ESG Strategy and Tax Avoidance: Insights from a Meta-Regression Analysis

Abstract

This research examines the relationship between environmental, social, and governance (ESG) criteria and tax behavior, with a particular focus on tax avoidance (TA). Despite the extensive literature on ESG and tax behavior, there remains a research gap concerning their interaction in the financial sector. The study is based on a dataset of 125 observations from 33 articles covering the period 2012–2023. The results of the meta-regression suggest that both ESG and TA indicators account for the different findings of the primary studies. Part of the observed heterogeneity can also be explained by the diversity of data samples and econometric approaches. Using the results of the meta-regression, we attempt to predict the association between ESG and TA in hypothetical and plausible study designs. The findings show no or small-to-moderate association between the two, suggesting that companies tend to separate ESG strategies from TA and underscoring the need for more consistent measurement practices. Notably, the link between the main variables appears to be strengthened in environments with extreme behaviors, both in terms of ESG and tax strategy. Distinct from prior meta-studies that centered on CSR and taxation, our analysis isolates the ESG/TA nexus by accounting for measurement heterogeneity (different ESG and TA proxies) and demonstrates that extreme behaviors largely drive the observed association. By examining the determinants of the heterogeneity of primary research into the ESG/TA relationship, this meta-analysis provides valuable insights that can guide future research, practical implementation, and regulatory policies. In particular, researchers should rely on long-run measures of TA (e.g., multi-year ETRs) and harmonized ESG indicators to reduce bias and enhance comparability across studies, thereby providing policymakers with more robust and consistent evidence.

1. Introduction

Good tax behavior is often associated with socially responsible corporate behavior. For instance, according to the Global Reporting Initiative Standard on Tax (GRI), good tax practices are expected of organizations that are committed to fulfilling their responsibility to their stakeholders (Kumar et al., 2024). From a social perspective, corporate taxation plays a central role in the relationship between business and society, as it reflects not only compliance with legal obligations but also a company’s broader commitment to responsible conduct and sustainability. Delaying or avoiding tax obligations can reduce public revenues and undermine transparency, social cohesion, and sustainable development (Mohanadas et al., 2020). From a corporate perspective, aggressive tax activities tend to cause animosity, damage the company’s reputation, and may even lead to a cessation of business activities (Heimberger, 2021). So, corporate tax aggressiveness can be perceived as socially irresponsible (Abdelfattah & Aboud, 2020). On the other hand, several businesses view high corporate taxes as discouraging innovation and investment and potentially harming job creation. In this case, socially responsible firms may consider not paying corporate taxes as the optimal means to achieve their social responsibility goals (Alsaadi, 2020). Recent findings underscore the complex associations of tax behavior with financial development and the broader macroeconomic context. Allam et al. (2024) find that financial development is associated with diminished tax evasion levels in 156 nations, though the findings diverge between OECD and non-OECD countries. These findings show that companies’ tax behavior largely depends on the wider financial and institutional context.

This calls into question the link between tax behavior and ESG (environmental, social, and governance) practices, which seek to integrate environmental and social issues into business processes, in line with the principles of responsible entrepreneurship (Pollman, 2021, 2022). In Europe, awareness of the significance of ESG criteria has grown significantly in response to recent sustainable finance regulations and international commitments to combat climate change (Oliver Yébenes, 2024; Redondo Alamillos & de Mariz, 2022). Moussa et al. (2024) concluded that a company’s ESG performance is directly linked to its financial health, observing a positive correlation between UK companies’ ESG performance and their market capitalization, with corporate governance mechanisms further enhancing this effect. The growing belief in the positive impact of ESG on corporate performance (Morgan Stanley, 2023) is also reflected in the fact that sustainability-focused funds continue to outperform other investment categories, although investment in the European markets has recently declined to an extent.

As ESG criteria gain prominence (Chytis et al., 2024; Chytis & Mitroulia, 2025), their relationship with corporate tax behavior and TA has become a critical area of scrutiny (Fonseca, 2020). TA refers to the methods companies use to reduce their tax liabilities and does not constitute a law violation (Avi-Yonah, 2008; Huseynov & Klamm, 2012). Fonseca (2020) investigated the relationship between a firm’s ESG performance and its propensity to engage in TA and suggested that aggressive tax avoidance may negatively affect the firm’s ESG ratings and reputation. Reputational risks are often the reason why companies with high ESG performance avoid aggressive TA (Baudot et al., 2020).

Various theories examine the connection between ESG and TA. The most widely accepted view of ESG links it to firms’ ethical and responsible behavior in connection to production. From a broad perspective, all legal decisions should reflect moral principles, and a firm’s long-term value is linked to maintaining good relations with all stakeholders (Haija, 2024), balancing the needs of all parties involved (Azzam et al., 2020). While some studies view aggressive TA practices as being inconsistent with ESG principles and potentially harmful to society (C. Zhang et al., 2021), this perspective may not apply to all forms of TA. The extent of the discord between TA and ESG is contingent on the manner in which TA is delineated and evaluated, given that specific tax planning strategies may be regarded as legitimate financial management as opposed to socially irresponsible conduct (Jallai & Gribnau, 2018).

However, the relationship between ESG and TA is complex. While some studies suggest that companies with high ESG scores tend to avoid aggressive TA strategies (Yoon et al., 2021), others find no clear relationship between the two (López-González et al., 2019). Corporate social responsibility (CSR) theory argues that socially responsible companies choose more conservative tax practices to protect their reputation (Ortas & Gallego-Álvarez, 2020). In contrast, agency theory suggests that managers may adopt TA strategies in order to maximize shareholder returns, even if this conflicts with ESG principles (Z. Chen & Xie, 2022). Some companies may use ESG as a means of enhancing their public image while simultaneously implementing aggressive tax strategies (Luo et al., 2023). The literature suggests that the role of ESG in corporate tax strategy is not unidimensional and requires further exploration. Understanding this relationship is crucial for investors, tax regulators, and stakeholders seeking to assess the long-term sustainability and responsibility of companies (Carolina et al., 2023).

Various theoretical approaches provide a range of predictions on the relationship under study. However, the current understanding of the multifaceted and complex aspects of this relationship remains limited. The mixed results of existing studies are the main motivation behind this meta-analysis.

Research findings on the correlation between ESG performance and TA have been conflicting. Some studies observed a negative correlation, indicating that firms with better ESG performance tend to engage less in aggressive tax practices (Hoi et al., 2013; Lanis & Richardson, 2013; Yoon et al., 2021), while others have found a positive correlation, suggesting that a company’s propensity to avoid taxes increases as its ESG performance improves. In other words, companies that score higher on ESG criteria are more likely to implement aggressive tax planning strategies to reduce their tax liabilities (Huang et al., 2017; Lee, 2020). Contextual parameters are significant in this nexus. According to Acosta Garcia et al. (2024), the positive association between Corporate Social Responsibility (CSR) and TA is weakened in business environments with high economic freedom (EF). Furthermore, the moderating effect of EF is intensified in societies with low power distance.

These inconclusive findings of the previous literature regarding the direction and magnitude of the relationship between ESG and TA, at least at an empirical level (Mayberry & Watson, 2021), suggest a research gap that needs to be addressed. Given the discrepancies emerging from variations in dataset composition and econometric specifications that make generalizations based primarily on traditional literature reviews challenging, a meta-analysis could synthesize research findings and help clarify the research landscape.

Against this backdrop, the goal of this paper is to use meta-analysis to synthesize and elucidate the findings from the literature on the ESG/TA relationship. It addresses the following two research questions: (1) Which factors account for the variation observed in the relationship of corporate TA and ESG found in prior studies? (2) What would be the overall effect size of the relationship between ESG and TA? To address the first research question, we use the meta-analysis method, which consists of a comprehensive survey and quantitative analysis of the important econometric evaluations of the relationship between ESG and TA. The dataset contains 125 observations sampled from 33 empirical studies carried out between 2012 and 2023. To answer the second research question, we use meta-regression to determine how the relationship between ESG and TA varies according to the design of individual studies.

Our main findings reveal significant heterogeneity in the relationship between ESG and TA, as reported in previous studies. This heterogeneity is explained by several factors, a significant one being the methodological choices of the studies. In particular, we found significant differences in the estimates of the primary studies regarding TA indicators (cash effective tax rates, or CETR, tax rate differences, extreme tax planning activities). Furthermore, the way ESG is measured also seems to influence the results. Studies using binary ESG measures or CSR proxies tend to find a positive ESG/TA relationship. Finally, the data samples and econometric specifications chosen in the studies also play an important role. For example, studies examining past losses and time-specific effects report higher effects of ESG on TA.

Using the estimates from our meta-regression, we predicted the association between ESG metrics and TA under several alternative and plausible hypothetical study designs in order to minimize specification bias. In most of these scenarios, no association was found between ESG and TA. However, when exploring individual ESG sources and depending on the TA proxy examined, several positive and statistically significant associations were observed. These associations are consistent with the opportunistic reporting argument, but in most cases, they are of low to moderate practical importance. Overall, the findings provide evidence that companies often decouple ESG activities from TA practices, a result which is consistent with the literature (Mayberry & Watson, 2021). Furthermore, using an alternative hypothetical research design, we found a moderate-scale positive association between CSR and extreme tax aggressiveness. This suggests that the relationship between CSR and tax aggressiveness may be more pronounced in certain contexts, especially when companies adopt extreme behaviors in either their ESG management or tax strategy.

To the best of our knowledge, this study contributes to the international literature by presenting one of the first quantitative surveys on the relationship between ESG and corporate taxation that aims to elucidate the influence of various indicators on discrepancies in the results of primary studies. It follows the methodology of Marques et al. (2023), who analyzed 117 estimates from 23 studies examining the relationship between CSR and corporate TA and found that both CSR practices and TA indicators contribute to explaining the differences in the results of the primary studies. In doing so, first, it complements previous studies conducted in different settings and following different methodologies, such as the recent meta-analysis by Mitroulia et al. (2025), who synthesized the empirical–quantitative findings on the link between corporate sustainability disclosure (CSD) and tax behavior (TB). On a final dataset of 50 articles spanning the period 2012–2022, the authors concluded that the link between CSD and TB is not strong enough to be of practical importance. Second, through the analysis of moderator variables, which could explain the differences in the findings of previous studies, the present study reveals either a nonexistent or very limited relationship. Third, while previous systematic literature reviews (SLRs) focused exclusively on taxation in the context of CSR, the present study broadens its scope to responsibility (ESG) in the field of taxation in general. For instance, the SLR conducted by Scarpa and Signori (2023) offered a structured overview of the existing knowledge on the relationship between CSR and taxation. The authors analyzed and discussed four dimensions of corporate tax responsibility (instrumental, political, comprehensive, and ethical) and examined the content of these dimensions, including issues such as compliance with the letter and spirit of the law, paying a fair share of tax, stakeholder management, and tax transparency. Additionally, the lack of a meta-analysis that focuses on tax accounting is also highlighted by Velte (2019), who conducted an SLR of 63 meta-analyses.

The present meta-analysis, thus, attempts to fill this gap in the literature and tries to identify potential sources of bias in prior research that help explain the heterogeneous findings of previous research. Furthermore, the study reveals critical factors that influence the relationship between ESG criteria and tax avoidance. This not only provides new knowledge but also lays the foundation for the development of more comprehensive theoretical frameworks and methodological approaches in the future. We suggest that the dynamics between ESG and TA need to be further explored, incorporating multidimensional factors in the examination that may influence the findings.

The results of the study are of interest to policymakers, regulators, and international organizations since they can contribute significantly to discussions about programs and actions related to corporate TA. First, it appears that the benefits of ESG do not come at a significant cost concerning TA, as no statistically significant relationship is found between ESG and TA. While our findings challenge the transparent reporting assumptions according to which ESG and TA are not linked, there is also evidence to suggest that extreme ESG behaviors and tax aggressiveness are inextricably linked. This helps to inform regulatory and tax authorities, as, for example, there are tax and non-financial disclosures that may allow them to identify significant tax deficiencies, as well as firms with uncertain tax positions (Mayberry & Watson, 2021). In addition, the results of this research will significantly contribute to informing third-party rating providers, analysts, sustainable investors, and sustainability standard setters, drawing attention to the possibility that companies with extreme behaviors may manipulate non-financial related disclosures.

The rest of the paper is structured as follows. The second section describes the research methodology and data selection and analyzes the meta-analysis procedure, which involves controlling for the influence of various factors that may cause heterogeneity in the results, with particular emphasis on meta-variables. The third section presents and analyzes the results of the meta-analysis, examines the potential publication bias in the literature, and presents the robustness tests. Finally, the fourth section summarizes the conclusions of the research.

2. Materials and Methods

2.1. Dataset Construction

A systematic and detailed literature review was conducted in January 2024, with the aim of analyzing the relationship between the two core variables. We searched the Scopus and Google databases to identify the relevant scientific works using the following keywords related to the subject of the research: “tax planning”, “tax avoidance”, “tax management”, “tax aggressiveness”, “tax evasion”, and “tax transparency”, as well as the phrases “ESG”, “environmental social governance”, “CSR”, “corporate social responsibility’, “IR”, and “integrated reporting*”. These keywords had to appear in any field or in the full text of the articles. For certain keywords like “environ*social*govern*” or “report*”, an asterisk was added to allow for different forms of the terms, thus widening the range of results.

The study period of 2012–2023 was selected to capture a decade of intensified research into ESG and tax behavior. This time frame encompasses the most pertinent and up-to-date research, exemplifying ESG methodologies, tax planning tactics, and the latest regulatory updates. It also coincides with governments and regulators’ increasing integration of ESG criteria into tax policies, as highlighted by the European Banking Taxation Forum (European Banking Taxonomy Forum [EBTF], 2022). The wealth of data and published research in the specific time period renders our meta-analysis reliable and comprehensive.

On the day of data collection, 30 January 2024, the search process resulted in 342 publications. Despite the fact that most of them had a relevant theme, a significant percentage was not based on an empirical approach. Therefore, additional filters were used to limit the sample to papers using empirical data, with the help of two methodological keywords, “empirical” and “statistics”, which had to appear anywhere in the article.

With the aim of creating a reliable dataset for our meta-analysis, we selected studies that met strict criteria of scientific accuracy and relevance. These criteria related to the methodology used, the quality of the data, and the clarity of the results. Specifically, the inclusion criteria of an article in the meta-analysis were the following: (i) it had to include an econometric model that estimates the link between ESG and TA. That restriction excluded several documents that only dealt with theoretical and qualitative surveys of the literature under study; (ii) it should be available in English; (iii) it should be published within the period 2012–2023; (iv) the dependent variable used should be a measure of TA; (v) the independent variables in the model had to include a variable for ESG; and (vi) it should provide the estimated ESG coefficient, as well as the standard errors or t-statistics.

The publications were carefully evaluated to determine their relevance to the topic under consideration. The criteria applied included the selection of papers that explicitly address ESG and TA issues. Initially, the titles, abstracts, and keywords of the articles were examined to exclude those that were not relevant to the subject of this study and did not follow criteria (i), (ii), and (iii). These stages are considered essential in the context of meta-analysis (Meca et al., 2013). In cases where the wording of the abstract was unclear, we examined the whole article. To ensure consistency, two of the authors independently screened all records and coded moderators. Disagreements were resolved by means of discussions with a third author, and this process resulted in 157 articles deemed suitable for further analysis. Approximately 30% of records were double-screened, with 82% raw agreement. Inter-rater reliability was assessed by means of the blind selection of 10 articles (≈30% of the 33 included studies). The level of agreement between the raters was calculated by means of Cohen’s Kappa κ (Cohen, 1960), the values of which are as follows: ≤0 (no agreement), 0.01–0.20 (none to slight), 0.21–0.40 (fair), 0.41–0.60 (moderate), 0.61–0.80 (substantial), and 0.81–1.00 (perfect) (McHugh, 2012). Our analysis yielded a κ = 0.78 confidence interval (CI) [0.66, 0.89], indicating substantial agreement. Following best practices in meta-analysis, a 25–30% double-screening rate is generally considered sufficient to ensure coding reliability while maintaining feasibility (Cooper, 2017; Petticrew & Roberts, 2006).

The selected articles were then assessed individually to confirm their compliance with inclusion criteria (iv), (v), and (vi). In this process, 152 articles were excluded for the following reasons: (1) although they examined the relationship between ESG and TA, they did not provide sufficient quantitative estimates, such as the necessary statistics for the meta-analysis (e.g., standard errors, t-statistics) and the dependent variable was ESG; (2) the dependent variable was not a measure of TA; and (3) the independent variable ESG entered the model in interaction with another variable.

Of all studies reviewed, 33 were compatible with the criteria described above, yielding a total of 125 estimates for the dataset. The characteristics of the studies (time period, geographical areas) are included in

Table A1. Summary of studies examining the link between ESG and tax avoidance, as included in this meta-analysis.

Articles	Journal	Location of Companies	ESG Database	Period	Sample Size
(Lanis & Richardson, 2012)	Journal of Accounting and Public Policy	Australia	Aspect-Huntley	2008–2009	408
(Hoi et al., 2013)	The Accounting Review	USA	KLD Research & Analytics	2003–2009	9147
(Landry et al., 2013)	Journal of Accounting, Ethics & Public Policy	Canada	CSID	2004–2008	551
(Laguir et al., 2015)	Journal of Cleaner Production	France	Diane	2003–2011	83
(Lanis & Richardson, 2015)	Journal of Business Ethics	USA	KLD Research & Analytics	2003–2009	434
(Davis et al., 2016)	The Accounting Review	USA	KLD Research & Analytics	2006–2011	5588
(Sari & Tjen, 2016)	International Research Journal of Business studies	Indonesia	Indonesia Stock Exchange	2009–2012	752
(J. Kim & Im, 2017)	Sustainability	South Korea	KEJI (Korea Economic Justice Institute)	2005–2007	491
(K. Z. Lin et al., 2017)	International Journal of Accounting	China	RKS (Rankins CSR Ratings)	2008–2012	1438
(Kiesewetter & Manthey, 2017)	Corporate Governance: The International Journal of Business in Society	Global	ASSET4	2005–2014	4309
(Gulzar et al., 2018)	Sustainability	China	RKS	2009–2015	3481
(X. Lin et al., 2019)	Managerial Auditing Journal	France	MSCI	1995–2013	15,960
(Salhi et al., 2020)	Social Responsibility Journal	UK & France	DataStream	2005–2017	6500
(Jarboui et al., 2020)	Journal of Financial Crime	UK	ASSET4	2005–2017	3900
(Fourati et al., 2019)	Journal of Applied Business and Economics	Global	ASSET4	2002–2015	8596
(Zeng, 2019)	Social Responsibility Journal	Global	Thomson Reuters	2011–2015	8993
(Hasan et al., 2019, 2025)	Munich Personal RePEc Archive; European Accounting Review	Global	Thomson Reuters	2009–2016	26,752
(Fallan & Fallan, 2019)	Scandinavian Journal of Management	Norway	Oslo Stock Exchange	2009/2010/2012	92
(Li et al., 2019)	Sustainability	China	RKS	2013–2016	1993
(J.-M. Kim & Choi, 2019)	Journal of International Trade & Commerce	South Korea	Korea Economic Justice Institute Index (KEJI)	2011–2016	1747
(Kristiadi et al., 2020)	Journal of Theory and Applied Management	Indonesia	Thomson Reuters Eikon	2008–2019	583
(Alsaadi, 2020)	Journal of Financial Reporting and Accounting	Global	ASSET	2008–2016	3205
(Dakhli, 2021)	Corporate Governance	France	Thomson Reuters-ASSET	2007–2018	2400
(Awad, 2021)	Arab Journal of Management	Egypt	Egyptian Corporate Responsibility Index	2007–2016	179
(W. L. Lin, 2021)	Corporate Social Responsibility and Environmental Management	USA	Fortune’s America’s Most Admired Companies (AMAC)	2008–2017	1129
(Montenegro, 2021)	Sustainability	Global	ASSET4	2004–2010	165
(Timbate, 2021)	Business Research Quarterly	USA	Thomson Reuters	2007–2016	2997
(Ding et al., 2022)	Frontiers in Psychology	China	Shanghai Stock Exchange (SSE) and Shenzhen Stock Exchange (SZSE)	2002–2017	23,317
(Khan et al., 2022)	Cogent Economics & Finance	Pakistan & Nigeria	Nigerian and Pakistani Stock Exchanges	2011–2020	1331
(Hajawiyah et al., 2022)	Cogent Business & Management	Indonesia	Indonesia Stock Exchange (IDX)	2013–2020	328
(Junaidi et al., 2023)	Journal of Governance and Regulation	Indonesia	Indonesia Stock Exchange (IDX)	2018–2020	117
(Aliani & Bouguila, 2023)	International Journal of Sustainable Economy	USA	US Securities and Exchange Commission	2020	100
(Kuo, 2023)	Review of Pacific Basin Financial Markets and Policies	Taiwan	Taiwan Economic Journal (TEJ)	2015–2020	5982

Notes: The table lists the studies included in the meta-analysis, providing information on the locations of the companies, the sample size, the source of the ESG, and the time period considered.

of Appendix A. The flow chart (Figure A1), according to the PRISMA guidelines (Moher et al., 2009), presents the results of the review and is included in Appendix A.

2.2. Studies Coding and Data Analysis

By reading the full text of the selected studies, the authors collected data based on the predefined criteria as to the information that should be recorded. Each paper was then assigned to a different author for cross-validation of the extracted information. Any discrepancies identified during this stage were resolved through collaborative discussions. One of the authors coded the data according to an initial set of moderators, and a second author then reviewed the process. The coded information fell into three major groups: (1) sample characteristics (e.g., sample size, time-period, country or countries where companies are located); (2) econometric model specifications (e.g., model, control variables, fixed effects control); and (3) characteristics of ESG and TA measures. Much of the codified information was in the form of a dummy variable. For example, studies in which the ESG measure consisted of a continuous variable were coded as 1 and 0 otherwise. Most of the codified information served as moderators for the meta-regression analysis. In the following section, we define and examine in detail the meta-variables included in the meta-regression analysis. Our effect size is represented by the partial correlation coefficient calculated from each regression estimate of ESG in primary studies.

For precision weighting, we used the inverse of the variances of the partial correlation coefficients in the meta-regression analysis, as the precision of the estimates varies across studies. From the 33 studies, we obtained 125 effect sizes that form our meta-dataset.

Table 1. Detailed description of the partial correlation statistics resulting from the individual studies.

Authors	Number of Effect Sizes	Mean	Median	Min	Max	Std.Dev	Standard Error
(Lanis & Richardson, 2012)	4	0.01370	0.01456	0.00747	0.01819	0.00449	0.00225
(Hoi et al., 2013)	3	0.00016	0.00001	0.00001	0.00045	0.00025	0.00015
(Landry et al., 2013)	2	−0.00027	−0.00027	−0.00063	0.00009	0.00051	0.00036
(Laguir et al., 2015)	5	0.00161	0.00187	−0.00011	0.00232	0.00100	0.00045
(Lanis & Richardson, 2015)	2	0.00183	0.00183	0.00173	0.00193	0.00015	0.00010
(Sari & Tjen, 2016)	2	−0.00175	−0.00175	−0.07327	0.06977	0.10115	0.07152
(Davis et al., 2016)	3	0.00003	0.00002	0.00002	0.00004	0.00001	0.00001
(J. Kim & Im, 2017)	6	0.00087	0.00014	0.00009	0.00236	0.00115	0.00047
(K. Z. Lin et al., 2017)	8	0.00066	0.00058	0.00020	0.00163	0.00046	0.00016
(Kiesewetter & Manthey, 2017)	2	−0.07610	−0.07610	−0.15863	0.00643	0.11671	0.08253
(Gulzar et al., 2018)	2	−0.03412	−0.03412	−0.05195	−0.01630	0.02521	0.01783
(Fallan & Fallan, 2019)	2	0.01356	0.01356	0.01223	0.01489	0.00188	0.00133
(Hasan et al., 2019, 2025)	2	0.00006	0.00006	0.00006	0.00006	0.00000	0.00000
(J.-M. Kim & Choi, 2019)	10	0.00542	0.00299	0.00092	0.02891	0.00834	0.00264
(X. Lin et al., 2019)	1	0.00005	0.00005	0.00005	0.00005	-	-
(Li et al., 2019)	3	0.00058	0.00074	0.00007	0.00092	0.00045	0.00026
(Fourati et al., 2019)	2	−0.01361	−0.01361	−0.02938	0.00215	0.02229	0.01576
(Zeng, 2019)	2	0.00004	0.00004	0.00002	0.00007	0.00003	0.00002
(Alsaadi, 2020)	18	−0.05190	−0.05698	−0.11021	−0.00728	0.02821	0.00665
(Jarboui et al., 2020)	1	0.00598	0.00598	0.00598	0.00598	-	-
(Salhi et al., 2020)	2	0.00350	0.00350	0.00090	0.00609	0.00367	0.00259
(Kristiadi et al., 2020)	1	−0.05169	−0.05169	−0.05169	−0.05169	-	-
(Awad, 2021)	2	0.17485	0.17485	0.16447	0.18523	0.01468	0.01038
(Dakhli, 2021)	2	0.00054	0.00054	0.00037	0.00072	0.00025	0.00018
(W. L. Lin, 2021)	5	−0.02890	−0.01373	−0.10387	0.01416	0.04828	0.02159
(Timbate, 2021)	4	0.00842	0.01930	−0.05279	0.04788	0.04401	0.02201
(Montenegro, 2021)	2	−0.07354	−0.07354	−0.15352	0.00644	0.11311	0.07998
(Hajawiyah et al., 2022)	1	0.45121	0.45121	0.45121	0.45121	-	-
(Khan et al., 2022)	12	0.04988	0.04048	0.00199	0.11004	0.04758	0.01373
(Ding et al., 2022)	2	0.00006	0.00006	0.00004	0.00009	0.00004	0.00003
(Kuo, 2023)	8	0.00344	0.00354	0.00256	0.00412	0.00050	0.00018
(Junaidi et al., 2023)	1	−0.15494	−0.15494	−0.15494	−0.15494	-	-
(Aliani & Bouguila, 2023)	3	0.01102	0.03636	−0.13366	0.13035	0.13382	0.07726

Notes: The table classifies the number of effect sizes included in the documents in our meta-analysis. Additionally, it lists partial correlation coefficient statistics, such as mean, median, standard deviation, and outliers (maximum and minimum) by study. It should be noted that the primary studies generally did not winsorize outliers. To address this issue, we conducted robustness checks by removing the top and bottom 5% of partial correlations, following best practices in meta-analysis (Stanley & Doucouliagos, 2012), and confirmed that our main results remained unaffected.

presents summary statistics for the partial correlation coefficients, showing that their number, as extracted from each study, ranges from 1 to 18. Furthermore, the average values of the partial correlation coefficients are between −0.15494 and 0.45121. As can be seen in Table 1, the relationship between the two main variables is diverse: in 10 out of the 33 studies, a negative correlation is observed, suggesting that higher ESG performance is accompanied by lower TA tendencies. In contrast, in the remaining studies, a positive correlation is found, i.e., a tendency for increased TA in companies with higher ESG performance. The absence of consistent findings provided the main motivation for the present meta-analysis, which aims to investigate the reasons for this heterogeneity.

2.3. Meta-Regression Approach

Research efforts to elucidate the relationship between ESG criteria and corporate TA practices have intensified over the past decades. However, the empirical findings of the relevant studies are inconclusive, with estimates of the magnitude and sign of the association varying considerably. A common research framework adopted in the literature is the estimation of regression models, whereby corporate TA (TaxAvoid) is explained as a function of ESG criteria and a set of control variables (Controls). With this approach, researchers seek to isolate the effect of ESG factors on corporate tax behavior, considering other explanatory variables that may influence this relationship.

TaxAvoid = α + β ESG + γ Controls + ε

(1)

Empirical studies examining the relationship between ESG criteria and corporate TA are characterized by significant heterogeneity in their methodological choices, concerning the data and variables used and the design of the empirical models. This heterogeneity is reflected in the wide dispersion of empirical results. We argue that some of it is due to study characteristics and not just conceptual differences. To investigate this issue, we propose the use of meta-analysis, a statistical technique that allows the systematic synthesis of the results of multiple studies, with the aim of drawing more general conclusions and evaluating the heterogeneity of findings. Meta-analysis has gained considerable traction in various scientific disciplines, such as medicine, psychology, and economics, and has also been successfully applied in the field of tax research (Marques et al., 2023).

The successful application of meta-analysis requires the existence of comparable estimates from primary studies. However, the diversity of measures used for TA and ESG criteria in the different studies makes it difficult to directly compare the regression coefficients. To overcome this obstacle, we adopt partial correlation coefficients (e.g., Knaisch & Pöschel, 2024) as a more appropriate measure for comparing relationships between variables. Partial correlation coefficients are independent of the units of measurement and provide a standardized measure of the strength and direction of the relationship between two variables (Stanley & Doucouliagos, 2012). Their estimation is based on the reported statistics of the regression models of the primary studies, thus allowing meta-analysis of the results.

Partial correlation coefficients can be calculated from the reported regression statistics, as follows:

$${\mathrm{r}}_{\mathrm{i}\mathrm{j}}={\displaystyle \frac{{\mathrm{t}}_{\mathrm{i}\mathrm{j}}}{\sqrt{{\mathrm{t}}_{\mathrm{i}\mathrm{j}}^2+{\mathrm{d}\mathrm{f}}_{\mathrm{i}\mathrm{j}}}}}$$

(2)

We then apply an alternative standardized effect size: semi-elasticity. Semi-elasticity expresses the percentage change in the corporate tax variable when the explanatory variable (weighted average of neighbors’ tax choices) changes by one unit. The calculation of semi-elasticities serves as a method of checking the reliability of the results obtained from the basic analysis, which was based on the use of partial correlation coefficients. Detailed information on this procedure and its results is provided in Section 3.3. This approach allows us to assess the robustness of our findings, ensuring the validity of the conclusions.

In each study (j), the beta coefficient is calculated and compared to a reference value (t-test) to determine if it is statistically significant. The result of this comparison is the value t (t_ij), which indicates confidence that beta differs significantly from zero. The number of degrees of freedom (df) helps us estimate the precision of this t-value. The partial correlation coefficient indicates how strong the relationship between two variables is after considering the effect of other variables. It is important to note that the sign of the partial correlation coefficient is the same as the sign of the beta coefficient. To construct the meta-analysis model we describe next, we also need a measure of how accurate the partial correlation coefficient is. This measurement is the standard error, and we can calculate it using the statistical method described by R. A. Fisher (1992), as shown in Equation (3).

$${\mathrm{s}\mathrm{e}}_{\mathrm{i}\mathrm{j}}={\displaystyle \frac{1-{\mathrm{r}}_{\mathrm{i}\mathrm{j}}^2}{{\mathrm{d}\mathrm{f}}_{\mathrm{i}\mathrm{j}}}}$$

(3)

In Equation (4), we apply a fixed-effects post-regression model. In this model, Y represents a set of partial correlation coefficients derived from various studies, while the matrix M includes variables characterizing these studies (moderators). The parameter u estimates the average effect of each moderator on the relationship between ESG and TA, relative to a specific benchmark.

Y = Mφ + ε

(4)

In the primary studies, TA was measured by a variety of metrics such as effective tax rate (ETR) and cash effective tax rate (CETR), book tax difference (BTD), tax rate difference (DifTax), and extreme tax planning activity (ExtAct). To standardize the interpretations of the results, the correlations of the partial coefficients based on ETRs and CETRs are multiplied by −1. Thus, all estimates positively reflect the relationship between ESG practices and TA.

Our estimation, as described in Equation (4), uses weighted least squares (WLS) with inverse variance weights. This method is widely accepted as the most suitable for the analysis of meta-regression models (Hansen et al., 2022). According to Zhai and Guyatt (2023), using the precision of estimates from Equation (3) as weights enhances study precision and, therefore, influence on the final results.

The primary studies we reviewed typically provided more than one estimate of the ESG coefficient. Following the method suggested by Feld and Heckemeyer (2011), we included all these estimates in our analysis to increase the reliability of the results due to the larger sample size. However, there is the risk that using multiple estimates from the same study may violate the assumption of independence of observations. To address this issue, we clustered standard errors at the study level as suggested by Marques et al. (2023).

As a first sensitivity check, we first removed 5% of the outliers from the partial correlations (column 2) and found that the results remained essentially the same as in the original analysis (column 1). This suggests that the outliers did not significantly affect the overall findings. Next, we examined various corporate tax variables (columns 3–6), such as the effective tax rate (ETR), the effective cash tax rate (CERT), and the tax difference (DifTax). While estimates varied when we used different variables, confidence intervals consistently excluded the null effect, supporting the hypothesis that corporate tax has an effect. These preliminary results suggest that the corporate tax variable significantly influences the outcomes of ETR and CETR, particularly in studies examining the interaction between ESG performance and tax behavior. Further analysis is required to ascertain whether this variable fully accounts for the observed variation when considering the additional factors explored in this research.

In addition, the magnitude of the effect of corporate tax behavior was estimated by calculating the average of all estimates collected from the relevant primary literature. At the same time, 95% confidence intervals were created around the calculated mean. To standardize the magnitude of the effect across different studies, partial correlation was applied. In our analysis, partial correlations capture the relationship between firms’ ESG practices and their TA behavior while controlling for firm-specific characteristics. Indeed, considering this relationship at the firm level is important, as recent evidence from Elgharbawy and Aladwey (2025) shows that higher ESG performance and greater board diversity can significantly reduce corporate TA in UK firms. This approach does not involve any inference of governmental policies. This method provides two main advantages. First, partial correlations are comparable across studies, as they are based on a dimensionless measure ranging from −1 to 1. Second, they allow for the calculation of a broader set of estimates compared to other effect size measures (Aloe & Thompson, 2013; Stanley et al., 2024).

The meta-analysis of studies on corporate tax behavior uses partial correlations, which are weighted according to the “precision” of each estimate. This precision is calculated as the reciprocal of the standard error or variance of the estimate. Based on all relevant primary studies that meet the selection criteria, the weighted average partial correlation is considered the most reliable estimate of the effect of tax competition, according to the literature. The weighted average correlation is calculated from the partial correlations of each study, weighted according to their precision.

The weighted average partial correlation is expressed by Formula (5):

$$\mathrm{w}\mathrm{r}={\displaystyle \frac{\sum {[\mathrm{P}}_{\mathrm{i}\mathrm{j}}{\mathrm{r}}_{\mathrm{i}\mathrm{j}}]}{\sum \left[{\mathrm{P}}_{\mathrm{i}\mathrm{j}}\right]}}$$

(5)

The partial correlation coefficient r is equal to the coefficient of i from study j, as shown in Equation (2). The precision of the partial correlations is represented by P. The weighted partial correlation coefficient indicator is used to answer two questions: (1) Is there evidence of strategic interactions in corporate taxation across firms? (2) How large is this effect? According to H. Doucouliagos’ (2011) guidelines, the absolute value of the partial correlation is evaluated as follows: below 0.07, the effect is small; between 0.07 and 0.17, the effect is moderate; above 0.33, the effect is large. Positive values of the partial correlation indicate that governments’ tax policies move in the same direction as those of their neighboring jurisdictions, which constitutes “competitive behavior”. Conversely, negative values indicate “reverse behavior”, where a government’s policies act as a “counterweight” to the policies of its neighbors (Overesch & Rincke, 2009; Parchet, 2019). The partial correlation index is, therefore, a useful tool for assessing the degree and direction of tax behavior in an international environment.

In our analysis, we chose to use the unrestricted WLS estimator, a method widely recommended in the meta-analysis literature (Stanley & Doucouliagos, 2015), because it gives greater weight to estimates that are considered more reliable; that is, those with smaller errors. The logic behind this choice is simple: the more precise an estimate is, the more valuable the information it provides. In addition, since many of the studies we examined presented more than one estimate, we grouped the standard error estimates at the study level, aiming to reduce the potential dependence between different estimates coming from the same study.

Table 2. Descriptive statistics of partial correlation coefficients on tax variables.

	All Estimates	Excluding Top and Bottom 5%	ETR	CETR	DifTax
#Estimates	125	112	91	19	15
Median	0.0009	0.0009	0.0019	0.00002	0.0003
Unweighted Average	−0.0001	−0.0023	−0.0025	−0.0042	0.0193
Unr. WLS	−0.0056	−0.0044	−0.0076	−0.0043	0.0036
95% Confidence	−0.0111	−0.0082	−0.0139	−0.0152	−0.0243
Interval (unr. WLS)	−0.0001	−0.0006	−0.0013	0.0065	0.0315

Notes: Unr. WLS: unrestricted weight least squares, using the inverse of the variances as precision weights of partial correlation coefficients; ETR: effective tax rate; CETR: cash ETR; DifTax: differential tax rate. Effect sizes reflect the strength and direction of the association between ESG and TA and should not be interpreted as direct changes in tax rates or as evidence of intergovernmental tax competition.

summarizes the results of the meta-analysis, including the median and unweighted average of the partial correlations. Column (1) presents the results for the entire sample, while the estimation of the unrestricted WLS model provides further evidence for the analysis. Table 2 also presents the descriptive statistics of the partial correlation coefficients (pcc) for various tax variables, from which the following conclusions emerge. The median value for all estimates is very close to zero (0.0009), which suggests that most estimates do not exhibit a strong positive or negative correlation. The independently weighted mean is slightly negative (−0.0001 to −0.0042), indicating that, in general, the effects of tax are either very small or negligible.

Negative or near-zero averages of the pcc (e.g., −0.0056 for ETR, −0.0076 for CETR) indicate that, on average, the choice of the corporate tax variable does not, on its own, produce significant differences in the estimated effective tax rates across the studies. This is critical for ESG strategy, as companies with high scores are assumed to engage in more ethical tax behavior, avoiding aggressive TA practices (Elgharbawy & Aladwey, 2025). The high variability in values (e.g., the 95% confidence interval ranges from −0.0111 to −0.0001) suggests that there is no consistent pattern in tax behavior. The lower bound of the confidence interval for all estimates is negative (−0.0111 to −0.0243), while the upper bound is close to zero or slightly positive. This suggests significant heterogeneity in the estimates, which can be attributed to differences in the methodology of the primary studies (e.g., use of different tax variables or selection of different country samples). The analysis of Table 1 shows that tax is not strongly associated with increased tax rates. This means that ESG strategy does not directly affect the tax behavior of companies (Yoon et al., 2021). Companies that adopt ESG practices may still use TA strategies to reduce their tax burden, despite the ethical commitments that are supposed to accompany ESG (Y. Zhang et al., 2025).

The next section analyzes the results of the current research. The R language was the main tool for performing all statistical analyses required.

Meta-analysis is a systematic approach that compares the results of previous studies, enabling the evaluation of quantitative research findings (Paul & Barari, 2022). Specifically, we expect that the following characteristics of the primary studies will influence the observed effects of ESG on TA: (1) the types and scope of the TA variables; (2) the ESG measures (including their individual dimensions); (3) the characteristics of the data sample; and (4) econometric specifications. Next, we analyze the different moderator variables that are associated with each characteristic (as coded in

Table 3. Summary statistics of variables used in the meta-regression analysis for n = 125 observations.

Variable Name	Variable Description	Mean	S. D
Partial Correlation	Partial correlation coefficient of ESG with tax variables	−0.00012	0.06534
Tax Variables
ETR	BD = 1: Effective tax rates used as dependent variable	0.728	0.447
CETR	BD = 1: Cash effective tax rates used as dependent variable	0.152	0.361
DifTax	BD = 1: Book tax rates used as dependent variable	0.120	0.326
Proxy Measures for ESG
ESG	BD = 1: ESG measure is based on three dimensions (environmental, social, governance)	0.576	0.496
CSRC	BD = 1: CSR measure is a continuous variable	0.200	0.402
CSRR	BD = 1: CSR proxy is based on a responsible view	0.312	0.465
Data Characteristics
MeanYearData	Logarithm of the mean year of the time sample of the data	7.607	0.002
Controls	Number of regression control variables	9.760	5.060
S_Country	BD = 1: Data for only one country used	0.768	0.424
Country Composition
Europe	BD = 1: Only European countries included in the data	0.280	0.451
MixOfCountries	BD = 1: Mix of European and non-European countries included in the data	0.048	0.215
NonEuropean	BD = 1: Only non-European countries included in the data	0.648	0.479
Econometric Characteristics
FirmFixedEffect	BD = 1: Firm fixed effect included	0.520	0.502
YearFixedEffect	BD = 1: Year fixed effect included	0.728	0.447
CountryFixedEffect	BD = 1: Country fixed effect included	0.176	0.382
Ols	BD = 1: OLS estimation used	0.552	0.499
Prob_Log_Tob_it	BD = 1: Probit, Logit, or Tobit estimation used	0.144	0.353
OtherEstimator	BD = 1: Estimator other than OLS, Probit, Logit, or Tobit used	0.216	0.413
Publication Characteristic
StandardError	Standard error of the partial correlation coefficient	0.0374	0.0281
Additional Control Variables
Size	BD = 1: Firm size included as control	0.848	0.361
Invint	BD = 1: Inventory intensity as control	0.288	0.455
ROA	BD = 1: Return on assets as control	0.768	0.424
BIG4	BD = 1: BIG 4 auditing firms as control	0.160	0.368
Ppe	BD = 1: Gross property plant and equipment control	0.688	0.465
Lev	BD = 1: Firm leverage as control	0.784	0.413
Ocf	BD = 1: Operating cash flow as control	0.248	0.434

Notes: BD (binary dummy) takes the value 1 if the condition is fulfilled and zero otherwise.

).

2.4. Variables in the Dataset

Tax avoidance (TA) and tax aggressiveness are often used as alternative terms in the literature, although they may express different concepts (Blouin, 2014). In line with existing empirical research in the tax field (S. Chen et al., 2010; Lanis et al., 2017; Lanis & Richardson, 2012), we broadly define tax aggressiveness as the deliberate reduction of taxable income through tax planning. While TA is an acceptable practice, tax aggressiveness encompasses a wider range of behaviors, including legal activities, practices that are in the grey zone of legality, as well as illegal tax evasion (Simser, 2008). The main objectives TA aims to achieve are these three: reducing the tax paid below the amount required by the laws of a country, transferring tax liabilities from the country where the profits were earned to another country, and other tax benefit strategies (Chytis et al., 2020; J. Fisher, 2014).

Concerning the categories of variables used in the primary studies, we distinguish several measures of tax aggressiveness: ETR (Davis et al., 2016), CETR (Hoi et al., 2013; Timbate, 2021), and DifTax (Hasan et al., 2019, 2025; Salhi et al., 2020; Zeng, 2019). It should be noted that the CETR is a distinct form of the ETR, in conjunction with other metrics such as GAAP ETR and current ETR. While all these indicators aspire to encapsulate elements of a firm’s tax obligation, the CETR concentrates on the long-term average tax rate, thus providing a consistent basis for comparison across studies and contexts. ETRs are the most commonly used measures of TA by academic researchers.

The reliability of the metrics adopted in empirical studies to quantify TA has been the subject of intense scientific debate. In particular, the use of financial statements to calculate taxable income has been found to be inaccurate by Hanlon and Heitzman (2010), as financial statements often do not provide sufficient information on actual taxable income or taxes paid. At the same time, CETR, which is widely used to measure TA, has been criticized by Watson (2015) for its inaccuracy, as its numerator may include taxes relating to periods other than the current one, which can lead to measurement distortions (Hanlon & Heitzman, 2010). For this reason, several studies use multi-year ETRs (e.g., three-, five-, or ten-year averages) to iron out fluctuations and improve the reliability of the indicator. In our meta-analysis, we distinguish between studies that use annual ETRs, multi-year ETRs, or both to ensure the comparability of the results (Dyreng et al., 2019).

The lack of a single and unambiguous metric for TA has led to the development of a plethora of alternative ones. Studies such as those by Hasan et al. (2019, 2025) and Zeng (2019) have proposed the use of tax rate differentials as indicators of TA, while other studies, such as Davis et al. (2016), have focused on extreme tax planning activities. Extreme tax planning activities are considered more direct measures to capture highly aggressive TA practices (Lanis & Richardson, 2015; Rego & Wilson, 2012; Shams et al., 2022). Overall, the choice of methods and surrogate measures in the literature affects the comparison of results between studies. The lack of harmonization in the measurement of TA may lead to heterogeneity in the findings, creating the need for careful evaluation of the methodologies adopted.

In the meta-analysis, the dependent variable corresponds to a measure of corporate tax rates as captured in each primary study. However, measuring corporate tax rates is not straightforward, as the relevant literature uses three different approaches. As discussed in the previous section, we take these differences into account by distinguishing estimates based on the ETR, CETR, and DifTax measures. Table 3 presents the mean and standard deviation for each of these different corporate tax variables. The mean values of the dummy variables are expressed in percentage form, providing useful descriptive information that facilitates data interpretation. In relation to tax variables, ETR emerges as the preferred choice for analyzing tax rates, used as the dependent variable in 72.8% of the cases, while CETR and DifTax measures cover 15.2% and 12%, respectively. Thus, the literature examined suggests that the relationship between ESG criteria and TA may be influenced by the way we measure TA. To obtain a more accurate picture, we examine additional factors that may explain the results of the original studies.

2.5. Proxy Measures of the Independent Variable

To date, the most widely used framework for assessing the sustainability of companies is that of ΕSG criteria. These criteria, often associated with ethical and socially responsible investment (Galbreath, 2013), have evolved into key indicators of managerial competence, effective risk management, and non-financial performance of companies. In contrast to concepts such as corporate social performance (CSP) or corporate social responsibility (CSR), the ESG framework covers a wider range of issues, including (i) environment (issues such as climate change, energy consumption, and carbon emissions), (ii) social responsibility (issues related to human rights, product safety, and employee welfare), and (iii) governance (issues such as board independence, anti-corruption, and shareholder protection) (Pollman, 2022).

However, accurately measuring ESG performance involves a number of challenges. (Martiny et al., 2024) observed considerable variability in definitions and measurements of ESG criteria, even in historical data from the same providers. Therefore, it remains unclear which method is most appropriate for measuring ESG indicators. At the same time, CSR is recognized as an important tool for enhancing competitiveness, creativity, and innovation, as well as for improving corporate reputation. CSR helps to create positive relationships with customers, employees, partners, government agencies, and non-governmental organizations (Nimani et al., 2022). Despite its value, CSR also presents conceptual ambiguity, as the literature includes a plethora of definitions and interpretations (Giannarakis, 2014; Kraus & Britzelmaier, 2012). This inconsistency often leads to confusion, with CSR being confused with related concepts such as corporate social responsiveness, corporate social performance, corporate citizenship, business ethics, and stakeholder management. Carroll (1999) points out that CSR also overlaps with ESG principles, shared value creation, and sustainability, and he stresses the need to separate these concepts to avoid confusion. These conceptual challenges have recently led to the tendency in the corporate world and international organizations to use the concepts of CSR and ESG interchangeably, treating them as almost the same in all contexts (Awad, 2021; L. Chen, 2024; Nikolaeva-Aranovich, 2023).

Previous empirical studies used different dimensions, such as environmental, social, corporate governance, and economic, to assess CSR and apply various types of variables to measure it. In some cases, CSR is calculated as a continuous measure, which may be based on the sum of concerns, benefits, scores, or the weighted average of these. In other cases, CSR is assigned as a dummy variable, which leads to different methodological approaches. Taking these variations into account, we define the meta-variables ESG and CSRC (Corporate Social Responsibility Reporting) to control for variability and adjust the results to the differences in evaluation methods. A further complication emerges from the fact that some studies interpret CSR in terms of social irresponsibility, a view which focuses exclusively on negative scores and behaviors (Hoi et al., 2013). To address the confounding effects resulting from heterogeneity in the views and approaches to CSR in the primary regressions, we consider the meta-variable CSRR as a moderator to ensure a more consistent and reliable analysis of the results (Marques et al., 2023).

Another source of conceptual inconsistency in the literature stems from the fact that it often fails to adequately distinguish between CSR performance and CSR disclosure, even though they are distinct conceptual entities (H. B. Christensen et al., 2021; Xu et al., 2022). At the same time, empirical differentiation between the two concepts is made difficult by the CSR indicators used in primary studies. de Villiers and van Staden (2011) point out that the relationship between CSR information and CSR activity varies depending on the source of the information. In our meta-analysis, many studies calculate CSR relying on indicators provided by independent third-party databases, such as KLD, ASSET4, KEJI, and RKS. These databases are among the most widely used sources in the literature on ESG and corporate TA and are generally regarded as reliable and credible (Berg et al., 2022; D. M. Christensen et al., 2022), although each presents certain limitations (e.g., varying methodologies, coverage, and scoring criteria). By acknowledging these limitations, our analysis accounts for potential heterogeneity arising from differences across data providers. ESG ratings are provided by KLD (now MSCI ESG) and ASSET4 (Thomson Reuters). These ratings are based on publicly available information and company disclosures. CSR indices are also offered by KEJI and RKS, and these have been validated in prior research, ensuring comparability and robustness across studies. To address the heterogeneity arising from various data sources, we applied fixed effects controls in our meta-regression. This approach allowed us to examine the potential impact of the data origin on the estimation of the relationships between CSR disclosure and CSR performance.

2.6. Data Sample Characteristics of Primary Studies

Data composition and characteristics: The dataset enhances the external validity of the meta-analysis as it reflects both dominant contexts (the United States) and a broader international perspective. Countries in the dataset include the United States—the most frequently analyzed context—followed by China, the United Kingdom, France, Germany, South Korea, Australia, and other European and Asian countries (Table A1). While several studies employ multi-country samples, 76.8% use data from a single country, which may limit generalizability. Generalizability may also be limited by the different policy and regulatory frameworks, as 28% of studies draw on data from exclusively European countries, while 64.8% use data from non-European regions. The mean value for the year of the data, calculated as a logarithmic value, is 7.607, indicating that the sample includes relatively recent studies. Furthermore, on average, each analysis uses 9.76 control variables, highlighting the importance of methodological rigor.

Εconometric methods used: Ordinary least squares (OLS) is the most common approach (55.2%), while other methods, such as Probit, Logit, or Tobit, are applied in 14.4% of cases. A significant percentage (52%) includes firm fixed effects to improve accuracy. At the same time, year fixed effects are applied in 72.8% of cases, while country fixed effects are applied in 17.6%, which may limit comparisons between different countries (Stanley & Doucouliagos, 2012).

Additional control factors: Firm size (Size) is used in 84.8% of cases, highlighting its importance in interpreting tax strategy. Leverage (Lev) is included in 78.4%, indicating the role of financial structure. Property, plant, and equipment (PPE) is included in 68.8%, highlighting the impact of physical capital on tax policy. The use of these variables enhances the reliability of the estimates, offering a richer analytical framework for the examination of the relationship between taxation and ESG (Griffith & Klemm, 2004).

Table 3 reveals the main trends in methodological choices concerning the study of the relationship between taxation and the ESG index. Most analyses are based on the ETR index, while an average of 10 control variables is used to improve accuracy. The data are mainly from non-European countries, and the OLS method dominates the econometric estimates. These results indicate a complex and multi-level analysis of the factors influencing the relationship between tax rates and ESG, with significant implications for research methods and geographical conditions.

Table 3 describes the summary statistics used in the meta-regression analysis for 125 observations. We focus on the relationship between ESG and TA variables, examining partial correlation and standard deviation (SD). It appears that the relationship between the main variables in this sample is either extremely weak or negligible, as indicated by the almost zero partial correlation (−0.00012). This may either mean that the effects of ESG on TA are limited or that they are significantly influenced by other factors, such as national policy or the financial characteristics of the company. The small variance (SD = 0.06534) indicates stability in the ESG data, strengthening the reliability of the study results. However, further analysis is required to investigate whether there are interactions that enhance the significance of this relationship under specific conditions, which are discussed in the results section.

3. Results

3.1. Main Meta-Regression Results

In this section, we examine alternative specifications. Our estimated coefficients of the variables (

Table 4. Multivariate meta-regression results.

	WLS (1)		WLS (Cluster-Robust) (2)		IV (3)		Fisher’s z (4)		Robust, Statistical Significance (5)
(Intercept)	0.1137829	*	0.113783		0.100470		0.113783		No
	(0.0525979)		(0.12513)		(0.11593)		(0.12513)
StandardError	0.1693606		0.169361		0.203881		0.169361		No
	(0.3158892)		(0.25889)		(0.25466)		(0.25889)
Tax Avoidance Measures
ETR	−0.0122306		−0.012231		−0.01181		−0.012231		Νο
	(0.0098527)		(0.00734)		(0.00711)		(0.00734)
CETR	−0.0133669		−0.013367		−0.01289		−0.013367		Νο
	(0.0111646)		(0.00719)		(0.00696)		(0.00719)
DifTax	0.01250		0.01250		0.012067		0.01250		Νο
	(0.009647)		(0.00719)		(0.0068)		(0.00719)
ESG Proxies Employed
CSRC	−0.0021012		−0.00210	**	−0.00257		−0.002101		Νο
	(0.0181794)		(0.01003)		(0.00961)		(0.01003)
ESG	0.0045163		0.004516		0.004436		0.004516		No
	(0.0168947)		(0.00878)		(0.00852)		(0.00878)
CSRR	−0.0009637		−0.000964		−0.00111		−0.000964		No
	(0.0117576)		(0.00411)		(0.00393)		(0.00411)
Country Composition
S_Country	0.0735179	***	0.073518	**	0.070913	**	0.073518	**	Yes
	(0.0197276)		(0.02422)		(0.02255)		(0.02422)
Europe	−0.1652227	***	−0.165223		−0.15121		−0.165223		Νο
	(0.0454197)		(0.13013)		(0.12089)		(0.13013)
MixOfCountries	−0.1101372	*	−0.110137		−0.09721		−0.110137		Νο
	(0.0482161)		(0.12455)		(0.11547)		(0.12455)
NonEuropean	−0.1770462	***	−0.177046		−0.16207		−0.177046		Νο
	(0.0454244)		(0.13586)		(0.12634)		(0.13586)
Additional Control Variables
BIG4	0.0242636		0.024264	*	0.023449	*	0.024264	*	Yes
	(0.0156598)		(0.00997)		(0.00935)		(0.00997)
Board_Independence	−0.0177567		−0.017757		−0.01617		−0.017757		Νο
	(0.0174503)		(0.01693)		(0.01584)		(0.01693)
ROA	0.0180286		0.018029		0.018800	*	0.018029		No
	(0.0193579)		(0.00964)		(0.00921)		(0.00964)
MB_ratio	−0.0003018		−0.000302		−0.00018		−0.000302		No
	(0.0096440)		(0.00674)		(0.00656)		(0.00674)
Ppe	0.0053588		0.005359		0.003843		0.005359		No
	(0.0155344)		(0.01195)		(0.01102)		(0.01195)
Lev	−0.0348101		−0.034810		−0.03002		−0.034810		No
	(0.0205860)		(0.01780)		(0.01681)		(0.01780)
Ocf	0.0023683		0.002368		0.001101		0.002368		No
	(0.0163862)				(0.01289)		(0.01379)
Size	0.0059121		0.005912		0.004495		0.005912		No
	(0.0253432)		(0.01833)		(0.01710)		(0.01833)
Growth	0.0006983		0.000698		0.000878		0.000698		No
	(0.0123660)		(0.00884)		(0.00861)		(0.00884)
Intagible	−0.0108530		−0.010853		−0.00908		−0.010853		No
	(0.0131517)		(0.01546)		(0.01435)		(0.01546)
R.D.Intensity	0.0096877		0.009688		0.010087		0.009688		No
	(0.0097801)		(0.00579)		(0.00553)		(0.00579)
Invint	0.0141019		0.014102		0.012953		0.014102		No
	(0.0120892)		(0.01211)		(0.01139)		(0.01211)
Prob_Log_Tob_it	0.0042192		0.004219		0.003420		0.004219		No
	(0.0131609)		(0.01068)		(0.01026)		(0.01068)
OtherEstimator	−0.0063300		−0.006330		−0.00632		−0.006330		No
	(0.0083403)		(0.00563)		(0.00555)		(0.00563)
Observations	125		125		125		125
R2	0.484		0.461		0.463		0.461

Notes: Dependent variable: Partial correlation coefficient. Fisher’s z: transformation of the partial correlations (e.g., Dunn & Clark, 1969) used only in column (4). IV: instrumental variable used for StandardError. ***, **, * denote statistical significance at the 1%, 5%, and 10% levels, respectively. Details on the variables included are available in Table 2. Standard errors (in parentheses) are clustered at the study level (exception: column (1)).

) show the average effect of ESG criteria on TA, especially if the original studies deviate from the benchmark we have defined. For example, a higher correlation between ESG and TA is observed when expressed through extreme tax practices (the difTax variable), compared to the approach through the ETR index. Similarly, the meta-coefficient of ~0.0046 for the ESG variable suggests that the effect of ESG on TA is more positive in studies using an ESG measure based on four dimensions than those based on fewer dimensions.

Table 4 presents the results of the multivariate meta-regression analysis, starting with general observations and concluding with specific findings. Columns (1) and (2) are based on the weighted least squares (WLS) method. The full meta-regression model results, with all moderator variables, are presented in

Table A3. Meta-regression results, full model: WLS with the inverse of variances as precision weights–standard errors clustered at the study level.

	Dependent Variable:
	PartialCorrelationCoefficient
StandardErrorPartialCorrelation	0.124
	(0.096)
CETR	−0.013
	(0.008)
ETR	−0.013
	(0.008)
CSRC	0.002
	(0.01)
ESG	0.005
	(0.01)
CSRR	0.001
	(0.005)
S_Country	0.092 **
	(0.0375)
Europe	−0.126 *
	(0.106)
MixOfCountries	−0.059
	(0.098)
NonEuropean	−0.147 **
	(0.11)
BIG4	0.031 *
	(0.016)
Board_Indipendence	−0.018
	(0.017)
Roa	0.012
	(0.009)
MB_ratio	0.0002
	(0.007)
Ppe	0.018
	(0.016)
Lev	−0.020 *
	(0.011)
Ocf	−0.012 *
	(0.019)
Size	−0.005
	(0.011)
Growth	−0.004
	(−0.01)
Intagible	−0.011
	(−0.014)
R.D.Intensity	−0.001
	(0.01)
Invint	0.016
	(0.014)
FirmFixedEffects	0.006
	(−0.008)
YearFixedEffects	−0.023 *
	(0.014)
CountryFixedEffect	0.006
	(0.019)
Ols	0.012
	(0.012)
Prob_Log_Tob_it	0.018
	(0.019)
OtherEstimator	0.004
	(0.013)
Constant	0.064
	(0.096)
Observations	125
R-squared	0.484
Adjusted R-squared	0.333

Note: * p < 0.1; ** p < 0.05.

of Appendix A. Column (2), which is considered the baseline model, applies clustering of standard errors at the study level, while column (1) presents unclustered standard errors. The interpretation of the findings is focused on the baseline model in column (2). An important finding is that the coefficient on StandardError is not statistically significant. Additionally, according to

Table A2. Summary of studies examining the link between ESG and TA, as included in this meta-analysis.

	(1)	(2)	(3)	(4)
	(ALL)	(ETR)	(CERT)	(DifTax)
β1(SE)	0.0004	−0.037	−0.827	0.968
{FAT}	(0.405)	(0.4)	(1.3)	(1.37)
βo(interc.)	−0.006	−0.007	0.006	−0.013
{PET}	(0.005)	(0.006)	(0.01)	(0.018)
N	125	91	19	15
R-squared	0.0001	0.0002	0.056	0.053

Notes: The dependent variable is the partial correlation, with the exception of column (2), which uses Fisher’s z-transformed partial correlation. Numbers in brackets represent standard errors clustered at the study level. N is the number of observations. All reported estimates are based on Equation (4). Columns (3) and (4) restrict the sample to specific types of tax rate variables, respectively. All results were obtained by using weighted least squares (weights based on the inverse of the variances). FAT tests for the presence of publication selection bias. PET estimates the average effect corrected for publication selection bias.

of Appendix A, no reliable evidence emerges, indicating selectivity in publications, even when only the bivariate relationship between the partial correlation coefficients and their standard errors is considered. This indicates that there is no evidence of publication selection bias when other sources of heterogeneity are considered. As this coefficient has no units, it is often used to estimate the magnitude of bias. According to C. Doucouliagos and Stanley (2013), when the coefficient is statistically insignificant or has an absolute value of less than 1, selectivity can be classified as “small to moderate”. In our analysis, the coefficient of StandardError is statistically insignificant and less than one, suggesting that the selectivity is small to zero and does not practically affect the results.

None of the variables related to corporate tax show statistically significant coefficients. This suggests that differences in corporate tax strategies, as derived from different ways of measuring tax rates, are, on average, not statistically significant when other moderating variables are considered. The descriptive statistics mentioned earlier in Table 1, based on these variables, appear to reflect mainly the effect of underlying data and specification choices rather than substantive differences.

Table 4 presents an analysis of results from different estimators (WLS, cluster-robust WLS, IV, and Fisher’s z) and the findings concerning various factors related to TA, ESG indicators, country composition, and additional control variables. Four types of estimators are included: (i) WLS, used to weight observations based on their accuracy and displaying significant values for some variables, such as CSRC and S_Country; (ii) cluster-robust WLS, controlling for dependence of observations within groups, providing more conservative statistical values; (iii) IV (instrumental variables), used to address endogeneity, although it does not report significant deviations from the other models; and (iv) Fisher’s z, indicating robustness of findings. Analyzing the main categories of the table, we can conclude that the intercept is positive in the WLS (0.131) and not statistically significant with p-val = 0.37 (p < 0.05), while no significance is shown in the other models. This suggests that the mean of the dependent variable is positive when all independent variables have zero values. Similar standard errors are recorded in all models, which indicates stability in the estimates.

Tax avoidance measures: In our analysis, we use the term “tax burden” to refer to proxies of firms’ tax liabilities, which are based primarily on effective tax rate measures (ETR and CETR). These proxies are commonly used in primary studies to indicate the extent to which firms reduce their tax payments relative to their statutory obligations. Accordingly, the three TA indices examined in this meta-analysis suggest a correlation between TA and these measures of tax burden. However, the magnitude and direction of the association vary across contexts. Specifically, ETR shows a negative effect (~−0.0122), but none of the estimates are statistically significant. This suggests a weak link between ETR and tax strategies. CETR depicts a negative effect (~−0.013), also non-significant, indicating that the effect of tax compliance through CETR is minimal. DifTax presents a slightly positive effect (~0.0012), but it is not statistically significant. No strong effect of the difference in tax rates is observed.

ESG indicators: The CSRC index (−0.0021) is statistically significant (p < 0.05). The negative coefficient indicates that a company’s commitment to CSR is associated with lower levels of TA. This means that companies that invest substantially in social responsibility are more likely to implement transparent tax practices and avoid aggressive tax reduction strategies. The ESG index (0.0045) is not statistically significant. The positive but not statistically significant coefficient indicates that the ESG index does not have a clear or consistent relationship with TA. Therefore, companies with high ESG scores do not differ significantly in their tax behavior. CSRR (−0.00096) is not statistically significant. The negative correlation suggests that ESG reporting may be associated with slightly lower TA, but this relationship is not statistically significant. This means that publishing ESG reports does not guarantee ethical tax behavior (Sadjiarto et al., 2024). Companies may publish ESG reports without having a substantial commitment to transparency or avoiding aggressive tax practices.

Country composition: The variables related to the geographic regions of the companies provide information on how location may affect corporate tax behavior. The S_Country index (0.0735) is statistically significant: (p < 0.01). A positive and statistically significant coefficient indicates that companies operating in certain countries may have higher tax rates or lower levels of TA. This suggests that the country in which a company is headquartered plays an important role in its tax practices, likely due to differences in tax policy regulation and regulatory oversight. National rules and cultural norms can affect how a company acts in terms of the environment, social issues, and taxes. This applies to the individual firms included in the datasets, whose behavior is shaped by the regulatory and cultural context of their home country.

The Europe index (−0.1652) is statistically significant (p < 0.01). A negative and statistically significant coefficient indicates that studies conducted on European samples report stronger ESG/TA associations, which may reflect measurement conventions, enforcement intensity, or sample composition rather than underlying TA levels.

This can be explained by several factors: (i) According to Szołno-Koguc and Ołówko (2019), low tax rates may act as a factor that favors TA. However, this finding does not imply that high rates necessarily discourage TA; on the contrary, the relevant literature shows that in high-tax environments, firms often have stronger incentives to implement aggressive tax planning strategies (Slemrod & Gillitzer, 2013); (ii) the presence of many multinational companies in Europe, which use complex profit shifting structures to reduce their tax burden (Huizinga & Laeven, 2008); (iii) differences in tax regulations between European countries, which allow companies to choose countries with more favorable tax regimes (Jesus et al., 2024).

The MixOfCountries indicator (−0.1101) is statistically significant (p < 0.10). The negative correlation indicates that studies that include samples from many countries record more TA. Some studies use multi-country samples, which comprise firms based in different countries within the same dataset. This design allows for cross-country comparisons and captures the diversity of institutional tax regimes and ESG practices. Additionally, including multinational firms highlights how opportunities for tax planning can arise from differences in different jurisdictions, thereby enriching the analysis of the relationship between ESG and tax behavior. Finally, the NonEuropean indicator (−0.1770) is also statistically significant (p < 0.01). The negative coefficient shows that companies outside Europe exhibit higher levels of TA. This suggests that tax strategies are more widespread in other regions, such as North America or Asia, where many multinationals utilize low-tax jurisdictions (e.g., Singapore, Hong Kong).

Additional control variables: These variables examine various financial and administrative factors that may affect a company’s tax strategy. The BIG4 (Big 4 Audit Firm) indicator (0.0243) is statistically significant (p < 0.10). Studies using samples of firms audited by Big 4 (PwC, Deloitte, EY, KPMG) report weaker ESG/TA associations, consistent with the interpretation that audit quality may moderate how ESG relates to reported tax outcomes. This can be explained by the fact that larger audit firms tend to adhere to stricter regulatory requirements and limit aggressive tax practices. The Board Independence indicator (−0.0178) is not statistically significant. The negative coefficient suggests that companies with independent boards may choose more responsible tax strategies, but this relationship is not statistically significant. The ROA (return on assets) ratio (0.0180) is statistically significant (p < 0.10). Companies with higher profitability (ROA) show less TA. This may mean that profitable companies do not need to resort to aggressive tax strategies, as they can absorb the higher tax burden without it affecting their competitiveness. Lev (Leverage) ratio (−0.0348) is not statistically significant. The negative coefficient indicates that highly leveraged companies may implement more tax strategies to reduce their tax burden (e.g., through tax deductions on loan interest).

Overall, geographic location and audit by large audit firms emerge as important factors that influence tax behavior. In addition, high profitability is associated with lower TA. However, other financial indicators, such as leverage and board independence, do not show a strong statistical relationship with TA.

The results highlight the complexity in linking ESG, tax compliance, and country composition. In particular, ESG indicators are negatively associated with TA, while country composition has mixed effects. Additional control variables provide a partial explanation, but not all are statistically significant. The explanatory power of the models (R²: 0.46–0.48) indicates that the data fit moderately well, while the use of multiple estimators yields a more complete picture.

3.2. Distribution of Partial Correlation and Publication Selection Bias

This section analyzes the distribution of partial correlation coefficients (pccs) and tests for publication selection bias, revealing significant heterogeneity in the estimated effects of ESG on firms’ tax behavior. First, we examine the distribution of the partial correlation coefficients and the precision of these estimates. We provide descriptive information about the data and then present the results of the multivariate regression models. Figure 1 presents a funnel plot, following the method of Sutton et al. (2000). The horizontal axis depicts the 125 estimated ESG/TA associations reported in studies based on European versus non-European samples, while the vertical axis shows the precision (the inverse of the standard error). The mean value of the pcc is −0.00012, and the median is 0.0009, indicating small effects (see also Table 1). The plot also reveals significant variance in the estimates: the minimum pcc is ~−0.16, while the maximum is ~0.45; The standard deviation is 0.065, indicating high heterogeneity. The majority of estimates are based on various tax behavior variables. In total, 72.8% use the ETR index, and 15.2% use the CETR index, while smaller percentages (12%) are related to other variables. The most accurate estimates (top of the diagram) show positive correlations, which reinforces the view that tax policies are influenced by strategic behaviors.

The symmetry of the funnel is important for assessing the presence of publication bias: a symmetric funnel indicates that the estimates are balanced, with the most accurate ones being close to the true outcome and the less accurate ones being widely dispersed. In Figure 1, we observe asymmetry, as the right part of the funnel is “heavier” than the left. This may indicate that positive associations dominate. Although visual inspection may provide clues, formal statistical tests are necessary to assess potential publication selection bias. The interpretation of the funnel plot suggests the existence of positive associations, but the asymmetry may reflect publication selectivity. To ensure the accuracy of the conclusions, we continue with multivariate regression models, which allow for the simultaneous analysis of other heterogeneity factors.

3.3. Robustness Checks

The reliability of the results presented in column (2) of Table 4 was examined through two approaches: the use of an instrumental variable (IV) estimator and the application of a transformation of the dependent variable. In the model in column (3), we use the StandardError variable as the dependent variable, applying an IV estimator with the number of observations as an instrument. Larger studies are usually characterized by higher precision compared to smaller ones, but the number of observations is not strongly correlated with the choices of methods or data. Specifically, the inverse of the square root of the number of degrees of freedom is calculated, an estimate that is analogous to the standard error (Cazachevici et al., 2020).

As can be seen from the results in column (3) of Table 4, the use of variable IV confirms the stability of the results for the moderator variables, while the StandardError variable becomes statistically significant. However, the findings change again in column (4).

A potentially important issue with using the partial correlation coefficient and standard error is the non-normality of their distribution, particularly when the correlation values approach the limits of −1 or 1 (Stanley & Doucouliagos, 2012). To address this issue, we applied Fisher’s z transformation (Dunn & Clark, 1969), a widely accepted method for improving the distribution.

In column (4) of Table 4, the results remain robust to the use of Fisher’s z, with no evidence of publication selection bias. Furthermore, the data choices and specifications discussed mitigate the effects of tax competition. The only exception is the Prior variable, which loses statistical significance compared to the basic model in column (2). The application of the IV estimator and Fisher’s z transformation confirms the stability of the post-regression results and strengthens the certainty of the absence of publication selection bias. The data suggest that specification choices play an important role in measuring the effects of corporate tax competition.

Furthermore, results were tested for robustness using different standardized effect sizes, in addition to the partial correlation coefficient. As a further sensitivity check, we calculated semi-elasticities from the regression data in the primary studies. Semi-elasticities indicate the percentage change in the corporate tax rate when the explanatory variable changes by one unit. It is worth noting that our sample was limited to only 61 estimates, which is about half the original sample size when using the partial correlation coefficient. This was due to the fact that many primary studies did not provide all the necessary information for constructing semi-elasticities. In particular, several studies failed to include descriptive statistics, which are crucial for calculating semi-elasticities (for more details, see Leibrecht & Hochgatterer, 2012). Using the WLS estimator, the mean semi-elasticity was estimated to be −0.00002 (95% confidence interval: −0.0002 to 0.0002). Since the confidence interval includes zero, this relationship is not statistically significant.

Examining semi-elasticities instead of partial correlations presents a challenge, as semi-elasticities are not always directly comparable due to their dependence on the unit of measurement of the independent variable. For comparisons, semi-elasticities based on the same corporate tax variable should be used instead.

Table A4. Semi-elasticity descriptive statistics.

	All Estimates	ETR	CETR	DifTAx
Number of estimates	62	44	11	7
Median	0.0001	0.00024	−0.00005	−0.00002
Unweighted average	−0.1906	−0.1291	−0.0042	−0.9271
Inr.WLS	−0.00002	−0.00005	0.03464	0.000008
95% confidence	−0.0002	−0.0002	0.01481	−0.0005
Interval 9unr.WLS)	0.0002	0.00008	0.05448	0.0006

Notes: Unr. WLS: unrestricted weighted least squares, using the variances as precision weight of semi-elasticity coefficients; ETR; CETR; DifTax.

in Appendix A shows that the weighted average semi-elasticity is negative for ETR and positive for CETR and DifTax. The WLS values are almost zero, suggesting that the tax response to changes in neighboring countries’ tax rates is negligible. In particular, the CETR and DifTax values (0.03464 and 0.000008, respectively) indicate that in some cases there may be a positive tax response, that is, countries increase taxes when neighboring countries’ tax rates increase. However, the average values are too low to be considered statistically significant. The descriptive statistics of the semi-elasticities broadly confirm the existence of tax competition effects, as documented in Appendix A.

3.4. Additional Robustness: General-to-Specific Modeling

As an additional robustness check, we applied the General-to-Specific (Gets) modeling strategy (Hendry, 1995), starting from the unrestricted specification and sequentially eliminating statistically insignificant regressors through backward elimination using weighted least squares (WLS). This procedure yielded a parsimonious model retaining only moderators with robust explanatory power.

The Gets robustness check, implemented with backward elimination and REML estimation, produced a mixed-effects specification explaining over 80% of the observed heterogeneity (R² = 81.8%, I² = 34.9%) with low residual variance (τ² = 0.0002). Several moderators remain highly robust across specifications, including S_Country, Europe, MixOfCountries, NonEuropean, BIG4, Leverage, R&D Intensity, and Investment Intensity. The variable ROA is only marginally significant (p ≈ 0.051). These results confirm the stability of the baseline findings and indicate that firm-level and regional characteristics capture the majority of variation in reported effects (see

Table 5. General-to-Specific (Gets) robustness check—final specification (cluster-robust SEs).

Predictor	CR_Estimate	CR_SE	CR_Pval	ME_Estimate	ME_SE	ME_Zval	ME_Pval	Sig
(Intercept)	0.1234	0.1395	0.3829	0.1152	0.0338	3.411	0.0006	***
S_Country	0.0610	0.00937	<0.001	0.0755	0.0122	6.191	<0.0001	***
Europe	−0.1743	0.1415	0.2271	−0.1741	0.0327	−5.321	<0.0001	***
MixOfCountries	−0.1235	0.1399	0.3841	−0.1195	0.0340	−3.518	0.0004	***
NonEurope	−0.1839	0.1432	0.2081	−0.1884	0.0331	−5.694	<0.0001	***
BIG4	0.0186	0.00300	<0.001	0.0172	0.0066	2.612	0.0090	**
Roa	0.0184	0.00966	0.0655	0.0200	0.0102	1.953	0.0508	.
Lev	−0.0334	0.01088	0.00434	−0.0360	0.0109	−3.299	0.0010	***
R.D.Intensity	0.0135	0.00318	<0.001	0.0134	0.0063	2.142	0.0322	*
Invint	0.0163	0.00520	0.00361	0.02141	0.0081	2.986	0.0028	**

Notes: Mixed-effects model selected via General-to-Specific (Gets) backward elimination. Cluster-robust standard errors account for within-study correlation. CR = Cluster-robust SEs (naive-t, df = 32), ME = mixed-effects model (REML, k = 125), Sig codes: *** p < 0.01, ** p < 0.05, * p < 0.1, residual heterogeneity (REML model): τ² = 0.0002, I² = 34.9%, R² = 81.8%.

).

4. Discussion and Conclusions

Over the past decade, research on the interplay between the two main variables (i.e., ESG and TA) has focused on the extent to which companies’ social and environmental actions include compliance with paying their fair share of taxes as part of their corporate responsibility. However, empirical results remain contradictory. On the one hand, the existing literature does not provide a clear consensus on whether ESG activities are associated with TA. On the other hand, where an association is found, there are differences in both the direction and magnitude of this nexus, making the findings inconsistent.

Our study used meta-analytical techniques to investigate the heterogeneity of empirical findings on the relationship between ESG and TA, focusing on design features of primary studies. Our database included 125 estimates from 33 studies, providing a rich dataset for analysis. Varying results are produced by different proxies for TA, as they capture distinct aspects of firms’ tax behavior. From our perspective, long-term ETR measurements (for example, three- to five-year averages) (Dyreng et al., 2008; Gupta & Newberry, 1997) and tax reserves such as FIN48/UTP (Blouin, 2014) are preferable since they decrease timing effects and provide a more precise reflection of ongoing tax avoidance strategies in comparison to yearly ETR. In particular, researchers should rely on long-run measures of TA (e.g., multi-year ETRs) and harmonized ESG indicators to reduce bias and enhance comparability across studies, thereby providing policymakers with more robust and consistent evidence. Moreover, studies using the CETR or DifTax present different results compared to those based on ETR. This suggests that the choice of TA measure significantly influences the conclusions. Studies based on comprehensive CSR measures, which include social, environmental, governance, and economic dimensions, showed a stronger effect of ESG on TA. Similarly, studies that used binary variables or examined irresponsible CSR practices also reported larger effects.

It was observed that econometric models significantly influence the relationship between ESG and TA. Models that do not include critical controls (e.g., control variables or time-constant effects) are biased, overestimating the relationship studied. Our meta-analysis highlights the need for careful selection of both TA avoidance indicators and ESG measures to produce reliable findings. Furthermore, the choice of econometric estimation strategy has a significant impact on the results. In particular, the application of an OLS estimator appears to produce higher tax estimates compared to the use of an instrumental variable (IV) estimator. A possible reason for this difference is that IV estimators, as a rule, concern underlying endogeneity issues, which may arise when governments coordinate their tax policy. In contrast, the OLS method cannot adequately address these issues, which may lead to different estimates.

Most of the predicted partial correlation coefficients were found to be not statistically significant. This implies that, on average, there is no systematic association between ESG and TA, which is in line with the findings of Mayberry and Watson (2021) and Mitroulia et al. (2025). Mayberry and Watson (2021), using a natural experiment, concluded that there is no association between the institution of statutory constituencies, which caused an exogenous shock to CSR and TA. They argue that firms decouple CSR from their tax policy. Following the research design of Hunter and Schmidt (1990), Mitroulia et al. (2025) conducted a standard meta-analysis that synthesizes empirical and quantitative data on the relationship between corporate sustainability and tax behavior. The results revealed the absence of a statistically significant effect on an important topical issue while highlighting a strong identification method common in management and accounting. This lack of a statistically significant correlation may reflect a discrepancy between actual corporate tax behavior and the communication or institutional signaling associated with ESG initiatives. This highlights the difference between real actions and projected commitments. In other words, companies’ ESG disclosures and strategies may not necessarily result in changes in their tax behavior. This suggests that ESG efforts could be more symbolic or communicative than directly impactful in terms of TA. Our predictions on the relationship between ESG and TA are fully aligned with their findings, supporting the general validity of their conclusions.

The analysis also revealed that the source of ESG measurement can influence the estimated relationship. Especially in scenarios that examine corporate practices through different assessment methods, the findings vary. In order to better understand the nature of this relationship, we examined the behavior of companies in extreme stability scenarios. When we adapted the design of the hypothetical study to include a measure assessing corporate social irresponsibility (CSI), we observed a moderate positive correlation between CSI and extreme tax aggressiveness. This result is supported by previous studies (Davis et al., 2016; Hoi et al., 2013; Marques et al., 2023), which reported a statistically significant correlation under conditions of extreme corporate practices.

The findings highlight the importance of context and methodology in studying the relationship under investigation. On average, a strong correlation is not observed, while in extreme scenarios of corporate irresponsibility, tax aggressiveness appears significantly increased. This finding highlights the need for differentiated approaches to ESG measurement and the interpretation of results in different economic and institutional contexts (Delegkos et al., 2022; Matsali et al., 2025). Recognizing the importance of publication bias in meta-analyses, the present study investigated its impact on overall effect sizes. The findings suggest that, in the context of the present analysis, selective publication is not a significant source of heterogeneity.

Several avenues remain open for refining and extending this work since some limitations may have affected our conclusions. First, our study is constrained by the inherent limitations of the primary literature, including the use of heterogeneous ESG and TA measures and the fact that meta-regression cannot entirely resolve endogeneity in primary studies. Second, our own methodological choices present constraints: the assumption of homogeneity may misrepresent heterogeneous phenomena; we included only English language studies; and our broad moderator analysis may have increased the risk of overfitting. Finally, generalizability may be affected by limited database coverage and the evolving nature of ESG and tax regulations during our sample period (2012–2023).

Future research could enhance methodological robustness by pre-specifying a smaller set of theoretically motivated moderators and applying advanced techniques (e.g., false-discovery rate controls, shrinkage estimators like LASSO meta-regression, robust estimators, and influence analysis) to confirm the stability of findings. Measurement precision can be improved by employing more precise and consistent proxies (e.g., FIN48/UTP for TA, multi-year measures, ensuring consistent ESG constructs) and conducting more granular analyses by stratifying results by ESG provider, ESG dimension (E, S, G), type (disclosure vs. performance), and TA proxy. Theoretical framing can be strengthened by explicitly linking competing perspectives (e.g., stakeholder vs. agency theory) to specific moderators (e.g., disclosure-based scores, Big 4 auditors). Future research should also investigate how institutional factors (e.g., regulatory frameworks, enforcement strength, cultural norms, industry dynamics, ownership structure) shape the ESG/TA relationship through cross-country studies and quasi-experimental designs on tax reforms. Finally, future studies could assess the unique influence of the individual dimensions of ESG, and qualitative analyses could reveal managerial motives behind tax strategies.