5.1. Graphical and Correlation Analysis
Before delving into a comprehensive discussion of our empirical findings within the framework of regression analysis, it is prudent to initially explore the basic relationship between CSR and BTD through a simple scatter plot, as depicted in
Figure 3. The analysis reveals a positive relationship between the company’s engagement in CSR initiatives and BTD. Notably, this association is particularly pronounced in total BTD and permanent differences, albeit less conspicuous in the case of temporary differences. Such observations lend credence to the notion that a considerable proportion of companies in Indonesia perceive CSR as a strategic mechanism to secure tax deductions, a sentiment echoed in extant literature [
65,
66]. Consequently, it fortifies the contention that the surge in overall BTD during the pandemic was primarily propelled by permanent differences rather than temporary differences (see
Figure 1).
Furthermore, a nuanced analysis of
Figure 3 suggests a prevalence of strategic incentives over normative considerations among firms engaging in CSR initiatives, thereby aligning with theoretical frameworks such as risk management and moral licensing theories. Hence, it is plausible to posit that companies grappling with financial constraints resort to accounting and financial manipulations, thereby fostering an inclination towards proactive CSR involvement to obfuscate such deviant behavior. This strategic maneuver, whether as a means of risk mitigation over sanctions and penalties or as a compensatory measure for perceived transgressions through benevolent actions, underscores the complex interplay between CSR endeavors and BTD dynamics in corporate landscapes, specifically in the Indonesian context.
A preliminary overview of our study’s outcomes is also supplied by the Pearson correlation matrix, which provides a rigorous formulation of the degree of linear relationship between two variables, as outlined in
Table 5. Consistent with the patterns depicted in
Figure 1, the economic turmoil stemming from the pandemic demonstrates a statistically significant positive correlation with both total BTD and permanent differences, albeit not with temporary differences. Moreover, a notable positive correlation between CSR and adverse economic conditions is evident, as previously captured in
Figure 2. However, it is crucial to emphasize that this correlation does not signify the presence of multicollinearity between the two variables (
and
), as the correlation coefficient falls well below the commonly accepted threshold of 0.7 (i.e., 0.154), as recommended by Gujarati [
151] and adopted by numerous subsequent studies. Therefore, the joint inclusion of these variables in regression analysis does not introduce bias to the estimation model. Lastly, echoing the insights gleaned from
Figure 3, our analysis unveils a significant positive correlation between corporate engagement in CSR initiatives and two dimensions of BTD, namely total and permanent differences. Nevertheless, the correlation with temporary differences is negligible. Taken together, the correlation analysis in this subsection reaffirms the dominance of strategic motives over normative considerations in propelling corporations to engage with CSR agendas, particularly in times of crisis.
Table 5 also offers insights into the correlation coefficients concerning the control variables. For instance, there exists a negative correlation between firm size and BTD, suggesting that entities with larger asset accumulations prioritize precision and transparency in financial reporting to mitigate potential penalties and uphold their reputation. An extensively embraced notion posits that larger enterprises are inherently more visible [
152], consequently rendering them susceptible to closer scrutiny from regulatory bodies, investors, and various stakeholders, including the press, independent non-governmental organizations, and social movement entities [
153]. The pronounced negative correlation, particularly evident with temporary differences, underscores the validity of this proposition. As previously discussed, temporary differences undergo reconciliation in subsequent periods, a process that reveals any inconsistencies during audits or reviews conducted by authorities. Given the current emphasis on tax compliance and the regularity of audits, temporary differences emerge as prominently discernible to authorities, rendering them more readily exposable compared to permanent differences.
On the other hand, profitability presents contrasting findings compared to company size, exhibiting a positive and statistically significant correlation with the three measures of BTD. Enhanced profitability means greater tax liabilities, which may motivate firms to pursue more proactive tax minimization strategies within legal frameworks to maximize after-tax profits [
154]. Practically, this involves leveraging tax loopholes, deductions, credits, and other tax savings tactics, consequently widening the rift between reported financial income and taxable income. Lastly, the remaining control variables, including leverage, tangible assets, intangible assets, and inventory, exhibit mixed coefficient signs, where their correlation with temporary differences consistently presents anomalies.
Though the scatter plot diagram and correlation coefficient matrix provided above offer valuable insights into the nature of relationships between variables, these analyses do not account for the influence of other variables or control for confounding factors, potentially leading to incomplete or misleading findings [
155]. Regression analysis, on the other hand, allows researchers to explore the relationship between a dependent variable and one or more independent variables while controlling for other pertinent factors. By simultaneously integrating multiple predictors, regression analysis facilitates a more nuanced understanding of the factors influencing the dependent variable, enabling the identification of significant predictors and the quantification of their effects. Thus, regression analysis offers a more comprehensive and rigorous approach to examining relationships between variables in a quantitative study. Detailed interpretation and discussions on the results of the regression analysis will ensue in the following subsections.
5.2. Baseline Results
Table 6 presents the baseline results concerning the influence of the economic crisis and CSR on BTD. In the current analysis, three individual measures of BTD, namely total, temporary differences, and permanent, were examined separately, resulting in three distinct specifications. In specifications 1 and 3, the Hausman test consistently rejects the null hypothesis of no FE as the
p-values are below the 1% threshold. Consequently, over the RE model, the FE model is chosen as a suitable approach for estimating Equation (3). In contrast, the Hausman test results for specification 2 indicate that the coefficient estimates from the RE model are consistent, leading to the decision to employ the RE model for this specification.
Our findings indicate that the economic crisis exerts a positive and significant influence on BTD in aggregate (
= 0.054,
p-values < 1%) and permanent differences (
= 0.052,
p-values < 1%) but not on temporary differences (
= 0.001,
p-values > 10%). Therefore, these results partially support the established hypotheses, namely H1a and H1c. CSR has a significant positive effect on total BTD (
= 0.190,
p-values < 1%), indicating that higher CSR disclosure in annual or sustainability reports increases a company’s overall BTD level. This impact primarily stems from an increase in permanent differences in financial reporting. As illustrated in
Table 6, the coefficient for the CSR-permanent differences relationship (
= 0.170,
p-values < 1%) is more pronounced compared to the coefficient for CSR-temporary differences (
= 0.018,
p-values < 10%). More substantively, an increase in the ratio of permanent differences and temporary differences relative to total assets by 0.17% and 0.02%, respectively, is concurrent with a 1% increase in the CSR disclosure index within a given fiscal year. After all, these findings provide support for any hypotheses related to the influence of CSR on outcomes of interest, namely H2a, H2b, and H2c.
Surprisingly, when the interaction between these two primary independent variables is examined to assess the moderating impact of CSR (i.e., ), the estimated coefficient shows an unexpected negative direction and fails to achieve statistical significance even at the 10% level, especially in specifications 1 and 3. It suggests that CSR does not moderate the relationship between the economic crisis and BTD—whether in total or permanent differences. H3a and H3c are thus rejected. These findings underscore that the economic crisis and CSR stand individually in their influence on total BTD and permanent differences. However, the moderating effect of CSR on the relationship between the economic crisis and temporary differences in such a way that the effect is negative is proven to be significant at the 5% level and thus supports H3b. This result may be unexpected, but it can be explained by several reasons. For instance, companies may strategically adjust their CSR spending or operational strategies to maintain financial stability during an economic crisis, thereby potentially aligning financial reporting practices with taxable income due to tighter financial oversight. Additionally, CSR initiatives that increase transparency and compliance with regulatory requirements during a crisis can reduce the variability of temporary differences that can arise in stressful economic conditions. Therefore, although CSR generally increases temporary differences, its moderating effect during economic crises can paradoxically reduce the magnitude of these differences by encouraging more conservative financial reporting practices.
The baseline findings of this study indicate that the economic crisis and CSR may lead to elevated levels of BTD among listed companies in Indonesia. Such positive effects appear predominantly driven by the augmentation of permanent differences rather than temporary ones. With respect to the economic crisis, the result supports the notion that in challenging periods like pandemics, businesses are motivated to employ tax planning and earnings management as tactics to alleviate economic strain. By effectively managing taxes and earnings, businesses can allocate resources more efficiently, invest in innovation, or adjust pricing strategies to remain competitive in their respective markets [
156,
157,
158]. Practically, it could involve deferring the recognition of revenue or accelerating the recognition of expenses in financial statements to smooth out taxable income during volatile economic periods [
159,
160]. Such actions can create temporary differences between the timing of financial accounting income and taxable income [
161,
162]. Alternatively, this objective can be accomplished through leveraging tax incentives or credits, which effectively lower the effective tax rate applied to taxable income [
163]. This reduction is a permanent difference because it affects the actual amount of tax paid, rather than timing differences that reverse over time [
164].
Given that the economic crisis significantly affects permanent differences rather than temporary ones, public companies in Indonesia have probably predominantly utilized tax credits rather than the timing of revenue and expenses to navigate financial challenges during the pandemic period. This argument is underscored by the Indonesian government’s introduction of tax credits as part of its strategy to stimulate economic activity amidst the crisis. According to Article 12, paragraph 1 of PMK number 9 of 2021, the government implemented a policy offering a 50% reduction in income tax installments due for the period from January 2021 to June 2021, which was then extended to December 2021. This incentive applies to 1018 business classifications, companies designated as KITE, and those holding permits, such as bonded zone operators or bonded zone entrepreneurs. Companies benefiting from this tax installment reduction are required to submit monthly realization reports using designated forms no later than the 20th of the month following the tax period. During the crisis, over 58,000 taxpayers, including individuals and entities, utilized these tax incentives, resulting in a total absorption of IDR 25.23 trillion. According to Ayers et al. [
137], such incentive policies typically impact permanent differences more significantly than temporary ones. In essence, our findings concerning the link between the economic crisis and BTD are well-grounded from practical and scientific standpoints.
Our research findings, which highlight the positive impact of CSR on permanent differences, reinforce the notion of CSR being utilized tactically for minimizing tax liabilities [
65,
66]. Considering that not all CSR expenditures qualify as deductible donations according to tax regulations, when the non-deductible ones are erroneously reported in tax returns, it will be considered an overstatement on tax deductions, thereby exacerbating permanent differences. Errors in reporting non-deductible CSR expenditures for tax purposes can stem from mechanical mistakes or taxpayers’ lack of understanding regarding tax regulations [
165]. Furthermore, in jurisdictions with intricate tax systems, such as Indonesia, variations in interpretation between tax authorities and taxpayers regarding tax rules may contribute to such discrepancies [
6].
Moreover, consistent with our findings on the link between the economic crisis and temporary differences, which is less prominent, firms may be less interested in implementing further managerial strategies by manipulating the timing of recognition of deductible CSR spending for both tax and market purposes. There are some plausible reasons behind these results. Firstly, during periods of economic uncertainty, companies may experience a downturn in profitability, diminishing the benefit of deferring expenses since there may be insufficient future taxable income to offset the deductions [
166]. Secondly, heightened oversight from tax authorities and stakeholders typically intensifies during a crisis [
100,
101]. Consequently, attempts to smooth income are more likely to be detected as these practices undergo scrutiny and reconciliation in subsequent periods, potentially undermining public confidence in the quality and informativeness of companies’ financial reporting. Thirdly, corporate engagement in earnings management practices, particularly through income smoothing, generally diminishes following the full convergence of IFRS in various countries [
167], including Indonesia [
168]. Fourthly, despite companies engaging in income smoothing through adjustments in the timing of deductible CSR expenditures, the resulting impact on temporary differences may exist but is typically less substantial. It is due to regulatory constraints that CSR expenditures are capped at 5% of net income.
Overall, our empirical findings provide strong evidence that strategic motives are the primary catalysts prompting companies to engage more actively in CSR programs. This phenomenon is particularly pronounced during periods of crisis, as evidenced by a significant surge in the CSR disclosure index compared to stable economic periods (see
Figure 2). In this context, companies utilize CSR initiatives to enhance their corporate image in response to adverse events within their internal environment during the pandemic, such as aggressive tax planning and earnings management. This strategic approach not only aims to mitigate scrutiny from regulatory authorities but also to minimize the risk of sanctions and penalties, thereby aligning with the notion of risk management theory. An empirical study by Rudyanto & Pirzada [
169] seems to provide support for this inference. Using panel data for companies listed on IDX between 2014 and 2016, they conclude that environmentally insensitive companies require sustainability reporting to reduce the reputational costs of tax avoidance. Also, the cross-country study supplied by Mutuc et al. [
170] using panel data for companies in eleven Asian countries (including Indonesia) revealed the opportunistic and concealed role of CSR on earnings management. Additionally, companies may sometimes consider pausing their strict adherence to accounting and tax regulations under the guise of having done something noble by engaging in CSR activities during a crisis, which is in line with insights drawn from moral licensing theory. List & Momeni [
171], with their experimental study involving 1500 workers from Amazon’s Mechanical Turk (Mturk), posit that the altruistic nature of CSR encourages workers to commit various frauds, which, if related in the context of our study, could include fraud in the field of accounting and taxation, which in fact is decided by the manager as a worker in a company. Eventually, our findings regarding the positive effects of CSR on total BTD can be both empirically and practically justified.
Surprisingly, when the economic crisis and CSR are combined into an interaction variable, their influence on all BTD measures becomes statistically insignificant, indicating that CSR does not moderate the relationship between the economic crisis and BTD. There is no clear justification for this perplexing result, especially considering that both factors individually play substantial roles in widening BTD. Nevertheless, implicit lessons can be drawn from these findings. For instance, it can be interpreted that firms strategically utilize CSR as a tactic to manage perceptions and mitigate the impact of economic turmoil on financial reporting practices. It is acknowledged that the pandemic has prompted firms to aggressively intensify tax planning and earnings management strategies, thereby increasing BTD. Within this context, CSR initiatives serve as a tool for firms to mitigate reputational risks associated with such deviant managerial behaviors during periods of turbulence. By implementing CSR programs that benefit communities, firms aim to portray a positive corporate image and signal adherence to ethical standards, allowing them to manage stakeholder perceptions regarding the credibility of financial disclosures.
This argument aligns with the concept of the ‘halo effect’ [
172] and Mercer’s [
173] model of financial disclosure credibility, where investors’ perceptions of management’s trustworthiness—shaped by companies demonstrating positive CSR performance [
174]—influence evaluations of the biased or unbiased nature of earnings reports. Wang & Tuttle [
175] empirically confirm this ‘halo effect’ through experiments involving business graduate students in the United States as proxies for non-professional investors. Their study indicates that CSR significantly influences how investors assess the credibility of financial disclosures. Given that BTD is a relevant indicator of credible disclosure [
9], CSR expenditures during the pandemic have the potential to mitigate negative perceptions associated with high BTD, thereby diminishing the explanatory power of the economic crisis on BTD, especially on temporary differences.
Turning to the control variables, profitability (
) consistently exhibits a positive and significant coefficient across all three BTD measures, as evidenced in the correlation analysis (see
Table 5). Thus, our conclusion remains intact: companies with higher profitability are more inclined to employ managerial strategies that optimize after-tax profits. Intangible assets (
) show a similar pattern in their effect on BTD as economic crises do—positively influencing total BTD and permanent differences, albeit orthogonal to temporary differences. These findings suggest that the accounting treatment of intangible assets may afford managers substantial discretion [
176] and receive comparatively less scrutiny from auditors [
177]. As noted by Gao et al. [
178], the productivity of a company’s intellectual property, such as patents, relates positively to its tax avoidance strategies. Additionally, Kimouche [
179] highlights that managers in France strategically use intangibles and goodwill to manipulate earnings. Inventory (
) yields contrasting results compared to return on assets, exhibiting a negative and significant influence on all BTD constructs. This finding aligns with Taylor & Richardson’s [
180] assertion that companies with high inventory levels should exercise caution to avoid excessive tax burdens. Apart from that, leverage (
) demonstrates a positive and significant effect on temporary differences. Lastly, company size (
) and tangible assets (
) are not significantly associated with the three BTD measures.
5.3. Robustness Check: Dealing with Endogeneity Bias
A standard practice in empirical studies involves conducting ‘robustness’ checks, where researchers explore how baseline regression coefficients react when the regression specification is altered. If the signs and magnitudes of the coefficients remain stable and plausible, it is typically considered evidence that these coefficients are ‘robust’ and can be interpreted reliably as the true causal effects of the associated regressors [
181]. Accordingly, Leamer [
182] strongly advocated for such investigations, contending that the ‘fragility’ of the coefficients may signal specification errors and that sensitivity analyses (i.e., robustness checks) should be routinely performed to help diagnose misspecification.
Bernard [
183] and Van Lent [
184] argue that misspecification can arise due to endogeneity within the estimation models. Endogeneity poses a critical methodological challenge in numerous realms of accounting research that utilize regression analysis to infer causality [
185]. Many studies in these fields struggle to effectively address endogeneity issues [
186]. Hamilton & Nickerson [
187] and Antonakis et al. [
188] highlighted a striking finding: approximately 90% of papers published in prestigious journals fail to adequately mitigate endogeneity bias. Moreover, Huang et al. [
189] conducted a meta-analysis, revealing that out of 437 empirical studies examining the link between corporate social and financial performance, only 54 applied specialized methods to tackle endogeneity issues. Encouragingly, 72% of these endogeneity-considered studies supported the prevailing consensus in their findings.
Drawing on insights from the literature above, we conduct a robustness check on our baseline results (see
Table 6) using instrumental variable (IV) methods designed to mitigate potential endogeneity bias in Equation (3). While such methods are commonly employed by accounting researchers in their primary analyses, their utilization for robustness testing purposes is approximately equivalent [
190]. Various IV methods exist to cope with endogeneity bias, including two-stage least squares (2SLS) and the system-generalized method of moments (system-GMM). Researchers must identify the root cause of the issue before selecting the most appropriate methods to effectively mitigate the adverse effects of endogeneity [
191,
192].
Generally, endogeneity occurs when one or more explanatory variables violate the exogeneity assumption by becoming correlated with the error term in a regression model [
193]. Two main sources create such a situation: omitted variable bias and simultaneity bias [
194]. Omitted variable bias occurs when relevant variables are not included in the regression model to estimate the variance of a particular dependent variable. Such omitted variables will be captured by the error term in the model and potentially related to any independent variables being the focal point of an analysis, thereby giving rise to endogeneity issues [
195]. Simultaneity bias arises when the independent and dependent variables mutually affect each other, operating reciprocally [
194]. Since the error term encapsulates all unobserved factors affecting the dependent variable, if there is simultaneity—where the dependent variable affects the independent variables—the error term can also correlate with the independent variables, exacerbating endogeneity concerns.
In the current study, we focus on two primary regressors:
and
, both potentially treated as endogenous. However,
measured as a period of the COVID-19 pandemic—a sudden global economic downturn—is totally considered an exogenous shock and thus unlikely to suffer from endogeneity issues stemming from both sources. Several studies exploring the relationship between the COVID-19 pandemic and managerial decisions share similar perspectives on this matter. e.g., [
75,
196].
Therefore, the endogeneity problem likely only applies to
, which directly pertains to corporate behavior. As presented in
Table 6, we incorporated pertinent financial indicators for BTD as a battery of control variables (i.e.,
,
,
,
,
, and
), called by Antonakis et al. [
188], as a best practice to avoid omitted variable bias. Nonetheless, we are overshadowed by a key point in Clarke [
197], p. 349: “It is impossible to include all the relevant variables in a regression equation.” Thus, certain omitted variables may affect the variance in
. For instance, firm value, a factor recognized in determining the extent of BTD [
198,
199], has not been incorporated into our analysis. At the same time, this variable has the potential to significantly influence corporate CSR activities. Higher firm value not only provides greater financial resources to firms but also heightens stakeholder expectations, thereby prompting firms to exhibit more pronounced commitments to CSR initiatives [
200]. Concerning the simultaneity bias, we stick to existing empirical results. On the one hand, CSR influences the magnitude of BTD through diverse causal mechanisms elucidated comprehensively in the previous sub-sections. On the other hand, companies exhibiting high BTD engage more in CSR to legitimize their actions and mitigate potential repercussions from their tax practices. Ling & Wahab [
28] provide support for the positive impact of BTD on CSR adoption.
Based on the identification outlined above, the endogeneity issue in this study appears to stem from both of those primary sources rather than just one. In such scenarios, employing 2SLS, which is recognized as the most used IV estimator, is highly recommended [
191]. Also, the statistical theory underpinning this estimator has been extensively developed [
201]. However, researchers often encounter significant challenges associated with the application of this estimator since it relies heavily on external source exogenous instruments, which in practice are not easy to find [
202,
203], sometimes even impossible [
188,
204].
As Angrist et al. [
205] and Bollen [
206] note, the instruments should meet at least two conditions: they must be highly correlated with the endogenous regressors (‘strong’), and they cannot be correlated with the error term (‘valid’). Maddala [
207], p. 154, provocatively discusses the ‘miracle’ of finding these instruments, cynically asking, “Where do you get such a variable?” Similarly, Larcker & Rusticus [
190] argue that in accounting research, strong and valid instruments are akin to a ‘holy grail’—highly desirable yet illusive. Proxies often prove imperfect, being correlated with one or more explanatory variables in the regression equation, while obtaining additional data suitable for panel data techniques remains challenging in management accounting settings [
184]. At the same time, if the chosen instruments fail to meet these stringent criteria, as cautioned by Bound et al. [
208] and Samadeni et al. [
209], the ‘cure’ may exacerbate the ‘disease,’ introducing bias into both estimates and standard errors. Consequently, researchers must grasp the implications of employing the 2SLS estimator, which does not precisely adhere to the essential assumptions regarding instrumental variables.
Given the circumstances, it is prudent to acknowledge the existence of alternative models that are logically plausible [
192]. In this case, Arellano and Bover [
210] and Blundell and Bond [
211] have introduced the system-GMM, which addresses various sources of endogeneity and is particularly suitable for panel data analysis [
202]. Unlike 2SLS, system-GMM does not rely on external exogenous instruments but instead employs two sets of equations, each with its own set of instruments. The first set of equations operates in levels with lagged differences between the dependent and independent variables serving as instruments. The second set involves equations in the first differences, using lagged levels of the dependent and independent variables as instruments. These instruments, derived from existing econometric models, are commonly referred to as ‘internal instruments’ [
212] and have demonstrated greater efficiency in mitigating endogeneity issues [
213]. Additionally, other advantages make a compelling case for adopting the system-GMM estimator over 2SLS. Firstly, system-GMM offers a straightforward framework for achieving asymptotically efficient estimators even in the presence of heteroscedasticity [
214]. Secondly, system-GMM tends to be more robust against weak instruments compared to 2SLS [
215], where the latter’s consistency and precision are contingent on strong instrument conditions [
216]. In short, the system-GMM estimator stands out as more efficient and consistent than 2SLS across a broad range of empirical conditions.
The characteristics of this study also align well with several assumptions that underpin the use of system-GMM. As noted by Roodman [
212], this estimation technique is tailored for dynamic panel data, implying that causal relationships among underlying phenomena evolve over time. For instance, it may not be the current year’s CSR influencing BTD but rather the previous year’s BTD that plays a significant role. This assumption proves pertinent, given that tax planning and earnings management strategies often exhibit cumulative effects. Successful managerial practices in one period can set a precedent for achieving similar outcomes in subsequent periods. To capture this dynamic, lagged values of the BTD measures are therefore used as explanatory variables. Moreover, the first set of system-GMM equations remains susceptible to omitted variable bias [
185], as does Equation (3). However, these omitted variables are assumed to be time-invariant, a plausible assumption if short panel data characterized by a short period (
) and a large number of companies (
) are used in the analysis. Fortuitously, our dataset aligns with this profile, featuring
and
.
Table 7 presents the estimation results related to the effects of the economic crisis and CSR on BTD using the system-GMM econometric model. It is worth noting that only the two-step estimator is reported due to several limitations associated with the one-step estimator. For instance, when a recent value of a variable is missing, employing a first-difference transformation (i.e., a variable’s recent value is adjusted by subtracting its previous value) may lead to significant data loss [
212]. To mitigate this issue, the two-step estimator adopts ‘forward orthogonal deviation.’ This method substitutes the subtraction of the previous value from the recent value of a variable with the subtraction of the mean of all available future values of a certain variable [
212]. Consequently, the two-step model offers more efficient and consistent coefficient estimates [
210].
Table 7 comprises nine specifications. In specifications 1, 4, and 7, industry FE is omitted from the regression model. For some reasons that we are unable to identify, the software we are using for statistical analysis (STATA 17.0) consistently rejects our attempt to include the predefined industry FE in the estimation model by giving an ‘invalid syntax’ message. As an alternative effort, we redefine industry FE as an indicator to mitigate differences in variables of interest across the ‘hard-to-tax’ and ‘easy-to-tax’ industries. Specifically, the industrial categorization from
Table 2 is narrowed down into such two divisions by drawing on empirical insights. For instance, Rajaraman [
218] identifies agriculture as the most challenging sector for taxation. According to Daniel et al. [
219], a substantial share of extractive industry revenues comes from multinational corporations, posing difficulties in taxing them effectively to ensure adequate revenue without impeding investment. Conceptually, Bird [
220] posits that service activities generally present difficulties for tax collection. However, Cevik et al. [
221] demonstrate that enhancing trade and telecommunications services improves tax collection efficiency, while healthcare—classified as non-market services—hinders tax efficiency. Bahl [
222] and Frijters et al. [
223] conclude that manufacturing is widely acknowledged as the easiest sector to tax compared to the other ones. Accordingly, firms operating in the agriculture, mining, and healthcare sectors are categorized as hard-to-tax, while the rest are deemed easy-to-tax. Thanks to this industrial division, STATA 17.0 unexpectedly enables us to incorporate industry FE into the estimation model, as evidenced in specifications 2, 5, and 8. Lastly, given our dataset’s nature, which involves multiple entities, we acknowledge the potential presence of heteroscedasticity. Therefore, we address this issue by employing the Windmeijer [
217] adjustment to correct for such a small sample bias, providing us with robust SEs for the two-step estimator. The corresponding regression results are presented in specifications 3, 6, and 9.
To ensure the appropriateness of the econometric model, two tests are reported: the Sargan test for the exogeneity of the instruments and the Arellano–Bond test for second-order autocorrelation (
) in residuals. The first-order autocorrelation (
) is less concerning as the equations are in first differences [
185,
192]. In all specifications reporting the results of the Sargan test, except specification 4, the
p-values of the test exceed the 10% significance level, indicating non-rejection of the null hypothesis. It suggests that the instruments employed in the system-GMM model are exogenously determined and then correctly specified. It is important to note that the Sargan test cannot be calculated when robust SEs are used in the model. Therefore, the results of the test are not reported in specifications 3, 6, and 9. Furthermore,
is found to be statistically insignificant in all specifications, implying no second-order correlation in the error terms between different periods. It supports the strong exogeneity assumption, indicating that lagged variables are uncorrelated with the error term.
Turning to the main results of the robustness testing, qualitatively, we find that our conclusions remain intact with the employment of the system-GMM as an estimation model. The economic crisis and CSR exert a positive influence on total BTD and permanent differences, though the impact on temporary differences is not statistically significant. This consistency is especially visible in specifications 2 and 8, where most of the estimated coefficients are significant at the 1% level. Although the magnitude of the estimated coefficients in specifications 3 and 9 is the same as in specifications 2 and 8, the use of Windmeijer’s [
217] correction inflates the SEs and thus reduces statistical significance to the 10% (for the economic crisis) and 5% (for CSR) levels. It indicates the presence of heteroscedasticity in the estimation model, which the use of robust SEs tolerates.
These findings, compared with the baseline findings, show a substantial reduction in statistical significance, considering the impact of the key explanatory variables on BTD, and the permanent differences in
Table 6 are significant even at the 1% level. However, the changes in the magnitude of coefficients are negligible, particularly for the crisis variable. The differences in coefficient size compared to the baseline results are 0.020 (from 0.054 to 0.034) and 0.022 (from 0.052 to 0.030) for the effects on total BTD and permanent differences, respectively. The magnitude of the coefficient representing the influence of CSR on total BTD fell very slightly, around 0.003 (from 0.190 to 0.187), but for the CSR-permanent differences relationship, the coefficient sizes are totally similar. Lastly, the moderating effect of CSR on the relationship between the crisis and the three BTD measures was consistently found to be insignificant across all specifications of the current robustness analysis. Despite some changes, which do not undermine the substance of our empirical inferences, we confidently assert that our robustness test has been successfully validated.
5.4. Robustness Check: Exploiting Industry Heterogeneity
Empirical analyses in finance often encounter the challenge of dealing with potentially heterogeneous data whose structure is not fully understood. Despite this complexity, estimators for key parameters remain pertinent and valuable even in the presence of such heterogeneity. However, accurately estimating their sampling variability becomes notably more difficult. To address these challenges, we can employ robust large-sample inference strategies. One approach, outlined by Ibragimov & Müller [
224], involves partitioning the dataset into several distinct groups. For each group, the model of interest is estimated. This method allows for robust inference even in datasets with heterogeneous structures. In cases where panel data is relatively short, such as in the current study with
and
, we may find it advantageous to assume some level of independence across the cross-section. For instance, in finance applications, it is common to assume minimal correlation between firms belonging to different industries, e.g., [
225]. We can leverage this assumption by grouping firms within the same industry together. This grouping strategy results in multiple distinct groups corresponding to different industries.
Building on the insights above, the sample for this study was divided into two distinct groups based on industry characteristics. The first group comprises companies operating within sectors most severely affected by the economic crisis, i.e., the COVID-19 pandemic, while the second group consists of companies operating in the least affected sectors, as indicated in
Table 2. Extensive empirical research has demonstrated the significant impact of the COVID-19 pandemic on firm performance, yet this impact varies considerably across different business sectors, e.g., [
226,
227]. Variations in market share and access to capital among industries play pivotal roles in shaping these findings [
228]. Larger industries tend to exhibit greater resilience during crises due to their multiple sources of funding that support operational continuity, both internally and externally. Consequently, exploring industry dimensions is crucial for understanding the heterogeneity of the impacts of economic crisis and CSR on BTD levels within corporate financial reports across industries.
To conduct this analysis, we re-ran our regression models separately for each industry group. The results are presented in
Table 8, which is divided into two panels (Panels A and B) comprising a total of 12 specifications. Panel A displays the regression outcomes for the industry group least impacted by the pandemic, while Panel B shows the results for the industry group most affected by it. Each panel includes six specifications. Three of these specifications utilize traditional panel data regression models (FE or RE models), while the remaining three employ system-GMM regression models. We opted for the system-GMM model to maintain consistency in addressing endogeneity bias throughout our analysis. Broadly speaking, estimation results involving data from the most affected industries tend to be comparable to our baseline findings. This conclusion is especially relevant for the direct influence of CSR and the moderating effect of CSR on BTD measures. The subsequent paragraphs provide a more detailed interpretation of these results.
The findings from
Table 8 indicate that the economic crisis led to an increase in both the size of aggregate BTD and permanent differences among companies in industries least impacted (Panel A) and most impacted (Panel B). These results align with our overall research conclusions regarding the relationship between the crisis and BTD. However, it is crucial to note the disparity in the coefficients’ magnitude, where the positive effects of the economic crisis on these BTD measures appear more pronounced in severely affected industries. These findings contrast with Kobbi-Fakhfakh & Bougacha [
21], who reported that the COVID-19 pandemic affected corporate tax avoidance uniformly across industries, regardless of potential industrial failure for BTD. Nevertheless, our results are consistent with those of Shen et al. [
229], who, using financial data from listed Chinese companies, found that the negative impact of COVID-19 on company performance was more pronounced in industries severely affected by the pandemic. Our findings thus underscore the importance of considering industry-specific conditions when analyzing the relationship between crises and corporate financial behaviors. Studies that could delve deeper into the underlying mechanisms driving these disparities are certainly needed to provide more nuanced guidance for policymakers and corporate strategists navigating uncertain economic environments.
As shown in
Table 8, our findings related to the positive impact of CSR on two BTD indicators, namely total BTD and permanent differences, remain intact when using a subsample of listed companies from the most affected industries. In contrast, when our investigation was carried out by employing data from the least affected industries, the opposite results were observed: CSR was found to have a negative effect on both total BTD and permanent differences. These negative effects were predominantly observed in the two-step system-GMM regression models, reaching a significance level of 5%. The contrasting results observed regarding the impact of CSR on BTD indicators in the least affected industries can be partly explained by differences in corporate culture and strategic responses to economic crises. Notably, in industries experiencing a less severe impact from the economic crisis, companies may perceive fewer immediate pressures to engage in CSR activities for crisis management purposes. Consequently, their CSR efforts might not be as strategically integrated with tax planning and earnings management strategies. These findings are understandable, given that the least affected industries typically have greater market share and access to capital [
228]. Furthermore, these industries naturally exhibit a negative correlation with temporary differences, as demonstrated in
Table 5.
Lastly, when the analysis focuses on the most affected industries, we observe that the moderating effect of CSR on the relationship between the crisis and BTD measures continues to show a negative sign, consistent with the baseline findings. However, these effects are significant only in estimation models that do not address endogeneity bias—namely, FE and RE—and are not significant in the two-step system-GMM estimation model, which corrects this bias. Thus, we choose to believe that CSR does not moderate the relationship between the economic crisis and BTD among the most affected industries, which is in line with the baseline results.
Conversely, the moderating effect of CSR turns positive when analyzing data from the least affected industries. These estimation results are significant at the 1% level for all BTD measures after addressing endogeneity bias using the two-step system-GMM model. This finding is unexpected given our earlier observation of the direct negative impact of CSR on BTD, highlighting the reluctance of industries least affected by the pandemic to engage in aggressive tax planning and earnings management strategies associated with CSR activities. At this point, normative motives and corporate culture theory justify these results. On the other hand, the significant positive moderating effect of CSR suggests that in the least affected industries, where financial stability might mitigate immediate crisis pressures, CSR initiatives could act as a strategic buffer. Companies may strategically leverage CSR to enhance public perception, build resilience, or differentiate themselves amid economic uncertainty, thus potentially influencing BTD positively in crisis contexts. These results are in line with strategic motives and risk management theory. Considering the duality, we suggest that among industries less affected by the pandemic, firms’ engagement in CSR generally reduces the size of BTD. However, when linked to the crisis context, CSR stimulates an increase in BTD among less affected industries, which indirectly provides support to our general conclusion on the dominant role of strategic motives in driving CSR practices by Indonesian public companies during times of crisis.