Firm Innovation in Hazy Days: Chinese Evidence

Abstract

This paper estimates the causal effect of air pollution on firm innovation based on a nationwide sample of Chinese manufacturing firms. Using thermal inversion as instrument variable (IV), we demonstrate that air pollution significantly hurts firm innovation performance: a 1% increase in air pollution (PM_2.5) in the past year, from its mean value, amounts to a 1.5% decrease in the number of invention patents compared to the average between 1999 and 2013. We also provide the casual estimate using regression discontinuity design (RDD) by exploiting the exogenous variation in pollution level caused by China’s Huai River policy: the decrease in firms’ invention patents is approximately 2.2% in the regions north of the Huai River Line. The underlying mechanism analyses reveal that poor air quality has detrimental effects on labor productivity, firms’ human capital accumulation, and financing ability. These findings highlight the substantial costs associated with air pollution, which are greater than previously thought, as it hinders innovation, a crucial driver of long-term economic growth.

1 Introduction

Innovation plays a crucial role in driving economic growth and ensuring the long-term success of firms (Freeman and Soete 1997; Lundvall 1992; Porter 1990; Schumpeter and Backhaus 2003). Not surprisingly, a large body of literature has tried to pinpoint firm innovation determinants from various dimensions, including, but not limited to the following: firm characteristics (e.g., corporate insider characteristics (Balsmeier et al. 2017; Chang et al. 2015; Galasso and Simcoe 2011; Hirshleifer et al. 2012)), analyst coverage (He and Tian 2013), institutional ownership (Aghion et al. 2013; Brav et al. 2018; Chemmanur and Tian 2018; Guadalupe et al. 2012), and stakeholders (Flammer and Kacperczyk 2016)), market factors (e.g., product market competition (Aghion et al. 2005; Autor et al. 2020; Yung 2016), investment cycle (Nanda and Rhodes-Kropf 2013, 2017), and financial sector development (Amore et al. 2013; Benfratello et al. 2008)), and social and economic forces (e.g., legal systems (Acharya et al. 2014; Fang et al. 2017; Williams 2013) and investor protection (Brown et al. 2013; Levine et al. 2017)).

While previous studies have enhanced our comprehension of the determinants of firm innovation, there is a notable lack of exploration into the influence of the natural environment, an essential component of firms’ operational contexts, on innovation performance. This research seeks to address this gap by investigating the impact of air pollution on firm innovation within the Chinese context.

Poor air quality has a detrimental impact on firm innovation for several reasons. Firstly, it hampers labor productivity, a crucial component of the innovation process (Chen et al. 2016; Hall 2002). Scientific literature extensively documents how air pollutants can adversely affect both physical and mental health by impairing lung and cardiovascular functions, blood flow, and circulation (Pope III 2000; Pope III et al. 2009; Seaton et al. 1995). Building on this evidence, various studies have demonstrated that air pollution can decrease worker productivity across physically demanding and highly skilled roles (Chang et al. 2016, 2019; Fu et al. 2021; Graff Zivin and Neidell 2012), a phenomenon that may also apply to inventors (Kapoor and Lim 2007). Secondly, air pollution presents challenges for firms in attracting and retaining skilled labor, including inventors. A recent study by Xue et al. (2021) reveals that companies located in areas with high air pollution levels are more likely to lose key employees who relocate to cleaner areas. Moreover, air pollution can constrain the financial resources available for innovative endeavors within firms. This is because air pollution escalates the economic costs associated with waste disposal, particularly for industries subject to stringent environmental regulations, potentially diverting funds away from innovation investments.

China’s well-documented air pollution issue makes it an ideal location for our research inquiry. Between 1999 and 2013, the average annual PM_2.5 concentration at the county level in China rose from 40.19 μg/m³ to 67.44 μg/m³, marking a 68% increase. Even in recent years, China’s PM_2.5 levels have remained approximately four times higher than those of most developed economies.¹ This unique situation allows us to explore the impact of severe air pollution. Furthermore, a patenting surge occurred in China during our sample period (Hu et al. 2017), which effectively contributes to our identification strategy.

However, a significant challenge in assessing the impact of air pollution lies in addressing potential endogeneity issues, as air pollutants are often linked to local economic activities that may also influence firms’ innovation performance. To mitigate this challenge, we primarily utilize an instrumental variable (IV) method. In this study, thermal inversion, a meteorological phenomenon unaffected by economic activities, is employed as the IV for air pollution, as previously done in studies by Arceo et al. (2016), Chen et al. (2022), Deschênes et al. (2020), and Fu et al. (2021). Thermal inversion serves as a valid instrument for air pollution for two reasons: Firstly, during a thermal inversion, warmer air moves above cooler air in the atmosphere, which is contrary to the normal pattern where temperature decreases with higher altitude. Thus, thermal inversions lead to an increase in air pollution as it restricts the diffusion of soot, air pollutants, and condensations. Secondly, thermal inversion is a complex meteorological phenomenon that occurs due to certain weather conditions, making it completely exogenous to local economic activities. Additionally, our paper utilizes the regression discontinuity design (RDD) approach by leveraging the exogenous variation in PM_2.5 levels resulting from China’s Huai River winter heating policy to assess the causal effect of air pollution, as explored by Chen et al. (2013), Ebenstein et al. (2017), and Ito and Zhang (2020).

Focusing on a nationwide sample of manufacturing firms, our two-stage least squares (2SLS) estimator indicates that air pollution has a significantly negative impact on firms’ invention patents, which serve as our primary measure of firm innovation. Specifically, a 1% increase in air pollution (PM_2.5) in the previous year, relative to its mean value, leads to a 1.5% decrease in the number of invention patents compared to the average value. Throughout 1998–2013, the rise in average PM_2.5 concentration results in a total loss of 3,281 invention patents for the sample firms, averaging 234 patents per year.² Our study’s results are robust across various tests, including robustness checks where we modify the independent variable, adjust for clustering standard errors and fixed effects, consider environmental policies, account for firms’ sorting decisions, and explore alternative patent types and air pollutants. These measures enhance the reliability and validity of our conclusions. Furthermore, consistent findings are observed in a different research context using an RDD approach, revealing a 2.2% decrease in invention patents of firms located north of the Huai River Line.

After establishing the impact of air pollution on firms’ innovation performance, we delve into the underlying mechanisms. Our initial focus is on the effect of air pollution on labor productivity, revealing a notable decline in output per worker in the presence of high levels of air pollutants. To further investigate whether air pollution impedes innovation by affecting firms’ ability to retain skilled employees, we examine the proportion of highly educated staff and the heterogeneity in industry concentration levels across different regions. The findings indicate that air pollution indeed hinders the firms’ accumulation of human capital, with industries characterized by low regional concentration experiencing more pronounced losses in innovation performance. Subsequently, we explore the hypothesis that air pollution diminishes firms’ innovation performance by constraining their financial resources allocated for innovation. As anticipated, our analysis demonstrates that local air pollution raises firms’ economic costs related to emitting pollutants, leading to greater innovation setbacks in polluting industries subjected to stringent government regulations.

Moreover, we investigate the non-linear impact of air pollution, in line with existing literature suggesting that the adverse effects intensify as air quality worsens (Chen et al. 2020; Schlenker and Walker 2016). Our results validate this pattern, showing that the detrimental effect of air pollution on firms’ innovation performance escalates disproportionately with deteriorating air quality.

Our study contributes to the literature in the following ways. First, to the best of our knowledge, this study is the first to comprehensively and rigorously investigate the impact of air pollution on micro-individuals’ (firms’) innovation, contributing to our knowledge of the factors influencing firms’ innovation outcomes. In addition to firm characteristics, market factors, and social and economic forces, environmental factors like air pollution may play a significant role in shaping innovation performance. While our research stands out for multiple dimensions, it is important to compare relevant literature within a narrow field. For instance, Xue et al. (2021) studies the effect of air pollution on firms’ human capital accumulation in China. We shift the focus from human capital to innovation performance in both the short and long term, thereby extending the research chain. Recently, Cui et al. (2023) discussed the effect of air quality on innovation activities at the city level in China, which is the closest to ours. However, our study offers unique contributions. First, our main specification and identification source differ from Cui et al. (2023), as we analyze the lag effect of air pollution on innovation by studying its impact in the previous year. It implies different interpretations of the coefficient of interest. Second, we further uncover the long-term (cumulative) effects of air pollution on firm innovation through an RDD analysis based on the Huai River heating policy in China. Finally, we investigate the influencing channels, which is highly important since the link between air pollution and economic outcomes such as innovation is not that clear. Overall, our paper contributes to deepening a more comprehensive understanding of how inventors respond to air pollution.

Second, our study contributes to the existing literature on the socioeconomic impacts of air pollution by examining its effects on firms’ innovation performance. While previous studies have primarily concentrated on productivity (both labor and capital) and human capital (such as health status, including physical and mental health), our findings highlight that the indirect costs of air pollution may be greater than previously believed. This is significant as innovation plays a crucial role in driving long-term economic development.

Finally, this paper highlights the significant policy implications of understanding the impact of air pollution on economic outcomes, particularly in developing countries. It underscores the importance of implementing anti-air pollution policies, such as those in China, and emphasizes the need to include innovation loss in further cost–benefit analyses for environmental policies. Additionally, the paper offers valuable insights for policymakers on promoting innovation as a key driver of long-term economic growth, especially in developing nations.

The following sections outline the structure of this paper: Section 2 provides details on the data and sample used. Section 3 explains the primary empirical design of the study. Section 4 discusses the results obtained. Section 5 delves into the mechanism analysis and further study. Finally, Section 6 offers concluding remarks.

2 Data and Sample

2.1 Firm-level Data

The State Intellectual Property Office of China (SIPO) has established an extensive patent database that includes information on all patents granted since the implementation of China’s Patent Law in 1985. This database offers detailed insights into inventors, distinguishing between individuals and firms, providing the name of the patent applicant, the type of patent, and the year of patent approval. The SIPO has been a valuable resource for numerous studies examining innovation activities in China (Fang et al. 2020; Hu et al. 2017; Liu et al. 2021). There are three main types of patents granted by the SIPO in China: invention patents, utility patents, and design patents. Invention patents in China represent novel technical proposals for products, methods, or improvements, possessing the highest technical complexity among all patents (see Article 2 in Patent Law of China),³ akin to utility patents in the United States. Utility patents in China encompass new technical solutions for product shape, structure, or their combination, designed for practical applications (see Article 2 in Patent Law of China). Design patents in China involve new designs of shape, pattern, or color combinations that are visually appealing and suitable for industrial use (see Article 2 in Patent Law of China).

In this study, we specifically examine invention patents due to their reputation for being the most innovative and technologically advanced, as highlighted in prior research by Cheng et al. (2019), Rong et al. (2017), and Wei et al. (2017).⁴ To address the lack of information on the firms behind these patents, we utilize firm-level data from the annual surveys of industrial firms (ASIF) conducted by the National Bureau of Statistics (NBS) in China. The surveys, carried out between 1998 and 2013, encompassed non-state-owned enterprises (non-SOEs) with annual sales exceeding 5 million CNY and all state-owned enterprises (SOEs). By matching the patent dataset with ASIF at the firm level, we are able to access firm characteristics such as book assets, industrial output value, total liability, total profits, sewage charges, main business income, location, employee information (including number of employees and educational background), and industry affiliation.

2.2 Air Pollution Data

To assess air pollution, PM_2.5 is commonly used as it has more severe impacts on human health compared to other pollutants. This fine particulate matter can deeply penetrate the body, leading to significant health risks, as highlighted by Anderson et al. (2012), Hill et al. (2023), and Southerland et al. (2022). Data on PM_2.5 is obtained through satellite-based Aerosol Optical Depth (AOD) retrievals, managed by NASA.⁵ These data have been employed in recent air pollution studies (Chen et al. 2022; Deschênes et al. 2020; Fu et al. 2021). Unlike ground-based pollution data limited to major cities, AOD data covers the entire country, reported monthly since 1980 at a detailed 50 × 60-km grid, spanning our study period.⁶ Furthermore, satellite-derived air pollution data are considered more trustworthy than ground-based data due to their immunity to human interference, a concern raised by Andrews (2008) and Ghanem and Zhang (2014). Following Buchard et al. (2016), we aggregate monthly grid-level PM_2.5 to the county level (2,839 counties in total) and calculate the annual average as our metric for yearly PM_2.5 levels. This county-level data is then matched with firm characteristics based on their geographical locations.

2.3 Thermal Inversion Data

As mentioned earlier, we instrument PM_2.5 with thermal inversion. The data is sourced from the MERRA-2 database managed by NASA, offering air temperature details across 42 atmospheric layers spanning from 110 to 36,000 m. This data is refreshed every six hours and is accessible for each 50 × 60-km grid. County-level data is derived by aggregating the grid-level information within each layer. Thermal inversion is characterized by higher air temperature in the upper layer compared to the lower layer every six hours. Next, we construct our IV by counting the total number of days where a thermal inversion occurs at least once every six hours in each county. Subsequently, the county-level data is integrated with firm-specific attributes based on the geographical location of each firm.

2.4 Weather Data

To consider the potential impact of local weather conditions on both air quality and firms’ innovation performance, we include variables such as precipitation, temperature, relative humidity, wind speed, sunshine duration, and air pressure. These variables are sourced from the National Meteorological Information Center of China and gathered at the station level. Following Deschênes and Greenstone (2011), we utilize the inverse distance weighting method with a 200-km radius to convert these variables from the station level to the county level. Subsequently, we compute the yearly averages of wind speed, sunshine duration, air pressure, and relative humidity, with the exception of precipitation and temperature. Instead of annual averages, we calculate the cumulative precipitation within a year (Deschênes et al. 2020; Fu et al. 2021). Regarding temperature, in line with Fu et al. (2021), we create 8 temperature bins (1 below 0°Celsius, 1 above 30°Celsius, and 6 in between with a 5°Celsius interval) to capture the potential non-linear impact of temperature. The count of days in which the average daily temperature falls within each bin is then tallied on an annual basis for each county. Finally, we link these county-level weather data with firm-level patent data based on firm locations.

2.5 Sample and Summary Statistics

After merging the above data sources, a final sample of 539,400 unique manufacturing firms in 2,783 counties across China from 1999 to 2013 is obtained, resulting in 2,643,178 firm-year observations.⁷ Table 1 provides detailed definitions of the main variables, while Table 2 presents their summary statistics. In Panel A of Table 2, the focus is on firm-level characteristics, with the mean of the primary measure of firm innovation, invention patents, being 0.19. It is worth noting that a considerable number of firm years report zero invention patents, as indicated by the median value of 0. This observation aligns with prior research indicating that innovation, especially in the form of invention patents with high novelty and technological inventiveness, is rare (Brav et al. 2018; He and Tian 2013). Moreover, the average firm in the sample has a logged total asset of 9.98, and around 14.6% of employees are highly educated.⁸ Panel B displays the summary statistics of air quality and weather-related variables at the county level. For instance, the main measure of air pollution, PM_2.5, has an average of 57.8 μg/m³. In terms of the instrument, thermal inversion, a county experiences an average of 110 days with at least one inversion per year, with a standard deviation of 61.

Table 1 Variable definitions

Table 2 Summary statistics

3 Empirical Strategy

To assess the influence of air pollution on firm innovation, our main econometric model is outlined as follows:

$${Innovation}_{ict}={\beta }_{0}+{\beta }_{1}{Air Pollution}_{c(t-1)}+f\left({W}_{c(t-1)}\right)+{X}_{ic(t-1)}+{\theta }_{i}+{\mu }_{jt}+{\varepsilon }_{ict}$$

(1)

where i, j, c and t denote firm, industry, county and year, respectively. ${Innovation}_{ict}$ indicates the logarithm of one plus invention patents granted of firm i in county c in year t. ${Air Pollution}_{c(t-1)}$ represents the logarithm of one plus the PM_2.5 levels from the previous year in county c (where firm i is situated). ${W}_{c(t-1)}$ is a vector containing weather variables in county c (where firm i is situated) in the previous year, such as temperature (in bins), precipitation, relative humidity, wind speed, sunshine duration, and air pressure. ${X}_{ic(t-1)}$ includes firm-specific controls that could positively or negatively impact innovation performance, such as firm size, leverage rate, and return on assets. ${\theta }_{i}$ accounts for firm fixed effects, capturing time-invariant attributes that influence firms’ innovation performance. We also incorporate industry-year fixed effects (${\mu }_{jt}$) to consider industry-specific trends, recognizing that different industries may be affected by air pollution in varying ways and have distinct innovation cycles. Additionally, we cluster the standard error at the county level. The coefficient of interest in the baseline model, ${\beta }_{1}$, represents the lagged effect of air pollution on firms’ innovation performance over one year. A significantly negative value of ${\beta }_{1}$ would indicate that air pollution has a detrimental impact on firms’ innovation performance.

However, the validity of the OLS estimation can be undermined due to the endogenous nature of air pollution, even with the use of a lagged term. This is because air pollutants are often a result of economic activities, and more developed regions tend to attract productive and skilled workers, as well as have better financial resources that can drive innovation. Nevertheless, if these regions also release more pollutants into the air, the OLS estimator may be biased upward, potentially exceeding zero. To address this issue and accurately measure the causal impact of air pollution on firm innovation, we utilize an IV approach. Consistent with previous studies (Arceo et al. 2016; Chen et al. 2022; Deschênes et al. 2020; Fu et al. 2021), we use thermal inversion as our instrument for air pollution. Figure 1 illustrates the time trend of air pollution and thermal inversion from 1990 to 2013. The annual occurrences of thermal inversion do not show any significant time trend, unlike the clear upward trend of air pollution (PM_2.5 and SO₂), indicating the exogeneity of our instrument variable. Figure 2 illustrates the yearly average air pollution trends from 1990–2013 among three groups categorized by their IV percentile: counties with IV below the 33rd percentile, between the 33rd and 66th percentile, and above the 66th percentile. The data indicates that air pollution tends to be more pronounced in regions with higher thermal inversion percentiles, thus confirming the validity of the IV methodology. Consequently, we employ 2SLS with the first-stage equation:

$${Air Pollution}_{c(t-1)}={\alpha }_{0}+{\alpha }_{1}{Thermal Inversion}_{c(t-1)}+f\left({W}_{c(t-1)}\right)+{X}_{ic(t-1)}+{\theta }_{i}+{\mu }_{jt}+{\sigma }_{ict}$$

(2)

where ${Thermal Inversion}_{c(t-1)}$ denotes the annual number of days with at least one thermal inversion occurrence in county c (where firm i is located) in the previous year. The weather controls in the second stage equation are incorporated to satisfy the exclusion restriction of the IV.

Fig. 1

Time trend of air pollution and thermal inversion (1990–2013). Notes: This figure shows the time trend of yearly average air pollution level (PM_2.5 and SO₂) and thermal inversion in China from 1990 to 2013. The vertical reference line represents year 1998

Fig. 2

Time trend of air pollution by occurrences of thermal inversion (1990–2013). Notes: This figure shows the time trend of yearly average air pollution level (PM_2.5 and SO₂) in China from 1990 to 2013 among three groups categorized by the number of days with thermal inversion: counties below the 33rd percentile, between the 33rd and 66th percentile, and above the 66th percentile. The vertical reference line represents year 1998

4 Results

4.1 Baseline Results

We first evaluate the validity of the instrument. The results in Table 3 indicate a strong first-stage relationship from the IV regressions. The coefficient of thermal inversion is statistically significant at the 1% level. When focusing on PM_2.5 as the primary measure of air pollution, one standard deviation increase in thermal inversion results in a notable 0.01 standard deviation increase in logged PM_2.5. Additionally, the KP F-statistic (120.36) greatly exceeds the critical value of 16.38 (Hausman 1978), suggesting that thermal inversion is a strong instrument. Consistent results are obtained when SO₂ is used as the measure of air pollution. Prior to presenting the results from the second-stage IV regressions, initial estimates are provided without accounting for any endogeneity bias for comparative analysis. The first column of Table 4 showcases the OLS estimates derived from Eq. (1), revealing a statistically significant negative coefficient of -0.0175. While the OLS findings imply that air pollutants impede firms’ innovation performance, their impact is attenuated by the endogenous nature of air pollution. This aligns with the results observed in the second-stage IV regressions. Furthermore, we conduct the heteroskedasticity robust Wu-Hausman test to assess the necessity of employing an IV approach, as suggested by Hausman (1978). The p-value displayed in Table 4 leads us to reject the null hypothesis that the explanatory variable of interest (air pollution) is exogenous.⁹

Table 3 First stage regression results

Table 4 The impact of air pollution on invention patents

We first present the significantly negative coefficient of interest without any controls in Column 3 of Table 4. In our preferred baseline regression, the result in Column 4 (accounting for firm size, leverage rate, return on assets, and other weather variables) indicates that the coefficient for the one-year lag of PM_2.5 is -0.2442, which is statistically significant at the 1% level.¹⁰¹¹ Economically, a 1% increase in air pollution (PM_2.5) in the previous year, compared to its mean value, leads to a 1.5% decrease in the number of invention patents compared to the average between 1999 and 2013. This highlights the significant economic impact of air pollution on firms’ innovation performance.¹² Over the study period, the average annual PM_2.5 concentration increased by 27.25 μg/m³, climbing from 40.19 μg/m³ in 1999 to 67.44 μg/m³ in 2013. Consequently, the total number of invention patents for the firms in our sample declined by around 3,281, with an average annual loss of 234 patents.¹³

To further validate our baseline findings, we conduct placebo tests to analyze the impact of one or two-year ahead air pollution (PM_2.5 and SO₂) on firms’ innovation performance. If our baseline results are unlikely to arise by chance, we would anticipate that future air pollution levels do not detrimentally affect current innovation performance. As expected, the results in Table 5 indicate that the coefficients of interest are statistically insignificant and notably smaller in magnitude compared to the baseline. Overall, our key finding is strongly reinforced by the results presented in both Tables 4 and 5, demonstrating that heightened air pollution levels have a detrimental effect on firms’ innovation performance, a finding that holds both statistical and economic significance.

Table 5 Placebo tests: The impact of one and two-year ahead air pollution on invention patents

4.2 Robustness Checks

4.2.1 Robustness Check I: Alternative Independent Variable, Clustered S.E., Controls, and Fixed Effects

To validate our primary conclusion that air pollution adversely affects firms’ innovation performance, we conducted various rigorous tests. By analyzing the impact of current, two-year lag, and cumulative air pollution on firms' innovation performance, we aim to assess both short-term and longer-term responses. The findings presented in Columns 1–2 of Table 6 demonstrate a notable negative influence of both the current and two-year lag air pollution on firms’ innovation performance. Given a potential argument that the detriment of air pollution to people may be a lasting and accumulated process, we then switch to studying the cumulative impact of air pollution. Specifically, we aggregate the average PM_2.5 concentration over the past three years using three weighting methods: exponential moving average, inverse weighted, and moving average. We apply these methods to the current period, as well as one-year and two-year lags of yearly average PM_2.5. The findings, presented in Columns 3–5, reveal that accumulated air pollution has a notably negative effect on firms’ innovation performance as well. The coefficient of cumulative air pollution is almost double the baseline result, indicating a higher level of harm. To address serial correlation within each county-year unit, standard errors are clustered at the county-year level in Column 6, resulting in a decrease in standard error while the coefficient of interest remains highly significant at the 1% level.

Table 6 Robustness check I: Alternative independent variable, clustered S.E., controls, and fixed effects

Our study further addresses the potential bias introduced by region-level environmental policies on our baseline results. To mitigate this concern, we carefully go through the environmental policy documents during the sample period, identifying two key regulations in China. The first is the “Key Cities Air Pollution Prevention and Control Cities (KCAPPC)” policy, implemented by the Ministry of Ecology and Environmental Protection (MEEP) in China at the end of 2002.¹⁴ The second regulation stems from the provincial pollution reduction targets outlined in China’s 11th Five-Year Plan.¹⁵ To account for the impact of these environmental policies, we incorporate them into our analysis and re-estimate the coefficient of interest. In Column 7 of Table 6, we introduce the KCAPPC as an additional control variable, represented by the interaction term ${KCAPPC}_{(i)kt}$, which equals one if firm i is located in the key city k after 2003 (defined as the first year of environmental regulation) and zero otherwise. The estimated coefficient of interest remains robust in magnitude and significance compared to the baseline results. Furthermore, we perform robustness checks by including provincial pollution reduction targets as an additional control variable, interacting with year dummies. The estimated coefficient in Column 8 remains consistent. Finally, we account for region-year fixed effects to address any time-varying confounders at the regional level. The results in Column 9 also demonstrate the robustness of our findings.¹⁶ Overall, the coefficients in Columns 7–9 of Table 6 suggest that environmental regulations or other policies at the regional level do not introduce significant bias in our analysis.

4.2.2 Robustness Check II: Alternative Type of Patent, Instrument, Different Sample and Air Pollutant

The SIPO reports on three types of patents that vary in subject matter, examination procedure, patenting and maintenance costs, and maximum protection period. To delve deeper into other types of patents, we present the findings in Columns 1–3 of Table 7. In Column 1, we substitute the term “logged one plus invention patent” with “logged one plus utility patents”. The estimated effect size is -0.2047, which is statistically significant at the 5% level. However, we do not observe any significant impact of air pollution on design patents. This finding keeps company with our previous argument that invention patents, characterized by high novelty and technological inventiveness, are the most susceptible to air pollution due to the intensive cognitive effort involved in their creation. Furthermore, when we aggregate all three types of patents for analysis, we find that the coefficient of interest is -0.3013, also statistically significant at the 5% level. In terms of economic significance, a 1% increase in air pollution (PM_2.5) in the previous year, relative to its mean value, leads to a 0.8% decrease in total patents compared to the average from 1999 to 2013. Although this impact remains significant, it is smaller than the effect of PM_2.5 on invention patents. Additionally, we substitute the instrumental variable with the number of one-year lag thermal inversion occurrences at each county-year unit. This adjustment yields a similarly negative and highly statistically significant result. Overall, these results collectively strengthen the argument that air pollution has a detrimental effect on firms’ innovation output.

Table 7 Robustness check II: Alternative type of patent, instrument, different sample, and air pollutant

Firms may base relocation decisions on local air quality levels, potentially leading innovative firms to move away from high-pollution areas. This could mean that our results reflect relocation intentions rather than the impact of air pollution. To address this issue, we exclude firms that changed locations across different cities during the sample period. The analysis in Column 5 reveals a slight reduction in sample size from 2,643,178 to 2,636,366 (a 0.26% decrease) after removing relocating firms, indicating that relocation is uncommon for manufacturing firms. With this smaller sample, the estimated effect remains close to the baseline at -0.2402, suggesting that our findings are not driven by relocation behaviors.¹⁷ There is also a concern that firms’ entry and exit behaviors may bias our baseline results. Firstly, new firms may choose locations to enter the market based on their characteristics and local air pollution levels, making it difficult to exclude all factors determining their choice despite firm fixed effects. Secondly, firms may exit the market if the harmful effect of air pollution is strong enough. To address this concern, we calculate the fraction of firms’ entry and exit at each county-year level based on Fu et al. (2021) to test for potential firm sorting problems.¹⁸ The results of regressing the fraction of firms’ entry and exit on the instrumented air pollution are presented in Table 14. Our findings indicate that the coefficient of interest is insignificant and close to zero in magnitude, suggesting that firms’ entry and exit is not a major concern in our paper.¹⁹

As evidenced by the summary statistics, a considerable portion of the firms in our study lack patenting experience within the specified time frame. In order to mitigate potential biases stemming from zero values, we have excluded firms that have never reported non-zero patents (in other words, firms without patenting experience) from our analysis. The outcome is detailed in Column 6 of Table 7. The estimated effect for this reduced sample is -0.9021 and remains statistically significant at the 1% level. In terms of economic significance, the result indicates that a 1% rise in air pollution (PM_2.5) over the past 12 months, relative to its average, leads to a 1.8 percentage point decrease in invention patents compared to the mean for firms with patenting experience. This impact mirrors the baseline effect. Additionally, to assess the robustness of our findings to different air pollutants, we explore alternative measures of air pollution. Specifically, we substitute PM_2.5 with SO₂, another significant air pollutant with a strong correlation to thermal inversion (as demonstrated in Column 4 of Table 3). The outcome in Column 7 reveals that a 1% increase in SO₂ from its average over the past year results in a 1.3% decline in the number of invention patents from their mean. This implies that our key finding is not limited to a specific air pollutant.

4.3 Regression Discontinuity Design: The Huai River Policy

To further check the validity of our baseline results and investigate the accumulated effect of air pollution, we examine the impact of air pollution by exploiting the exogenous variation in PM_2.5 levels caused by China’s Huai River winter heating policy. As documented by Ebenstein et al. (2017), China’s heating policy was implemented in the 1950s during a time when China’s economy was centrally planned. The northern region of the Huai River Line, where average temperatures in January were around 0 °C, was equipped with a coal-based heating system. However, no such system exists in the southern provinces of China, even near the line. Because the incomplete coal combustion injects massive pollutants (including particulate matter, SO₂, NO₂, etc.), this policy led to a discontinuity in air pollution levels around the Huai River Line, thus, providing a good opportunity to investigate the impact of air pollution on firms’ innovation performance using RDD (see, for example, Almond et al. 2009; Chen et al. 2013; Ebenstein et al. 2017; Ito and Zhang 2020; Xue et al. 2021).²⁰ Specifically, we collapse all time-variant variables to the mean value during our entire period (1998–2013) in the analysis of RDD methodology by exploiting the Huai River heating policy. This suggests that our improved RDD results reflect the accumulated effect of air pollution on firms’ innovation performance over the whole period. We estimate the following parametric equation by introducing fourth polynomials of the latitude distance between the firm location and the Huai River Line:

$${Innovation}_{ij}={\varphi }_{0}+{\varphi }_{1}{North\_South}_{ij}+{f(Latitude\_Distance)}_{ij}+{f(Latitude\_Distance)}_{ij}\times {North\_South}_{ij}+{W}_{c}+{X}_{ij}+{\mu }_{j}+{longtitude}_{d}+{\varepsilon }_{ij}$$

(3)

${Innovation}_{ij}$ is the logged one plus the mean value of invention patent of firm i (located in county c) in industry j during the sample period. ${North\_South}_{ij}$ is an indicator of firm i of industry j located in the northern part of the Huai River Line. ${f(Latitude\_Distance)}_{ij}$ is a polynomial (linear, square, cubic and quartic) term in latitude distance to the Huai River Line of firm i in industry j, which is also interacted with the ${North\_South}_{ij}$ to account for the possibility that the functional forms are different on the two sides of the threshold. ${X}_{ij}$ is a vector of firm characteristics mean value and ${W}_{c}$ is a vector of the weather controls mean value during our sample period. ${\mu }_{j}$ denotes the industry fixed effects. ${longtitude}_{d}$ represents the longitude decile fixed effects (including 10 quantiles according to its maximum and minimum values).²¹ To ensure consistency with the baseline model, we also cluster the standard error at the county level. A negative ${\varphi }_{1}$ indicates that firms in the polluted heating regions perform poorly in innovation output.

We first use the global polynomial regression in the parametric method. The regression estimates of interest from Eq. (3) are presented in Columns 1–4 of Table 8. The coefficient of ${North\_South}_{ij}$ is negative and statistically significant at least at the 5% level in Table 8. This suggests that firms located on the heating (north) side of the Huai River Line experience a decrease in firms’ innovation output. Taking the result in Column 1 as an example, we find that firms that are located on the heating side suffer a 4% decrease in innovation over our sample period. Second, we also conduct a non-parametric local linear regression (LLR) analysis following the optimal bandwidth selection criteria of Imbens and Kalyanaraman (2012) to relax the assumption about the functional form in Eq. (3). The result utilizing a triangular kernel (thus, placing greater weights on locations near the Huai River Line), the most commonly used kernel (Calonico et al. 2014; Imbens and Kalyanaraman 2012; Ito and Zhang 2020), is reported in Column 5. It shows a statistically significant decrease of 2.2% in the number of invention patents in the north of the Huai River Line. When we switch to another commonly used kernel, Epanechnikov, the result is highly consistent (Column 6).

Table 8 Robustness check III: RDD estimates of the Huai River policy

To visually represent the difference in innovation performance between firms across the Huai River Line, all RDD results are plotted in Fig. 3. The dots in the figure indicate the average logged invention patents in each distance bin along the heating line (10 bins on each side). A fitted line is created using a polynomial regression of order 1–4 and kernel-weighted LLR on both sides of the line. As shown in Fig. 3, there is a notable decrease in firms’ invention patents across the Huai River Line. The RDD results validate our main findings and importantly, provide compelling evidence of the long-term detrimental impact of air pollution on firms’ innovation performance. By leveraging the exogenous variation caused by the Huai River heating policy, the RDD results also contribute valuable insights for policy-makers in conducting benefit–cost analysis, as the loss of innovation on the heating side poses a significant cost to overall socioeconomic development.

Fig. 3

RDD plots of innovation across the Huai River Line. Notes: This figure shows the RDD estimator of invention patents across the Huai River Line. “Polynomial order of 1–4” means that we gradually add the linear term, quadratic term, cubic term and quartic term of the latitude distance between the firms’ locations and the Huai River Line. The last two images show the RDD estimators of local linear regression using the kernel functions of Triangular and Epanechnikov, respectively

5 Mechanism Analysis and Further Study

In this section, following the confirmation of the detrimental effects of air pollution on firm innovation, we intend to investigate three potential mechanisms: productivity loss, reduced human capital retention, and a decline in financing ability. Furthermore, we propose to conduct a further study on the nonlinear effects.

5.1 Mechanism Analysis on Labor Productivity Loss

To assess the impact of air pollution on firms’ innovation output, direct measurement of inventors’ labor productivity would be ideal, but this data is unavailable. Therefore, we adopt an indirect approach by examining average labor productivity, assuming a strong correlation with inventors’ productivity. Specifically, we investigate whether the documented reduction in firms’ labor productivity due to air pollution, as found by Fu et al. (2021), holds in our sample. By regressing the natural logarithm of output per employee on lagged PM_2.5 levels instrumented with thermal inversion and other control variables, we find a statistically significant coefficient of -2.5 in the second-stage regression. This result confirms the validity of Fu et al.’s (2021) findings in our study. Additionally, the observed loss in average labor productivity at the firm level suggests a potential dampening effect on inventors’ labor productivity.

5.2 Mechanism Analysis on Human Capital Retention Decrease

In the next step, we investigate the impact of air pollution on innovation by analyzing how it may affect a firm’s ability to attract and retain highly skilled employees, such as inventors. Due to the lack of detailed employee composition data, we focus on employees’ educational backgrounds, a common approach in previous research (Lise and Robin 2017; Lopes de Melo 2018). We operate under the assumption that skilled labor typically requires a longer period of education to acquire the necessary knowledge for high-skilled tasks. Using data from ASIF, which provides information on employees’ educational backgrounds at the firm level, we analyze a cross-section of 126,955 firms by further leveraging the 2004 National Economic Census in China. To address potential confounding factors, we control for industry fixed effects and include province fixed effects to consider local economic, social, and political influences on employees’ educational levels and air quality. We define the proportion of highly skilled labor in a firm as the ratio of employees with a college degree or higher to total employees. After accounting for other variables, our regression analysis, presented in Column 1 of Table 6, reveals a significantly negative coefficient (-0.20) between air pollution and the proportion of highly educated employees. This suggests that poor air quality may impede firms’ ability to attract and retain highly educated human capital, which is crucial for innovation. Furthermore, the economic impact is noteworthy, with a one-year lagged PM_2.5 increase of one standard deviation leading to a 0.5 standard deviation decrease in the percentage of highly educated employees.

In addition, we also examine the heterogeneous impact of air pollution on human capital accumulation in different firms, taking into account industry concentration across different regions. If our argument above is true, we expect that air pollution leads to a larger and more significantly negative effect on industries that are more dispersed across regions since workers in geographically dispersed industries (low regional concentration industries), particularly those who are highly skilled or educated, are easier to move to regions with better air quality (in other words, job-hopping behaviors), which contributes to the human capital loss for the original air-polluted places. To test our hypothesis, a method akin to the Herfindahl Index is employed to calculate industry concentration across counties. This involves calculating the total main business revenue across industries (denoted as A), and then calculating the total main business revenue across each industry-county unit (denoted as B). We then calculate the values of B/A and its square and add up the two values (denoted as C) to obtain the median of C value at the industry level. Based on this index, we divide our sample into two subsamples: industries with a C value higher (lower) than the median are placed in the high (low) regional concentration subsample. Columns 3 and 4 of Table 9 report the results for two subsamples. The coefficient of -0.24, statistically significant at the 1% level, supports the hypothesis for firms in low regional concentration industries, but loses significance for those in high-regional concentration industries. These findings reinforce the argument that air pollution hampers firms’ ability to access crucial human capital, thereby impeding innovation.

Table 9 Mechanism analysis I: Productivity and human capital retention loss

5.3 Mechanism Analysis on Financing Ability Decline

Air pollution not only reduces workers’ productivity and the likelihood of retaining skilled workers, but it may also increase firms’ environmental investment. This is particularly evident in industries subject to stringent government oversight or regulations. However, this heightened focus on environmental investment could potentially detract from innovation investment, ultimately impacting firms’ innovation development. To test this hypothesis, a cross-sectional analysis is conducted, regressing logged sewage charges on instrumented one-year lag logged PM_2.5.²² The results offer suggestive evidence in support of this argument. The coefficient of interest, presented in Column 1 of Table 10, indicates a statistically significant positive impact of PM_2.5 from the previous year on firms’ sewage charges at a 10% significance level. In economic terms, a 1% increase in air pollution (PM_2.5) from the previous year, relative to its mean value, leads to a 2.2% increase in sewage charges compared to the average, highlighting a noteworthy effect.

Table 10 Mechanism analysis II: Financing ability decline

Additionally, if air pollution affects firms’ innovation performance by decreasing innovation investment, we might see a more significant decline in innovation in polluting industries compared to non-polluting industries as a result of stricter environmental regulations. To categorize the industries, we follow the classification standard outlined by He et al. (2020) and split them into polluting and non-polluting sectors. The findings in columns 3–4 support our argument, showing that firms in polluting industries suffer a greater innovation loss due to air pollution. This is evident from the higher magnitude and significance of the coefficient of interest in comparison to non-polluting industries. These results indirectly imply that the reduction in firms’ financing capability due to air pollution is another crucial factor contributing to the negative impact on innovation.

5.4 Further Study on Non-linear Effect of Air Pollution

In this section, we delve into the non-linear impact of air pollution on firm innovation, building on previous studies that indicate the negative effects of air pollution on economic outcomes become more pronounced as air quality worsens (Chen et al. 2020; Schlenker and Walker 2016). To examine this non-linearity, we incorporate a spline regressor into the baseline econometric model, as per Chen et al. (2020), enabling us to assess how firms’ innovation performance changes in response to different levels of air pollution. The regression model used is as follows:

$${Innovation}_{ict}={\beta }_{0}+{\beta }_{1}{Air Pollution}_{c(t-1)}+{\beta }_{2}\left[{Air Pollution}_{c\left(t-1\right)}-S\right]\times I\left[{Air Pollution}_{c\left(t-1\right)}>S\right]+f\left({W}_{ct(t-1)}\right)+{X}_{ic(t-1)}+{\theta }_{i}+{\mu }_{jt}+{\varepsilon }_{ict}$$

(4)

$S$ refers to the specific value where the non-linear effect of air pollution is assumed to occurs. Then, $\left[{Air Pollution}_{c\left(t-1\right)}-S\right]$ represents the one-year lag PM_2.5 concentration minus the $S$ value and $I\left[{Air Pollution}_{c\left(t-1\right)}>S\right]$ denotes a dummy variable that equals one if one-year lag PM_2.5 concentration is larger than $S$ value (zero otherwise). We do the same for our IV as well. The rest is the same as baseline regression setting. We now interpret the ${\beta }_{1}$ as the marginal effect of one-year lag PM_2.5 on firm innovation when one-year lag PM_2.5 value is smaller than or equal to $S$. ${\beta }_{1}$ plus ${\beta }_{2}$ is the marginal effect of air pollution on firm innovation if one-year lag PM_2.5 concentration is larger than $S$ value. Therefore, ${\beta }_{2}$ is the coefficient of interest in this section as it provides evidence of non-linearity of air pollution’s impact as long as it is significantly different from zero. We set up 5 $S$ values with an interval of 25 and obtain the 2SLS estimators, respectively. The results are reported in Table 11. It is worth noting that each coefficient of interest (${\beta }_{2}$) is negative and statistically significant at least at the 5% level. This confirms our inference that the response of firm innovation to air pollution is also non-linear in our paper.

Table 11 Non-linear effect of air pollution

6 Conclusion

This study examines the impact of air pollution on firms’ innovation performance in China, using thermal inversion as an instrumental variable to address endogeneity concerns. Analyzing data from 539,400 manufacturing firms, our paper reveals a negative causal relationship between air pollution and firms’ innovation performance, as indicated by the number of invention patents, in both the short and long term. The findings are robust across various tests and suggest that air pollution impedes innovation performance through reduced employee (including inventor) productivity, decreased retention of skilled workers, and limited financing for innovation activities. Additionally, a non-linear effect of air pollution is observed. These results shed light on the factors influencing firms’ innovation performance, underscoring the significance of environmental conditions like air quality. The study advocates for environmental regulations and innovation incentives in China, offering valuable insights for policy-making and cost–benefit analyses in this context.

More broadly speaking, escaping the middle-income trap is often more challenging than overcoming absolute poverty. Developing countries can navigate this by focusing on effective industrial upgrading through innovation. While many nations have aimed to shift from traditional production-driven models to innovation-led economies in the past decade, some have struggled with this transition. Meanwhile, environmental challenges such as climate risks and air pollution are persistent issues in the economic development of developing countries. Therefore, our study can provide valuable insights for other nations in similar stages of economic development as China was at that time.

Appendix

Fig. 4

Smoothness tests of covariates across the Huai River Line. Notes: This figure shows smoothness tests of all covariates across the Huai River Line

4

Table 12 List of materials required for three types of patents in China

12

Table 13 The impact of air pollution on invention patents (Plus the coefficients of temperature bins)

13

Table 14 Firms’ entry and exit

14