3.1. Model environment

An economy consists of $N$ regions, each governed by one local government and served by a representative firm. The representative firm produce output with a Cobb-Douglas technology:

\begin{equation*} Y_{it}=A_{it}Z_{it}K_{it}^{\alpha }\left(L_{it}G_{it}\right)^{1-\alpha }, \end{equation*}

where $A_{it}$ is local productivity, $Z_{it}$ is the flow emission input for production that is controlled by central government’s environmental policy, $K_{it}$ is capital used for production, $L_{it}$ is local labor input, and $G_{it}$ is infrastructure created by the local government.

In each period, the representative firm first observes local productivity $A_{it}$ , emission $Z_{it}$ , infrastructure $G_{it},$ and then hires capital $K_{it}$ , and labor $L_{it}$ to maximize its profit:

\begin{equation*} max_{\left\{{Z_{it}},{K_{it}},{L_{it}}\right\}}A_{it}Z_{it}K_{it}^{\alpha }L_{it}^{1-\alpha }G_{it}^{1-\alpha }-RK_{it}-\Phi _{it}L_{it}, \end{equation*}

where $\Phi _{it}$ is the competitive wage, $R$ is the rental price of capital that equals to the interest rate. $R$ is globally determined since each region is a small open economy. Suppose the labor is immobile and each region has a fixed labor supply $L_{it}=1$ . The first order conditions are

\begin{equation*} \Phi _{it}=\left(1-\alpha \right)A_{it}Z_{it}K_{it}^{\alpha }G_{it}^{1-\alpha }, \end{equation*}

\begin{equation*} K_{it}=\left(\frac{\alpha A_{it}Z_{it}}{R}\right)^{\frac{1}{1-\alpha }}G_{it}. \end{equation*}

By substituting $K_{it}$ and $L_{it}$ into the production function, we have

\begin{equation*} Y_{it}=\left(\frac{\alpha }{R}\right)^{\frac{\alpha }{1-\alpha }}\left(A_{it}Z_{it}\right)^{\frac{1}{1-\alpha }}G_{it}. \end{equation*}

In each region, there are overlapping generation of households, with each generation living for two periods. Each individual born at period $t$ has identical preferences given by

\begin{equation*} ln\left(C_{it}^{t}\right)+\beta ln\left(C_{it+1}^{t}\right), \end{equation*}

where $C_{it}^{t}$ and $C_{it+1}^{t}$ denote consumption chosen by the individual across his/her lifetime at $t$ and $t+1$ .Footnote ³ The budget constraints for each individual is

\begin{equation*} C_{it+1}^{t}=\left(1-\tau \right)RS_{it}^{t}, \end{equation*}

where $S_{it}^{t}$ is individual’s saving and $\tau$ is the tax rate on both labor and capital income.

Local government’s tax revenue in period $t$ is

\begin{equation*} \tau \left(\Phi _{it}L_{it}+RK_{it}\right)=\tau Y_{it}, \end{equation*}

which contributes to its budget at the end of period $t$ :

\begin{equation*} W_{it}=\tau Y_{it}+\left(1-\delta _{G}\right)G_{it}-RD_{it-1}, \end{equation*}

where $W_{it}$ is the budget resources owned by local government, $\delta _{G}$ is the depreciation rate of infrastructure, $D_{it-1}$ is local government debt issued at $t-1$ . Local governor can issue new debt $D_{t}$ , in addition to $W_{t}$ , to fund government employees’ consumption $E_{it}^{G}$ in the current period and the infrastructure for next period $G_{it+1}$ :

\begin{equation*} W_{it}+D_{it}=G_{it+1}+E_{it}^{G}. \end{equation*}

3.2. Role of career incentives amidst environmental regulation

There exists an agency problem between the central and local governments. We adopt the following specification of the productivity of region $i$ :

\begin{equation*} A_{it}=e^{{f_{t}}+{a_{it}}+{\varepsilon _{it}}}. \end{equation*}

where $f_{t}$ denotes a countrywide common shock, $a_{it}$ represents the local governor’s ability in developing the local economy, $\varepsilon _{it}$ is an idiosyncratic noise component. These components are independent of each other, and all are not publicly observable. Central government promotes local governors based on their anticipated economic abilities. The central government forms posterior prediction on local governors’ ability based on the realized economic outcome of all local governors:

\begin{equation*} \hat{a}_{it+1}=E\left[a_{it+1}|\left\{Y_{it+1}\right\}_{i=1,2,\ldots, M}\right]. \end{equation*}

To derive $\hat{a}_{it+1}$ above, we need to get an expression for $Y_{it+1}$ , which, in log-scale, is given by:

\begin{equation*} y_{it+1}\equiv ln\left(Y_{it+1}\right)=ln\left[\left(\frac{\alpha }{R}\right)^{\frac{\alpha }{1-\alpha }}\left(A_{it}Z_{it}\right)^{\frac{1}{1-\alpha }}G_{it+1}\right] \end{equation*}

\begin{equation*} =\frac{1}{1-\alpha }\left(f_{t+1}+a_{it+1}+\varepsilon _{it+1}\right)+\frac{1}{1-\alpha }ln\left(Z_{it+1}\right)+\frac{\alpha }{1-\alpha }ln\left(\frac{\alpha }{R}\right)+ln\left(G_{it+1}\right). \end{equation*}

Following Xiong (Reference Xiong2019), there exists a general sufficient statistic for governor’s economic ability:

\begin{equation*} m_{it+1}\equiv \left(1-\alpha \right)\{y_{it+1}-\left[\frac{1}{1-\alpha }ln\left(Z_{it+1}\right)+\frac{\alpha }{1-\alpha }ln\left(\frac{\alpha }{R}\right)+ln\left(G_{it+1}^{\mathrm{*}}\right)\right]. \end{equation*}

The central government expects that local governor will choose $G_{it+1}$ equaling to $G_{it+1}^{\mathrm{*}}$ . Hence, from the central government’s perspective in interpreting the information content of this statistics, $m_{it+1}$ becomes

\begin{equation*} m_{it+1}^{c}=f_{t+1}+a_{it+1}+\varepsilon _{it+1}. \end{equation*}

Hence, the learning rule for central government becomes

\begin{equation*} \hat{a}_{it+1}=E\left[a_{it+1}|\left\{m_{it+1}^{c}\right\}_{i=1,2,\ldots, M}\right]. \end{equation*}

From local governor’s perspective, the sufficient statistic becomes

\begin{equation*} m_{it+1}^{l}=f_{t+1}+a_{it+1}+\varepsilon _{it+1}+\left(1-\alpha \right)\left[ln\left(G_{it+1}\right)-ln\left(G_{it+1}^{\mathrm{*}}\right)\right], \end{equation*}

which is increasing in local government spending $G_{it+1}$ . Under this setup, the maximization problem for the local governor becomes

\begin{equation*} V\left(W_{it}\right)=max_{{G_{it+1}},{D_{it}}}\{\rho \left[ln\left(C_{it}^{t}\right)+ln\left(C_{it}^{t-1}\right)\right]+\gamma ln\left(W_{it}+D_{it}-G_{it+1}\right) \end{equation*}

\begin{equation*} +\beta E_{t}V\left(\tau Y_{it+1}+\left(1-\delta _{G}\right)G_{it+1}-RD_{it}\right)+\kappa _{i}\left[lnG_{it+1}-lnG_{it+1}^{\mathrm{*}}\right]\}. \end{equation*}

where the parameter $\rho$ and $\gamma$ are weights of the households’ and government employees’ consumption assigned by local government. $\kappa _{i}$ is a function of deep parameters in the signal-noise setting derived in Xiong (Reference Xiong2019) that captures the career incentive.Footnote ⁴ If the local governor is benevolent, $\kappa _{i}=0$ . If he/she cares about future promotion, $\kappa _{i}\gt 0$ . Issuing more debt increases the likelihood of being promoted, which increases the local governor’s value function, and the strength of this “marginal private benefit” is dictated by $\kappa _{i}$ . This benefit is classified as “private” since it is excessively reaped by the local governor rather than the broader society. The additional term $\kappa _{i}[lnG_{it+1}-lnG_{it+1}^{\mathrm{*}}]$ , which captures the difference between “benevolent” and “politically incentivized” governors, remains unchanged even if households experience disutility from flow emissions $Z_{it}$ , and the local governor takes this into consideration.

We now formally derive our testable hypotheses. Define two ratios:

\begin{equation*} g_{it+1}=\frac{G_{it+1}}{W_{it}}, d_{it}=\frac{D_{it}}{G_{it+1}}. \end{equation*}

After solving the local governor’s maximization problem following the method in Xiong (Reference Xiong2019), we can haveFootnote ⁵

\begin{equation*} g_{it+1}=\frac{\beta \gamma +\kappa _{i}}{\gamma +\kappa _{i}}\frac{1}{1-d_{it}}. \end{equation*}

Based on this equation, we can further derive the response of $d_{it}$ to $g_{it+1}$ as

\begin{equation*} \frac{\partial d_{it}}{\partial g_{it+1}}=\frac{\beta \gamma +\kappa _{i}}{\gamma +\kappa _{i}}\left(g_{it+1}\right)^{-2}. \end{equation*}

Given the parameter values and $\kappa _{i}\geqq 0$ , we know $\frac{\partial d_{it}}{\partial g_{it}}\gt 0$ . If there is a persistent environment shock to restrict emission $Z_{it}$ , then current output $Y_{t}$ decreases, so does the local government budget wealth $W_{it}$ . The local governor will choose to increase infrastructure $G_{it+1}$ in order to stabilize the economy. Hence, $g_{it+1}$ increases, then $d_{it}$ also goes up given $\frac{\partial d_{it}}{\partial g_{it}}\gt 0$ . A higher $d_{it}$ implies that local governor chooses to issue more government debt $D_{it}$ to mitigate the negative effect of environment policy on government budget. This implication holds for benevolent governor or governor with career incentive. Hence, we have the first hypothesis.

Hypothesis 1: If an environmental policy shock occurs that restricts emissions, it will lead to a decline in economic output. In response, both a benevolent governor and a governor motivated by career incentives are likely to choose to issue more debt.

Additionally, we can derive

\begin{equation*} \frac{\frac{\partial d_{it}}{\partial g_{it+1}}}{\partial \kappa _{i}}=\frac{\left(1-\beta \right)\gamma }{\left(\gamma +\kappa _{i}\right)^{2}}\left(g_{it}\right)^{-2}. \end{equation*}

Since the discount factor $\beta \lt 1$ , $\frac{\frac{\partial d_{it}}{\partial g_{it+1}}}{\partial \kappa _{i}}\gt 0$ . Given the same environmental policy shock that restricts emissions and reduces economic output, a governor with greater career incentives will choose to issue more debt to counteract the economic downturn. However, if the governor is benevolent, the decision to issue debt in response to the environmental shock will not be influenced by career incentives.

Hypothesis 2: If a governor is benevolent, debt issuance in response to an environmental shock is not influenced by career incentives. However, if the governor has career incentives, the risk of debt accumulation (future debt repayment) in response to environmental policy increases with the strength of career incentives.

Intuitively, A benevolent local governor determines the optimal debt level by equating the marginal social benefits of debt issuance, which comes from short-term GDP stabilization, and the marginal social costs, which come from a higher debt repayment in the future.Footnote ⁶ A politically incentivized local governor, in contrast, additionally considers the marginal private promotion benefits in the decision-making process, resulting in excessive debt accumulation facing the economic loss brought by the environmental policy.

4.1. The action plan on air pollution prevention and control

Rapid industrialization since the Reform and Opening-up has led to both unprecedented economic growth and environmental deterioration in China. In response to environmental challenges, China passed the Environmental Protection Law in 1989 and developed a multifaceted regulatory system.Footnote ⁷ However, air pollution concentration still reached a record high in 2013 due to lax enforcement (Greenstone et al. Reference Greenstone, He, Li and Zou2021).

To address the serious air pollution problem, the central government announced the Action Plan, a national air quality improvement project, in September 2013.Footnote ⁸ Similar to previous environmental policies, the Action Plan sets air quality goals for each province. In practice, provincial governments further break their allocated targets down to prefectures in their jurisdictions. Overall, the Action Plan required nationwide particular matter ( $\mathrm{PM}_{2.5}$ and $\mathrm{PM}_{10}$ ) concentrations to be reduced by more than 10% by the end of 2017, compared to 2012 levels.Footnote ⁹ Given the spatially differentiated nature of local economic conditions, such as industrial structure and energy reliance, there was wide variation in the tightness of the Action Plan across prefectures. We discuss how we measure “tightness” in Section 5.1.

In China, local officials are responsible for various aspects of local affairs. Since 2006, to receive a promotion, local officials have been required to meet environmental protection goals assigned by higher-level governmentsFootnote ¹⁰ and the weight of local environmental performance in the evaluation system keeps rising (Zhang and Xiong Reference Zhang and Xiong2020).Footnote ¹¹ However, local economic growth still takes the first priority for local officials’ evaluation.Footnote ¹² Therefore, local officials need to carefully balance economic growth and environmental protection under the Action Plan.

4.2. The local government financing vehicles (LGFVs)

Unlike in most western countries, China’s Budget Law imposes strict prohibitions on official local government debt, thereby constraining local governments’ capacity to promote economic growth at the local level. However, in response to the global financial crisis, the central government initiated a fiscal stimulus package valued at 4 trillion RMB in 2008, with only 1 trillion RMB came from the central government (Bai et al. Reference Bai, Hsieh and Song2016) and local governments assumed most of the debt burden.

To align with the provisions of the Budget Law, which prohibits local governments from incurring official on-balance-sheet debts, the China Banking and Insurance Regulatory Commission (CBIRC) and Ministry of Finance permit local governments to effectively issue off-balance-sheet debts by establishing local government financing vehicles (LGFVs) to fund stimulus spending (Bai et al. Reference Bai, Hsieh and Song2016; Chen et al. Reference Chen, He and Liu2020). Practically, LGFVs are registered simply as business entities, thereby excluding their debt from local governments’ balance sheets. Nonetheless, given that these LGFVs are endorsed by the local governments, their debt becomes de facto local government debt. Shown by the dashed line in Figure 1, LGFV debts expanded rapidly after 2009 and act as an important fiscal instrument for local governments.

5.1. Treatment variable: Tightness of air pollution control under the action plan

In this section, we discuss how we measure prefecture-level regulation tightness under the Action Plan, which is used as the treatment variable in the regression presented in Section 5.4. In particular, this measure is intended to capture the potential burden imposed on the economy by the Action Plan.

At first glance, prefecture-specific targets to reduce $\mathrm{PM}_{2.5}$ concentration, if available,Footnote ¹³ appear to be good candidates for the tightness measure. However, we argue that this is not the best fit for our study for the following reasons. First, unlike pollution emissions, $\mathrm{PM}_{2.5}$ concentration measures ambient environmental qualities that are affected by not only local polluting activities but also pollutants transported from other locations by wind. Thus, a prefecture may have a high $\mathrm{PM}_{2.5}$ concentration and face a high concentration reduction target even if the prefecture itself has few polluting sources. In such a case, emission reductions in other prefectures of the province needs to be required to achieve the $\mathrm{PM}_{2.5}$ concentration reduction goal of this prefecture; therefore, the actual regulation tightness for this prefecture does not necessarily to be high.

Second, even if $\mathrm{PM}_{2.5}$ concentration can credibly reflect local polluting activities, the targets are not good measures of regulation tightness for our purposes. If a prefecture with a higher concentration reduction target is much less reliant on pollution- or energy-intensive industries than another prefecture with a lower target, the regulation tightness, or, in other words, the economic costs of the Action Plan, might be lower for the first prefecture compared to the second.

An ideal solution to the first problem above is to measure regulation tightness by $\mathrm{PM}_{2.5}$ generated or emitted by polluting sources in each prefecture, rather than $\mathrm{PM}_{2.5}$ concentrations. Unfortunately, we do not have access to this information; thus, we use information on the generation and emissions of $SO_{2}$ , a major precursor of $\mathrm{PM}_{2.5}$ (Baker et al. Reference Baker, Collins, Olivié, Cherian, Hodnebrog, Myhre and Quaas2015), as an alternative to construct the tightness measure.Footnote ¹⁴

To resolve the second problem, we incorporate a prefecture’s pollution reliance into the measurement of regulation tightness:

(1) \begin{equation} \textit{Tight}_{i}=\underset{\textit{abatement} \textit{intensity}}{\underbrace{\frac{SO_{2}~ \textit{abatement}_{i}}{GDP_{i}}}}|_{2012}=\underset{\textit{abatement~ ratio}}{\underbrace{\frac{SO_{2}~ \textit{abatement}_{i}}{SO_{2}~ \textit{generation}_{i}}}}|_{2012}\times \underset{SO_{2}\,\textit{generation} \textit{intensity}}{\underbrace{\frac{SO_{2}~ \textit{generation}_{i}}{GDP_{i}}}|_{2012},} \end{equation}

where $SO_{2}~ \textit{generation}_{i}$ is the amount of $SO_{2}$ generated by all polluting sources located in prefecture i. It is the sum of the amount of $SO_{2}$ ultimately emitted into the atmosphere ( $SO_{2}~ \textit{emission}_{i}$ ) and the portion that is abated by emission abatement technologies such as $SO_{2}$ scrubbers ( $SO_{2}~ \textit{abatement}_{i}$ ) before the waste gas goes into the atmosphere. All variables in equation (1) are measured at the pre-policy levels in 2012. Following Liu (Reference Liu2019), to eliminate the sensitivity of the choice of regulation tightness measures, we standardize $\textit{tight}_{i}$ to have mean 0 and standard deviation 1 for the ease of results interpretation.

Based on equation (1), regulation tightness is defined as prefecture-level $SO_{2}~$ abatement intensity in 2012, which can be decomposed to the product of the $SO_{2}$ abatement ratio and $SO_{2}~$ generation intensity. To mitigate the endogeneity issues, we employ individual-level characteristics prior to the Action Plan (i.e., in 2012) to measure regulation strength each individual unit face, which is widely adopted recently in the literature (Mian and Sufi, Reference Mian and Sufi2014; Giroud and Mueller, Reference Giroud and Mueller2019; Duval et al. Reference Duval, Hong and Timmer2020; Mao et al. Reference Mao, Wang and Yin2022). A higher pre-policy abatement ratio indicates that the prefecture had already abated a large share of emissions generated from fuel combustion before the policy took effect. Therefore, the cost associated with further increasing the abatement ratio is higher.Footnote ¹⁵ Furthermore, a higher $SO_{2}$ generation intensity implies that the local economy relies heavily on pollution-intensive industries, so emissions reduction for the economy is also likely to be tougher. Data required to calculate the toughness measure are drawn from the City Statistical Yearbook of China.

As shown in Figure 2, the geographic distribution of regulation tightness overlaps to some degree with that of the $\mathrm{PM}_{2.5}$ reduction targets set by the Action Plan. For example, the Beijing–Tianjin–Hebei area, which had the highest $\mathrm{PM}_{2.5}$ concentration prior to the Action Plan and was assigned the highest $\mathrm{PM}_{2.5}$ reduction targets, is also among the most tightly regulated areas, according to our tightness measure.

Figure 2.

Geographic distribution of regulation tightness.

Notes: This figure plots geographic distribution of regulation tightness in 2012, prior to the implementation of the Action Plan, which is computed by equation (1) and captures the toughness of achieving $\mathrm{PM}_{2.5}$ reduction goal. The higher the intensity, the more toughness in a prefecture. There are 261 prefectures in the sample. The intensity is sorted in an ascending order, and then classified into 4 groups. The intensity increases from group 1 to group 4. There is no data if a prefecture is out of sample.

As shown in Figure 2, Yunnan, Sichuan, and Guizhou also had the most tightly regulated prefectures, although these provinces had low $\mathrm{PM}_{2.5}$ concentration reduction targets. In particular, Qujing (in Yunnan) and Guang’an (in Sichuan) faced significant challenges because they rely heavily on the secondary industry.

A final message from Figure 2 is that there exists sufficient heterogeneity in our tightness measure even for prefectures in the same province. As we will discuss in more detail in Section 5.4, this allows us to use within-province variations to identify the responses of local governments to the Action Plan, which alleviates potential province-level selection problems that may bias empirical estimation. This is an additional benefit of using our tightness measure in lieu of the $\mathrm{PM}_{2.5}$ concentration reduction targets, which are province-specific rather than prefecture-specific.

5.2. LGFV debts

We follow the method in Bai et al. (Reference Bai, Hsieh and Song2016) and Huang et al. (Reference Huang, Pagano and Panizza2020) to estimate the total financial liabilities and categories of LGFVs from their disclosed balance sheets that are required by bond issuance procedure. Consistent with Bai et al. (Reference Bai, Hsieh and Song2016) and Huang et al. (Reference Huang, Pagano and Panizza2020), all LGFV balance sheets were drawn from Wind Information Co. (WIND), which is the Chinese version of Bloomberg and the leading vendor of financial and economic data in China.Footnote ¹⁶ The sample covers all firms that issued MCBs from 2011 to 2017, which accounts for a majority of all LGFVs in the country.Footnote ¹⁷

The information used to construct the LGFV debt-related variables in this study include LGFV firm identity, location, total debts, and “other payables” in each year. In particular, total LGFV debts include stock of bank loans, payable bonds, obligations to goods and services suppliers, and lending from affiliated firms. Facing different lending standards and regulations, these sub-items feature different risk levels. Based on the degree of transparency and embedded risk levels, we divide them into two categories: “other payables” and “non-other payables.” Other payables mainly comprise less transparent debts from affiliated or connected firms, which require lower lending standards, regulations, and (even undetermined) maturity but higher interest rates.Footnote ¹⁸ Although this type of debt relieves the financial constraints of LGFVs (i.e., more flexible) to some extent, a higher risk in the future is imposed in exchange. Finally, we define “non-other payables” as the difference between total debts and other payables, including debt items such as bank loans and MCBs that are subject to stricter regulations and lower risks.

We aggregate the outstanding total debts and outstanding other payables of all LGFVs within each prefecture by the end of each fiscal year to get prefecture-level total LGFV debts and other payables.Footnote ¹⁹

5.3. Other prefecture-level characteristics and summary statistics

In our mechanism analysis, we additionally supplement our main dataset with prefecture-level land sales data. We use parcel-level land transaction data collected from the website of China’s Ministry of National Land and Resources to construct these variables.Footnote ²⁰ In particular, the following information is used: land transaction date, address, buyer identity, buyer’s two-digital industry code, and sales revenue.Footnote ²¹ As discussed in Appendix B1, local governments categorize urban land to its usage purpose: residential, commercial or industrial. Industrial land sales only account for a small share of land sale revenue and can be supplied for other reasons, such as structural adjustment, job creation, and local tax revenue. To mitigate the influence of non-financing considerations, we only consider commercial and residential land sales conducted for financing purposes.

A key variable in our heterogeneity analysis is the strength of local officials’ political promotion incentive. We manually collected prefect-level party secretary information from 2011 to 2017, including name, age, and position starting date of the party secretary of each city for each year. The key variable is position starting date, based on which we construct a variable to capture local officials’ incentive to promote the local economy for career promotion.Footnote ²² The construction method is detailed in Section 6.4. Other prefecture-level information, such as GDP, revenue of industrial firms, tax revenue from industrial firms, total local government tax revenue, total population, and miles of paved road, are collected from the City Statistical Yearbook of China.

To make data comparable across prefectures, we exclude the data from Beijing, Shanghai, Tianjin and Chongqing, four province-level prefectures in China. We also drop prefectures in Xinjiang, Qinghai, Hainan, and Tibet from the sample since the prefectures with non-missing data is too few to generate enough variation within corresponding province.Footnote ²³ Thus, the number of prefectures reduces from 336 to 301. Additionally, 40 prefectures did not report $\mathrm{SO}_{2}$ abatement in 2012, so the final sample covers 261 prefectures in total. To minimize the impact of outliers in our study, we winsorize all variables at the 2.5% and 97.5% levels.Footnote ²⁴ Table 1 shows the summary statistics for the sampling period from 2011 to 2017. Around 80% of all observations are located within one standard deviation of the tightness of air pollution control. Approximately 57% of all observations are from 2014 to 2017 (i.e., post-Action Plan).

Table 1.

Summary statistics

Notes: This table reports summary statistics in this study. Tight is the $\mathrm{SO}_{2}$ abatement intensity of each prefecture in 2012, which is computed by equation (1) and standardized to have mean 0 and standard deviation 1. The dummy variable “Post2013” equals 1 if year >2013 and 0 otherwise. “pdebt1113” is the ratio of LGFV debts to GDP, averaged from 2011 to 2013, which is standardized to have mean 0 and standard deviation 1. The dummy variable “HighIn” equals 1 if a party secretary is in the 2nd–4th (or 5th–8th) year of the current position, i.e., having high promotion urgency, in a given year in a given prefecture and 0 otherwise (Guo, Reference Guo2009). Total LGFV debts and other payables are aggregated from the outstanding total debt and outstanding other payables of all LGFVs within each prefecture by the end of each fiscal year, respectively. LGFV non-other payables is the difference between total debt and other payables. Monetary units are all discounted to 2010 price level. Revenue of industrial firms, GDP, tax revenue from industrial firms, total local government tax revenue, total population, nonagricultural population, and miles of paved road, are collected from the City Statistical Yearbook of China. Firm-level LGFV debt data were drawn from WIND.

5.4. Empirical model

We adopt the standard difference-in-difference (DID) framework to identify local government responses to the Action Plan.Footnote ²⁵ The baseline regression model is shown in equation (2):

(2) \begin{equation} \mathit{\log } \left(y_{ijt}\right)=\beta _{0}+\beta _{1}\textit{Tight}_{i}\times Post2013_{t}+X_{it}^{'}\alpha +\delta _{i}+\tau _{jt}+\varepsilon _{ijt}, \end{equation}

where $y_{ijt}$ is a list of (logged) outcome variable of prefecture $i$ in province $j$ and year $t$ including total LGFV debts, ratio of non-other payable debt in total LGFV debts, the amount of non-other payable debts, and the amount of other payable debts. In the mechanism analysis in Section 6.3, $y_{ijt}$ is replaced by local industrial revenue, GDP, taxation, and land sale income. $\textit{Tight}_{i}$ , as defined in Section 5.1, is a continuous variable that measures treatment intensity, or, alternatively, the tightness of the Action Plan, facing prefecture $i$ . $Post2013_{t}$ is a dummy variable that equals 1 if $t\gt 2013$ , and 0 otherwise. $\beta _{1}$ is the parameter of interest, which measures the relative change in outcome variables due to the implementation of the Action Plan for prefectures facing higher treatment intensity compared to those facing lower treatment intensity. LGFV debts are possibly higher in areas with faster urbanization and economic development, where the tightness of the action plan may also be high due to the potential reliance on the dirty industry for the fast development. To control for this confounding effect, we include in ${X}_{it}^{'}$ prefecture-level lagged total population (in log) and lagged miles of paved road (in log), control variables that are highly correlated with urbanization and economic development. The error terms $\varepsilon _{ijt}$ are clustered at province-by-year level to account for potential correlations in the error structure.Footnote ²⁶ $\delta _{i}$ is a series of prefecture fixed effects that control for the time-invariant determinants of the outcome variables.

One might argue that prefectures in the same province may share similar unobservable socioeconomic and natural conditions that simultaneously influence the outcome variables and regulation tightness, making treatment assignment endogenous. Fortunately, by construction, $\textit{Tight}_{i}$ varies across prefectures in the same province. Therefore, we include in the regression $\tau _{jt}$ , a vector of province-by-year fixed effects that the impact of any time-varying concurrent policies that happen at province level, shielding our treatment effect estimation from being contaminated by these confounding policies. In other words, we are essentially utilizing within-province prefecture-level variations in treatment intensity to identify the effects of the Action Plan on the outcome variables, which allow us to eliminate the possible province-level treatment selection problem.Footnote ²⁷ The inclusion of $\tau _{jt}$ also teases out potential confounding effects from province-level concurrent policies, such as the establishment of carbon pilots in seven provinces in 2012. In Section 6.2, we provide further evidence that that prefecture-level selection may not be a serious concern that biases our treatment effect estimation. In that section, we also try our best to directly control for potential prefecture-level sample selection issue by including the interaction between a series of pre-policy control variables and a flexible function of time as additional regressors, following Gentzkow (Reference Gentzkow2006), Chen et al. (Reference Chen, Li and Lu2018) and Gollin et al. (Reference Gollin, Hansen and Wingender2021). It turns out that the regression results remain largely unchanged.