Advertisement

Journal of Population Economics

, Volume 31, Issue 1, pp 235–266 | Cite as

Run away? Air pollution and emigration interests in China

  • Yu QinEmail author
  • Hongjia Zhu
Original Paper

Abstract

This paper investigates the impact of air pollution on people’s interest in emigration. Using an online search index on “emigration” which is positively correlated with its search volume, we develop a city-by-day measurement of people’s emigration sentiment. We find that searches on “emigration” will grow by approximately 2.3–4.8% the next day if today’s air quality index (AQI) is increased by 100 points. In addition, such an effect is more pronounced when the AQI level is above 200, a sign of “heavily polluted” and “severely polluted” days. We also find that such effect differs by destination countries and by metropolitan areas.

Keywords

Emigration Air pollution China Online searches 

JEL Classification

Q53 Q56 R23 

1 Introduction

International migration plays an important role in the globalization process, imposing substantial influences on the labor markets, productivity, and economic growth. A growing body of literature has documented the profound consequences of emigration on the origination country, such as the “brain drain effect”.1 Therefore, in addition to measuring the socioeconomic consequences of international migration, it is no less important to understand the driving forces of emigration from various perspectives. Many previous works have documented important factors motivating emigration, such as a lower level of development/income/wages in the source country (Docquier and Rapoport 2012; Hatton and Williamson 2002; Hunt 2006), higher income/wages in the destination countries (Karemera et al. 2000), increasing income per capita gap between the origin and destination countries (Ortega and Peri 2009), and more relaxed immigration law in the destination countries (Ortega and Peri 2009; Mayda 2010).

Despite the high economic growth rate in China in the past few decades, the total number of emigrants in China continues to rise (Fig. 1). As one of the largest suppliers of international migration, the total number of Chinese emigrants (above age 15) was 3.86 million in OECD countries in 2010–2011; thus, China ranked as the second largest supplier of emigrants after Mexico.2 It is reported that approximately one million Chinese have obtained permanent-resident status in Canada or America in the past decade, placing Chinese migrants first in Canada and second in America behind Mexicans.3 A large share of Chinese emigrants are high-skilled or high-income. Of these, 1.66 million of the 3.86 million Chinese emigrants in OECD countries are highly educated.4 In addition, a survey by Barclays Bank in 2014 found that 47% of rich Chinese plan to emigrate in the next 5 years, compared with 23% of Singaporeans and 16% of Hong Kongers. These facts suggest that China is encountering increasing challenges in human capital and wealth flight due to international emigration.5
Fig. 1

China’s GDP per capita and Overseas Emigration. Data source: Emigration data: United Nation Department of Economic and Social Affairs http://www.un.org/en/development/desa/population/migration/data/estimates2/estimatesorigin.shtml; GDP per capita data: the World Bank

Emigration is a long-term decision and is attributed to many factors, such as income, politics, education, and quality of life. As an important component of the quality of life, the environmental quality may affect the emigration decision. Although China has been experiencing high economic growth, it also has suffered from severe environmental degradation in recent years. For example, air pollution is now recognized as an increasing concern that affects China’s public health, industrial development, and economic growth (Brandt 2008). However, within the recent literature that studies international emigration, there is limited research that has linked emigration incentives with the environmental degradation in the origin country. The major objective of this paper is to investigate the contemporary association between air pollution and the interest in international emigration in China.

In this paper, we test the hypothesis that a very short-term (daily) shock on air pollution levels may drive up the interest in emigration in a city. No papers have ever established a causal link between pollution and emigration possibly due to the scarcity of data on international migration. In addition, because emigration takes time to process, it is difficult to trace it to the time when the emigration decision was made. However, this is an important question to ask because pollution is likely to motivate emigration due to its negative impacts on health and subjective well-being.6 In fact, anecdotal evidence in Hong Kong suggests that 26% of surveyed Hong Kong adults have considered emigration due to poor air quality. More importantly, those people who considered leaving were among the most competitive individuals in Hong Kong, including those with undergraduate and post-graduate degrees and high-income earners.7

To address the data scarcity and inaccessibility of international migration, we develop a proxy for people’s intent to migrate by collecting a city-by-day search index on emigration (“yi min” in Chinese) via Baidu, the largest Chinese search engine. The value of the Baidu Index is positively correlated with people’s search volume on the key word “emigration” in a city in a day; thus, it captures the contemporaneous aggregate interests on emigration in a city. We merge the search index with the air quality index (AQI) of 153 major Chinese cities in 2014. Our regression results show that the one-day lagged air pollution level significantly increases people’s searches on emigration. On average, a 100 point increase in the air pollution level one day before significantly leads to an approximately 2.5% increase in the current search index. In addition, the estimated impacts illustrate nonlinear patterns. The impacts become more pronounced when the air pollution level achieves “heavily polluted” (AQI above 200 and below 300) and “severely polluted” (AQI above 300). We find that the magnitude of the air pollution’s impact on searches for “emigration” is approximately one third of its impact on searches for “masks” and one tenth of its impact on searches for “smog” and “PM2.5,” possibly because the top income earners alone could afford the option to emigrate.

In addition, we find heterogeneous impacts of air pollution on searches for emigration with different destination countries. Among the top four destinations of Chinese emigrates, i.e., Australia, Canada, New Zealand, and the USA, the impact is most pronounced for emigration to the USA, which is the most popular destination of Chinese emigrants. A 100 points increase in air pollution significantly increases people’s searches on “emigration to the U.S.” by 2.6%. Moreover, we find heterogeneous impacts across the four largest metropolitan areas in China, including Beijing, Shanghai, Shenzhen, and Guangzhou. In particular, the impact is the largest and most significant in Beijing, which is one of the most polluted large cities in China. Last, to address the possible concern regarding the potential manipulation problem of the official data of China, we verify the causal link between air pollution and search behavior using the hourly PM2.5 data released by the US Embassy and Consulates in five cities. Similar results are found by using data from a different source.

This paper contributes to the literature in two ways. First, to the best of our knowledge, this is the first paper that investigates the causal link between pollution and emigration, which adds to the literature studying the determinants of emigration (Docquier and Rapoport 2012; Hatton and Williamson 2002; Karemera et al. 2000; Hunt 2006; Ortega and Peri 2009; Mayda 2010). Second, this paper also adds to a growing body of literature that investigates the economic impacts of air pollution from a new perspective of lost human capital. In addition to the substantial literature that studies the long-term and short-term health impacts of air pollution, certain recent literature finds that air quality has significant impacts on workers’ productivity (Graff Zivin and Neidell 2012; He et al. 2016) and academic outcomes (Currie et al. 2009; Stafford 2015); this suggests that the impacts of air pollution may be limited to its direct impacts not only on the health, but also on a various socioeconomic outcomes. This paper attempts to associate air pollution to a new perspective, the interest in emigration. Although emigration is a long-term decision, our findings indicate that severe air pollution in the short run may increase emigration interest in the population intending to emigrate, particularly for the population who has not yet decided to migrate.8

2 Data description

2.1 Search index

A novelty of this study is that we collect the information of the daily variation of the sentiment to emigrate in all of the prefecture cities in China via Baidu. The Baidu search index, similar to Google Trends (GT), provides a measurement of the search volume of a key word in a given time period. There is a growing body of economic literature that uses GT to measure the searching interest of Internet users. For example, in a recent study of evaluating the impacts of the MTV show on teen childbearing, Kearney and Levine (2015) use data from GT to measure levels of search interest in birth control and abortion as the outcome variables; Tefft (2011) uses GT on the terms “depression” and “anxiety” to proxy the intent to seek treatment for psychological distress; Goel et al. (2010) and Choi and Varian (2012) find that GT has considerable prediction power on economic outcomes such as macroeconomic indicators, product sales, and consumer behavior. Askitas and Zimmermann (2015) use GT to capture the impact of the 2008 Financial and Economic Crisis on well-being.9

China has experienced very rapid growth in its internet penetration rate. The number of internet users increased from 0.33 billion in 2009 to 0.63 billion in 2014, which is approximately twice the U.S. population.10 However, because Google services are blocked in China by the Great Fire Wall, the conventional GT data used in the previous literature cannot reflect the fluctuation of the real search volume of Chinese Internet users. Instead, Baidu, which was founded in 2000, is the largest search engine in China and has more than 50% of the market share.11 The Baidu Index is a data product analogous to GT, which measures the search frequencies of the selected terms. According to the official explanation, the Baidu Index provides a weighted sum of the search volume on a key word in a given period. We have no information regarding the specific formula it adopts to transform search volume into an index because it is a trade secret. However, the search index is very likely to be linearly correlated with the search volume of a key word based on the small sample of search volume and search index data that we obtained online (refer to Appendix Fig. 7).

Because international emigration is typically a long-term decision process, searching the related information regarding it on the internet can be interpreted as a revelation of emigration interest. Therefore, we collect daily Baidu Index data on the Chinese keyword “yi min” (emigration) in the 153 matched cities with the air pollution data in 2014. Fortunately, there is no ambiguity of the term “yi min” in Chinese. Moreover, this term primarily indicates the meaning of international emigration in a single-word context. Therefore, we believe that the Baidu Index for this specific term provides a credible measurement of the overall interest of a city’s international emigration. Figure 2 presents the spatial disparity of the average Baidu Index of “emigration” in our selected cities in 2014. Compared with the rest of the cities, Beijing, Guangzhou and Shanghai have the highest values, which is partially due to their large population size and higher development levels.
Fig. 2

Spatial distribution of the average Baidu Index of “Emigration” in 2014. Data source: Emigration Index is collected from Baidu; map is provided by China Data Online

To investigate the potential heterogeneity of emigration interests by destination countries, we collect key word searches for the largest four destination countries of Chinese emigrants, including Australia, Canada, New Zealand, and the USA. Specifically, we collect Baidu Index data for the Chinese keywords search on “emigration to Australia,” “emigration to Canada,” “emigration to New Zealand,” and “emigration to the USA.” In addition, we also test the potential impacts of air pollution on certain other searching interests directly associated with air pollution. To do this, we collect Baidu Index data on “PM2.5”, “smog” (wu mai) and “mask” (kou zhao) covering the same sample period as the outcome variables in the estimations. A few recent literature studies such as Mu and Zhang (2014) and Sun et al. (2017) use the online purchase data and find that there is significant impact of air pollution on the purchasing of masks. Because these online searches on pollutants, pollution and preventative measures are likely to be triggered by the pollution episodes, our purpose is to compare the potential impacts, if any, of air pollution on emigration and on these other related searches.

Table 1 provides the summary statistics for Baidu Index on all of the above keywords. The average index on “emigration” is approximately 72, with the maximum value as 1006, and the minimum as 0. There is no direct means to interpret the economic meaning of the numbers because we cannot precisely match them to the corresponding search volume. However, it is possible to compare the relative magnitude across different indices. Askitas and Zimmermann (2015) use the number of searches in “football” as a benchmark to get a better intuition about the size of the search volume for “symptoms.” In our paper, we may use a popular keyword “job-hunting” to replicate this procedure. According to Table 1, we find that in China, for every 100 job-hunting searches in 2014 we had about 25 for emigration. Compared to the overall emigration index, the country-specific emigration indices are lower on average. Among all of these four destination countries, he search index for emigration to New Zealand has the highest mean value, whereas the index is relatively lower for keyword searches on emigration to the USA. It is noteworthy that the differences on search frequency may be due to many reasons, such as the general interest in a specific country, the respective immigration policies, and other socioeconomic factors. In addition, it is found that on average, there are more searches on “PM2.5” and “smog” in Chinese cities compared with searches on “emigration.”
Table 1

Summary statistics

 

Description

Count

Mean

Sd

Min

Max

AQI

Air Quality Index (MEP data)

52515

94.712

54.883

12

500

emigration

Baidu Index of “emigration”

52515

72.917

72.204

0

1006

us_migrate

Baidu Index of “emigrate to the U.S.”

52515

26.858

43.611

0

811

can_migrate

Baidu Index of “emigrate to Canada”

52515

37.241

50.945

0

899

au_migrate

Baidu Index of “emigrate to Australia”

52515

34.128

47.280

0

353

nz_migrate

Baidu Index of “emigrate to New Zealand”

52515

42.705

70.954

0

3588

PM2.5

Baidu Index of “PM2.5

52515

142.148

407.695

0

44444

smog

Baidu Index of “smog”

52515

97.474

147.343

0

12162

mask

Baidu Index of “mask”

52515

38.406

51.883

0

1812

socks

Baidu Index of “socks”

52382

54.066

53.286

0

475

cloths

Baidu Index of “cloths”

52382

146.083

100.763

0

997

job-hunting

Baidu Index of “job-hunting”

52382

280.021

301.870

0

8445

wdsp

mean wind speed (knots)

52382

4.720

2.343

0

26.1

rain

rainfall amount (inches)

52382

0.103

0.389

0

15.4

maxt_c

maximum temperature (Celsius)

52382

20.751

10.451

−20.8

43.1

mint_c

minimum temperature (Celsius)

52382

11.063

11.139

−35

32

ctemp

mean temperature (Celsius)

52382

15.914

10.467

−27.4

36.6

cdewp

mean dew point (Celsius)

52382

8.881

12.256

−33.7

27.7

pm2.5_mep

PM2.5 concentration level (MEP data)

55291

62.540

49.318

0

1030

pm10_mep

PM10 concentration level (MEP data)

55291

106.490

74.966

0

2394

so2_mep

SO2 concentration level (MEP data)

55291

35.103

34.380

0

406

no2_mep

NO2 concentration level (MEP data)

55291

37.903

20.542

0

221

co_mep

CO concentration level (MEP data)

55291

1.235

0.703

0

12.61

o3_mep

O3 concentration level (MEP data)

55291

88.085

51.793

1

1080

pm2.5_us_mean

mean PM2.5 (U.S. Embassy data)

1798

71.368

55.357

5.0

558.3

pm2.5_us_max

maximum PM2.5 (U.S. Embassy data)

1798

123.314

89.251

11

725

pm2.5_us_mean_wh

mean PM2.5 during working hours (U.S. Embassy data)

1798

66.805

55.234

2.3

532.2

pm2.5_us_max_wh

maximum PM2.5 during working hours (U.S. Embassy data)

1798

95.983

74.216

5

630

Note: (1) The full dataset covers 153 prefecture cities which have daily AQI readings in 2014. (2) AQI data is from the Ministry of Environmental Protection in China. (3) All the Baidu indices are collected from Baidu. (4) Pollutant-specific concentration levels from the MEP side are purchased from Qingyue Open Environmental Data Center (https://data.epmap.org), which are sourced from the air quality real-time monitoring system of the MEP (http://106.37.208.233:20035). (5) PM2.5 concentration level from the U.S. side is downloaded from the website: http://www.stateair.net/web/historical/1/1.html

2.2 Air pollution data

We collect daily air pollution data from the Ministry of Environmental Protection (MEP) for the empirical analysis of this paper. Our data source is the same as certain previous studies related to air pollution in China, such as Viard and Fu (2015) and Mu and Zhang (2014). Since 2001, the MEP has published daily air pollution data on its website. Before 2012, the air pollution index (API), rather than air quality index (AQI), was reported on the website. API is a composite index measuring air quality based on city’s concentration levels of sulfur dioxide (SO2), nitrogen dioxide (NO2), and particulate matter 10 (PM10). A major difference between API and AQI is that the latter considers the concentration level of PM2.5, which is one of the major pollutants in many cities that have caused the public’s increasing concern in recent years. In addition, AQI also incorporates the concentration levels of ozone (O3) and carbon monoxide (CO) in the index.12

MEP began to report the new measurement of AQI on its website in 2014. However, there have been 153 cities publicizing their daily AQI since January of 2014, whereas the remainder of the cities was gradually included in later months. To construct a balanced panel for our analysis, we collect the daily AQI information solely in these 153 cities and then matched the data with the emigration indices. The core information of our dataset is the daily AQI measurement of each city in 2014, which is the average of the daily AQI readings from different monitor stations in each city. Additionally, the levels of air pollution are also provided according to the reported AQI. In China, air quality levels are classified into six categories: excellent, good, lightly polluted, moderately polluted, heavily polluted, and severely polluted, with corresponding AQI cutoff points at 50, 100, 150, 200, 300, and 500.13

In our data, the daily AQI readings range from 12 to 500, with significant spatial heterogeneity across different regions. Figure 3 presents the average AQI of the 153 cities in 2014. Air pollution is more severe in northern China. Beijing, for example, had only 28 days of good air quality (AQI < 50) in 2014, but had 14 heavily polluted days and 28 severely polluted days. In contrast, Shenzhen, a southern high-technological city and the home of over 18 million people, had 155 good air quality days and there was no day with AQI value higher than 150 in 2014.
Fig. 3

Spatial distribution of the average AQI in 2014. Data source: AQI data are from the Ministry of Environmental Protection in China; map is provided by China Data Online

Figure 4a shows the time series of the daily average AQI and Baidu Index of the 153 cities in our sample. Air pollution is most severe in winter (the beginning and the end of a calendar year). It is also interesting to see that Baidu Index has a stronger positive association with AQI in those heavily polluted days. Such association is stronger for cities in northern China, which are more polluted regions (Fig. 4b) but weaker for cities in less polluted southern China (Fig. 4c). We also note that Baidu Index on emigration peaks in August–September, which is probably due to increasing emigration sentiments in graduating seasons. This can also be reflected from Fig. 5a where we plot the daily average of AQI in 153 cities against the daily average of emigration indices in those cities. A negative relationship of these two variables is present without adjusting for seasonality and other common factors. However, the correlation between AQI and emigration index becomes positive after removing common factors, including weather variables, day-of-week fixed effects, public holiday dummy, month fixed effects, and city fixed effects, as shown in Fig. 5b. We will control for the above-mentioned common factors in regression analysis later.
Fig. 4

Time series of AQI and Baidu Index. Data source: Ministry of Environmental Protection and Baidu Index

Fig. 5

Daily average AQI and Emigration Index in 2014. Note: AQI data are from the Ministry of Environmental Protection in China, Emigration Index is collected from Baidu. In (a), we plot the daily average of AQI and Emigration Index in the 153 cities in 2014. To rule out the potential problem of spurious relationship, in (b), we first obtain the residuals of AQI and Baidu Index after regressing them on the same set of variables, which include weather variables, day-of-week fixed effects, public holiday fixed effects, month fixed effects, and city fixed effects. We then plot the daily average of residuals of AQI and Emigration Index

A potential concern of the official AQI data is that it may be manipulated by the government. As an effort for pollution abatement, the air quality of a city is now associated with the promotion of the local government officers. Therefore, this promotion opportunity may provide incentives for government officers to manipulate the reported air pollution data. Indeed, using the daily data of API, Chen et al. (2012) find that there is evidence of downward manipulation at the cutoff point at 100, which was the threshold of defining a “blue-sky day.” Although there is no recent evidence on whether the newly adopted AQI also suffers from data manipulation, we address this concern by providing evidence using another data source from the US Embassy and Consulates in five cities in China, which are not manipulated by the local government.

The five US Embassy and Consulates in China measure and publicize the hourly reading of PM2.5 in Beijing, Shanghai, Guangzhou, Chengdu, and Shenyang since 2008. We create four measures from the hourly reading of PM2.5 for these five cities, including the mean and the maximum of PM2.5 in the day and in working hours (9 am to 6 pm) and report the summary statistics of the four measures in Table 1. However, it should be noted that, due to different measurements, the readings from MEP and the US Embassy and Consulates are not directly comparable. In addition to the credibility of the data quality, another advantage of using the US data is that it allows us to study the response of the search index with respect to the air pollution in different time periods. A simple statistics of the hourly air quality shows that the air is typically more polluted during the night time than the day time. Because the night-time air pollution is less observable than the day-time pollution, it may potentially lead to heterogeneous effects on searching behavior. Therefore, using the hourly data, we are able to test whether the daily change of the search index is more sensitive to the pollution during working hours.

2.3 Weather data

Weather may also affect people’s intention to migrate. For example, people are more likely to be depressed on rainy days than sunny days; thus, they are more likely to be dissatisfied with their current living conditions. The short-term psychological effects of weather conditions have been examined in the finance literature using security market data (Saunders 1993; Hirshleifer et al. 2003). It is also worth noting that weather conditions have considerable effects on the local air quality, as discussed in Viard and Fu (2015) and Mu and Zhang (2014). If weather conditions are omitted from our regressions, the correlations between weather and local air quality as well as the interests of emigration will lead to biased estimations.

We address this problem by including a series of daily weather measurements in our regressions. Specifically, we control for daily temperature, dew point, wind speed, and precipitation during the entire year of 2014, the summary statistics of which are available in Table 1. The weather data are from the National Climatic Data Center under the US National Oceanic and Atmospheric Administration (NOAA), which provides rich daily weather information at the monitor station level. To match the daily weather data with air pollution and Baidu Index data, we search for the nearest weather station for each geographic city center.

3 Empirical strategies

The major objective of this paper is to examine the potential short-term effects of air pollution on the interest in emigration in China. To begin with, we specify the following estimation function:
$$ \text{Baidu}_{i,t} = \text{Constant} + \text{FE} + \sum\limits_{n=-2}^{6} \beta_{n} \text{AQI}_{i,t-n} + \lambda \text{Weather}_{i,t} + \epsilon_{i,t} $$
(1)

As described in the data section, we use the search frequency of the keyword “emigration” in Chinese as a proxy for people’s emigration sentiments. The daily AQI is the core independent variable to measure the air quality. However, because it is not clear whether air pollution has both contemporaneous and lagged effects on searching behavior, we include both the current day as well as six lagged days of air quality in Eq. 1. The vector of parameters β n represents the potential effects of air pollution on searching behavior in different periods. Additionally, we also include two leading days’ air pollution index as a falsification test. If our model is correctly specified, the current searches are expected not to be significantly affected by the air pollution index tomorrow and the day after. Moreover, to capture the potential effects of weather conditions on both the emigration interests and local air quality levels, we control for a list of weather variables including wind speed, rainfall, maximum temperature, minimum temperature, mean temperature, and dew points. We also include different sets of fixed effects in different specifications. In particular, in the most conservative regression form, we control for city-week fixed effects which rule out all the city-week level omitted variables, public holiday fixed effects, and day-of-week fixed effects. Additionally, we employ robust standard errors clustered at the city level.

It is noteworthy that the values of the Baidu Index on “emigration” are intrinsically non-negative count numbers. Figure 6 shows a typically skewed distribution of the Baidu Index, which has 32% of the values at 0. Therefore, the normal distribution assumption of error terms tends not to hold in the ordinary least square (OLS) regression if there is no transformation on the dependent variable, which may subsequently result in invalid statistical inferences. Instead, the Poisson regression model is an appropriate alternative estimation method in this context, which is also applied in Mu and Zhang (2014) where the online sales index of mask is used as the outcome variable.14 The interpretation of β n in the Poisson model is analogous to the log-linear model, which implies that every one unit increase of A Q I i,tn will lead to β n % change of the search frequencies in Eq. 1.
Fig. 6

Histogram of Baidu Index of “Emigration” in 2014

The estimation of Eq. 1 assumes that there is a linear effect of air pollution on people’s searching behavior. That is, every 100 increase in the AQI values at different levels has the same impact on search frequency. However, recent literature finds that the effects of air pollution display significant nonlinear features. In particular, the average impacts of air pollution at higher levels are usually disproportionately larger than the air pollution at lower levels. To test whether there is a nonlinear pattern of the impacts of air pollution in this study, we collapse the continuous measurement of AQI i,tn into six dummy indicators, AQI m , to denote the pollution levels as defined by the government Eq. 2.15 Therefore, the estimated coefficient of β m captures the partial effect of a day in the m th pollution level on the local search frequency of keyword “emigration.” If there is a nonlinear pattern of these effects, it is expected that the estimates of these coefficients show particular trends either in statistical significance or in magnitudes.
$$ \text{Baidu}_{i,t}~=~\text{Constant}~+~\text{FE}~+~\sum\limits_{m=2}^{6} \beta_{m} \text{AQI}^{m}_{i,t} + \lambda \text{Weather}_{i,t} + \epsilon_{i,t} $$
(2)

4 Main findings

4.1 Dynamic effects of air pollution on Baidu index

Because we do not have a priori evidence regarding the dynamic effects of air pollution on the change of search interest on “emigration,” finding the effective timing of air pollution on search behavior is fundamental to our analysis. To answer this key question, we use the econometric specification of Eq. 1 in Section 3.

Table 2 presents the dynamic regression results. As discussed in the previous section, we adopt the Poisson regression method and control for a rich set of fixed effects in the estimations. To facilitate the interpretation of the coefficient, we rescale the value of AQI by dividing the original AQI values by 100. In addition to controlling for day-of-week fixed effects, public holidays, and weather conditions, we use different alternative levels of city and time fixed effects across columns 1 to 3 to test the sensitivity of regression results. Specifically, in the first column, we control for month and prefecture city fixed effects to capture the unobservable factors in both the temporal and spatial dimensions. In the second column, we control for city-by-month fixed effect, which eliminates biases due to potentially heterogeneous time patterns in different cities. Last, the third column provides evidence using the most conservative regression form, i.e., a city-by-week fixed effect, which controls for time-variant city level variables potentially affecting online search activities, such as population.
Table 2

Dynamic effects of air pollution on Baidu Index of “Emigration”

Dependent variable: Baidu Index of “Emigration”

 

(1)

(2)

(3)

6-day lag of AQI/100

0.016

0.007

0.006

 

(0.009)

(0.008)

(0.008)

5-day lag of AQI/100

−0.001

−0.008

−0.003

 

(0.008)

(0.008)

(0.008)

4-day lag of AQI/100

0.013

0.007

0.005

 

(0.008)

(0.007)

(0.007)

3-day lag of AQI/100

0.002

−0.005

−0.008

 

(0.007)

(0.007)

(0.008)

2-day lag of AQI/100

0.009

−0.000

0.001

 

(0.008)

(0.007)

(0.008)

1-day lag of AQI/100

0.022∗∗∗

0.012

0.010

 

(0.007)

(0.007)

(0.008)

Current day AQI/100

0.019∗∗

0.009

0.003

 

(0.008)

(0.007)

(0.008)

1-day lead of AQI/100

−0.001

−0.007

−0.005

 

(0.007)

(0.008)

(0.007)

2-day lead of AQI/100

0.004

−0.004

−0.011

 

(0.008)

(0.007)

(0.008)

Month FE

Yes

No

No

City FE

Yes

No

No

City-month FE

No

Yes

No

City-week FE

No

No

Yes

Public Holiday FE

Yes

Yes

Yes

Day-of-Week FE

Yes

Yes

Yes

Weather

Yes

Yes

Yes

N

38,905

38,514

36,857

Notes: (1) AQI/100 is measured by the daily AQI readings divided by 100. (2) Weather variables, including wind speed, precipitation, minimum temperature, maximum temperature, average temperature, and dew point, have been controlled in all the regressions. (3) Public holiday fixed effect and day-of-week fixed effect have also been controlled in all the specifications. (4) The first specification controls for month and city fixed effect; the second specification allows different patterns of emigration searches in different months of the same city by controlling for city-month fixed effect; the last specification further controls for city-week fixed effect to eliminate the omitted variables at the city-week level. (5) *** p ≤ 0.01; ∗∗ p ≤ 0.05; ∗p ≤ 0.1. (6) Robust standard errors clustered at the city level are reported in parentheses

The results in Table 2 suggest a positive impact of the lagged-one value of the AQI on online searches for “emigration.” The positive coefficients in columns 1 and 2 are 0.022 and 0.012, which indicates that a 100 point increase in the AQI leads to an approximately 1.2 to 2.2% growth in the Baidu Index one day after. The coefficient on the one-day lag of AQI is positive but marginally insignificant though in column 3 after controlling for city-week fixed effect, which is probably due to insufficient variation within such small cells. In addition, although the AQI has a significant and positive contemporary effect on the Baidu Index in the estimation of column 1, the effect becomes statistically insignificant in our preferred estimation in columns 2 and 3, which controls for the variations within city-month and city-week units, respectively. Last but not least, none of the coefficients of the two AQI leads is significant, which mitigates the concern of model mis-specification. In summary, according to the findings above, we conclude that the variation of the Baidu Index are primarily affected by the lagged one day AQI values; this nominates the lagged one day AQI as the key independent variable in our subsequent analysis.

Table 3 presents the results of regressing the Baidu Index on the lagged one day value of AQI. Again, different combinations of fixed effects are controlled for across three columns. The results suggest a positive impact of lagged air pollution on online searches for “emigration.” Specifically, a 100 point increase in the AQI level leads to an approximate 4.8% growth in terms of the Baidu Index on “emigration” the next day, as indicated in column 1. The magnitude of effect decreases to approximately 2.5% after imposing stronger fixed effects in column 2 and 3, but it remains statistically significant.
Table 3

Impact of air pollution on Baidu Index of “Emigration”

Dependent variable: Baidu Index of “Emigration”

 

(1)

(2)

(3)

1-day lag of AQI/100

0.048∗∗∗

0.023∗∗∗

0.025∗∗∗

 

(0.009)

(0.005)

(0.005)

Month FE

Yes

No

No

City FE

Yes

No

No

City-month FE

No

Yes

No

City-week FE

No

No

Yes

Public holiday FE

Yes

Yes

Yes

Day-of-week FE

Yes

Yes

Yes

Weather

Yes

Yes

Yes

N

51,763

51,941

49,671

Notes: (1) AQI/100 is measured by the daily AQI readings divided by 100. (2) Weather variables, including wind speed, precipitation, minimum temperature, maximum temperature, average temperature, and dew point, have been controlled in all the regressions. (3) Public holiday fixed effect and day-of-week fixed effect have also been controlled in all the specifications. (4) The first specification controls for month and city fixed effect; the second specification allows different patterns of emigration searches in different months of the same city by controlling for city-month fixed effect; the last specification further controls for city-week fixed effect to eliminate the omitted variables at the city-week level. (5) *** p ≤ 0.01; ** p ≤ 0.05; * p ≤ 0.1. (6) Robust standard errors clustered at the city level are reported in parentheses

Overall, our findings above indicate a short-term effect of air pollution on the international emigration interest in China. These findings are in accordance with recent evidence from Zhang et al. (2015), who find that the air pollution has immediate negative effects on shorter-term hedonic happiness. However, because we do not have data on real immigration, the evidence we have provided can be interpreted as a pulse response of emigration interest because of air pollution episodes.

4.2 Nonlinear effects

Certain previous studies find that the impacts of air pollution tend to be nonlinear (He et al. 2016; Currie et al. 2009; Graff Zivin and Neidell 2012; Schlenker and Walker 2015). In this section, we will test whether there is any nonlinear pattern of the estimated effects of air pollution on the Baidu Index. To do so, we categorize the AQI index into six groups based on the guidance by the Ministry of Environmental Protection of China: “excellent” (AQI ⊂[0,50]); “good” (AQI ⊂[51,100]); “lightly polluted” (AQI ⊂[101,150]); “moderately polluted” (AQI ⊂[151,200]); “heavily polluted” (AQI ⊂[201,300]); and “severely polluted”(AQI ⊂[301,500]).16 We convert the daily AQI value into the corresponding dummy variable. To avoid the perfect multi-collinearity problem in the regression, the dummy for “good” is absorbed as the baseline group.

Table 4 reports the regression results for the nonlinear effect of air pollution on the Baidu Index on “emigration.” Interestingly, the estimated coefficient increases in magnitude as AQI achieves higher levels. The impact of air pollution on emigration searches is most significant after AQI exceeds 200 (corresponding to “heavily polluted” and “severely polluted”). A “heavily polluted” day with AQI between 201 to 300 leads to approximately 5.5 to 11.1% growth in the emigration index on the subsequent day. Moreover, a “severely polluted” day with an AQI exceeding 300 leads to approximately 7.9 to 12.8% growth in the Baidu search index on emigration, which is approximately two times as high as the effects of “heavily polluted” days and six times as high as the effects of “moderately polluted” days.17
Table 4

Impact of air pollution on Baidu Index of “Emigration”: nonlinear effects

Dependent variable: Baidu Index of “Emigration”

 

(1)

(2)

(3)

1-day lag of AQI ⊂[51,100]

−0.001

0.008

0.008

 

(0.009)

(0.008)

(0.008)

1-day lag of AQI ⊂[101,150]

0.011

0.013

0.007

 

(0.011)

(0.008)

(0.009)

1-day lag of AQI ⊂[151,200]

0.040∗∗

0.019

0.015

 

(0.016)

(0.011)

(0.012)

1-day lag of AQI ⊂[201,300]

0.111∗∗∗

0.055∗∗∗

0.057∗∗∗

 

(0.023)

(0.015)

(0.015)

1-day lag of AQI ⊂[301,500]

0.128∗∗∗

0.079∗∗∗

0.096∗∗∗

 

(0.030)

(0.025)

(0.023)

Month FE

Yes

No

No

City FE

Yes

No

No

City-month FE

No

Yes

No

City-week FE

No

No

Yes

Public holiday FE

Yes

Yes

Yes

Day-of-week FE

Yes

Yes

Yes

Weather

Yes

Yes

Yes

N

51763

51941

49671

Notes: (1) 1-day lag of AQI ⊂[0,50] is the default category for all the regressions. (2) Weather variables, including wind speed, precipitation, minimum temperature, maximum temperature, average temperature, and dew point, have been controlled in all the regressions. (3) Public holiday fixed effect and day-of-week fixed effect have also been controlled in all the specifications. (4) The first specification controls for month and city fixed effect; the second specification allows different patterns of emigration searches in different months of the same city by controlling for city-month fixed effect; the last specification further controls for city-week fixed effect to eliminate the omitted variables at the city-week level. (5) *** p ≤ 0.01; ** p ≤ 0.05; * p ≤ 0.1. (6) Robust standard errors clustered at the city level are reported in parentheses

5 Discussions

5.1 Using air pollution data from the US Embassy

There is concern raised in the previous literature regarding the accuracy of official AQI data in China. For example, Chen et al. (2012) use data from the air pollution index (API) from 2000 to 2009 and find that, as a result of local government officers’ promotion incentives, the air pollution data may be manipulated to achieve the number of “blue sky days” in a year. Such manipulation typically affects the API values that are slightly above 100, which is the threshold for defining a “blue sky day.” However, because our results suggest that the AQI level does not lead to increased emigration searches until it hits 200, the manipulation for “blue sky” is not likely to affect our main findings.

To completely eliminate the possibility of biased results due to data manipulation, we also collect the PM2.5 data reported by the US Embassy in Beijing and its Consulates in four cities, including Shanghai, Guangzhou, Shenyang, and Chengdu. A salient feature of the US data is that it provides the hourly readings of the PM2.5 concentration level, thus allowing us to investigate the effects of pollution within different time periods. Therefore, we create four outcome variables using the hourly data: the mean and maximum values of the PM2.5 level during working hours (9 am to 6 pm) and the whole day, respectively.

Similar to the previous analysis, we conduct regressions by using the continuous measurement of air pollution as well as exploring the nonlinear patterns of it.18 However, it is noteworthy that although the PM2.5 is the major pollutant in certain cities, the PM2.5 data cannot be directly compared with the AQI data in levels because AQI is a composite index composed of six different pollutants. Therefore, it is not sensible to create dummy variables for PM2.5 levels based on the classification method of AQI levels. To mitigate the concern on the comparability of AQI and PM2.5 categories, we adopt the cutoffs of PM2.5 from the Technical Regulation on Ambient Air Quality Index published by the Ministry of Environmental Protection in 2012. Such cutoffs are used to define the individual air quality index (IAQI), i.e., the major input variables to calculate the AQI. Thus, we create dummy variables for cutoff points of PM2.5 at 35, 75, 115, 150, 250, and 350, which results in seven dummy variables (including the omitted category) that represent the pollution levels.19

Table 5 reports the findings using the continuous PM2.5 data in four various measurements. To save space, we only report the regression results using the finest city-by-week fixed effect for each outcome variable. we find that the one-day lag of average level of PM2.5 during working hours significantly affects the search index. Moreover, the nonlinear effects found in the AQI data also appear when the PM2.5 data are used. As shown in Table 6, the impact of lagged PM2.5 on the emigration index is most pronounced when PM2.5 achieves 350. The lagged one day PM2.5 value higher than 350, on average, leads to approximately 5.3 to 13.8% growth in the Baidu Index of “emigration.” Additionally, we also show the heterogeneous effects of different pollution level definitions. Overall, the dummy variables of the highest level of pollution category are all statistically significant across all the four columns. Particularly, the effects of the mean and maximum pollution levels during working hours are generally larger than the effects of the mean and maximum pollution levels of the entire day, suggesting that people may be more affected by the pollution during the time that they will interact in a day.20
Table 5

Impact of air pollution on Baidu Index of “Emigration” using U.S. embassy data

 

Dependent variable: Baidu Index of “Emigration”

 

(1)

(2)

(3)

(4)

PM2.5 definitions

Max

Working hour max

Mean

Working hour mean

1-day lag of PM2.5/100

0.003

0.010

0.012

0.014

 

(0.006)

(0.006)

(0.009)

(0.008)

City-week FE

Yes

Yes

Yes

Yes

Public Holiday FE

Yes

Yes

Yes

Yes

Day-of-Week FE

Yes

Yes

Yes

Yes

Weather

Yes

Yes

Yes

Yes

N

1798

1790

1798

1790

Notes: (1) 1-day lag of AQI ⊂[0,50] is the default category for all the regressions. (2) Weather variables, including wind speed, precipitation, minimum temperature, maximum temperature, average temperature and dew point, have been controlled in all the regressions. (3) City-week fixed effect, public holiday fixed effect and day-of-week fixed effect have also been controlled in all the specifications. (4) *** p ≤ 0.01; ** p ≤ 0.05; * p ≤ 0.1. (5) Robust standard errors are reported in parentheses

Table 6

Impact of air pollution on Baidu Index of “Emigration” using U.S. embassy data: nonlinear effect

 

Dependent variable: Baidu Index of “Emigration”

 

(1)

(2)

(3)

(4)

PM2.5 Definitions

Max

Working hour max

Mean

Working hour mean

1-day lag of P M 2.5 ⊂[36,75]

0.006

0.009

0.005

0.017

 

(0.015)

(0.011)

(0.011)

(0.010)

1-day lag of P M 2.5 ⊂[76,115]

−0.007

0.002

0.008

0.003

 

(0.016)

(0.013)

(0.015)

(0.014)

1-day lag of P M 2.5 ⊂[116,150]

0.012

0.008

−0.054∗∗∗

-0.022

 

(0.019)

(0.016)

(0.018)

(0.017)

1-day lag of P M 2.5 ⊂[151,250]

−0.030

−0.005

0.006

0.004

 

(0.018)

(0.015)

(0.020)

(0.019)

1-day lag of P M 2.5 ⊂[251,350]

0.002

0.008

0.064

0.043

 

(0.025)

(0.029)

(0.034)

(0.030)

1-day lag of P M 2.5 ⊂[351,500]

0.053∗∗

0.104∗∗∗

0.110∗∗

0.138∗∗

 

(0.027)

(0.029)

(0.054)

(0.054)

City-week FE

Yes

Yes

Yes

Yes

Public Holiday FE

Yes

Yes

Yes

Yes

Day-of-Week FE

Yes

Yes

Yes

Yes

Weather

Yes

Yes

Yes

Yes

N

1798

1790

1798

1790

Notes: (1) P M 2.5 ⊂[0,35] is the default category for all the regressions. (2) Weather variables, including wind speed, precipitation, minimum temperature, maximum temperature, average temperature and dew point, have been controlled in all the regressions. (3) City-week fixed effect, public holiday fixed effect and day-of-week fixed effect have also been controlled in all the specifications. (4) *** p ≤ 0.01; ∗∗ p ≤ 0.05; ∗p ≤ 0.1. (5) Robust standard errors are reported in parentheses

5.2 Heterogeneous effects by country of destination

Our main results suggest that a higher level of air pollution leads to more searches on emigration. In this section, we take a further step to investigate the potential heterogeneous effects of air pollution by different destination countries of emigration. Specifically, we use the Chinese keyword searches on “emigration to Australia,” “emigration to Canada,” “emigration to New Zealand,” and “emigration to the USA” as outcome variables. We select these four countries since they are the top choices for Chinese emigrants.21

Table 7 reports the results for the four popular destination countries. We only report the results of regression specifications with city-by-week fixed effects. Interestingly, the impact of air pollution on searches for emigration to the USA is the most significant among all of these four countries. A 100 point increase in the lagged AQI leads to a 2.6% growth in the Baidu Index on “emigration to the USA.” Table 8 presents the nonlinear effect of air pollution on emigration searches by destination countries. Again, the effect is more pronounced for searches on “emigration to the USA” when AQI hits 300. One explanation for these results is that the USA is the most popular destination country for Chinese emigrants potentially due to higher income, cleaner air, lower housing prices and better education. As suggested by the statistics, in 2012, 81,784 Chinese emigrants obtained permanent residence in the USA, compared with 33,018 Chinese emigrants in Canada, 25,509 in Australia, and 7223 in New Zealand.22
Table 7

Impact of air pollution on Baidu Index of “Emigration” by destination countries

Dependent variable: Baidu Index of “Emigration” by destination countries

 

US

Canada

Australia

New Zealand

 

(1)

(2)

(3)

(4)

1-day lag of AQI/100

0.026∗∗

0.003

0.005

-0.004

 

(0.012)

(0.008)

(0.009)

(0.009)

City-Week FE

Yes

Yes

Yes

Yes

Public Holiday FE

Yes

Yes

Yes

Yes

Day-of-Week FE

Yes

Yes

Yes

Yes

Weather

Yes

Yes

Yes

Yes

N

35558

40069

40025

41981

Notes: (1) AQI/100 is measured by the daily AQI readings divided by 100. (2) Weather variables, including wind speed, precipitation, minimum temperature, maximum temperature, average temperature and dew point, have been controlled in all the regressions. (3) Public holiday fixed effect and day-of-week fixed effect have also been controlled in all the specifications. (4) The first specification controls for month and city fixed effect; the second specification allows different patterns of emigration searches in different months of the same city by controlling for city-month fixed effect; the last specification further controls for city-week fixed effect to eliminate the omitted variables at the city-week level. (5) *** p ≤ 0.01; ∗∗ p ≤ 0.05; ∗p ≤ 0.1. (6) Robust standard errors clustered at the city level are reported in parentheses

Table 8

Impact of Air Pollution on Baidu Index of “Emigration” by Destination Countries: Nonlinear Effect

Dependent Variable: Baidu Index of “Emigration”

 

US

Canada

Australia

New Zealand

 

(1)

(2)

(3)

(4)

1-day lag of AQI ⊂[51,100]

−0.039∗∗

0.004

0.019

−0.004

 

(0.018)

(0.014)

(0.014)

(0.018)

1-day lag of AQI ⊂[101,150]

−0.028

0.017

0.031

0.032

 

(0.022)

(0.018)

(0.017)

(0.022)

1-day lag of AQI ⊂[151,200]

−0.002

0.012

0.019

−0.003

 

(0.027)

(0.021)

(0.020)

(0.023)

1-day lag of AQI ⊂[201,300]

0.052

0.002

0.001

−0.025

 

(0.032)

(0.023)

(0.024)

(0.025)

1-day lag of AQI ⊂[301,500]

0.093

−0.003

0.012

0.002

 

(0.055)

(0.040)

(0.044)

(0.043)

City-Week FE

Yes

Yes

Yes

Yes

Public Holiday FE

Yes

Yes

Yes

Yes

Day-of-Week FE

Yes

Yes

Yes

Yes

Weather

Yes

Yes

Yes

Yes

N

35558

40069

40025

41981

Notes: 1. 1-day lag of AQI ⊂[0,50] is the default category for all the regressions. 2. Weather variables, including wind speed, precipitation, minimum temperature, maximum temperature, average temperature and dew point. 3. City-week fixed effect, public holiday fixed effect and day-of-week fixed effect have also been controlled in all the specifications. 4. *** p ≤ 0.01; ∗∗ p ≤ 0.05; ∗p ≤ 0.1. 5. Robust standard errors clustered at the city level are reported in parentheses

5.3 Effect in four largest metropolitan areas

In this subsection, we test the effect of air pollution on emigration searches in the four largest metropolitan areas in China, namely Beijing, Shanghai, Guangzhou, and Shenzhen. These cities host large populations of highly educated, high income and wealthy individuals. For each city, we include a full set of control variables (including weather, day-of-week fixed effect and holiday dummy) and the weekly fixed effect.

Table 9 presents the regression results on these four respective cities. Interestingly, the impact is most significant in Beijing but primarily muted in other three cities. A 100 point increase in the lagged AQI level leads to an approximately 2.9% increase in emigration searches in Beijing. Such an effect is particularly large when the AQI level achieves “heavily polluted.” As shown in Table 10, the search index on “emigration” in Beijing grows by approximately 12.7% when AQI achieves 300. A possible explanation of such heterogeneity is that Beijing is on average more polluted than the other three cities. The average AQI in Beijing is 127 in 2014, whereas the number is approximately 80 in Shanghai and Guangzhou in the same period and 55 in Shenzhen.
Table 9

Impact of Air Pollution on Baidu Index of “Emigration” by Cities

Dependent Variable: Baidu Index of “Emigration”

 

Beijing

Shanghai

Guangzhou

Shenzhen

 

(1)

(2)

(3)

(4)

1-day lag of AQI/100

0.029∗∗

−0.025

0.078

−0.015

 

(0.011)

(0.017)

(0.049)

(0.049)

Week FE

Yes

Yes

Yes

Yes

Public Holiday FE

Yes

Yes

Yes

Yes

Day-of-Week FE

Yes

Yes

Yes

Yes

Weather

Yes

Yes

Yes

Yes

N

341

340

335

335

Notes: 1. AQI/100 is measured by the daily AQI readings divided by 100. 2. Weather variables, including wind speed, precipitation, minimum temperature, maximum temperature, average temperature and dew point, have been controlled in all the regressions. 3. Week fixed effect, public holiday fixed effect and day-of-week fixed effect have also been controlled in all the specifications. 4. *** p ≤ 0.01; ∗∗ p ≤ 0.05; ∗p ≤ 0.1. 5. Robust standard errors are reported in parentheses

Table 10

Impact of Air Pollution on Baidu Index of “Emigration” by Cities: Nonlinear Effect

Dependent Variable: Baidu Index of “Emigration”

 

Beijing

Shanghai

Guangzhou

Shenzhen

 

(1)

(2)

(3)

(4)

1-day lag of AQI ⊂[51,100]

−0.018

0.017

0.036

−0.007

 

(0.026)

(0.017)

(0.031)

(0.017)

1-day lag of AQI ⊂[101,150]

−0.017

0.016

0.070

−0.006

 

(0.033)

(0.019)

(0.044)

(0.047)

1-day lag of AQI ⊂[151,200]

0.023

−0.028

0.186∗∗

 
 

(0.033)

(0.028)

(0.085)

 

1-day lag of AQI ⊂[201,300]

−0.004

−0.032

  
 

(0.036)

(0.070)

  

1-day lag of AQI ⊂[301,500]

0.127∗∗∗

   
 

(0.046)

   

Week FE

Yes

Yes

Yes

Yes

Public Holiday FE

Yes

Yes

Yes

Yes

Day-of-Week FE

Yes

Yes

Yes

Yes

Weather

Yes

Yes

Yes

Yes

N

341

340

335

335

Notes: 1. 1-day lag of AQI ⊂[0,50] is the default category for all the regressions. 2. Weather variables, including wind speed, precipitation, minimum temperature, maximum temperature, average temperature and dew point, have been controlled in all the regressions. 3. Week fixed effect, public holiday fixed effect and day-of-week fixed effect have also been controlled in all the specifications. 4. *** p ≤ 0.01; ∗∗ p ≤ 0.05; ∗p ≤ 0.1. 5. Robust standard errors are reported in parentheses

5.4 Magnitude

To gain a better understanding of the magnitude of air pollution’s impact on emigration sentiment, we compare the impact of air pollution on emigration with its impact on other pollution-related keyword searches. Specifically, we select the Baidu Index of keyword searches on “PM2.5,” “smog,” and “mask” in Chinese as the outcome variables. We apply the same regression methods and present the findings in Table 11 (panels 1–3). In summary, a higher pollution level leads to significantly more searches for all three keywords the next day. A 100 point increase in the AQI level leads to 19.5–27.2% more searches on “smog,” 23.3–32.1 % more searches on “PM2.5,” and 10.3–12.7% more searches on “mask.”
Table 11

Impact of Air Pollution on Other Keyword Searches

Dependent Variable: Baidu Index

 

(1)

(2)

(3)

Panel 1: Keyword: Smog

1-day lag of AQI/100

0.272∗∗∗

0.253∗∗∗

0.195∗∗∗

 

(0.049)

(0.041)

(0.040)

N

51763

51945

50329

Panel 2: Keyword: P M 2.5

1-day lag of AQI/100

0.321∗∗∗

0.292∗∗∗

0.233∗∗∗

 

(0.068)

(0.058)

(0.047)

N

51763

52161

51615

Panel 3: Keyword: Mask

1-day lag of AQI/100

0.127∗∗∗

0.124∗∗∗

0.103∗∗∗

 

(0.029)

(0.025)

(0.023)

N

51763

50671

42248

Panel 4: Keyword: Socks

1-day lag of AQI/100

−0.012

0.005

0.009

 

(0.007)

(0.006)

(0.007)

N

51763

51671

48005

Panel 5: Keyword: Clothes

1-day lag of AQI/100

-0.002

0.000

−0.001

 

(0.005)

(0.003)

(0.002)

N

51763

52211

52135

Panel 6: Keyword: Job-hunting

1-day lag of AQI/100

0.001

−0.005

0.004

 

(0.006)

(0.005)

(0.003)

N

51763

52239

52158

Month FE

Yes

No

No

City FE

Yes

No

No

City-month FE

No

Yes

No

City-week FE

No

No

Yes

Public Holiday FE

Yes

Yes

Yes

Day-of-Week FE

Yes

Yes

Yes

Weather

Yes

Yes

Yes

Notes: 1. AQI/100 is measured by the daily AQI readings divided by 100. 2. Weather variables, including wind speed, precipitation, minimum temperature, maximum temperature, average temperature and dew point, have been controlled in all the regressions. 3. Public holiday fixed effect and day-of-week fixed effect have also been controlled in all the specifications. 4. The first specification controls for month and city fixed effect; the second specification allows different patterns of emigration searches in different months of the same city by controlling for city-month fixed effect; the last specification further controls for city-week fixed effect to eliminate the omitted variables at the city-week level. 5. *** p ≤ 0.01; ∗∗ p ≤ 0.05; ∗p ≤ 0.1. 6. Robust standard errors clustered at the city level are reported in respective columns

As all of those three keywords above are directly related to the air pollution, it is not surprising that we find positive effects of air pollution on the search frequency of these keywords. However, one may worry that people spend more time indoor in hazy days which lead to more online searches for everything. To mitigate such concern, it is necessary to carry out falsification tests using random keywords to make sure that the effect of air pollution on emigration searches is not due to the overall increase of online searches in polluted days. Therefore, we select the following keywords: socks, clothes, and job-hunting, which are neutral and less subject to seasonality of online searches. The keyword “socks” is also used in Sun et al. (2017) as a falsification test to understand the effect of pollution on online shopping of masks and air filters. We report these results in Table 11 (panels 4–6). None of the coefficients is statistically significant at the 0.05 level which suggests that the rises in emigration related searches are not likely to be driven by the overall increases of online search activities in polluted days.

In comparison, the impact on searches for “emigration” is approximately 2.3–4.7%, which is approximately one third of the impact of searching on “mask” and approximately one tenth of the impact of searching on “smog” and “PM2.5.” It appears that the air pollution effect is much smaller on emigration than keyword searches on PM2.5 and smog. However, this is an economically reasonable magnitude because a very small proportion of the overall population may consider emigration.23 In contrast, the majority of the population tends to obtain additional information on pollution as well as related self-protective measures as a rational response to air pollution shocks. Our finding is related to the study by Mu and Zhang (2014), who find that air pollution increases people’s purchases of masks as self-protection. However, the relatively smaller coefficient of air pollution on the Baidu Index on “emigration” than other terms does not imply that the real effect is negligible. In contrast, it is worth noting that people who consider emigration as an option are likely to be much wealthier and more educated compared with the general population. Thus, the economic consequences of emigration could be important and need further study in future research.

5.5 Controlling for anti-corruption campaign

Another concern on our estimation regards about the existence of other confounders such as the anti-corruption campaigns. China conducted waves of anti-corruption campaigns in 2014, targeting corrupt leaders from high ranks to low ranks. Therefore, such campaigns may motivate the corrupted leaders and their relatives to emigrate as the last option. Recent study by Schneider (2015) also suggest that corruption encourages high-skilled migrants to emigrate. To tease out the effect of an anti-corruption campaign on emigration searches, we collect the Baidu Index search data on “anti-corruption” (fan fu in Chinese) and use it as a control variable in the main regressions. Table 12 reports the impact of the lagged AQI on the emigration index using both the continuous measure and the dummies for AQI classifications. The results remain very similar to Tables 3 and 4, indicating that our results are robust to the exclusion of this significant movement.
Table 12

Robustness Check: Controlling for Searches on Anti-Corruption Campaign

Dependent Variable: Baidu Index of “Emigration”

 

(1)

(2)

(3)

 

(4)

(5)

(6)

1-day lag of

0.047∗∗∗

0.023∗∗∗

0.025∗∗∗

1-day lag of

−0.002

0.008

0.007

AQI/100

   

AQI ⊂[51,100]

   
 

(0.009)

(0.005)

(0.005)

 

(0.009)

(0.008)

(0.008)

    

1-day lag of

0.010

0.013

0.007

    

AQI ⊂[101,150]

   
     

(0.011)

(0.008)

(0.009)

    

1-day lag of

0.037∗∗

0.018

0.015

    

AQI ⊂[151,200]

   
     

(0.016)

(0.011)

(0.012)

    

1-day lag of

0.108∗∗∗

0.054∗∗∗

0.057∗∗∗

    

AQI ⊂[201,300]

   
     

(0.023)

(0.015)

(0.015)

    

1-day lag of

0.125∗∗∗

0.077∗∗∗

0.096∗∗∗

    

AQI ⊂[301,500]

   
     

(0.029)

(0.024)

(0.023)

Month FE

Yes

No

No

 

Yes

No

No

City FE

Yes

No

No

 

Yes

No

No

City-month FE

No

Yes

No

 

No

Yes

No

City-week FE

No

No

Yes

 

No

No

Yes

Public

Yes

Yes

Yes

 

Yes

Yes

Yes

Holiday FE

       

Day-of-Week FE

Yes

Yes

Yes

 

Yes

Yes

Yes

Weather

Yes

Yes

Yes

 

Yes

Yes

Yes

N

51739

51911

49635

 

51739

51911

49635

Notes: 1. The left panel replicates the three columns in Table 3 with the inclusion of search index on anti-corruption campaigns. 2. The right panel replicates the three columns in Table 3 with the inclusion of search index on anti-corruption campaigns. 3. *** p ≤ 0.01; ∗∗ p ≤ 0.05; ∗p ≤ 0.1. 4. Robust standard errors clustered at the city level are reported in respective columns in two panels

6 Conclusion

In this paper, we study the impact of air pollution on people’s interest in international emigration in China. Using the Baidu search index on emigration-related Chinese keywords as the measurement of overall expressed interest, we establish a causal relationship between air pollution and the interest in emigration at the city-by-day level. Specifically, we find that the searches on “emigration” will grow by approximately 2.3–4.8% the next day if the AQI today is raised by 100 points, which suggests that air pollution increases people’s interest in emigration. In addition, this effect is more pronounced when the AQI level is above 200, which indicates “heavily polluted” and “severely polluted” days. We also find that air pollution’s impact on the emigration interest differs by destination countries. Among all of the top four destination countries, the impact is most significant for the USA, which is the number one destination country of Chinese emigrants. Moreover, the effect is large and significant in Beijing but not in the other three largest metropolitan areas, i.e., Shanghai, Guangzhou, and Shenzhen, which are possibly driven by the low pollution levels in these cities.

Emigration is a long-term decision. However, our findings indicate that severe air pollution in the short run may switch on people’s interest in emigration, particularly for the marginal population who are not yet determined to emigrate. Because there is no literature documenting the relationship between the interest in emigration and the real behavior of emigration, we cannot estimate the impact of pollution on the number of new emigrants, which is a major caveat of this paper. We leave it for future research and await new data collection on emigration behaviors. However, given that a positive correlation between online searching behavior and a final decision usually exists, the policy makers may consider the potential human capital and wealth flight due to emigration in the cost benefit analysis of pollution abatement.

Footnotes

  1. 1.

    Regarding whether the net “brain drain effect” is detrimental or beneficial to the source country remains controversial. For example, Vidal (1998) suggests that emigration to a higher return to skills country may encourage people in the source country to invest in human capital; Chen (2006) argues that the relaxation of restrictions on the emigration of high-skilled workers will damage the economic growth of a source country in the long run, although a “brain gain” may happen in the short run. Beine et al. (2008) find that most countries combining low levels of human capital and low migration rates of skilled workers tend to be positively affected by the brain drain, whereas the brain drain appears to have negative growth effects in countries where the migration rate of the highly educated is above 20% and/or where the proportion of people with higher education is above 5%. Agrawal et al. (2011) indicate that the emigration of highly skilled individuals weakens local knowledge networks (brain drain) but may help remaining innovators access valuable knowledge accumulated abroad (brain bank).

  2. 2.
  3. 3.
  4. 4.
  5. 5.

    Another example is that Chinese buyers spent more than 221 billion U.S. Dollars on property in the U.S. alone, between April 2013 and March 2014. See http://www.rfa.org/english/news/china/flood-02122015104709.html

  6. 6.

    A growing body of literature studies the health impacts of air pollution. For example, Chay and Greenstone (2003) and Currie and Walker (2011) estimate the significant effects of air pollution on the infant mortality rate, premature births, and low birth weight using the U.S. data. Schlenker and Walker (2015) focus on a shorter time span of the impacts and show the contemporaneous health impacts of air pollution for various population cohorts. Using China’s data, Chen et al. (2013) find that the higher concentration levels of total suspended particulate (TSP) due to the winter heating policy in north China is responsible for approximately 5.5 years of lower life expectancy. Zhang et al. (2015) show that air pollution significantly reduces short-term hedonic happiness.

  7. 7.
  8. 8.

    We assume that emigration searches by people who have planned to move are less likely to be affected by temporary shocks, such as air pollution.

  9. 9.

    Although we do not use GT for our empirical analysis in this paper, it is worth noting that GT and Baidu Index have a very high correlation (0.84 and 0.78, respectively) when searching universities and companies in China as suggested by Vaughan and Chen (2015). In this paper, we assume that the search index algorithm is similar between Baidu Index and GT since Baidu does not publicize its methodology on constructing the Index.

  10. 10.
  11. 11.
  12. 12.

    We also collect data for six pollutants, namely PM2.5, SO2, NO2, PM10, CO, and O3. However, we still use the AQI as the major measurement of air quality in our analysis, because it scientifically considers the concentration levels of different pollutants. More importantly, in our data, we find that PM2.5 only accounts for about 45% of polluting days as a major pollutant. The other pollutants, such as PM10, O3, and NO2, accounts for about 27, 19, and 3.5% of polluting days as major pollutants, respectively. Therefore, pollutant-specific concentrations may not be representative enough as a measurement for overall air quality. Please refer to http://www.cnemc.cn/publish/106/news/news_25941.html for the new Ambient air quality standard (GB3095-2012).

  13. 13.

    Please refer to Table 2 of the Technical Regulation on Ambient Air Quality Index available at http://210.72.1.216:8080/gzaqi/Document/aqijsgd.pdf.

  14. 14.

    In our analysis, we find that the results from OLS regressions and Poisson regressions all point to similar conclusions.

  15. 15.

    To avoid the perfect multicollinearity problem in the regression, we omit the first category A Q I 1 in Eq. 2.

  16. 16.

    Please refer to Table 2 of the Technical Regulation on Ambient Air Quality Index available at http://210.72.1.216:8080/gzaqi/Document/aqijsgd.pdf.

  17. 17.

    The impact of AQI on people’s emigration decision could be either physiological or behavioral or both. We try to disentangle these two mechanisms by conducting regression discontinuity style regressions at the cutoffs of various AQI levels to examine whether search intensity changes significantly right above and below the cutoff points. We find that the estimated coefficient is sensitive to the econometric setting of the regression discontinuity design, therefore we do not find robust evidence for the potential behavioral effects at the cutoffs. The results are available upon request.

  18. 18.

    We cannot cluster the standard errors at the city level as in previous regressions because there are only five cities with PM2.5 measurement from the U.S. Embassy and the Consulates. Therefore, we adopt robust standard errors for these regressions.

  19. 19.

    Please refer to Table 1 of the Technical Regulation on Ambient Air Quality Index available at http://210.72.1.216:8080/gzaqi/Document/aqijsgd.pdf.

  20. 20.

    We check the robustness of our results using the concentration of each pollutant, including PM2.5, PM10, SO2, NO2, CO, and O3. The results are reported in Appendix Table 13. We find that the results are largely consistent with our main results in terms of both significance and magnitude. It should be noted that the coefficient on CO is much larger than the rest of the pollutants, because the mean value of the CO concentration is much lower than the other pollutants as shown in Table 1.

  21. 21.
  22. 22.
  23. 23.

    The accumulated number of emigrants is approximately 9.34 million by the end of 2013, of a total population of approximately 1.4 billion.

Notes

Acknowledgement

We are grateful to Sumit Agarwal, Yongheng Deng, Shuaizhang Feng, Shihe Fu, Xiaobo Zhang, Klaus F. Zimmermann (the Editor), two anonymous reviewers, and seminar participants at the National University of Singapore for their valuable comments and suggestions. Thanks to Qingyue Open Environmental Data Center (https://data.epmap.org) for support on Environmental data processing. Qin acknowledges financial support from the Academic Research Fund - Tier 1 (WBS: R-297-000-129-133). Zhu acknowledges financial support from the National Natural Science Foundation of China (Project 71603103). All remaining errors are ours.

Compliance with ethical standards

Conflict of interest

The authors declare that they have no conflict of interest.

Funding

This study was funded by Ministry of Education, Singapore (grant number R-297-000-129-133) and the National Natural Science Foundation of China (grant number 71603103).

References

  1. Agrawal A, Kapur D, McHale J, Oettl A (2011) Brain drain or brain bank? the impact of skilled emigration on poor-country innovation. J Urban Econ 69 (1):43–55CrossRefGoogle Scholar
  2. Askitas N, Zimmermann KF (2015) Health and well-being in the great recession. Int J Manpow 36(1): 26–47CrossRefGoogle Scholar
  3. Beine M, Docquier F, Rapoport H (2008) Brain drain and human capital formation in developing countries: Winners and losers*. Econ J 118(528):631–652CrossRefGoogle Scholar
  4. Brandt L, Rawski TG (2008) China’s great economic transformation. Cambridge university pressGoogle Scholar
  5. Chay KY, Greenstone M (2003) The impact of air pollution on infant mortality: Evidence from geographic variation in pollution shocks induced by a recession. Q J Econ 118(3):1121–1167CrossRefGoogle Scholar
  6. Chen H-J (2006) International migration and economic growth: a source country perspective. J Popul Econ 19(4):725–748CrossRefGoogle Scholar
  7. Chen Y, Jin GZ, Kumar N, Shi G (2012) Gaming in air pollution data? lessons from China. BE J Econ Anal Policy 12(3)Google Scholar
  8. Chen Y, Ebenstein A, Greenstone M, Li H (2013) Evidence on the impact of sustained exposure to air pollution on life expectancy from China’s huai river policy. Proc Natl Acad Sci 110(32):12936–12941CrossRefGoogle Scholar
  9. Choi H, Varian H (2012) Predicting the present with google trends. Econ Record 88(s1):2–9CrossRefGoogle Scholar
  10. Currie J, Walker R (2011) Traffic congestion and infant health: Evidence from e-zpass. Am Econ J Appl Econ 3(1):65–90CrossRefGoogle Scholar
  11. Currie J, Hanushek EA, Kahn EM, Neidell M, Rivkin SG (2009) Does pollution increase school absences? Rev Econ Stat 91(4):682–694CrossRefGoogle Scholar
  12. Docquier F, Rapoport H (2012) Globalization, brain drain, and development. J Econ Lit, 681–730Google Scholar
  13. Goel S, Hofman JM, Lahaie S, Pennock DM, Watts DJ (2010) Predicting consumer behavior with web search. Proc Natl Acad Sci 107(41):17486–17490CrossRefGoogle Scholar
  14. Graff Zivin J, Neidell M (2012) The impact of pollution on worker productivity. Am Econ Rev 102(7):3652–73CrossRefGoogle Scholar
  15. Hatton TJ, Williamson JG (2002) What fundamentals drive world migration? Technical report, National Bureau of Economic ResearchGoogle Scholar
  16. He J, Liu H, Salvo A (2016) Severe air pollution and labor productivity: Evidence from industrial towns in china. IZA Discussion Paper No. 8916Google Scholar
  17. Hirshleifer D, Shumway T, et al (2003) Good day sunshine: Stock returns and the weather. J Financ, 58(3)Google Scholar
  18. Hunt J (2006) Staunching emigration from east Germany: Age and the determinants of migration. J Eur Econ Assoc 4(5):1014–1037CrossRefGoogle Scholar
  19. Karemera D, Oguledo VI, Davis B (2000) A gravity model analysis of international migration to north america. Appl Econ 32(13):1745–1755CrossRefGoogle Scholar
  20. Kearney MS, Levine PB (2015) Media influences on social outcomes: the impact of mtv’s 16 and pregnant on teen childbearing. Am Econ Rev 105(12):3597–3632CrossRefGoogle Scholar
  21. Mayda AM (2010) International migration: a panel data analysis of the determinants of bilateral flows. J Popul Econ 23(4):1249–1274CrossRefGoogle Scholar
  22. Mu Q, Zhang J (2014) Air pollution and defensive expenditures: Evidence from particulate-filtering facemasks. Available at SSRN 2518032Google Scholar
  23. Ortega F, Peri G (2009) The causes and effects of international migrations: Evidence from oecd countries 1980-2005. Technical report, National Bureau of Economic ResearchGoogle Scholar
  24. Saunders EM (1993) Stock prices and wall street weather. Am Econ Rev, 1337–1345Google Scholar
  25. Schneider F (2015) Does corruption promote emigration? IZA World of LaborGoogle Scholar
  26. Schlenker W, Walker WR (2015) Airports, air pollution, and contemporaneous health. Rev Econ Stud 83(2):768–809CrossRefGoogle Scholar
  27. Stafford TM (2015) Indoor air quality and academic performance. J Environ Econ Manag 70:34–50CrossRefGoogle Scholar
  28. Sun C, Kahn ME, Zheng S (2017) Self-protection investment exacerbates air pollution exposure inequality in urban China. Ecol Econ 131:468–474CrossRefGoogle Scholar
  29. Tefft N (2011) Insights on unemployment, unemployment insurance, and mental health. J Health Econ 30(2):258–264CrossRefGoogle Scholar
  30. Vaughan L, Chen Y (2015) Data mining from web search queries: a comparison of google trends and baidu index. J Assoc Inf Sci Technol 66(1):13–22CrossRefGoogle Scholar
  31. Viard VB, Fu S (2015) The effect of beijing’s driving restrictions on pollution and economic activity. J Public Econ 125:98–115CrossRefGoogle Scholar
  32. Vidal J-P (1998) The effect of emigration on human capital formation. J Popul Econ 11(4):589–600CrossRefGoogle Scholar
  33. Zhang X, Zhang X, Chen X (2015) Happiness in the air: How does a dirty sky affect subjective well-being? IZA Discussion Paper No. 9312.Google Scholar

Copyright information

© Springer-Verlag Berlin Heidelberg 2017

Authors and Affiliations

  1. 1.Department of Real EstateNational University of SingaporeSingaporeSingapore
  2. 2.Institute for Economic and Social ResearchJinan UniversityGuangzhouChina

Personalised recommendations