cities aim to develop their manufacturing

sector and attract workers from non-pilot cities, which will lead to the inflow of migrants

and increase in housing rent. To get a more detailed picture of migration induced by the

policy, we also applied the China Migrants Dynamic Survey data, which is an yearly cross-

sectional data covering around 160,000 households in all provinces in China from 2010 to

2018.

In addition to workers’ sorting across locations, the behavior of firms is another channel

through which the policy could affect welfare. We gathered the registration information of

all firms from China’s Administrative Registration Database (CARD). These data include

all firms in China starting in 1980, including firm name, registration place, operation place,

registration asset, and registered industry (4-digit). Using these data, we can analyze the

effect of building a high-tech cluster on the inflow of new firms, which could be a main

mechanism behind the changes in labor demand.

4 Empirical Strategy

Our empirical analysis’ main objective is to estimate the causal effects of MIC25 on labor

demand and housing market in the targeted and neighboring cities. As described above,

30 cities were selected as pilot cities by the Chinese central government. The main obstacle

for the causal inference is that the pilot cities are not randomly selected, which generates

biases in a difference-in-difference strategy. Ideally we would have chosen cities that were

candidates to become pilot cities, but were not chosen. Unfortunately, such information is

not available on MIC25. Given that we do not have the details on how pilot cities are se-

lected, we rely on observed characteristics to create a comparable sample of control cities.

To create the control sample, we estimate propensity scores of being chosen as a pilot city

using logistic regression. The propensity score matching (PSM) has been widely used in ap-

plied research and Dehejia and Wahba (1999) find that the PSM approach can be powerful to

eliminate the biases coming from the observable differences between the treated and control

groups.

The propensity score, P(D = 1|X), gives the predicted probability of being a pilot city

(D = 1) under the assumption of selection on observables ((Y0, Y1) ⊥ D|X). The regres-

12

Figure 3: The Maps of the Cities of Interest

(a) Pilot and Neighboring Cities (b) Control and Neighboring Cities

Notes: The control cities in Panel (b) are chosen by the matching algorithm based on thecity-level characteristics in the base year 2015. The neighbor cities are the closest 5 citiesfor each pilot and control cities except for the pilot and control cities.

sors of the logit model include population, employment, GDP per capita, land area, number

of industrial firms, government tax income, GDP growth, and employment shares of main

industries in the base year 2015 from the Chinese City Statistical Yearbook. After getting

the predicted propensity scores from the logistic model, we pick the 30 cities with the clos-

est propensity scores to the treated cities to be in the control group. Note that we are not

selecting one control city for each pilot city for pairwise comparison. We compare the 30

pilot cities versus the 30 control cities. Based on our predicted scores, selecting 30 cities

corresponds to selecting the 90th percentile of the most similar cities in terms of propensity

scores.

The neighboring cities of pilot cities are not allowed to be chosen as control cities since if

treatment has positive or negative spillover effects to the nearby regions, then we would be

under- or over-estimating the impact of MIC25. We define neighboring cities as the closest

5 cities in terms of distance between the city centers. Since there are some clusters of pilot

cities, there are way less than 150 neighboring cities. Four different groups of interest are

depicted in Figure 3. On the left, we can see the pilot cities in red and the neighboring cities

of pilot cities in blue. On the right, you can see the control cities in red and the neighboring

cities of the control cities in blue. The maps imply that the pilot and control cities are not

concentrated in some particular region of China.

13

It is important to note that the results shown in Section 5 do not depend on our choice of

control cities. Choosing a control sample of 60 cities instead of 30, or using all other cities

as controls yield very similar results. As a further robustness check, we estimate propensity

scores using random forest rather than the logistic regression. There are several reasons why

we consider a random forest approach. Firstly, a parametric logistic model is prone to model

misspecification, which results in the inaccurate or imprecise estimation of the propensity

score (Zhao et al., 2016). Secondly, using a random forest algorithm allows for non-linearities

and interactions which are not captured by linear models. Finally, the literature suggests that

using logistic regression leads to subpar performance compared to random forests or other

machine learning methods (Zhao et al. (2016); Lee, Lessler, and Stuart (2010); Setoguchi et al.

(2008)).

Table A.1 shows the summary statistics of the treated and control samples from using 2015

data from the yearbook. Panel (a) presents the main outcomes in our empirical analysis

and panel (b) are the matched variables in our matching algorithms. Average values are

shown throughout the table. Column (1) shows the averages for treated cities. Columns

(2) and (4) show the averages for the control sample selected using logistic regression and

random forest, respectively. Columns (3) and (5) show the difference between the control

samples and the treated sample. We run a t-test to see if the differences are statistically

significant. The random forest algorithm does slightly better than the logistic regression

in balancing observables. There are still significant differences for a few variables, which

are to be expected given the small and limited number of cities. In general, both matching

samples are much similar to the treated sample than the sample where all non-pilot cities

are included, as shown in column (6).

To validate that all cities are evolving similarly if there were no MIC25 and gain confidence

in the causal effect of MIC25, we conduct an event study approach to compare the treated

pilot cities to the control cities constructed above. We exploit the variation in the timing of

the start date across the pilot cities. The data is aggregated to a city by quarter by industry

level. Our main specification is given by:

Yict = α +14

∑k 6=−1;k≥−4

βk · 1{t=t∗c+k,Dc=1} + γic + δit + εict. (1)

14

where Yict is the outcome (log of job openings, offered wage, rents, etc.) at city c and quarter

t and industry i. Dc is the indicator of being chosen as a pilot city, so 1{t=t∗c+k,Dc=1} = 1 if

city c is a pilot city and quarter t is k quarters before/after the start date (t∗c ) and 0 otherwise.

We control for industry-by-city (γic), and industry-by-time (δit). γic is included to control for

any time-invariant unobservable heterogeneity of city-specific industry and δit additionally

takes away the industry-specific time trends. We weight observations by the number of job

postings during 2015, the year before the policy was implemented. This weighting scheme

allows to up-weight cities with larger number of postings to help with the precision of our

results.

If our assumption of selection on observables are satisfied or the fixed effects well control for

the possible unobservable heterogeneity (1{t=t∗c+k,Dc=1} ⊥ εict|γic, δit), then our coefficients

of interest

βk = E[Yc,k −Yc,−1|Dc = 1]− E[Yc,k −Yc,−1|Dc = 0] (2)

should capture the causal effects of MIC25 on the outcome variables at k quarters after the

policy change. Based on our conceptual framework in Section 6, we also utilize βk on log

wage and log rent to investigate the welfare incidence of MIC25. The coefficients on firm

entry and population would give the evidence on the mechanism of MIC25.

As has been recognized by Goodman-Bacon (2021), the two-way fixed effects models and

event studies could lead to biased estimates if the policy had heterogeneous impacts across

cities (i.e., treatment heterogeneity) and the treatment effects evolved over time (i.e., dy-

namic treatment effects). To mitigate this concern, we use the modified approach following

Callaway and Sant’Anna (2020). We also compare to approaches following Borusyak and

Jaravel (2017) and Sun and Abraham (2020), the results are similar.

15

Tabl

e2:

Sum

mar

ySt

atis

tics

(1)

(2)

(3)

(4)

(5)

(6)

Var

iabl

esPi

lotC

itie

sM

atch

ing

(Log

it)

Mat

chin

g(R

ando

mFo

rest

)A

llN

on-P

ilotC

itie

sM

ean

Mea

nD

iffer

ence

Mea

nD

iffer

ence

Mea

n

Pane

lA:M

ain

outc

omes

(not

targ

eted

inm

atch

ing)

Num

ber

ofjo

bpo

stin

g(1

0,00

0)32

.67

22.5

3-1

0.14

22.6

6-1

0.01

4.80

Num

ber

ofjo

bop

enin

g(1

0,00

0)10

3.12

70.2

1-3

2.91

71.5

3-3

1.59

14.4

7#

Wag

e(m

onth

ly,y

uan)

5353

.69

5085

.52

-268

.17

4866

.12

-487

.645

79.4

6

Pane

lB:M

atch

edva

riab

les

Popu

lati

on(1

0,00

0)57

1.48

725.

9715

4.49

775.

0720

3.59

443.

97Em

ploy

men

t(10

,000

)12

9.67

130.

921.

2513

9.93

10.2

664

.00

GD

P(b

illio

nyu

an)

580.

3346

8.11

-112

.22

480.

71-9

9.62

247.

41La

ndar

ea(k

m2 )

1031

9.10

1923

1.77

8912

.67*

*13

662.

0033

42.9

016

460.

93N

umbe

rof

indu

stri

alfir

ms

2978

.10

3382

.33

404.

2320

51.9

3-9

26.1

7**

1425

.47

Tax

profi

t(bi

llion

yuan

)55

.09

61.7

46.

6448

.55

-6.5

523

.87

GD

Pgr

owth

rate

(%)

8.42

7.99

-0.4

38.

30-0

.12

7.58

Empl

oym

ents

hare

(sec

onda

ryse

ctor

,%)

55.2

047

.80

-7.4

0**

51.5

5-3

.65

45.9

2Em

ploy

men

tsha

re(t

hird

sect

or,%

)44

.47

51.7

77.

30**

47.4

52.

9851

.64

GD

Psh

are

(sec

onda

ryse

ctor

,%)

49.1

449

.92

0.78

48.0

2-1

.12

46.5

2G

DP

shar

e(t

hird

sect

or,%

)44

.84

42.0

3-2

.81

42.4

9-2

.35

41.0

3Em

ploy

men

tsha

re(m

anuf

actu

ring

,%)

38.4

29.5

1-8

.89*

*31

.05

-7.3

5**

25.1

0Em

ploy

men

tsha

re(I

T,%

)1.

591.

860.

271.

850.

261.

33Em

ploy

men

tsha

re(b

usin

ess,

%)

2.08

1.99

-0.0

92.

490.

411.

74Em

ploy

men

tsha

re(s

cien

ce,%

)1.

962.

500.

541.

91-0

.05

1.70

Num

ber

ofC

itie

s30

3030

290

Not

es:

This

tabl

epr

esen

tsth

eav

erag

esac

ross

the

pilo

tci

ties

and

the

two

cont

rol

sam

ples

crea

ted

usin

glo

git

and

rand

omfo

rest

mat

chin

g.D

iffer

ence

colu

mns

show

whe

ther

the

diff

eren

ces

betw

een

the

cont

rol

sam

ples

and

trea

ted

sam

ple

are

sign

ifica

nt.

Star

s**

show

sign

ifica

nce

atth

e5%

leve

l.A

llin

dust

ries

are

clas

sifie

din

toth

ree

broa

dse

ctor

s(p

rim

ary,

seco

ndar

y,an

dth

ird)

.Ea

chbr

oad

sect

orin

clud

esso

me

gran

ular

sect

ors,

for

exam

ple,

the

seco

ndar

yse

ctor

incl

udes

the

man

ufac

turi

ngin

dust

ry,t

heco

nstr

ucti

onin

dust

ry;t

heth

ird

sect

orin

clud

esth

ein

form

atio

ntr

ansf

er,c

ompu

ter

serv

ice

and

soft

war

ein

dust

ry(n

amed

asIT

inth

eta

ble)

.Mor

ede

tails

can

befo

und

here

.

16

It is also important for us to capture the announcement date of each pilot city along with

the start date considering that economic agents can respond right after the announcement

of policies if they have any incentive to do that (Gruber and Koszegi, 2001). For instance, a

firm in a neighboring city of a pilot city that decides to move to the pilot city or to construct

a branch in the pilot city may reduce labor demand at the neighboring city in advance to

minimize labor cost. Home owners in the pilot cities may expect inflow of migrants search-

ing for jobs to the pilot cities and raise the housing rent right after the announcement of

designation. The official announcement dates are not available for many pilot cities, so we

reasonably define two quarters before the start date as the announcement date for each pilot

city based on the available information. To be specific, we use the announcement dates to

analyze labor demand in the neighborhood cities and housing market in pilot and neighbor-

hood cities following the above logic and find that they do respond to the announcement

date before the start dates of MIC25.

5 Results

Our results unfold in four parts. First, we use the online job posting data to explore the effect

of MIC25 on labor demand and wages. We investigate labor demand and wage changes in

pilot cities. We provide support for our job posting results using yearly employment and

wage data from City Statistical Yearbooks. We discuss the inequalities across industries or

occupations, as well as the spillover effect to nearby cities. Secondly, we investigate the

impact of MIC25 on the housing market in pilot and neighboring cities. Thirdly, we dig into

the sorting behaviors of firms and workers, to instruct the analysis on the policy’s welfare

implications. Lastly, using the results above, we derive the welfare impact of MIC25.

5.1 Labor Demand and Wages

5.1.1 Pilot Cities

We begin by analyzing the results for pilot cities. Figure 4 plots the coefficients of interest

from Equation (1). The outcome on the left-hand side of the figure is the log of job openings

and on the right log wages. Wage information is attached to the job posting. 60 cities are in-

cluded in the regression, 30 pilot cities, and 30 control cities. The number of job openings for

17

Figure 4: The Impacts on Job Postings in the Pilot Cities

(a) Number of Job Openings (b) Monthly Wages

Notes: Figures present coefficients and 95% confidence intervals on the interaction be-tween the indicator of relative quarterly periods from the policy start date and the indi-cator of the pilot cities. Industry-by-city and industry-by-calendar quarter fixed effectsare controlled and each city-by-industry observation is weighted by the total number ofpostings in 2015.

pilot cities increase significantly after the policy compared to control cities. The lack of a pre-

trend is assuring that the impact on the job openings is coming from MIC25. The increase

is quite large and reaches a level of about 50% after 2 years since MIC25 was implemented.

Overall, we find an increase of 24% in job openings in pilot cities. The results are shown

in columns (1) and (2) of Table 3 where we run a difference in difference specification. This

results is comparable to literature estimates for the effect of place based policies on employ-

ment in China of about 35% (Lu, Wang, and Zhu, 2019), and the US and UK at about 10-15%

(Kline and Moretti (2014a); Criscuolo et al. (2019)). The event study shows that it takes about

a year for the positive effect to arise. Then it increases significantly to 25% for one year after

the policy and 50% after two years. We believe that this pattern is reasonable, considering

that it should take time to apply and receive funding, built the proper production facilities

and new firms to emerge.

Offered monthly wages, on the other hand, appear to be constant over time with an in-

creasing trend towards the latest periods. This pattern is partly consistent with the lagged

response in the number of job openings. Overall, we observe an increase of about 2% in

wages shown in columns (3) and (4) of Table 3. However, the effect is not statistically signif-

icant. We believe that the event study is more informative as it shows the increasing trend

towards the end of the sample period. The relatively muted effects of MIC25 on wages at-

18

Table 3: Estimation Results of Labor Demands in the Pilot Cities

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

Log(Openings) Log(Openings) Log(Wage) Log(Wage)

Post×Pilot Cities 0.244** 0.243*** 0.025 0.020

(0.122) (0.093) (0.026) (0.022)

Year FE Y Y

Industry-Year FE Y Y

Industry-City FE Y Y Y Y

Observations 20,160 20,160 20,160 20,160

Notes: The estimates are calculated based on the aggregated treatment effects version of figures 4. The controlcities are based on the propensity matching method. Each city-by-industry observation is weighted by thetotal number of postings in 2015.

tached to the online job postings could indicate two different patterns in the labor market of

the pilot cities. First, firms may not need to increase the wages if there are enough inflows of

job seekers from the other cities even though they need to hire more workers in response to

the policy change. Second, there could be composition changes in online job postings. The

increase in the number of postings in the early periods may be accounted for by increasing

labor demand for low-skilled jobs, which also results in static wages. Then, the increase in

labor demand more concentrates on high-skilled jobs, in turn generating the growing wage

effects in the later periods.

These results indicate that MIC25 generates regional inequality between pilot cities and

other cities. Inequality arises in two dimensions. First, the pilot cities are already cities

with good industrial and technological foundations. Thus, MIC25 further increases the in-

equality by creating these high-tech clusters. This is opposite to the place-based industrial

policies in the developed countries where the government selects disadvantaged areas to re-

duce spatial inequalities. Second, the migration of workers in China is restrained by Hukou

system. So, Chinese workers face much larger moving costs if they move out of their home

town to get better job opportunities. Given this restrictions, MIC25 generates inequality

between residents of the selected pilot cities versus other cities.

The main drawback of using job vacancy data is that one cannot see how many of the post-

ings turn into realized employment. To tackle this issue, we use formal employment data

from City Statistical Yearbooks in China. The yearbook for 2019-2020 is not yet available, so

the furthest we can go is 2018-2019. On the positive side, we have data available even before

2015 which is when our sample starts for the job postings data. Hence we use data going

19

Figure 5: The Impacts on Employment

(a) Employment (b) Monthly Wages

Notes: Figures present coefficients and 95% confidence intervals on the interaction be-tween the indicator of years and the indicator of the pilot cities. City and year fixedeffects are controlled and each city is weighted by population in 2015.

back until 2013. The data is yearly instead of quarterly. Since we have pilot cities being an-

nounced in 2016 and 2017, we decide to choose 2016 as the policy start year. The event study

for formal employment is shown in Figure 5. We observe a large and significant increase in

formal employment for pilot cities close to 20%. The lack of a pre-trend is assuring that what

we are capturing is the effect of MIC25. Similar to the job postings, we find an increase in

wages of about 3%. Overall, the formal employment results support the representativeness

of the job postings data. Having established robust effects on the pilot cities, we move on to

explore the impact of MIC25 in neighboring areas.

5.1.2 Spillover Effects

The main criticism of place-based policies is that they might just draw resources from neigh-

boring places, resulting in no overall effects in the end. Therefore, in the next exercise, we

compare job openings and wages between neighboring cities of the treated group and neigh-

boring cities of the control group. Visually, we are comparing the blue cities on the left of

Figure versus the blue cities on the right of Figure 3. The results are shown in Figure 6.

Neighboring cities of the treated group suffer a sharp decrease in job openings after the pi-

lot city announcement. The decrease is large and significant, reaching about 50% after two

quarters. After the first year, the decreasing effects start fading away and become close to

zero towards the end of our sample period. Similar effects can be observed on wages shown

20

Figure 6: The Impacts on Job Postings in the Neighboring Cities

(a) Number of Job Openings (b) Monthly Wages

Notes: Figures present coefficients and 95% confidence intervals on the interaction be-tween the indicator of relative quarterly periods from the policy announcement dateand the indicator of the neighborhood cities around the pilot cities. Industry-by-cityand industry-by-calendar quarter fixed effects are controlled and each city-by-industryobservation is weighted by the total number of postings in 2015.

in panel (b) of Figure 6. The neighboring cities of the treated group suffer a 5% decrease in

wages about a year from the announcement. The decrease in wages is persistent, though it

gets smaller towards the end of the sample period.

Therefore, MIC25 has negative spillover effects on labor demand at least in the short term.

Moreover, the spillover effect on wages remains negative throughout the sample period.

Overall, the results provide support that high-tech clusters generated by MIC25 are draining

resources from the neighboring cities. These results are particularly interesting in the context

of China. The Hukou system in China restrains the migration of workers who want to seek

better job opportunities in the other cities, while there are no particular restrictions on the

reallocation of firms.

We have shown that pilot cities experience large growth in job openings and an increasing

trend in wages and their neighboring cities suffer in the short term from the decrease in

job openings and wages. Which of the effects dominates in the short term? To provide an

answer, we compare the sample of pilot cities and their neighbors to the sample of control

cities and their neighbors. Visually, we are comparing the red plus blue cities on the left of

versus the red plus blue cities on the right of Figure 3. We call these two samples composite

cities. The aim of this exercise is to figure out if the positive effect on pilot cities outweighs

the negative effect on neighboring cities or vice versa. The results are shown in Figure 7. We

21

Figure 7: The Impacts on Job Postings in the Composite Cities

(a) Number of Job Openings (b) Monthly Wages

Notes: Figures present coefficients and 95% confidence intervals on the interaction be-tween the indicator of relative quarterly periods from the policy announcement date andthe indicator of the composite cities around the pilot cities. The composite cities of thepilot cities include both pilot cities and their neighboring cities and the composite citiesof control cities include both control cities and their neighboring cities. Industry-by-cityand industry-by-calendar quarter fixed effects are controlled and each city-by-industryobservation is weighted by the total number of postings in 2015.

find that the negative effects on neighboring cities dominate the positive effects of the pilot

cities. This is not surprising as there are 5 neighbor cities experiencing negative effects for

each pilot city experiencing large positive effects.

To get a sense of the overall effects after the policy, we run a difference in difference specifica-

tion. The results are shown at Table 4. Columns (1) and (3) of Table 4 compare neighbor cities

of pilot cities versus neighbor cities of control cities before and after the policy. Columns (2)

and (4) compare composite cities of pilot cities versus composite cities of control cities be-

fore and after MIC25. The coefficient in column (1) shows that neighboring cities suffer a

decrease of about 25% overall. Column (2) shows that the increase in pilot cities mitigates

slightly for the decrease. However, the overall effect is still negative at about 17% in the

composite cities. Similarly for wages. The decrease in wages for the neighboring cities dom-

inates the slight increase in pilot cities. The overall decrease in wage in composite cities is

about 2%. Thus, we conclude that in the short-term, the construction of high-tech clusters in

the pilot cities increases labor demand and wages in pilot cities, but the negative spillovers

in neighboring areas dominate the overall effects of MIC25.

22

Table 4: Estimation Results of Labor Demands around the Pilot Cities

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

Log(Openings) Log(Openings) Log(Wage) Log(Wage)

Post×Pilot Cities -0.248*** -0.171** -0.032** -0.020**

(0.086) (0.085) (0.013) (0.010)

Unit Neighborhood Cities Composite Cities Neighborhood Cities Composite Cities

Observations 98,612 119,997 95,398 116,368

Notes: The estimates are calculated based on the aggregated treatment effects version of figures 6 and 7. Thecontrol cities are based on the propensity matching method. The treatment group in columns (2) and (4)include both pilot cities and their neighboring cities and the control cities include both control cities and theirneighboring cities. Industry-by-city and industry-by-calendar quarter fixed effects are controlled and eachcity-by-industry observation is weighted by the total number of postings in 2015.

5.1.3 Heterogeneity Analysis

We have shown the dynamics between pilot and neighboring cities in terms of job openings

and wages. What about the heterogeneity within the high-tech clusters? The next step is to

investigate the heterogeneity of effects between industries and types of occupations. Given

MIC25 is establishing high-tech clusters, one might expect the benefits to be concentrated

on targeted industries and related occupations. Therefore, we run the event studies across

industries and occupations.

We plot the event study coefficients for target and non-targeted industries in Figure B.1.

Targeted industries are defined as manufacturing, IT, science, and business. The number of

job openings does not vary. The large increase in job openings is evident for targeted and

non-targeted industries. Thus, MIC25 enlarges job opportunities not only in the targeted

high-tech industries, but also the other industries which could have complementary with

the targeted industries. The other reason why the targeted and non-targeted industries have

the similar increasing patterns in job openings would be that the targeted industries are more

capital-intensive. Though the government’s investment subsidies would concentrate in the

targeted industries, the effective labor demand increase in those industries could be smaller

than that in the non-targeted industries. So, the direct effects in the targeted industries could

be similar to the spillover effects in the non-targeted industries.

On the right panel of Figure B.1, wages in targeted and non-targeted industries appear to co-

move perfectly around the policy period. Interestingly, after one year of the implementation

of MIC25, we observe a diverging pattern on wages for targeted and non-targeted industries.

While not significant, the point estimates hint that wages in targeted industries are 2-3%

23

Figure 8: The Impacts on Job Postings in the Pilot Cities by Industry

(a) Number of Job Openings (b) Monthly Wages

Notes: Figures present coefficients and 95% confidence intervals on the interaction be-tween the indicator of relative quarterly periods from the policy start date and the indi-cator of the pilot cities. The targeted industries of MIC25 include manufacturing, IT, sci-ence, and business sectors. Industry-by-city and industry-by-calendar quarter fixed ef-fects are controlled and each city-by-industry observation is weighted by the total num-ber of postings in 2015..

higher and on an increasing trend. The point estimate for wages in targeted industries in

the end of our sample suggests an 8% increase. The divergence in wages arises at the same

time when the job postings start increasing. Given this result, the next natural step is to

investigate whether there is heterogeneity across occupation types.

If building high-tech clusters induces routine-biased technological change as reported in

the developed countries, then we would observe the disproportionate impacts of MIC25 on

the demand for non-routine positions. So, we classify the occupations based on the tasks

they perform following Acemoglu and Autor (2011): routine versus non-routine. On the

left panel of Figure 9, after about 6 quarters of the policy, we observe that job openings for

non-routine jobs appear to be 10-15% higher than routine jobs. Thus, the hiring patterns are

consistent with the hypothesis that MIC25 results in routine-biased technological change

and upskilling as the policy targets technological upgrading of high-tech industries. The co-

movement before the policy across occupations and initially after the policy is convincing

that MIC25 is behind the divergence. Similar to all previous results, the divergence arises

after 1 year the policy has been announced.

On the right panel of Figure 9, wages for routine and non-routine jobs are quite similar up

to 1 years after the policy. After 1 years, we notice the same diverging pattern as before

24

Figure 9: The Impacts on Job Postings in the Pilot Cities by Occupation

(a) Number of Job Openings (b) Monthly Wages

Notes: Figures present coefficients and 95% confidence intervals on the interaction be-tween the indicator of relative quarterly periods from the policy start date and the in-dicator of the pilot cities. The classification of occupations into non-routine and routinejobs follows Acemoglu and Autor (2011). Industry-by-city and industry-by-calendarquarter fixed effects are controlled and each city-by-industry observation is weighted bythe total number of postings in 2015. Each line is estimated by a separate regression foreach occupational group.

where wages for non-routine jobs increase more. While the differences are not significant,

the coefficient estimates provide evidence that MIC25 is opening more jobs for non-routine

occupations with higher wages after 1 year of the implementation date. This means that the

wage gap among occupations are disproportionately raised within the pilot cities compared

to the control cities after building high-tech clusters.

To summarize the before and after effects of MIC25 by occupation we run a difference in

difference specification by occupation group. The results are shown in table 5. Columns (1)

and (2) show the results of job openings between non-routine and routine occupations. The

coefficients are both positive and significant. The job openings for non-routine occupations

increase by about 27% while the job openings for routine occupations increase by about

19%. There is 8% more job openings created by MIC25 for non-routine occupations. Similar

results for wages are shown in columns (3) and (4). Wages for non-routine occupations

show an increase of about 3% and is statistically significant. The effects on wage for routine

occupations are close to zero.

We conclude that aside from regional inequality, MIC25 is causing inequality in terms of

labor demand across occupations as well as wages across occupations and industries. Non-

25

Table 5: Estimation Results of Labor Demands in the Pilot Cities by Industry and Occupation

Panel A: By Industry

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

Log(Openings) Log(Openings) Log(Wage) Log(Wage)

Post×Pilot Cities 0.259** 0.235* 0.009 0.025

(0.128) (0.133) (0.027) (0.021)

Industry Targeted Non-targeted Targeted Non-targeted

Observations 5,025 15,135 5,025 15,135

Panel B: By Occupation

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

Log(Openings) Log(Openings) Log(Wage) Log(Wage)

Post×Pilot Cities 0.266*** 0.192*** 0.027*** 0.013

(0.042) (0.056) (0.009) (0.008)

Job Type Non-routine Routine Non-routine Routine

Observations 145,323 97,609 139,358 93,340

Notes: The estimates are calculated based on the aggregated treatment effects version of figures B.1 and 9. Thecontrol cities are based on the propensity matching method. In Panel A, the targeted industries of MIC25 in-clude manufacturing, IT, science, and business sectors and industry-by-city and industry-by-calendar quarterfixed effects are controlled. In Panel B, the classification of occupations into non-routine and routine jobs fol-low Acemoglu and Autor (2011) and occupation-by-city and occupation-by-calendar quarter fixed effects arecontrolled. Each observation is weighted by the total number of postings in 2015.

routine occupations and targeted industries observe increased wages as expected given

the high-tech nature of the policy. The divergence in wages is increasing more and more

throughout the sample. While we investigate short-run effects, inequality could potentially

be higher in the long-run given the trends. Equipped with the wage results, we turn our

attention to the highest cost of living in China, the housing market.

5.2 Housing Market

The importance of looking at the housing market is twofold. First of all, rent and housing

are some of the largest costs of living for China. It can give us an idea of whether the policy

is making it more expensive to live in the pilot cities. Secondly, increasing house and rent

prices might indicate higher demand and migration. Given we cannot observe between city

migration or seasonal migration, the housing market serves as a proxy for it. The results

from the event studies are shown in Figure 10. Note that the specification used here differs

slightly from the main specification. Firstly, we only have data available for around 100

cities. We have rent price data for 24 pilot cities. Since we do not have all of the control cities,

26

Figure 10: The Impacts on Monthly Rents

(a) Pilot Cities (b) Neighboring Cities

Notes: Figures present coefficients and 95% confidence intervals on the interaction be-tween the indicator of relative quarterly periods from the policy announcement dateand the indicator of the pilot cities (or the indicator of the neighborhood cities in Figure(b)). City and calendar quarter fixed effects are controlled and each city is weighted bypopulation in 2015.

we decide to use all other cities we have data on as control cities except for neighboring

cities. Ideally, we would want to have data on all cities, but since most of the pilot cities are

available the exercise is still informative.

Figure 10 shows that pilot cities experience significant increases on rent prices relative to

other cities. The effects jump quickly to around 5% immediately after the policy. After 1

year the effects remain persistent at 10% until the end of the sample period. The lack of

pre-trends is assuring that the housing market is responding to the announcement of pilot

cities. To investigate the overall effects, we also run a difference-in-differences specification,

the results are shown in Table A.3. The average effect of the policy is a 7% increase in rent

prices in pilot cities. The average monthly rent is around 360$ for our sample.7 Thus, the

policy results in 25$ increase in monthly rent (302$ per year). MIC25 causes the residents in

pilot cities to pay higher rent and housing prices. In contrast, the neighboring cities do not

experience any increase or decrease in monthly rent prices as shown in panel (b).

Given wages between pilot and control cities did not change for the first two years after the

policy, people in the pilot cities were worse off due to the housing market price increase.

In Figure 9 we see that wages increase after 2 years. The increase in wages for non-routine

occupations is steeper and statistically significant compared to routine occupations. Since

7We use 1 USD = 7 CNY as the approximate conversion rate.

27

Table 6: Estimation Results of Monthly Rent

(1) (2) (3)

Log(Rent) Log(Rent) Log(Rent)

Post×Pilot Cities 0.071** -0.018 0.051*

(0.032) (0.059) (0.029)

Unit Pilot Cities Neighborhood Cities Composite Cities

Observations 1,010 951 1,247

Notes: The estimates are calculated based on the aggregated treatment effects version of figures 10. Column(1) includes 24 pilot cities and 56 controls cities, which are not the neighborhood cities. Column (2) includes26 neighborhood cities and 56 controls cities, which are not the pilot cities. Column (3) includes 50 pilot andneighborhood cities and 56 controls cities, which are not the pilot cities.

the rent and houses are now more expensive, workers in routine occupations are worse off

compared to non-routine occupations. Both occupation groups are worse off in the first two

years as the housing market reacts as soon as the policy is enacted. Wages are slower to react

and they pick up around 2 years after the policy which corresponds to pilot cities becoming

National Demonstration Zones and gaining even more support from the government. Since

the policy is still on going, the increasing trend in wages would have to persist in order to

overcome the increase in living expenses.

In addition to living expenses, the housing market speaks to migration patterns. We know

from economic models of place based policies that if workers are mobile and housing supply

is inelastic workers are going to move and rent and housing prices will increase (Kline and

Moretti, 2014b). The event studies in Figure 10 show that the housing market responds

even earlier than the labor demand. We interpret this as evidence that workers are reacting

quickly to the policy change and moving to the pilot cities as soon as they are announced.

Similarly, Alder, Shao, and Zilibotti (2016) find that Special Economic Zones in China attract

large populations. Hence, the limited housing supply and increased demand can explain

the increase in the price of the housing market. This scenario would make landowners in

the area better off (Kline and Moretti, 2014b), while having a negative effect on the average

worker whose rent or house price would increase. In summary, we observe increasing rent

prices and flat wages in the short term with an increasing trend towards the end of the

sample period. Our conclusion is that in the short term, the average worker is worse off

as the living expenses increase and the wage does not. This is more severe for workers in

non-targeted industries and workers in routine occupations. The increasing trend in wages

for non-routine and target industries may worsen inequality in the long-run.

28

5.3 Mechanisms

We have shown evidence of regional and wage inequality caused by the high-tech clusters

formed by MIC25. The aim of this section is shed light on the mechanisms behind the results.

Specifically, we focus on firm and worker location choices. Firstly, we want to investigate

whether the large labor demand increase is due to incumbent firms expanding of new firm

creation. Secondly, the housing cost increase suggests potential migration in order to fill the

labor demand increase.

Using job postings data we cannot distinguish whether a firm posting for the first time is

a new firm or it might be an old firm posting online for the first time. Therefore to answer

the question we leverage new firm registration data. Similarly, we run event study and a

difference in difference specification shown in Figure 11 and Table 7 respectively.

Panel (a) of Figure 11 compares firm entry in pilot versus control cities. The pre-trend is

flat and the firms react as soon as the pilot city is announced as expected. The increase is

significant and larger right after the announcement period. The increase in the number of

firms early on could explain why labor demand is slower to pick up as it might take some

time for firms to establish and receive funding on their projects before hiring. Panel (b)

shows the results for neighboring areas. Right after the policy, neighboring cities suffer large

and persistent firm entry decreases. The timing is reasonable. As soon as an entrepreneur

learns about the high-tech clusters, they would rather locate there compared to neighboring

ares. This impact of MIC25 could be troublesome for neighboring cities in the long run

as firms would rather locate in the pilot cities. In order to explain the increasing trend in

labor demand in the neighbor cities found in section 6.1, it must be that incumbent firms are

hiring more in neighboring cities in the later periods. The intensive margin appears to play

an important role in affecting labor demand for neighboring cities.

After understanding the timing of the reaction of the new firm to MIC25, we investigate

the overall effects for pilot, neighboring and composite cities as before. Columns (1)-(3) of

Table 7 presents the difference in difference results for firm registration. Overall, pilot cities

experience about 10% increase in new firm entry after MIC25 compared to control cities.

That corresponds to about 8,000 more new firms per year in pilot cities versus the control

cities. For comparison, (Lu, Wang, and Zhu, 2019) find Special Economic Zones (around

2004-2008) experience an increase of 29% in the number of firms.

29

Figure 11: The Impacts on Firm Entry

(a) Pilot Cities (b) Neighboring Cities

Notes: Figures present coefficients and 95% confidence intervals on the interaction be-tween the indicator of relative quarterly periods from the policy start date and theindicator of the pilot cities (or the indicator of the neighborhood cities in Figure (b)).Industry-by-city and industry-by-calendar quarter fixed effects are controlled and eachcity-by-industry observation is weighted by the number of registered firms in 2015.

Even more staggering, the estimates for neighboring cities how a 32% decrease in new firm

entry. When we compare composite cities as before, the negative effects suffered by neigh-

boring areas outweigh the positive effects in the high-tech clusters. There are simply more

neighboring cities undergoing lack of firm entry. The overall effects are large and negative.

One caveat here is that we cannot distinguish how much of the labor demand increase is

from incumbents and how much from new firms. However, the results on firm entry sug-

gests a significant labor demand impact from the creation or lack of new firms. (Lu, Wang,

and Zhu, 2019) find that almost 80% of the increase in employment is due to the creation

of new firms when studying Special Economic Zones in China. Similarly, (Criscuolo et al.,

2019) find that most of the positive effect on employment in place based policy in UK comes

from small or new established firms. While we cannot conclude the degree of extensive ver-

sus the intensive margin, we believe that the establishment of new firms plays a key role in

increasing labor demand for pilot cities and decreasing in neighboring areas.

After establishing that MIC25 caused firm reallocation, we shift our attention to worker lo-

cation decisions. In order to do so, we run the main specification event study and difference

in difference using City Statistical Yearbook population variable. A caveat is that in the City

Yearbook the population measure corresponds to the end-of-year number of persons whose

Hukou (house-hold registration) are registered in the city. As a result, this measure poten-

30

Figure 12: The Impacts on Population

(a) Pilot Cities (b) Neighboring Cities

Notes: Figures present coefficients and 95% confidence intervals on the interaction be-tween the years and the indicator of the pilot cities (or the indicator of the neighborhoodcities in Figure (b)). The policy is implemented in 2016 at all the pilot cities. City andyear fixed effects are controlled and each city is weighted by population in 2015.

tially excludes migrant workers who work in the city but do not have their Hukou registered

in the same city. Moreover, the City Yearbook is not available for 2019-2020. On the other

hand, we do more data to observe pre-trends. The results are shown in Figure 12. After

MIC25 we observe an increase in population. The trend is also increasing, which would

lead to larger population increase in the future for pilot cities. The effect is not significant

and close to zero for neighboring ares.

Similar to other results, we show the difference in difference specification on columns (4)-(6)

of Table 7. The overall increase in population is significant at 2%. Given the measure for

population excludes a portion of migrants, we consider the result to be a lower bound. In

contrast, neighboring ares have close to zero effects. In column (5) we see that the composite

cities experience population growth driven by pilot cities. We conclude that workers are

moving into pilot cities along with new firms. This result is consistent with the higher la-

bor demand and wages in the pilot cities. The increase in population would also result in

higher rent prices due to increased demand. While our measure is not perfect, it shows evi-

dence that one of the mechanisms of MIC25 is worker reallocation even though the Hukou

restriction. This channel would potentially be even more important for countries with no

movement restrictions.

31

Table 7: Estimation Results of Firm and Worker Sorting

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

Log(FirmEntry)

Log(FirmEntry)

Log(FirmEntry)

Log(Pop) Log(Pop) Log(Pop)

Post×Pilot Cities 0.100* -0.317*** -0.298*** 0.021*** 0.006 0.009***

(0.059) (0.084) (0.080) (0.007) (0.004) (0.003)

Unit Pilot Cities NeighborhoodCities

CompositeCities

Pilot Cities NeighborhoodCities

CompositeCities

Observations 22,858 110,594 133,452 345 1,642 1,987

Notes: The estimates are calculated based on the aggregated treatment effects version of figures 11 and 5. Ob-servations in columns (1) through (3) are city-industry pairs and each city-by-industry observation is weightedby the number of registered firms in 2015. Industry-by-city and industry-by-calendar quarter fixed effects arecontrolled in columns (1) through (3). Observations in columns (4) through (6) are cities and each city isweighted by population in 2015. City and year fixed effects are controlled in columns (4) through (6). Thetreatment group in columns (3) and (6) include both pilot cities and their neighboring cities and the controlcities include both control cities and their neighboring cities.

6 Welfare Analysis

The empirical results indicate a variety of impacts of MIC25 on different economic variables

within and around the pilot cities where the high-tech clusters were built: (1) Labor de-

mand and offered wages increase in the pilot cities especially for non-routine jobs; (2) Labor

demand and offered wages decrease in the neighboring cities of the pilot cities; (3) Hous-

ing cost rises in the pilot cities; (4) Firms and workers are more likely to move to the pilot

cities.

6.1 Conceptual Framework

To combine all those estimates in Section 5 and learn the welfare implication of all those ef-

fects, we present a conceptual framework consistent with our reduced-form estimates.

Following the literature, we rely on three main channels driven by workers’ and firms’ be-

haviors. First, since the policy provides extra investment incentives to firms in pilot cities,

firms’ hiring decision (extensive margin) changes in a heterogeneous way among the pilot

and neighboring cities. Second, firms could change their labor demand across occupations

(i.e., intensive margin) within the pilot cities because the adoption of technologies induced

by high-tech investment has heterogeneous effects on the productivity of different occupa-

tions. Third, workers’ migration and sorting for jobs to pilot cities raise demand in the local

housing market, increasing the rent therein.

32

Following Kline and Moretti (2014b), we construct a parsimonious spatial equilibrium model

to highlight all the aforementioned channels by extending conventional spatial equilibrium

models of Rosen (1979) and Roback (1982). In the model, we allow for two types of la-

bor inputs and heterogeneity in capital-labor complementarity. We assume a continuum of

workers of measure one and each of them resides and works in a city j ∈ J = {1, 2, · · · , J},

which is either a pilot city (a) or a neighboring city (b). A typical worker demands a unit of

housing which is rent at local market rates. Each worker elastically supplies a unit of labor

of one type, routine (R) or non-routine (NR).8 To account for the migration constraint in

China (i.e., hukou or household registration system), we assume a moving cost if the worker

moves out of her hukou city. The indirect utility of worker i, from city j0 ∈ J (hukou city),

choosing to live in city j ∈ J and work in occupation s ∈ {R, NR}9 is:

Vsi,j = ws

j − rj − bs1(j0 6= j)− Asj + σsεi,j ≡ vs,j0

i,j + σsεi,j, (3)

where wsj is the nominal wage in city j and occupation s, rj is the rent of housing, As

j is a

measure of local amenities. bs is the moving cost, which depends on workers’ skills. The

heterogeneity in individual preference for a city and an occupation is important to generate

heterogeneous welfare effects of the policy change, as some workers are infra-marginal and

can take economic rents of policy changes. We assume that σsεi,j is independently and iden-

tically distributed across individuals and follows a type-I extreme value (T1EV) distribution

with mean zero and scale parameter σs (occupation specific). The scale parameter governs

the strength of idiosyncratic preferences for cities and occupations and in turn affects the

sorting of workers. The larger σs is, the larger shocks to wages and rents are required to

induce workers to switch.

Firms in each city j produce a single good Y using capital (K), routine labor (LR), and non-

routine labor (LNR) following a constant return to scale production technology that incorpo-

rates capital-skill complementarity

Yj = L1−βNR,j

(αRLµ

R,j + αkKµj

) βµ . (4)

8Note that here we assume workers are of two types: high skilled and low skilled. High-skilled workerschoose to work in non-routine occupations while low-skilled workers work in routine occupations.

9To characterize the welfare impact of MIC25, we assume that workers cannot switch their occupationacross cities. This assumption is consistent with our focus on short-term effects.

33

In this production function, capital is a complement to non-routine labor and a substitute

for routine labor.10 This form has been widely used in the literature to investigate the ef-

fect of the decrease in capital cost on the composition of labor force (Autor and Dorn, 2013;

Karabarbounis and Neiman, 2014). In our setting, the implementation of MIC25 dispropor-

tionately lowers capital cost for the pilot cities. The labor market is perfectly competitive

and firms set wages equal to the marginal product of labor. A friction-less capital market

supplies capital perfectly elastically at local-level price ρj.11

We can derive routine-non-routine wage gap in city j:

log

(wNR

j

wRj

)= log

(1− β

β

)+ log

(LR,j

LNR,j

)− log

(1−

αkKµj

αRLµR,j + αkKµ

j

)

' log(

1− β

β

)+ log

(LR,j

LNR,j

)+

αkKµj

αRLµR,j + αkKµ

j.

(6)

The second row of Equation (6) gives a linear decomposition of the sources of location-

specific policy changes in wage inequality incorporating two particular channels. First, in-

equality can increase as a result of an increase in the proportion of routine labor compared

to non-routine labor (i.e., LR,j/LNR,j). Second, the share of capital in the factor of produc-

tion that combines capital and routine labor raises the wage gap as shown in the third term,

which reflects routine-biased technological change induced by high-tech clusters. So, the ef-

fect of MIC25, a decline of capital cost in pilot cities, on the wage gap in pilot cities depends

on the composition change of inputs and worker’s distribution over the preference for a city

and an occupation.

For simplicity, we assume housing supply is upward sloping in each city to capture the land

10The elasticity of substitution between capital or routine labor and non-routine labor is β1−β and that be-

tween capital and routine labor is 11−µ with β < 1 and µ < 1.

11Firms’ inverse demands for capital and labor at city j are given by:

ρj = βL1−βNR,jQ

βµ−1j

(αkKµ−1

j

),

wNRj = (1− β)L−β

NR,jQβµ

j ,

wRj = βL1−β

NR,jQβµ−1j

(αRLµ−1

R,j

),

(5)

where Qj = αRLµR,j + αkKµ

j . Note that the capital cost is set at city level to account for the fact that the taxbenefits and subsidies of MIC25 lowers the capital cost for the firms in the pilot cities.

34

supply restrictions in China. The inverse supply function is given by equating local rents

as in Busso, Gregory, and Kline (2013), Kline and Moretti (2014b), and Liang, Song, and

Timmins (2020):

ln(rj) = αj + κj ln(Nj), (7)

where the aggregate housing demand is determined by the number of workers in the city

regardless of their occupation. We assume that the elasticity of housing supply (κj) is exoge-

nously determined by geography and local land use regulations.

We can write the expected utility of workers from city j0 and working in occupation s

as:

Vs,j0 = Eε

[max

j{vs,j0

i,j + σsεi,j}]= σsNs,j0 ln

(1 + ∑

j∈Jexp

(vs,j0

i,j /σs))

, (8)

where Ns,j0 is the population of workers working in occupation s and with hukou city j0.

From the T1EV assumption on the idiosyncratic term σsεi,j, we can derive the number of

workers of types (s, j0) who choose to work and live in city j as

Ns,j0j = Pr

(Vs,j0

i,j = maxj′∈J

Vs,j0i,j′

)=

exp(vs,j0i,j /σs)

1 + ∑j′∈J exp(vs,j0i,j′ /σs)

, (9)

where the first equality is from our assumption on a continuum of workers.

Note that MIC25 lowers the capital cost in pilot cities (ρa) by granting subsidies and tax

deductions to the firms in the city.12 So, we model the implementation of MIC25 as a de-

cline in ρa and the welfare incidence of MIC25 on the workers from city j0 and working in

occupation s can be represented by:

dVs,j0

dX=

dVs,j0

dvs,j0i,j

dvs,j0i,j

dX= Ns,j0

[∑j∈J

Ns,j0j

(dws

j

dX−

drj

dX− dbs1(j0 6= j)

dX−

dAsj

dX

)]

' Ns,j0

[∑j∈J

Ns,j0j

(dws

j

dX−

drj

dX

)],

(10)

where Ns,j0j represents the population of workers of type (s, j0) who work in city j prior

to the implementation of the policy. The second row of Equation (10) holds because we

12In estimation, we cannot directly observe the change in capital cost. So, the policy shock (dX) can beviewed as the one reducing the capital cost in the pilot cities (dρa/dX < 0)).

35

focus on short-run effect of the policy and migration cost and local amenities are not likely

to change dramatically in the short run. Thus, the welfare incidence of MIC25 on workers

in each occupation group boils down to the change in wages and rents weighted by their

population shares across hometown j0 and their choice probabilities across the cities before

the policy shock.13

As shown in Equation (6), the wage gap between routine and non-routine jobs is widened

as a result of decreasing capital cost in the pilot cities: d(wNRa /wR

a )dX = d(wNR

a /wRa )

dKadKadρa

dρadX > 0.

Because of in-migration of workers from other cities, the rents in the pilot cities increases:dradX > 0. These model-based predictions are well aligned with our estimation results in

Section 5.

Consequently, the welfare gap between routine and non-routine in the pilot cities is also

widened because the incidence of MIC25 on the rents is equally shared by both types of

workers in the pilot cities. However, the welfare of routine-job workers in the pilot cities

gets worse if the in-migration of workers from the neighbor cities is large enough to offset

the wage increase for them: d(wNRa −ra)dX > 0 > d(wR

a −ra)dX . It depends on how many workers are

induced to move to the pilot cities after MIC25, which depends on worker’s distribution of

preference across cities and the moving cost induced by hukou constraint.

We do not explicitly model firm’s resource allocation across cities, but we can investigate

the impact of MIC25 on the neighboring cities by assuming the substitutability between

the production facilities in the two places. This assumption is not implausible because the

benefits of MIC25 are conditional on investment in the pilot cities and are not shared within

the firms’ internal networks across regions. So, investment subsidies in the pilot cities reduce

investment in the neighboring cities: dKbdX < 1. From Equation (5), we can observe that

wages for both occupational groups can be lowered if the decline of labor force does not

fully absorb the capital decrease: dwsb

dX < 0 ∀s.

However, housing cost in the neighboring cities can be reduced in response to building

high-tech clusters in the pilot cities if the outflow of workers from the neighboring cities to

the pilot cities is large enough. Our empirical results confirm that the decreases in capital

13Of course, after the shock, some workers who were initially “marginal” workers will re-optimize andmove. However, the Envelope Theorem suggests agents will only have small gains in utility by switchingtheir choices in response to marginal changes in prices. Hence, these behavioral effects due to price changescould only have secondary effects on worker welfare (Qian and Tan, 2021).

36

Table 8: Migrants in Pilot Cities and Neighboring Cities

Occ. j0 j Ns,j0 (%) Ns,j0j (%) ws

j ($)dws

jdX (%) rj ($)

drjdX (%) Ns,j0

jdv

s,j0i,j

dX ($) dVs,j0dX ($)

1 NR a a 14.98 73.90 10241.64 2.7 4853.4 7 -46.71 -7.00

NR a b 26.10 9450.25 -3.2 4572.48 0 -78.93 -11.82

2 R a a 9.54 27.46 9069.46 0 4853.4 7 -93.30 -8.90

R a b 72.54 8359.92 -3.2 4572.48 0 -194.04 -18.51

3 NR b a 13.75 33.38 10241.64 2.7 4853.4 7 -21.10 -2.90

NR b b 66.62 9450.25 -3.2 4572.48 0 -201.46 -27.70

4 R b a 61.72 42.98 9069.46 0 4853.4 7 -146.03 -90.13

R b b 57.02 8359.92 -3.2 4572.48 0 -152.53 -94.14

Total 100 -261.11

Notes: The annual total welfare change in the table is calculated for one representative worker. All the sharesof workers are calculated from the 2017 wave of China’s migrants dynamic survey. People with college degreeor above are assumed to work in NR occupations. Only people in working age (16-65) and with hukou cityin either a pilot city (a) or a neighboring city (b) are included. The total number of migrants is 51,021, whichaccounts for 30% people in the survey (169,989 people are surveyed in the 2017 wave). The annual wages ws

jare calculated from averaged monthly wages in 2016, the annual housing rents rj are calculated from averagedmonthly rents for a 100 m2 house in 2020.

(Figure 11) is large enough to lower wages in the neighboring cities. On the contrary, the net

outflow of workers cannot be observed in Figure 12 and housing cost also does not respond

to MIC25 much as shown in Figure 10.

6.2 Welfare

From the model, we can predict the effects of MIC25 on wages and rents change for those

four subgroups (s, j0) of population: {(NR, a), (R, a), (NR, b), (R, b)}.14 As has been argued

in the previous subsection, the change in utilities of people switching their location choices

has only secondary effects on worker welfare, so we don’t need to specify people’s locations

after the policy. Then we can calculate the welfare incidence on those four subgroups of pop-

ulation by combining the shares of migrants and natives across the pilot and neighboring

cities.

We utilize the 2017 wave of China’s migrants dynamic survey to calculate the share of mi-

grants in each city, then we aggregate them to the four subgroups and calculate the welfare

change by combining Equation (10) and estimates from Section 5.

Table 8 shows the welfare change for workers in each subgroup and aggregated welfare

14Note that a represents pilot cities and b refers to neighboring cities.

37

change. Only people of high skills (i.e., working in NR occupations) and choosing to work

in pilot cities could gain from the policy. People of low skills (i.e., working in R occupa-

tions) and choosing to work in neighboring cities suffer the most from the policy due to the

significant wage drop and their large share of among all migrants.

The back-of-the-envelope calculation in Table 8 suggests that on average a migrant worker

from either a pilot city or a neighboring city would lose around $260 due to the policy,

which accounts for around 3% loss in annual wages. In 2017, there are 244 million migrants

in China, 30% of them are in pilot cities or neighboring cities, so the welfare loss for all

migrants accounts for around 0.15% of the annual GDP ($ 19 billion/$12.31 trillion).

7 Conclusion

We evaluate the impact of high-tech clusters on labor demand and inequality between re-

gions and occupations. The place-based feature and staggered implementation of the policy

in China provide a suitable setting to study the causal effects. We see a significant increase

in labor demand in the pilot cities using the detailed job posting data. Offered wages are

increasing for non-routine occupations, while there is no significant change in wage for rou-

tine occupations. These results suggest that high-tech clusters are causing inequality across

different occupations. On the other hand, labor demand and wage decrease significantly for

neighboring areas, leading to inequality across regions.

Living costs sharply increase in the pilot cities right after the implementation of MIC25, in-

dicating that the policy results in the inflow of workers. In the neighboring cities, we could

not observe any significant changes in rent. So, building a high-tech cluster has dispropor-

tionate effects in the housing markets across regions. To calculate the overall welfare effects

of MIC25 across subgroups of population, we construct a simple conceptual framework.

The overall welfare effects depend on the population shares across the pilot and neighbor-

ing cities and across non-routine and routine positions along with the effects of the policy on

wages and rents. The back-of-the-envelope from our conceptual framework calculation sug-

gests that on average a migrant worker from either a pilot city or a neighboring city would

lose around $260 per year due to the policy, which accounts for around 3% loss in yearly

wages, and the total welfare loss for migrants accounts for around 0.15% ($ 19 billion) of

China’s annual GDP in 2017. More importantly, the welfare loss concentrates in routine jobs

38

and neighboring cities.

As far as we know, this study is the first to evaluate the consequence of building national-

level high-tech clusters. Given several developed and developing countries are implement-

ing similar policies, we believe our findings would provide helpful policy insight. This is an

essential first step in understanding this new wave of industrial policies. Given that China

has restrictions on migration due to the Hukou system, the potential impact of labor real-

location and wage inequality might be even more pronounced in other countries. Since the

policy is still ongoing, we focus more on the short-term effects in this paper. Although it

might be hard to evaluate long-run effects due to the disruption caused by Covid-19 right

after our sample period, including post-pandemic periods could be a sensible next step and

constructive for assessing the long-term impact.

Building national-level high-tech clusters provide an exceptional tool for policymakers to

adopt modern technology and to upgrade production capability of a nation. Our findings,

however, imply that policymakers may need to be cautious in at least two dimensions. Cre-

ating high-tech clusters appears to have negative short term effects in neighboring areas as

well as longer lasting firm entry decrease. Therefore, undesired regional inequalities may

arise. Finally, high-tech clusters create wage inequality between occupations while increas-

ing living costs. We believe further research is required to see whether the long term pro-

ductivity effects of the high-tech clusters might outweigh the negative short term effects in

inequality.

39

References

Acemoglu, Daron and David Autor. 2011. “Skills, Tasks and Technologies: Implications for

Employment and Earnings.” Handbook of Labor Economics 4:1043–1171.

Acemoglu, Daron, David Autor, Joe Hazell, and Pascual Restrepo. 2020. “AI and Jobs: Evi-

dence from Online Vacancies.” Tech. rep., Mimeo. Massachusetts Institute of Technology.

Acemoglu, Daron and Pascual Restrepo. 2020. “Robots and Jobs: Evidence from US Labor

Markets.” Journal of Political Economy 128 (6):2188–2244.

Alder, Simon, Lin Shao, and Fabrizio Zilibotti. 2016. “Economic Reforms and Industrial

Policy in a Panel of Chinese Cities.” Journal of Economic Growth 21 (4):305–349.

Autor, David H and David Dorn. 2013. “The Growth of Low-Skill Service Jobs and the

Polarization of the US Labor Market.” American Economic Review 103 (5):1553–1597.

Bloom, Nicholas, Mark Schankerman, and John Van Reenen. 2013. “Identifying Technology

Spillovers and Product Market Rivalry.” Econometrica 81 (4):1347–1393.

Borusyak, Kirill and Xavier Jaravel. 2017. “Revisiting event study designs.” Available at

SSRN 2826228 .

Busso, Matias, Jesse Gregory, and Patrick Kline. 2013. “Assessing the Incidence and Effi-

ciency of a Prominent Place Based Policy.” American Economic Review 103 (2):897–947.

Callaway, Brantly and Pedro HC Sant’Anna. 2020. “Difference-in-differences with multiple

time periods.” Journal of Econometrics .

Carlino, Gerald and William R. Kerr. 2015. “Agglomeration and Innovation.” In Handbook

of Regional and Urban Economics, Handbook of Regional and Urban Economics, vol. 5. 349–404.

Criscuolo, Chiara, Ralf Martin, Henry G Overman, and John Van Reenen. 2019. “Some

Causal Effects of an Industrial Policy.” American Economic Review 109 (1):48–85.

Dehejia, Rajeev H. and Sadek Wahba. 1999. “Causal Effects in Nonexperimental Studies:

Reevaluating the Evaluation of Training Programs.” Journal of the American Statistical As-

sociation 94 (448):1053–1062.

Ehrlich, Maximilian and Tobias Seidel. 2018. “The Persistent Effects of Place-Based Policy:

40

Evidence from the West-German Zonenrandgebiet.” American Economic Journal: Economic

Policy 10 (4):344–374.

Fang, Hanming, Chunmian Ge, Hanwei Huang, and Hongbin Li. 2020. “Pandemics, Global

Supply Chains, and Local Labor Demand: Evidence from 100 Million Posted Jobs in

China.” Working Paper 28072, National Bureau of Economic Research.

Forsythe, Eliza, Lisa B. Kahn, Fabian Lange, and David Wiczer. 2020. “Labor Demand in

the Time of COVID-19: Evidence from Vacancy Postings and UI Claims.” Journal of Public

Economics 189.

Goodman-Bacon, Andrew. 2021. “Difference-in-differences with variation in treatment tim-

ing.” Journal of Econometrics .

Graetz, Georg and Guy Michaels. 2018. “Robots at Work.” Review of Economics and Statistics

100 (5):753–768.

Gruber, Jonathan and Botond Koszegi. 2001. “Is Addiction “Rational”? Theory and Evi-

dence.” The Quarterly Journal of Economics 116 (4):1261–1303.

Hanson, Andrew and Shawn Rohlin. 2013. “Do Spatially Targeted Redevelopment Pro-

grams Spillover?” Regional Science and Urban Economics 43 (1):86–100.

Hanson, Gordon H. 2020. “Who Will Fill China’s Shoes? The Global Evolution of Labor-

Intensive Manufacturing.” Working Paper 28313, National Bureau of Economic Research.

He, Chuan, Karsten Mau, and Mingzhi Xu. 2021. “Trade Shocks and Firms Hiring Decisions:

Evidence from Vacancy Postings of Chinese Firms in the Trade War.” Labour Economics 71.

Helleseter, Miguel Delgado, Peter Kuhn, and Kailing Shen. 2020. “The Age Twist in Em-

ployers’ Gender Requests: Evidence from Four Job Boards.” Journal of Human Resources

55 (2):428–469.

Hershbein, Brad and Lisa B Kahn. 2018. “Do Recessions Accelerate Routine-Biased

Technological Change? Evidence from Vacancy Postings.” American Economic Review

108 (7):1737–72.

Kantor, Shawn and Alexander Whalley. 2014. “Knowledge Spillovers from Research Uni-

versities: Evidence from Endowment Value Shocks.” The Review of Economics and Statistics

96 (1):171–188.

41

Karabarbounis, L. and B. Neiman. 2014. “The Global Decline of the Labor Share.” The

Quarterly Journal of Economics 129 (1):61–103.

Kline, Patrick and Enrico Moretti. 2014a. “Local Economic Development, Agglomeration

Economies, and the Big Push: 100 Years of Evidence from the Tennessee Valley Authority.”

The Quarterly Journal of Economics 129 (1):275–331.

———. 2014b. “People, Places, and Public Policy: Some Simple Welfare Economics of Local

Economic Development Programs.” Annual Review of Economics 6 (1):629–662.

Kuhn, Peter and Kailing Shen. 2013. “Gender Discrimination in Job Ads: Evidence from

China*.” Quarterly Journal of Economics 128 (1):287–336.

Lee, Brian K, Justin Lessler, and Elizabeth A Stuart. 2010. “Improving propensity score

weighting using machine learning.” Statistics in Medicine 29 (3):337–346.

Li, Ling. 2018. “China’s Manufacturing Locus in 2025: With a Comparison of “Made-in-

China 2025” and “Industry 4.0”.” Technological Forecasting and Social Change 135:66–74.

Liang, Wenquan, Ran Song, and Christopher Timmins. 2020. “Frictional Sorting.” Tech. rep.,

National Bureau of Economic Research.

Lu, Yi, Jin Wang, and Lianming Zhu. 2019. “Place-based policies, creation, and agglomer-

ation economies: Evidence from China’s economic zone program.” American Economic

Journal: Economic Policy 11 (3):325–60.

Malkin, Anton. 2018. “Made in China 2025 as a Challenge in Global Trade Governance:

Analysis and Recommendations.” .

Moretti, Enrico. 2021. “The Effect of High-Tech Clusters on the Productivity of Top Inven-

tors.” American Economic Review 111 (10):3328–75.

Murphy, Kevin M., Andrei Shleifer, and Robert W. Vishny. 1989. “Industrialization and the

Big Push.” Journal of Political Economy 97 (5):1003–1026.

Qian, Franklin and Rose Tan. 2021. “The Effects of High-skilled Firm Entry on Incumbent

Residents.” .

Roback, Jennifer. 1982. “Wages, rents, and the quality of life.” Journal of political Economy

90 (6):1257–1278.

42

Rosen, Sherwin. 1979. “Wage-based indexes of urban quality of life.” Current issues in urban

economics :74–104.

Rosenstein-Rodan, P. N. 1943. “Problems of Industrialisation of Eastern and South-Eastern

Europe.” The Economic Journal 53:202–211.

Setoguchi, Soko, Sebastian Schneeweiss, M Alan Brookhart, Robert J Glynn, and E Francis

Cook. 2008. “Evaluating uses of data mining techniques in propensity score estimation: a

simulation study.” Pharmacoepidemiology and Drug Safety 17 (6):546–555.

Sun, Liyang and Sarah Abraham. 2020. “Estimating dynamic treatment effects in event

studies with heterogeneous treatment effects.” Journal of Econometrics .

Zenglein, Max J and Anna Holzmann. 2019. “Evolving Made in China 2025.” Mercator

Institute for China Studies 12.

Zhao, Peng, Xiaogang Su, Tingting Ge, and Juanjuan Fan. 2016. “Propensity score and prox-

imity matching using random forest.” Contemporary Clinical Trials 47:85–92.

43

Appendix A Appendix Tables

44

Tabl

eA

.1:B

alan

ceC

heck

:Nei

ghbo

rsof

Pilo

tCit

ies

vsN

eigh

bors

ofC

on-

trol

Cit

ies

Var

iabl

esN

eigh

bors

ofPi

lotC

itie

sN

eigh

bors

ofC

ontr

olC

itie

sD

iffer

ence

P-V

alue

s

Num

ber

ofjo

bpo

stin

g(1

0,00

0)2.

844.

67-1

.83

0.48

62N

umbe

rof

job

open

ing

(10,

000)

7.85

11.9

6-4

.11

0.53

85W

age

(mon

thly

,yua

n)45

22.5

945

59.0

4-3

6.45

0.70

46Po

pula

tion

(10,

000)

497.

4841

5.44

82.0

40.

0969

Empl

oym

ent(

10,0

00)

63.7

266

.15

-2.4

30.

8784

GD

P(b

illio

nyu

an)

238.

5325

0.44

-11.

910.

8072

Land

area

(km

$2$)

1510

3.89

1117

2.16

3931

.74

0.02

93N

umbe

rof

indu

stri

alfir

ms

1160

.22

1519

.24

-359

.03

0.08

49Ta

xpr

ofit(

billi

onyu

an)

26.9

325

.47

1.46

0.84

82G

DP

grow

thra

te(\

%)

7.62

7.57

0.05

0.89

94Em

ploy

men

tsha

re(s

econ

dary

sect

or,\

%)

45.8

450

.67

-4.8

30.

0179

Empl

oym

ents

hare

(thi

rdse

ctor

,\%

)52

.72

48.2

64.

460.

0237

GD

Psh

are

(sec

onda

ryse

ctor

,\%

)47

.90

48.5

0-0

.59

0.63

57G

DP

shar

e(t

hird

sect

or,\

%)

40.7

839

.88

0.90

0.48

22Em

ploy

men

tsha

re(m

anuf

actu

ring

,\%

)23

.61

28.9

2-5

.31

0.00

43Em

ploy

men

tsha

re(I

T,\%

)1.

361.

210.

150.

2880

Empl

oym

ents

hare

(bus

ines

s,\%

)1.

691.

640.

060.

7901

Empl

oym

ents

hare

(sci

ence

,\%

)1.

651.

350.

290.

0614

Not

es:T

his

tabl

epr

esen

tsth

eav

erag

esac

ross

the

neig

hbor

sof

pilo

tcit

ies

vers

usth

ene

ighb

ors

ofco

ntro

lcit

ies

in20

15be

fore

MIC

25ha

dbe

enim

plem

ente

d.C

olum

n4

show

sth

edi

ffer

ence

betw

een

the

aver

ages

,and

colu

mn

5sh

ows

the

p-va

lues

fora

t-te

stde

sign

edto

test

sw

heth

erth

eav

erag

esof

the

two

grou

psar

est

atis

tica

llydi

ffer

ent

from

one

anot

her.

Not

eth

atne

ighb

orci

ties

are

chos

enpu

rely

onge

ogra

phic

alch

arac

teri

stic

san

dva

riab

les

show

she

rear

eno

tta

rget

edby

usto

besi

mila

rbe

twee

nth

egr

oups

.

45

Table A.2: Average Treatment Effects with the Random Forest Matching Algorithm

Panel A: Pilot Cities

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

Log(Openings) Log(Wage) Log(Firm Entry) Log(Pop)

RF100 0.249*** 0.020 0.110* 0.019***

(0.089) (0.026) (0.063) (0.007)

RF200 0.250*** 0.018 0.111* 0.016**

(0.086) (0.026) (0.063) (0.007)

RF500 0.250*** 0.018 0.114* 0.016**

(0.086) (0.028) (0.062) (0.007)

RF1000 0.248*** 0.018 0.120** 0.017**

(0.084) (0.026) (0.061) (0.007)

RF5000 0.247*** 0.018 0.117* 0.018***

(0.083) (0.026) (0.061) (0.007)

Panel B: Neighbor Cities

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

Log(Openings) Log(Wage) Log(Firm Entry) Log(Pop)

RF100 -0.224** -0.034** -0.359*** 0.000

(0.113) (0.014) (0.082) (0.003)

RF200 -0.177 -0.026* -0.342*** 0.000

(0.118) (0.014) (0.086) (0.003)

RF500 -0.172 -0.026* -0.349*** 0.000

(0.121) (0.014) (0.082) (0.003)

RF1000 -0.174 -0.026* -0.343*** 0.000

(0.123) (0.013) (0.083) (0.003)

RF5000 -0.225* -0.033** -0.348*** 0.000

(0.116) (0.013) (0.082) (0.003)

Panel C: Composite Cities

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

Log(Openings) Log(Wage) Log(Firm Entry) Log(Pop)

RF100 -0.143 -0.019* -0.291*** 0.004

(0.114) (0.011) (0.077) (0.003)

RF200 -0.127 -0.019* -0.276*** 0.004

(0.111) (0.011) (0.080) (0.003)

RF500 -0.117 -0.019* -0.281*** 0.003

(0.111) (0.010) (0.076) (0.003)

RF1000 -0.119 -0.019* -0.276*** 0.004

(0.112) (0.010) (0.077) (0.003)

RF5000 -0.148 -0.020* -0.282*** 0.004

(0.113) (0.010) (0.077) (0.003)

Notes: The aggregated treatment effects are calculated following Callaway and Sant’Anna (2020). The obser-vations, controls, and weights are corresponding to the parallel parts of the main analysis. To make sure ourresults do not depend on the parameters of the random forest model, we use a different number of trees ineach row of the table. i.e. RF100 means that a random forest with 100 trees is used.

46

Table A.3: Average Treatment Effects for Neighboring Cities by Distance

(1) (2) (3)

Distance Log(Openings) Log(Wage) Log(Firm Entry)

100km -0.321** -0.055*** -0.288***

(0.140) (0.011) (0.107)

200km -0.268* -0.017** -0.168***

(0.144) (0.009) (0.053)

300km -0.247** -0.011* -0.056

(0.102) (0.006) (0.040)

400km -0.187** -0.010* -0.027

(0.090) (0.006) (0.026)

500km -0.149* -0.007 -0.012

(0.079) (0.005) (0.020)

Notes: The aggregated treatment effects are calculated following Callaway and Sant’Anna (2020). The observa-tions, controls, and weights are corresponding to the parallel parts of the main analysis. Distances are betweencity centers and a non-pilot city within each distance of a pilot or a control city is defined as a neighboring city.

47

Appendix B Appendix Figures

48

Figure B.1: The Impacts on Job Postings in the Pilot Cities by TargetedIndustry

(a) Job Openings in IT (b) Monthly Wages in IT

(c) Job Openings in Science (d) Monthly Wages in Science

(e) Job Openings in Manufacturing (f) Monthly Wages in Manufacturing

(g) Job Openings in Business (h) Monthly Wages in Business

Notes: Figures present coefficients and 95% confidence intervals on the interaction be-tween the indicator of relative quarterly periods from the policy start date and the indi-cator of the pilot cities. City and calendar quarter fixed effects are controlled and eachobservation is weighted by the total number of postings in 2015.

49

Figure B.2: Leave One Pilot City Out: The Impacts on Job Postings in thePilot Cities

(a) Job Openings

(b) Monthly Wages

Notes: Figures present the impact of MIC25 on job postings in pilot cities with each timeleaving out one pilot city, indexed from 1 to 30. The solid horizontal line represents theimpact without leaving any city out. The dashed lines show the 95% confidence intervalsfor each estimation where one city is left out.

50

Figure B.3: Leave One Pilot City Out: The Impacts on Job Postings in thePilot Cities (Targeted Industries)

(a) Job Openings

(b) Monthly Wages

Notes: Figures present the impact of MIC25 on job postings in targeted industries in pilotcities with each time leaving out one pilot city, indexed from 1 to 30. The solid horizontalline represents the impact without leaving any city out. The dashed lines show the 95%confidence intervals for each estimation where one city is left out.

51

Figure B.4: Leave One Pilot City Out: The Impacts on Job Postings in thePilot Cities (Non-targeted Industries)

(a) Job Openings

(b) Monthly Wages

Notes: Figures present the impact of MIC25 on job postings in non-targeted industriesin pilot cities with each time leaving out one pilot city, indexed from 1 to 30. The solidhorizontal line represents the impact without leaving any city out. The dashed linesshow the 95% confidence intervals for each estimation where one city is left out.

52