A dual dimension

In constructing the ESI, we focus on evaluating two crucial dimensions of environmental regulations. Firstly, we aim to quantify nations’ commitment to implementing environmental policies and fostering international cooperation. Secondly, we assess the effectiveness of enforcement mechanisms and the level of dedication toward achieving established objectives, especially those related to climate change mitigation. This focus directly stems from the status of climate change as one of the major challenges of the 21st century. The Paris Agreement (2015), which constitutes the primary action taken to address this challenge, urges countries to develop long-term strategies of low greenhouse gas emissions. Thus, by centering our indicator on climate change mitigation, we align with the goals of the Paris Agreement, making it a crucial tool for assessing countries’ progress in implementing environmental policies.

To collect the data, we rely on the University of Gothenburg’s Quality of Government Environmental Indicators Dataset (Povitkina et al. Reference Povitkina, Pachon and Dalli2021). This open resource includes various environmental indicators, such as policy presence, ecological footprint, level of emissions, and public sentiment.

Implementation

Our approach to policy implementation focuses on two key aspects, encompassing both national and international commitments to combat climate change. At the national level, we quantify the number of climate-related policies currently in place, drawing on the Climate change mitigation law or policy in place ( $CCL$ ) variable (GRICCE and SCCCL, 2021). The legislative and policy landscape target the reduction of greenhouse gas emissions, employing various strategies such as emission limits, permit trading systems, and investment in cleaner technologies. Additionally, we consider legislation related to forests and land use if they specifically contribute to emission reduction and carbon capture efforts. On the international front, our analysis centers on International Environmental Agreements ( $IEA$ ) specifically linked to climate change. Using data from the International Environmental Agreements Database Project, we construct a variable, ranging from 0 to 6, that captures whether a country has signed or ratified a selected set of six key agreements. Three of these agreements were initially used by Javorcik and Wei (Reference Javorcik and Wei2003), the Convention on Environmental Impact Assessment in a Transboundary Context, the Convention on Long-range Transboundary Air Pollution, and the Convention on the Transboundary Effects of Industrial Accidents. To reflect more recent developments, we also include the Paris Agreement, the Montreal Protocol on Substances that Deplete the Ozone Layer, and the United Nations Framework Convention on Climate Change (UNFCCC). All of these agreements collectively aim to mitigate climate change through diverse approaches. The Paris Agreement and the Montreal Protocol directly target greenhouse gas emissions and ozone-depleting substances, respectively, aligning with broader climate change mitigation goals. Similarly, the UNFCCC seeks to stabilize greenhouse gas concentrations in the atmosphere. Additionally, agreements addressing environmental impacts and industrial accidents indirectly support climate change mitigation by promoting sustainable development practices and reducing pollutants that contribute to global warming.

Enforcement

Regarding countries’ enforcement, we base our approach on the study by Brunel and Levinson (Reference Brunel and Levinson2013) and develop a country emissions-based indicator. We argue that the most effective way to measure a country’s abatement efforts is by examining their actions to reduce greenhouse gas emissions, as outlined in the Paris Agreement (2015). To that extent, we compute the enforcement aspect $Enf_{yt}$ as the ratio of predicted emissions intensity of country y ( $\hat e_{yt}$ ) to actual emissions intensity of country y at time t ( $e_{yt}$ ) considering $z$ sectors:

(1) \begin{equation} Enf_{yt}=\frac {\hat e_{yt}}{e_{yt}} \quad {\rm with}\quad \hat e_{yt} = \sum _{z=1}^4 \frac {V_{yzt}}{V_{yt}} \frac {E_{zt}}{V_{zt}} \quad {\rm and}\quad e_{yt}=\frac {E_{yt}}{V_{yt}} \end{equation}

The actual emission intensity $e_{yt}$ reflects the total emissions $E_{yt}$ of country $y$ relative to its total value added $V_{yt}$ , serving as a measure of the emissions generated per unit of output. Conversely, the predicted emission intensity $\hat {e}_{yt}$ is computed as a weighted average of sector-specific emission intensities ( $\frac {E{zt}}{V_{zt}}$ ), where the weights correspond to the sectoral composition of country $y$ , that is, the share of each sector $z$ in the country’s total value added ( $\frac {V_{yzt}}{V_{yt}}$ ). Countries exhibiting higher $Enf_{yt}$ levels are regarded as stringent nations as $\hat {e}_{yt} \gt e_{yt}$ suggests that the predicted emissions intensity within their industries exceeds their current emission intensity. It is a signal of pollution reduction efforts undertaken under stricter regulations. We source sectoral emissions data ( $E_{yzt}$ ) from Climate Watch (Reference Watch2022), and sectoral value added ( $V_{yzt}$ ) from the United Nations database. To obtain country-level aggregates ( $E_{yt}$ , $V_{yt}$ ), we sum all sector values within each country while sector-level aggregates ( $E_{zt}$ , $V_{zt}$ ) are derived by summing values across all countries for each sector. Sectors are grouped into four broad categoriesFootnote ¹ to ensure consistency with emissions data. All monetary values are converted into U.S. dollars to allow cross-country comparability. Descriptive statistics for both the implementation and enforcement indices are presented in Table A1.

Z-score approach

Inspired by Kheder and Zugravu (Reference Kheder and Zugravu2012), who applied the Z-Score strategy to variables such as multilateral environmental agreements, international non-governmental organization and emission efficiency, we adopt a similar approach and standardize the data by calculating individual Z-scores for each of the three variables $s$ . This process involves subtracting the mean $\mu _{st}$ from each data point ( $X_{jst}$ ) and then dividing by the standard deviation $\sigma _{st}$ , ensuring that the variables are on a comparable scale.

(2) \begin{equation} Z_{jst}=\frac {X_{jst} - \mu _{st}}{\sigma _{st}} \end{equation}

The Z-scores for the number of climate-related policies, international environmental agreements, and the enforcement component are then summed with equal weights to create a comprehensive Z-score index ranging from 0 to 100, with 100 representing the highest level of stringency.

Descriptive statistics

With a panel of more than 160 countries, our analysis reveals notable heterogeneity in environmental policy stringency across the globe, as shown on the world map in Figure 1. Darker shades signify more stringent environmental policies, particularly evident in regions such as Europe. North America similarly demonstrates a dedication to robust regulatory frameworks. In contrast, lighter shades predominate in numerous African nations, suggesting lower levels of stringency. Similar results regarding heterogeneity were observed in the first year under analysis, 2001, as documented in Figure A1.

Regarding the evolution over the past twenty years and considering the developmental stage of each country, Figure A2 indicates a global trend towards more stringent environmental regulations. Developed economies consistently demonstrate the highest levels of environmental stringency, beginning at around 30 in 2000 and steadily increasing to approximately 55 by 2020. In contrast, emerging and developing economies show a gradual rise in their mean ESI, but the gap between them and developed economies has widened over the study period.

Lastly, it is crucial to assess whether increases and disparities in environmental stringency are driven by implementation or enforcement. Figure A3 compares these aspects across countries for the year 2018. Developed nations, especially in Europe and America, exhibit higher adoption rates and enforcement of environmental measures. In contrast, developing countries encounter challenges in both implementing and adopting stricter environmental policies. These observations underscore a critical point: by examining both implementation and enforcement, we gain a more comprehensive understanding of environmental rigor and its implications.

Figure 1. ESI by country for 2018.

Identification strategy

The gravity model, originally introduced by Tinbergen (Reference Tinbergen1962), has become a foundational tool for analyzing bilateral trade and, subsequently, FDI flows (Eaton and Tamura, Reference Eaton and Tamura1994; Wei, Reference Wei2000). It now represents the dominant empirical framework for studying FDI (Blonigen et al. Reference Blonigen, Davies, Waddell and Naughton2007, p. 1309). Our analysis builds on a theoretical adaptation of the gravity model developed by Anderson et al. (Reference Anderson, Larch and Yotov2019), tailored to bilateral FDI and incorporating FDI-specific multilateral resistance terms. Specifically, we adopt the framework of Kox and Rojas-Romagosa (Reference Kox and Rojas-Romagosa2020), who extend the model in partial equilibrium using a knowledge-capital view of FDI, where firms “lease” their proprietary knowledge across borders. Their model allows for FDI frictions that are distinct from trade frictions and relies on CES preferences to derive the structural gravity equation. The system of equations is:

(3) \begin{align}&\quad FDI_{ij}^{stock} = \omega _{ij} \frac {\alpha Y_i}{P_i} \frac {\beta Y_j}{\Pi _j} \end{align}

(4) \begin{align}& P_i = \left [ \sum _{j=1}^{N} \left (\frac {z_{ij}}{\Pi _j}\right )^{1-\sigma } \frac {Y_j}{Y}\right ]^{\frac {1}{1-\sigma }} \end{align}

(5) \begin{align}& \Pi _j = \left [ \sum _{i=1}^{N} \left (\frac {z_{ji}}{P_i}\right )^{1-\sigma } \frac {Y_i}{Y}\right ]^{\frac {1}{1-\sigma }} \end{align}

With, $Y_i$ , and $Y_j$ , economic masses of country $i$ and $j$ , $\sigma$ , the elasticity of substitution, and $P_i$ and $\Pi _j$ , multilateral resistance terms of country $i$ and country $j$ . Thus, bilateral FDI stocks depends positively of economic mass of origin and host countries, which can be proxied by their GDP levels and negatively by relative FDI friction costs (transportation and communication costs, costs related to different legal regimes or costs related to policy-made barriers).

Based on Eqs. (3)–(5), we proceed to estimate an augmented structural gravity model for inward FDI stocks utilizing panel data, expressed as follows:

(6) \begin{equation} FDI_{ijt}^{stock} = exp\left [\gamma _1 + \gamma _2 ln \left (GDP_{jt}\right ) + \gamma _3 ESI_{jt} + \beta _k\sum _{k=8}^K X_{jt} + \eta _{it} + \eta _{ij} \right ] + \epsilon _{ijt} \end{equation}

where $FDI_{ijt}$ , is related to total FDI stock of country $i$ in country $j$ at time $t$ , $ESI_{jt}$ represents the level of environmental stringency imposed by country $j$ at time $t$ , $GDP_{jt}$ is the GDP of country $j$ respectively, in time $t$ . The vector $X_{jt}$ includes additional controls that are identified as determinants of inward FDI.Footnote ⁴ $\eta _{it}$ represents origin-time fixed effects that control for multilateral resistance term of country $i$ , and thus, absorb origin size variable ( $GDP_{it}$ ). $\eta _{ij}$ are origin-host country fixed effects controlling for unobserved heterogeneity specific to each pair of countries.

Endogeneity concerns

The estimation of Eq. (6) could be subject to two primary limitations. The first limitation pertains to the “Omitted Variable Bias” (OVB), which has the potential to induce inconsistency in the estimated coefficient associated with the ESI. While the inclusion of origin-year and host country fixed effects helps mitigate this issue, time-varying host country variables may still influence inward FDI. The second limitation concerns reverse causality bias, wherein governments may adjust their environmental regulations to attract foreign investments. This creates a simultaneity bias that do not allow for a causal interpretation. To address these challenges and test for the presence of endogeneity in our setting, we rely on an instrumental variable (IV) strategy.

Our identification strategy is based on the assumption of a strategic interaction between environmental regulations among countries. Indeed, Fredriksson and Millimet (Reference Fredriksson and Millimet2002) reveals that the stringency of environmental policies of U.S. states is influenced by those of both contiguous and regional neighbors. Other studies have demonstrated a “race to the bottom” strategic interaction in environmental regulations among provinces in China (Zhang et al. Reference Zhang, Zhang and Liang2017; Song et al. Reference Song, Zhang and Zhang2021). Thus, our identification assumption is that environmental regulation tends to diffuse regionally with a delay, due to institutional convergence, political emulation, or shared regulatory frameworks, and that lagged regional policy does not affect current inward FDI flows except through its influence on domestic policy. Thus, our IV strategy is based on the average ESI in a particular region, which leaves out the country, $j$ , under scrutiny. This variable is defined as follows:

(7) \begin{equation} I_{jt}=\frac {1}{N_r - 1} \sum _{k \in r, k \neq i} ESI_{kt} \end{equation}

where $r$ represents one of the eight regions retained,Footnote ⁵ and $N_r$ denotes the number of countries in that region. Such measure has been introduced by Acemoglu et al. (Reference Acemoglu, Naidu, Restrepo and Robinson2019) to instrument regional waves of democracy and used by Chiappini and Gaglio (Reference Chiappini and Gaglio2024) for waves of digitalization. We employ the second lag of the variable $I_{jt}$ to instrument the ESI in Eq. (6) following the findings of Fredriksson and Millimet (Reference Fredriksson and Millimet2002), which suggest that the impact of neighboring environmental policies operates within a five-year timeframe.Footnote ⁶

The IV strategy requires that the instrument relevance assumption is fulfilled, meaning that the instrument must be correlated with the endogenous variable. This assumption can be tested using an F-statistic test. Furthermore, the IV strategy imposes the exclusion restriction assumption, which stipulates that the ESI of country $j$ ’s neighboring countries affects country $j$ ’s level of inward FDI exclusively through its effect on country $j$ ’s level of ESI. While this assumption is not directly testable, the strategic interaction between environmental regulations implies that a change in the stringency of environmental policies in economies in the same region directly influences country $j$ ’s environmental regulation. However, it is plausible that an external shock in country $j$ ’s neighboring countries within the same region impacts both inward FDI and ESI in country $j$ . To mitigate concerns that regional macroeconomic conditions might confound the instrument’s validity, we include region fixed effects and control for time-varying regional GDP. This specification absorbs both time-invariant regional characteristics and contemporaneous regional shocks, while year fixed effects capture global regulatory and investment trends. A possible limitation of this strategy is that multinational firms may respond directly to the broader regional regulatory environment, anticipating future harmonization or reacting to perceived policy spillovers. If such responses occur independently of domestic regulation, the exclusion restriction may be imperfectly satisfied, and thus results cannot be interpreted as causal but rather as correlation. However, to the extent that our fixed effects and controls account for common regional and global shocks, we argue that the remaining variation in the instrument is plausibly exogenous.

Note that even if the two previous assumptions are verified, the IV strategy only identify a local average treatment effect (ATE)(Imbens and Angrist, Reference Imbens and Angrist1994) rather than an ATE. Indeed, IV regressions only identify an ATE for complying countries, i.e., countries that are affected by the instrument. Although it is highly unlikely that the effect in complying countries is different from the average effect, we cannot test this.

Estimation method

Equation (6) is derived from an adaptation of the theoretical model of Anderson et al. (Reference Anderson, Larch and Yotov2019). This theoretical model holds both for intensive and extensive margins. Thus, zero FDI stocks are accounted for. Consequently, following standard practice in the trade literature, we rely on the PPML estimator developed by Santos Silva and Tenreyro (Reference Santos Silva and Tenreyro2006) to estimate Eq. (6). This estimator has the advantage of handling zero-value observations without the need for data transformation, thereby avoiding potential biases that may arise from omitting or transforming them. Given that our database contains more than 100,000 zero FDI values, this methodology is the most suitable.

Note that the IV strategy involves relying on a two-stage least squares (2SLS) estimation. In the first stage, the variable $ESI_{jt}$ is regressed on the excluded instrument $I_{jt-1}$ , the other covariates $X_{jt}$ , and the fixed effects using the OLS estimator. Then, in the second-stage regression, the fitted values of $ESI_{jt}$ obtained in the first stage are used to estimate Eq. (6). However, this method cannot be used with the PPML estimator as it is subject to the incidental parameter problem in this case (Anderson and Yotov, Reference Anderson and Yotov2020). To address this issue, Lin and Wooldridge (Reference Lin, Wooldridge, Lin and Wooldridge2019) proposed the use of a control function. This procedure involves obtaining the residuals from the first-stage estimation and introducing them into the second-stage regression, i.e., the gravity model described in Eq. (6), and estimating it using the PPML estimator with bootstrapped standard errors. Subsequently, Lin and Wooldridge (Reference Lin, Wooldridge, Lin and Wooldridge2019) argue that a robust Wald test of the null hypothesis that the coefficient on the first-stage residuals is equal to zero can be conducted to evaluate the absence of idiosyncratic endogeneity.

Foreign direct investment data

Navigating the intricate dynamics of global economic interactions presents significant hurdles when striving for accurate FDIs measurements, particularly when analyzing bilateral data. The FDI dataset used in this study originates from the meticulous efforts of Broner et al. (Reference Broner, Didier, Schmukler and von Peter2023), who curated data from the Bilateral FDI Statistics provided by the United Nations Conference on Trade and Development (UNCTAD). In their work, Broner et al. (Reference Broner, Didier, Schmukler and von Peter2023) employed a methodological approach aimed at enhancing the dataset’s comprehensiveness. This approach involved leveraging assets reported by the source country and liabilities reported by the destination country. By adopting such a strategy, maximum coverage was ensured, even in cases where source countries did not report their asset holdings. Moreover, in situations where both source and destination countries reported data, Broner et al. (Reference Broner, Didier, Schmukler and von Peter2023) implemented a meticulous reconciliation process, elaborated upon in the Appendix A of their study. In our specific analysis, we focus on bilateral FDI stocks, expressed in 2011 million U.S. dollars, to capture the magnitude of foreign investments in a comprehensive manner. The period from 2001 to 2018 corresponds to the timeframe for which bilateral FDI data are available in Broner et al. (Reference Broner, Didier, Schmukler and von Peter2023) dataset. This period is also relevant to our research question, as it coincides with a phase of accelerated growth in global FDI flows (see Figure B1) and the reinforcement of environmental regulation (see Figure A2), marked by major international agreements such as the Kyoto Protocol (2005), the Copenhagen Accord (2009), and the Paris Agreement (2015). Ending the analysis in 2018 also avoids the structural break introduced by the COVID-19 pandemic, which could have generated confounding effects unrelated to environmental policy.

Descriptive statistics on inward FDI are presented in Appendix B, including information on the sectoral composition of FDI and average emissions by sector. These data are intended to address potential concerns that the observed pollution haven effect in aggregate FDI may be disproportionately driven by a small number of highly polluting industries. Although our primary analysis relies on aggregate bilateral FDI flows, we supplement it with sector-level data from a separate World Bank dataset covering the period 2008–2019 (Steenbergen et al. Reference Steenbergen, Liu, Latorre and Zhu2022), which is based on comparable sources. This dataset reveals that inward FDI is distributed across both low- and high-emission sectors (see Figures B3 and B4). For instance, financial and insurance activities account for a substantial share of total FDI, while more pollution-intensive manufacturing sectors—such as chemicals, rubber, and plastics—also receive a notable portion. At the same time, relatively cleaner industries, including food processing and electronics, likewise attract significant investment.

Agreements

We rely on the Electronic Database of Investment Treaties to account for both bilateral and multilateral investment treaties (BITs) that have come into force during the study period. Investment treaties are critical in the context of FDI determinants as they provide a stable and predictable legal framework for investors, reducing the risks associated with investing in foreign markets. These agreements often include provisions for the protection of investments, dispute resolution mechanisms, and guarantees of fair treatment, which can enhance investor confidence and encourage cross-border investment flows.

Relative endowments

To measure relative endowments, we employ the capital-to-labor ratio (in logarithm). This metric is computed using data sourced from the World Bank’s World Development Indicators, specifically drawing upon gross capital formation expressed in dollars and the total labor force. The findings align with theoretical expectations, as evidenced by developed countries displaying higher ratios, indicative of their comparatively greater abundance of capital.

Natural resources rents

We use total natural resources rents, as well as specific measures for oil and mineral rents, all expressed as percentages of GDP and sourced from the World Bank database. By incorporating these variables, we aim to analyze how the dependence on natural resources affects affects FDI patterns, through the resource-seeking motive.

Quality of institutions

The quality of institutions is assessed using the Control of Corruption index from the Worldwide Governance Indicators database. This index evaluates how much public power is misused for personal benefit, including small-scale corruption, large-scale corruption, and the manipulation of state policies by powerful individuals and groups. Scores range from -2.5, indicating high levels of corruption, to 2.5, indicating very clean governance.

Wages

Within our analysis, we use the “Labor Regulations and Minimum Wage” variable sourced from the Fraser Institute’s Economic Freedom Index. This variable is assessed on a scale of 1 to 10, where higher scores indicate fewer regulatory barriers related to hiring procedures and minimum wage legislation, reflecting greater labor market flexibility. This tends to reduce labor costs and enhance a country’s attractiveness for investment.

Tariffs

To measure the trade openness of each country, we rely on the Tariffs section of the Fraser Institute’s Economic Freedom Index, which combines trade tax revenue, mean tariff rate, and the standard deviation of tariff rates. This section provides a comprehensive assessment ranging from 0 (indicating high protectionism) to 10 (indicating high openness).

Market access

To assess the freedom to enter markets and compete within each country, we rely on the corresponding component from the Fraser Institute’s Economic Freedom Index. This component evaluates variables such as market openness, ease of obtaining business permits, and the level of distortions in the business environment. Scores on this index range from 0 (indicating a closed market with high entry barriers) to 10 (indicating an open market with minimal barriers).

The detailed description of all variables used in this analysis can be found in Table C2, while summary statistics are provided in Table C3.

ESI and EPS

In columns (1) and (2) of Table D4, we compare the estimation results using the EPS as an alternative to the ESI to measure environmental stringency. This index, described in Section 3.3, is widely used in the empirical literature. As with Table D1, we report standardized coefficients in Appendix Table D5 to ensure comparability across specifications. We therefore estimate the augmented structural gravity model using both indicators on the same sample of countries. Column (1) presents the results for the ESI, while column (2) presents the results for the EPS. The results are quantitatively and qualitatively similar in both regressions, confirming the reliability of our new indicator as a measure of environmental stringency.

Alternative measures of enforcement

In column (3), we address the enforcement aspect of environmental policies using a different measurement approach. Specifically, we apply the EPI scoring methodology to recalculate the climate change mitigation category using the weights assigned in the 2022 Technical Annex. Key variables in this analysis include the growth rate of $CO_2$ , projected GHG emissions in 2050, and the growth rate of F-gases. As with the previous indicators, we standardize this enforcement index to ensure comparability (see Appendix Table D5). The results of this alternative method, reported in column (3), are consistent with our previous findings (see column (1) of Table 3), further confirming the insignificance of government enforcement efforts. In column (4), we refine our enforcement index by calculating average emissions per sector for different groups of countries, taking into account regional differences in emission intensity. This adjustment confirms the robustness of our results, as no significant changes in the results are observed.

Leads and lags

We also test the robustness of our results to the inclusion of leads and lags on the ESI variable in the structural gravity equations. We add two leads and two lags and estimate the structural gravity model as an event study design. Figure D1 shows the estimation results. We find no significant impact of the two leads of the ESI variable on FDI, suggesting no anticipatory effects of environmental policies for FDI. Furthermore, the two lags are significant and negative, confirming our previous results and revealing a persistent effect of environmental policies on FDI.