2.1. Explainability and predictability of ESG ratings

Several studies examine the determinants, divergence, and prediction of ESG scores. Berg et al. (Reference Berg, Kölbel and Rigobon2022) examined the divergence in ESG ratings across six leading providers (KLD, Sustainalytics, Moody's ESG/Vigeo-Eiris, S&P, Refinitiv, and MSCI), showing that divergence mainly stems from how ESG is measured, with additional contributions from scope and weighting. It also identifies a ‘rater effect’, in which each agency's overall assessment of a firm influences its scoring of specific ESG categories, underscoring the need for greater transparency in ESG data and methodologies. Del Vitto et al. (Reference Del Vitto, Marazzina and Stocco2023) applied machine-learning techniques to replicate and interpret the ESG rating process of Refinitiv, focusing on firms from the United States, European countries, and China. They employed both fully explainable (white-box) models and more obscure (black-box) models; the authors demonstrated that the proprietary rating methodology could be approximated with high accuracy. The use of interpretability tools further identified the most influential variables driving ESG ratings, providing insights into the otherwise opaque assessment process.

A series of studies using machine learning techniques investigated the predictability of ESG ratings using fundamental data and past ESG track record. García et al. (Reference García, González-Bueno, Guijarro and Oliver2020) employed a novel modeling approach to forecast ESG ratings. They investigated the relationship between ESG scores, fundamental and market-based financial performance indicators, and industry sectors for publicly traded European firms over 2013–2018. By conducting 500 simulations across varying discretization parameters and grouping scenarios, their analysis evaluated the predictive capacity of financial and industry variables for ESG classification. The results demonstrated that these variables effectively differentiate firms when ESG scores are clustered into broad categories, but their predictive accuracy declines as the granularity of ESG groupings increases.

In a similar vein, D'Amato et al. (Reference D'Amato, D'Ecclesia and Levantesi2021) assessed the extent to which fundamental financial variables predicted Bloomberg ESG scores for firms in the Euro Stoxx 600 over the past decade. Utilizing a machine-learning framework based on the random-forest algorithm, the authors demonstrated that financial statement variables possessed substantial predictive power, indicating that traditional accounting data can serve as effective determinants of ESG ratings. This work on predicting ESG events using financial variables casts a causal shadow over the literature that holds that ESG is driving financial performance. Taskin et al. (Reference Taskin, Sariyer, Acar and Cagli2025) examined the prediction of future ESG performance using machine learning methods for firms in the BIST, the main Turkish stock exchange, sustainability index in Turkey. Applying decision tree, random forest, k-nearest neighbor, and logistic regression models, the authors demonstrated that past ESG scores and grades served as reliable predictors of future ESG ratings. The findings highlighted that simple score-based indicators can effectively forecast ESG performance, offering practical tools for sustainable finance in emerging markets and broader applications in international financial contexts.

2.2. Prediction of ESG incidents across and within sectors

The bulk of studies around ESG incidents is centered around the impacts on asset prices (Groen-Xu & Zeume, Reference Groen-Xu and Zeume2021; Kölbel et al., Reference Kölbel, Busch and Jancso2017; Krüger, Reference Krüger2015; Orpiszewski & Thompson, Reference Orpiszewski and Thompson2025), financial performance (DasGupta, Reference DasGupta2022; Jucá et al., Reference Jucá, Muren, Valentinčič and Ichev2024), or integration of ESG controversies into risk management and investment strategies (Iwata & Weibel, Reference Iwata and Weibel2024). We have identified several studies on the prediction of ESG controversies across sectors, and in this section we refer to the most relevant ones.

Zheng and Yan (Reference Zheng and Yan2022) proposed a logit-based incident prediction model to test whether ESG ratings can reduce headline risk. Using RepRisk data from 2007 to 2021, they showed that past incidents and firm fundamentals strongly predicted future ESG controversies, while policy- and disclosure-based ratings often failed, suggesting greenwashing concerns. Their model outperformed third-party benchmarks with considerably higher accuracy, offering investors a more reliable tool for ESG risk management. Svanberg et al. (Reference Svanberg, Ardeshiri, Samsten, Öhman, Neidermeyer, Rana, Svanberg, Öhman and Lowe2023) implemented a series of machine learning–based methods for predicting ESG controversies using fundamental, market, and ESG rating data as model inputs. The set of models predicted which companies were likely to face ESG controversies with 70–84% accuracy and identified the most influential risk indicators.

Shakil et al. (Reference Shakil, Pollestad and Kyaw2025) employ a machine-learning approach to show that ESG controversies significantly increase firms’ systematic risk, reinforcing the view that controversy incidence is a financially material risk factor. Seow (Reference Seow2025), in a systematic literature review of machine-learning applications in ESG analytics, similarly emphasizes that controversy prediction and incident-based risk measures are becoming central themes in ESG research.

Dipierro et al. (Reference Dipierro, Barrionuevo and Toma2025) investigated ESG-related controversies (ESGC) in the banking sector using a panel of 140 banks from 2011 to 2020. Applying machine learning feature selection methods to 28 variables, the authors showed that governance–employee dynamics were the strongest predictors of ESGC risk, surpassing operational, and financial factors. The findings highlighted the value of ML for ESG risk prediction and offered implications for regulatory technology and risk governance. Lheureux (Reference Lheureux2024) took a similar approach, exploring the use of Twitter-based sentiment and machine learning to predict ESG controversies.

Sun et al. (Reference Sun, Luo, Yiu, Yu and Ding2024) examined the relationship between ESG scores and ESG scandal probability. They employed low-frequency ESG Ratings from MSCI ESG, Refinitiv ESG, Sustainalytics, and S&P Global, as well as ESG Incidents collected by RepRisk. The authors attempted to predict the probability of occurrence of an ESG incident as a function of ESG Score, divergence of ESG Scores between agencies, and a set of company fundamental ratios. They also calculated the cumulative stock abnormal returns around event dates using the Capital Asset Pricing Model. Their empirical evidence shows that higher ESG scores are linked to lower probabilities of ESG scandals and weaker negative stock reactions when scandals occur, suggesting that ESG scores provide meaningful signals for future reputational and financial risk.

Further empirical work has examined ESG controversies through specific institutional and technological channels. Agnese et al. (Reference Agnese, Cerciello, Giacomini and Taddeo2023) analyze European banks and show that board structures and governance arrangements are systematically related to banks’ ESG controversy scores, highlighting the role of internal governance in shaping incident risk. Ayadi and Hunjra (Reference Ayadi and Hunjra2026) document that, in periods of economic uncertainty, greater digital integration reduces information asymmetry and is associated with fewer ESG controversies, suggesting that digital disclosure can mitigate controversy risk. Related strands of research link institutional and informational environments to environmental and governance outcomes, including work on economic freedom and green growth or carbon emissions (Hou et al., Reference Hou, Jin and Feng2025), and the role of financial reporting in governance following scandals (Iqbal, Reference Iqbal2025).

There are numerous domain-specific studies on severe incidents that are worth mentioning, though we do not intend to delve further into each theme. Beckmann and Rogmann (Reference Beckmann and Rogmann2024) analyzed ESG incidents in the energy sector across 150 countries. They found that the number of ESG controversies was associated with the level of wealth in a country and with policy-uncertainty metrics, economic growth, and fossil fuel use.

In a meta-study on the oil and gas sector, Al-Sabaeei et al. (Reference Al-Sabaeei, Alhussian, Abdulkadir and Jagadeesh2023) employed machine-learning techniques to assess oil and gas pipeline failures. In the mining sector, several studies compiled historical datasets and explored the mine tailings dam failures using local dam parameters (Concha Larrauri & Lall, Reference Concha Larrauri and Lall2018; Piciullo et al., Reference Piciullo, Storrøsten, Liu, Nadim and Lacasse2022). Finally, in a recent meta-study of over 125 empirical studies, Bel-Oms (Reference Bel-Oms2025) undertook a systematic literature review of ESG controversies. The analysis yielded a conceptual framework identifying the origins and consequences of ESG controversies. Their findings highlight that explainability and predictability of ESG incidents regarding country and cultural differences remain areas for future research.

Research gap: Interestingly, the majority of empirical studies employed global ESG controversy data and, logically, controlled for country fixed effects or clustered the countries, however, to date, none of the studies explored the role of the country of incident occurrence, or country of headquarters, in the prediction, considering both company and country characteristics in a uniform setup. This deficit is glaring when one considers that each country has its own institutional particularities, backed by a monopoly of violence on its own territory to enforce rules and regulations. For this reason, we try to respect this reality by incorporating both sets of institutional rules into the baseline hazard rate for ESG events. That said, the relationship between state institutions and ESG events is theoretically ambiguous: states with more rules may lead to more ESG events; states with poor enforcement or few rules may lead to fewer adverse events if there are fewer norms to violate.

3.1. Data

3.1.1. ESG incident data collected by reprisk

For over twenty years, RepRisk has systematically collected information on companies’ adverse incidents from diverse sources such as local and international media outlets, government and Non-gonvernmental Organization (NGO) reports and websites, newsletters, blogs, and, to a limited extent, social media. Machine-learning algorithms identify these incidents, which are then reviewed and validated daily by human analysts. Each incident is linked to potentially impacted companies and categorized under the relevant ESG dimension, with some cases spanning multiple ESG areas. Incidents are further classified by their novelty, severity, and reach. All entries are timestamped according to the date provided by the source, which may differ from the actual date of occurrence.

In this study, we employed the global dataset between 2008 and 2023. The mining and materials companies covered are headquartered in 67 countries. Table 1 presents the number of companies per sector. There seem to be different reasons for the event counts. Financials seem to be involved in a disproportionate number of events, perhaps due to the number of loans and deals and general banking regulation, whereas real estate is large in terms of value in the economy, but the operations tend to be very low-salience and local. The industry-specific attributes probably drive more of the event reporting than how irresponsible any industry actually is.

Table 1.

Summary statistics of ESG incidents collected by RepRisk

Table 1 Long description

Sector-level counts pair the number of companies with the number of recorded ESG incidents. Missing or Other is the largest category, with 7,092 companies and 82,604 incidents. Financials has 838 companies and 63,651 incidents, the second-highest incident total despite far fewer companies than Missing or Other. Mid-range incident totals include Materials at 31,192, Industrials at 28,523, and Consumer staples at 28,161. Utilities and Energy follow with 15,458 and 14,169 incidents, while Healthcare and Information technology are around eleven thousand incidents each. Consumer discretionary and Communications are lower at 8,514 and 6,035 incidents, and Real estate is minimal with 14 companies and 52 incidents. Overall totals are 10,687 companies and 300,288 incidents; comparisons should be read with caution because the Missing or Other grouping is broad and sector mapping may be incomplete.

Note: We mapped the ISINs for US, European, and Japanese stocks that are included in the major indices. RepRisk's internal classification was mapped to sectors where available.

The meta-information registered for each ESG incident includes the location of the incident, the language, and the number of companies involved in the incident. In terms of location of incidents, Table 2 illustrates that the bulk of ESG incidents are reported in the US, followed by China, Britain, India, and the large European countries. Surprisingly, Switzerland has the same number of incidents as Japan, although Japanese Gross Domestic Product (GDP) in USD is substantially larger than the Swiss GDP. The number of events for a country or sector is idiosyncratic, depending on a myriad of factors such as the size of the country, press coverage, or the size of the capital market; for example, Russia has a relatively low number of incidents compared with Australia given its size. This means that event counts are a product of both the actual incidents and various forces like investors’ or national institutions’ interest in equity.

Table 2.

Top 10 countries with the largest count of ESG incidents

Table 2 Long description

Countries and their total count of ESG incidents are listed, along with each countrys share of all recorded incidents. The United States ranks first with 23,217 incidents, representing 7.73 percent of all incidents. China is second with 10,211 incidents and 3.40 percent, followed by the United Kingdom with 7,678 and 2.56 percent. India and Brazil are close, with 6,454 and 2.15 percent versus 6,332 and 2.11 percent. Canada, Indonesia, France, Germany, and Australia each contribute between 1.56 percent and 1.86 percent, with totals from 4,693 to 5,598. Additional entries shown include Russia at 4,323 and 1.44 percent, Japan at 3,648 and 1.21 percent, and Switzerland at 3,500 and 1.17 percent. Overall, the distribution is top heavy, with the United States contributing more than double Chinas incident count and a noticeably larger share than any other country. Counts and shares reflect recorded incidents in the underlying dataset and may be influenced by reporting coverage and monitoring intensity across countries.

Note: RepRisk incident data.

It is noteworthy that RepRisk registers the incidents in the database using the date of publication of the report and not the date when the actual incident occurred. As a result, at times there are several incidents registered per day. Also, one registered incident can involve more than one company; at times, several companies from the sector are involved. Several studies, such as (Orpiszewski et al., Reference Orpiszewski, Thompson and Schwendner2024), provided a detailed description of this phenomenon and the adjustments necessary to correctly calculate the cumulative abnormal returns around the ESG event dates. In this study, we are interested in the count and rate of occurrence of ESG incidents and hence do not intend to undertake any adjustments to the raw data. Figure 1 shows that the early years were about including more events, ramping up to a peak number of events in 2017, and since then it looks like RepRisk may be truncating coverage in some manner, with some variability around the ∼19,000 events.

Figure 1.

Events by year.

2. Explanatory variables

Table 3 defines the variables used below. The country-level indicators are based on recognized global benchmark variables. The indicators were selected to align with the internal classification from RepRisk. Some of these are used by Axelrod and Leitner (Reference Axelrod and Leitner2016), who provided an empirical, peer-reviewed proof that sovereign-level social and governance datasets contain financially material signals for stock market performance. The twist here is that we are looking at the actual ESG news events for those equities, in part to ascertain whether the positive impact on equity performance is mediated by these news flows – i.e. whether ESG is the causal channel.

Table 3.

Country institution indicators

Table 3 Long description

Eleven country institution indicators are defined to assess human rights, labor conditions, governance, corruption risk, environmental health services, democracy, and press freedom. For each indicator, an expected sign is provided, indicating whether higher values are generally treated as better, worse, or context-dependent, plus a short description of the data source and what it measures. Most indicators are described as rescaled to a zero-to-one range, though the corruption measure notes that higher scores indicate less corruption before scaling. Expected signs vary: negative for child labor, rule of law, bribery and corruption, labor rights issues, and water stewardship; positive for forced labor prevalence, money-laundering risk, workplace safety injuries, and press freedom; and ambiguous for human rights and liberal democracy. Several measures are composite indices built from multiple inputs, such as corruption perceptions, money-laundering vulnerability, and labor rights scoring. Interpretation should be cautious because indicators differ in methodology, year, and whether they capture outcomes, risks, or perceptions rather than directly observed events.

Note: Human Rights Measurement Initiative https://humanrightsmeasurement.org/, The Global Slavery Index https://www.walkfree.org/global-slavery-index/, Children's Rights in the Workplace Index https://www.unicef.ch/en/what-we-do/national/publications-teaching-materials/childrens-rights-workplace-index, World Justice Project https://worldjusticeproject.org/rule-of-law-index/, Transparency International https://www.transparency.org/en, Basel Anti–money laundering, https://baselgovernance.org/resources/basel-aml-index/, Labor Rights Index https://wageindicator.org/work/labour-law/index/, V-Dem https://www.v-dem.net/, Reporters Sans Frontiers https://rsf.org.

The choice of these country-level determinants is consistent with institutional and information-based perspectives on ESG controversies. Institutional theory suggests that legal quality, labor protections, corruption control, democratic accountability, and media freedom shape both firms’ incentives to comply with social and environmental norms and the probability that violations are detected and publicly reported. This is particularly relevant for variables such as rule of law and press freedom: stricter rule of law should improve enforcement and deter misconduct, whereas greater press freedom should increase the visibility of adverse events and reduce selective underreporting.

However, in our context, there is no empirical prior on whether strong democracy leads to more rules to violate, or better behavior by the companies that operate. There are no strong empirical priors on the variables because they can imply two opposite effects: on the one hand, strong institutions encourage better corporate behavior like in Switzerland, but on the other hand, those institutions also induce more critique of companies where expectations of corporate propriety are set higher, so by definition more violations occur. In theory, one could bridge the work here to develop priors if one has very strong beliefs about returns and ESG. That is to say, if better ESG behavior causes higher returns, the return literature can be used conversely to establish a prior metric. Axelrod and Leitner (Reference Axelrod and Leitner2016) find a positive correlation between equity returns and the same Liberal Democracy Indicator we use; thus, we would expect fewer ESG events if and only if better ESG behavior leads (i.e. fewer adverse RepRisk events) to higher equity returns.

Because the dependent variable is the number of reported ESG incidents, the expected signs reflect both underlying misconduct and the likelihood that incidents are detected and reported. Stronger institutions may reduce misconduct, but they may also increase transparency and public scrutiny. We therefore expect rule of law, lower corruption, stronger labor protections, child-labor safeguards, and water stewardship to be negatively associated with reported incidents. By contrast, press freedom is expected to have a positive sign because it increases visibility and reporting. Liberal democracy and human-rights protection are treated as theoretically ambiguous because they may both improve corporate conduct and increase the public detection of violations.

Like the variables above, the variables in Table 4 below are country-based. We separated them conceptually because they are intended to be control variables, independent of any institutional assumptions. For example, population and GDP control for size, in that one would expect that bigger countries have more events.

Table 4.

Control variables

Table 4 Long description

Control variables used in an analysis are listed, giving each variables expected direction of association and its data source with a brief definition. Variables expected to have a positive association are population, GDP, and market capitalization; each is taken from World Bank or Refinitiv data and transformed using the natural logarithm. Event country and event year come from RepRisk and identify where and when an event occurred, with no expected direction stated. S and P sector and headquarter country come from Refinitiv and classify the firm by industry and legal headquarters location, also with no expected direction stated. The content is descriptive rather than reporting values, so it does not show magnitudes, statistical significance, or time trends. Expected signs are assumptions for modeling and may differ from estimated results depending on the outcome definition and sample.

We also expect some baseline industry differences but are largely agnostic about which industries and why they might have more adverse effects. We therefore expect control for industry to account for potential heterogeneity in the distribution of firms across countries. For example, a large share of real estate companies in our sample is located in Switzerland, a country with strong institutions. Without industry controls, the relatively low incidence of ESG events among real estate firms could be incorrectly attributed to Switzerland's institutional environment rather than to industry-specific characteristics.

In Figure 2, we show Pearson's correlation between the major country-level indicators that might help predict events in color, and their significance, in p-values, is likewise shown. The loadings, while related, do not seem to exhibit excessively high correlations (>0.9) that would indicate they are measuring the same latent country factor. Money laundering is a type of rule of law issue, and hence has a high correlation, but given that some events are violations of domestic legislation, it did not seem wise to proxy rule of law with money laundering. Furthermore, in the gold space, money laundering is a specific concern that warrants its own metric for communicating with stakeholders. All the correlations on the diagonal are significant by construction, and those with no absolute correlation tend also toward no significance.

Figure 2.

Correlation matrix for explanatory variables.

3.2. Methodology

The RepRisk incident data were aggregated by event country (c), company (i), and year (t). This aggregation yields a count variable of events related to a particular company having occurred in a given country and year. The most common and appropriate distribution to model count data within a given period is the Poisson distribution, for example, Passas et al. (Reference Passas, Ragazou, Zafeiriou, Garefalakis and Zopounidis2022) analyze sociopolitical determinants of ESG controversies in EU firms using Poisson-based specifications.

Given that the mean event count is about 2 per year, but the sample variance is about 17, and a Poisson model is only appropriate when the first and second moments are equal, a quasi-Poisson distribution is more appropriate. Accounting for t is done by extending the model with a dispersion parameter. To identify the country-level and firm-level effects simultaneously, we use a hierarchical generalized Poisson model where each firm is active in a set of {c_i}. Each firm has a vector of events occurring in different countries, with both local-event etiology and firm-level effects.

In the generalized linear Poisson model, the expected number of events (Y) occurring in a country (c) for a firm (i) in year (t) is modeled using the firm attributes (F), and country indicators (I) using the following expression:

(1)\begin{equation}\ln \left( {Ex\left\{ {{Y_{i,c,t}}\,\left| {{F_{i,t,}}\,\left. {{I_{c,t}}} \right|} \right.} \right\}} \right)\, = \,\alpha + {\varphi ^T}{F_{i,t\,}}\, + \,{\iota ^T}{I_{c,t}}\end{equation}

Our generalized linear Poisson specification relies on standard generalized least squares assumptions for count data. First, conditional on the covariates, firm–country–year observations are assumed to be independent, implying no remaining serial or cross-sectional autocorrelation in the residuals once the model terms are included. Second, the log-link function is assumed to be correctly specified such that the conditional mean of the ESG incident count is an exponential function of the linear predictor. Third, a conventional Poisson model assumes equidispersion (equality of the conditional mean and variance); in our data, the conditional variance substantially exceeds the mean, so we explicitly account for overdispersion through a quasi-Poisson specification with an estimated dispersion parameter. Finally, we assess overall model fit and overdispersion by comparing residual deviance to the residual degrees of freedom and by inspecting Pearson residuals.

In the second model, we estimate a Bayesian hierarchical negative binomial count model, where the observed number of events generated in the equation above is censored by press freedom. After the underlying process generates an event, the press reports on this event, hence, there is a truncation of this reporting in countries where freedom of the press is limited.

(2)\begin{equation}ln(\mu )\, = \,\ln (\cdot )\,-\,\pi \,pres{s_{c,t}}\,\end{equation}

where µ = exp(∙ −π press_c,t) and y = negativeBinomial (µ, ϕ).

The Bayesian hierarchical negative binomial model extends this framework by allowing the mean–variance relationship to depart flexibly from equidispersion and by introducing random-effects terms that capture unobserved firm- and country-level heterogeneity. The key assumptions in this setting are that, conditional on the covariates and random effects, counts are independent draws from a negative binomial distribution (conditional independence), that the specified hierarchical structure appropriately captures residual clustering, and that the chosen prior distributions for regression coefficients and dispersion parameters are weakly informative but not overly restrictive. As in the quasi-Poisson case, we examine posterior predictive checks and residual diagnostics to verify that no major autocorrelation or systematic lack of fit remains. In the next section, we shall review the results from these two models.

4.1. Test of metrics

Table 7 below shows the results of a negative binomial model estimating the event counts by category. Each generalized linear regression includes country-level indicators for the event category.

Table 8 shows the primary event regression broken down by S&P sector. (Real estate regression is not shown because it is barely covered in the RepRisk dataset.) While country-level fixed effects have been maintained, headquarters-country effects have been elided for concision. What we learn from Table 8 is that, at the country level, effects like liberal democracy or the RSF index show little difference in the industry's ESG response. Almost all the coefficients are near 0 for the RSF index. Though if we were to adjust the standard errors for multiplicity, repeated tests of an industry would indicate that most of the differences are spurious, albeit in the case of real estate both the GDP and RSF index would reject the two-sided Bonferroni-adjusted null of 0.005; one hypothesis might be that real estate companies are intrinsically tied to their local and thus more heavily influenced by the country-level variables, but small sample bias is a more likely theory why real estate is an outlier.

Table 8.

Event counts by sector. Dependent variable: annual number of ESG controversies per firm–country–year

Table 8 Long description

Sector-specific regressions relate the annual count of ESG controversies per firm to country to year to calendar year, firm size (log market cap), country population and GDP, liberal democracy, and the RSF Index. The year coefficient is positive and statistically significant for industrials, consumer discretionary, consumer staples, health care, communications, and information technology, indicating rising controversy counts over time in those sectors; it is not significant for materials, financials, or energy, and is only weakly positive for utilities. The largest positive time association appears in consumer discretionary and communications, while industrials and consumer staples also show clear increases. Firm size is associated with fewer controversies in industrials and health care, but with more controversies in materials, consumer discretionary, consumer staples, and utilities; it is not clearly related in financials, energy, communications, or information technology. Country population and GDP show mixed patterns, with notably positive and significant effects in consumer staples and some positive associations in communications and utilities, while information technology shows a negative GDP association. Liberal democracy is not consistently related to controversy counts across sectors, with no clear statistically significant pattern. The RSF Index is positively associated with controversy counts in consumer discretionary, consumer staples, and information technology, and is weakly positive in financials, but shows no clear relationship in several other sectors. Sample sizes vary widely by sector, from about six thousand observations in communications to over sixty-three thousand in financials, so precision and comparability differ across columns.

Note: Standard errors are reported in parentheses and clustered by ISIN and country; Signif. codes:

* p < 0.10,

** p < 0.05,

*** p < 0.01.

That larger firms face more recorded ESG incidents is consistent with recent global evidence that business complexity and scale are positively associated with ESG controversies, partly because more complex firms operate in a greater number of jurisdictions and face more intensive analyst and media scrutiny. Studies of bank ESG controversies similarly find that governance structures at large financial institutions are tightly linked to controversy risk.

In this last Bayesian hierarchical model presented in Table 9, we try to model the observed events as conditional on the amount of press freedom in the country, again as measured by the RSF Index. We use an uninformative normal prior for all the parameters, except for the negative binomial outcome where we used a gamma prior for the success probability hyperparameters, which were parametrized with an initial distribution to obtain the mean number of events as the unconditional null.

Table 9.

Conditional event count model (negative-binomial). Dependent variable: annual number of ESG controversies per firm–country–year

Table 9 Long description

Negative-binomial regression results relate firm and country factors to the annual count of ESG controversies per firm-country-year. The RSF Index has a negative coefficient of about minus 0.016 with strong statistical support, indicating fewer controversies as the index increases. Year has a small positive coefficient of about 0.01 with marginal statistical support, suggesting a slight upward time trend. Firm size, measured as log market capitalization, is negative at about minus 0.24 but not statistically clear. Country economic size, measured as log GDP, is positive at about 0.09 and not statistically clear. Liberal Democracy is positive at about 0.20 and not statistically clear. The intercept is about 0.33. Country and headquarters-country dummy effects are included but not displayed, so some cross-country differences are absorbed by those controls.

Note: N_eff = 8; (s.e. with implied p-value significance codes:

* p < 0.10, **p < 0.05,

*** p < 0.01).

This model shows the influence of censoring and press coverage as evidenced by the country, which enters as a multiplicative term. The effect is minimal, reducing the number of events by about 1.6%. In theory, more press freedom ought to lead to more reporting, but the model may just be picking up on some latent institutional aspects of the country not otherwise captured by GDP or democracy, which (non-)coincidentally is zero.

Last but not least, our use of quasi-Poisson and Bayesian hierarchical negative binomial models directly addresses the main robustness concerns for count data, namely overdispersion and unobserved firm- and country-level heterogeneity. We also estimate alternative specifications with different covariate sets and fixed-effects structures and obtain very similar coefficients for the core institutional and firm-level variables. In light of these built-in specification checks and the large sample size, additional ad-hoc robustness tests would primarily replicate the same results rather than provide new insight.