Does ESG protect profitability in a crisis? Evidence from US industrial firms during COVID-19

1 Introduction

COVID-19 triggered the world’s largest simultaneous recession since the Great global gross domestic product fell by 3.1% in 2020 (International Monetary Fund (IMF), 2021), and stock returns fell by 34% within a matter of weeks (MSCI, 2020). The COVID-19 crisis was global, but its effects were not. Some firms suffered dire market and financial consequences, while others did not. This has stirred keen interest in identifying firm-level factors that may explain resilience. The most prominent, and most controversial, candidate is a firm’s environmental, social, and governance (ESG) performance. Two clarifications frame the present design. First, the pandemic is treated here as a common macroeconomic shock to which firms were differentially exposed, rather than as a conventional treated-versus-untreated experiment: all firms were affected, and pre-crisis ESG standing is a pre-existing firm characteristic rather than an assigned treatment. The difference-in-differences (DiD) interaction therefore identifies whether firms with different pre-crisis ESG profiles experienced different changes in return on assets (ROA); it does not by itself establish that ESG caused those differences. Second, the relationship between ESG and resilience remains empirically contested, and the study is designed to adjudicate between two competing directional predictions.

The first perspective, labeled the insurance-value hypothesis (H1a), holds that ESG engagement accumulates relational capital and stakeholder trust that limit opportunistic defection during crises, thereby protecting operational performance. Lins et al. (2017) provide foundational evidence: firms with greater social capital outperformed peers by 4–7 percentage points during the 2008–2009 financial crisis. The same logic has been applied to COVID-19: Albuquerque et al. (2020) find that firms with higher environmental and social scores experienced lower equity volatility, and Broadstock et al. (2021) find that ESG performance reduced financial risk in Chinese firms. These results are sometimes used to argue that ESG functioned as “insurance” during the pandemic’s first wave. The second perspective, labeled the short-run cost hypothesis (H1b), holds that the costs of ESG engagement, including compliance expenses, resource commitments, and governance structures, cannot be avoided in the short run and squeeze the margins of high-ESG firms disproportionately during a demand shock (Fatemi et al., 2018). Consistent with this view, pre-crisis ESG engagement was not found to shield US firm performance during COVID-19 by Demers et al. (2021) or Bae et al. (2021). The proposed causal chain under H1b runs from ESG orientation to a higher share of contractually and reputationally rigid expenditures—compliance programs, workforce commitments, and environmental and governance infrastructure—which increases operating leverage and compresses margins when revenue contracts. Because these intermediate quantities (ESG-related expenditures, cost rigidity, and financing of sustainability programs) are not observed in this study, the analysis identifies the resulting profitability pattern; it cannot, and does not claim to, demonstrate the mechanism itself. Which mechanism dominates in the short run, particularly for accounting-based profitability, remains unresolved.

The existing literature is saturated with market-based evidence: the resilience premium documented for high-ESG firms is overwhelmingly a premium in stock returns and return volatility (Albuquerque et al., 2020; Broadstock et al., 2021; Ding et al., 2021), outcomes driven substantially by investor expectations and sentiment that do not directly reveal whether ESG enables a firm to remain profitable. This produces an accounting-versus-market-valuation decoupling that remains theoretically unresolved: equity markets can reward ESG as a signal of quality and long-run value, even as the operating commitments underlying ESG compress accounting margins. The critical unresolved question, therefore, is how the structural rigidity of sustainability commitments (compliance programs, workforce guarantees, and environmental and governance infrastructure that cannot be scaled down at the speed at which demand collapses) affects core accounting fundamentals during a severe demand shock. ROA, a direct measure of how efficiently firms use their resource base to generate accounting profits, is the appropriate outcome variable for this question, and evidence on ESG’s association with ROA during COVID-19 remains limited and mixed (Bae et al., 2021). Addressing this decoupling is the manuscript’s central contribution; it is developed theoretically in Section 2.1 as a sustainability trade-off mechanism and evaluated with a within-firm difference-in-differences design.

Another issue is that ESG is a composite of environmental, social, and governance practices that are not conceptually equivalent. Each pillar may affect firm performance through distinct channels. Environmental practices can improve efficiency and reduce regulatory risks; social practices can improve employee productivity and stakeholder relationships; and governance practices can enhance decision-making, risk mitigation, and capital allocation (Ahmad et al., 2023). Aggregating ESG may, therefore, mask underlying pillar-level differences. We disaggregate ESG to evaluate whether any specific pillar drives the relationship with profitability. We also assess ESG effects against the financial factors that Ding et al. (2021) identify as central to crisis-period resilience, leverage, and liquidity to establish whether ESG explains incremental variation in ROA beyond these structural determinants. Recent global evidence that the financial value of ESG is heterogeneous across pillars, industries, and regions (Soltanisehat and Sibdari, 2026) reinforces the case for disaggregation.

In this context, this study adopts both difference-in-differences and panel data methods, which allow us to account for year-level shocks and time-invariant firm characteristics, respectively, while examining whether the change in ROA differed systematically between high-ESG and low-ESG firms during the crisis period. The study makes three contributions to the ESG–performance literature, ordered by theoretical novelty. First, and most distinctively relative to Bae et al. (2021) and Demers et al. (2021), which examine composite corporate social responsibility (CSR) or ESG scores against market outcomes, it disaggregates ESG into its environmental, social, and governance pillars within a single unified crisis specification, providing a direct horse-race between the three ESG channels. This pillar-level design is the study’s primary theoretical contribution because the sustainability trade-off mechanism developed in Section 2.1 makes different predictions for capacity-like environmental and social commitments than for governance. Second, it places financial leverage alongside ESG as a competing resilience mechanism in the same model, building on Ding et al. (2021) and theorizing the tension between debt-financed sustainability programs and balance-sheet conservatism (Section 5.3); this joint treatment, rather than a control-variable role for leverage, further distinguishes the study from the existing ESG–COVID literature. Third, it evaluates the insurance-value (H1a) and short-run cost (H1b) hypotheses against ROA rather than equity returns, distinguishing the operational channel from the investor-sentiment channel, and it does so with an identification strategy that includes a dynamic event-study specification, a triple-difference (DDD) design exploiting industry pandemic exposure, industry-by-year fixed effects, and controls for research-and-development (R&D) and advertising intensity. The findings also carry direct implications for environmental investment and sustainability policy, developed in Section 5: a crisis-window accounting penalty for high-ESG firms speaks to the timing profile of returns to environmental and social investment, rather than its long-run value, and should inform how regulators, investors, and managers interpret ESG ratings under stress.

2 Literature review

2.1 Theoretical foundations: stakeholder theory, the resource-based view, and the sustainability trade-off mechanism

The theoretical basis of the study is derived from two complementary frameworks: stakeholder theory and the resource-based view (RBV) of the firm. Stakeholder theory, in its classic formulation by Freeman (1984), posits that firms that actively manage their responsibilities to a wide array of stakeholders, including employees, suppliers, creditors, communities, and the natural environment, acquire relational capital and trust that insulate organizational performance when external shocks threaten viability. The hypothesis is that ESG investment accumulates this social capital in calm times, and the capital is drawn down during crises by limiting opportunistic defection of key stakeholder groups (Lins et al., 2017). This is the basic logic of the ESG-resilience hypothesis examined in this study.

The resource-based view conceptualizes ESG performance as a source of sustainable competitive advantage (Barney, 1991). Valuable, rare, inimitable, and non-substitutable resources, such as distinctive environmental-management capabilities, human-capital quality, and governance systems, are developed over time. During crises, these resources may take the form of operational flexibility, greater stakeholder loyalty, and better risk management, which collectively should support ROA resilience (Wernerfelt, 2016).

The advantages of ESG-based resources may take a long time to accrue; stakeholder theory and the RBV both therefore imply that the associated outlays (compliance expenditures, governance infrastructure, and environmental investment) are short-run costs of long-run capability building. Fatemi et al. (2018) formalize the trade-off, suggesting that ESG practices create value via lower financial risk and improved reputation over multi-year horizons but carry transitional costs that depress near-term accounting profitability. This temporal trade-off is a key motivation for focusing on ROA during a crisis window (2020–2021) rather than long-run market valuation. It is empirically corroborated by Alessi et al. (2021), who show that composite ESG scores lacked substantial explanatory power for stock returns during the COVID-19 pandemic once financial characteristics were controlled, arguing that aggregate ESG scores may be too blunt an instrument to capture crisis-period benefits through accounting-based channels. Demers et al. (2021) similarly find that ESG did not immunize US stocks during the COVID-19 crisis, while investments in intangible assets did, underscoring that the source of resilience matters as much as its presence. The theoretical expectation for this study is thus qualified: although ESG potentially confers long-run strategic value, accounting resilience effects during an acute demand-side shock may be attenuated or even reversed by the cost commitments that ESG engagement embodies.

Building on these two frameworks, this study advances a sustainability trade-off mechanism as its organizing theoretical proposition. Environmental and social capital is best understood as a fixed capacity investment, rather than as variable operating inputs. Emissions-abatement equipment, energy-efficiency retrofits and environmental management systems, workforce safety, welfare, training and retention commitments, and supplier and community engagement structures are largely irreversible, contractually or reputationally locked in, and dimensioned to normal-period activity. In calm periods, these investments build relational value: stakeholder trust (stakeholder theory) and inimitable capabilities (RBV) that are expected to pay off over multi-year horizons. In the short run, however, they increase operating leverage: a larger share of the cost base is fixed relative to revenue, so that a given percentage decrease in sales produces a larger percentage decrease in operating profit. When an exogenous shock removes demand faster than these commitments can be scaled down, high-ESG firms face a mechanical compression of accounting margins and a constraint on financial flexibility that low-ESG firms, with more variable cost structures, do not experience to the same degree. The mechanism yields three observable implications. First, the differential should appear in accounting profitability (ROA), where operating leverage bites, but need not appear in market valuation, where investors may continue to price the long-run relational value; this accounting–market decoupling motivates the study. Second, the differential should be concentrated in the environmental and social pillars, whose commitments are most capacity-like, rather than in governance, whose direct costs are comparatively limited and whose benefits operate through decision quality and capital-structure discipline. Third, the differential should be larger where the demand collapse was deepest and where fixed financing charges compound fixed operating commitments, that is, in the most pandemic-exposed industries and among more highly leveraged firms. The empirical design evaluates the first implication directly (Models 1–2 and the event-study specification), the second through pillar disaggregation (Model 3), and the third through the triple-difference and leverage analyses (Sections 4.7 and 5.3).

2.2 ESG performance and corporate financial performance

Over the past 2 decades, the link between corporate financial performance and ESG performance has been a topic of intense empirical scrutiny, and yet, consensus remains elusive. Eccles et al. (2014) indicate that companies that voluntarily adopt sustainability policies achieve higher long-run stock market and accounting performance relative to matched companies, and Servaes and Tamayo (2013) find that stakeholder visibility strengthens the returns to ESG investment. On the materiality question, Khan et al. (2016) show that firms investing in sustainability issues that are financially material to their industry significantly outperform those that do not, with no such evidence for immaterial issues. Meta-analytically, Margolis et al. (2009) combine the results of more than 200 empirical studies and conclude that there is a small but significant positive correlation between corporate social and financial performance across samples, contexts, and periods. Although meta-analytical syntheses consistently find a net-positive association (Xie et al., 2019), Friede et al. (2015) aggregate evidence from more than 2,000 empirical studies to conclude that roughly 90% report a non-negative ESG–CFP relationship, yet substantial heterogeneity persists across measurement instruments, time horizons, and institutional contexts. A fundamental methodological problem plagues much of this literature: ESG ratings differ substantially across providers. Berg et al. (2022) document inter-rater correlations as low as 0.38 across six major rating agencies, a phenomenon they term aggregate confusion, attributable to differences in scope, measurement choices, and weighting schemes. This divergence is not merely technical; it means that composite ESG scores carry substantial construct-validity uncertainty and that cross-study comparisons of ESG effects on profitability frequently compare incommensurable constructs.

This heterogeneity is not random, and a critical reading of the evidence suggests four systematic sources of divergence. First, the outcome measure matters: market-based outcomes (returns and volatility) embed investor expectations and sentiment and respond to ESG partly through valuation channels, whereas accounting outcomes such as ROA reflect realized operations; positive market-side findings therefore coexist with null or negative accounting-side findings for the same period. Second, the horizon matters: sustainability investments are theorized to pay off over multi-year windows, so short crisis windows mechanically load on the cost side of the trade-off (Fatemi et al., 2018). Third, the measurement instrument matters: provider divergence (Berg et al., 2022) and the mixing of financially material and immaterial components within composites (Khan et al., 2016) dilute and destabilize estimated effects; relatedly, El Khoury et al. (2023) document concavity–convexity patterns in the ESG–performance relationship, implying that linear specifications estimated over different score ranges can yield different signs; Soltanisehat and Alizadeh (2026) reach the same conclusion for global supply-chain firms, showing that the functional form of the ESG–CFP relationship changes across ESG levels. Fourth, context matters: industry composition, country institutions, and crisis type condition both the exposure of firms to shocks and the value of stakeholder capital; in a 50,000 firm-year global sample, Soltanisehat and Sibdari (2026) find that the financial return to ESG and its individual pillars varies substantially across regions and institutional settings. The present study is positioned within this landscape as an accounting-outcome, short-horizon, single-provider, and single-country design, and its results should be read with those boundary conditions in view; they also speak to the ESG materiality debate since a composite-score penalty during a crisis is consistent with composites blending material and immaterial commitments.

Despite this, Refinitiv-sourced ESG scores are employed in the present study for three reasons that mitigate the aggregate-confusion problem. First, Refinitiv provides explicit pillar-level decomposition: environmental, social, and governance sub-scores are published separately alongside the composite, enabling the disaggregated analysis conducted in Model 3. Second, Refinitiv’s methodology is among the most transparent available, with a publicly documented weighting scheme (Refinitiv, 2022), reducing the construct-validity ambiguity that afflicts providers with opaque aggregation rules. Third, Refinitiv offers consistent longitudinal coverage for US industrial firms across the 2012–2021 window required by the DiD design. Because rating providers disagree substantially (Berg et al., 2022), the results reported here are conditional on the Refinitiv measurement system and may differ under MSCI, Bloomberg, Sustainalytics, or other providers; replicating the design with an alternative provider is an important robustness extension that current coverage constraints place beyond this sample, and this single-rater constraint is treated as a first-order limitation (Section 5.4) rather than a footnote. Using data from S&P 500 firms, Alareeni and Hamdan (2020) and Kim and Li (2021) find positive associations between composite ESG scores and key performance metrics including return on equity but note that governance was the only dimension producing consistent effects across all model specifications. Chi et al. (2024) provide cross-national panel evidence revealing ROA improvement and lower earnings volatility following positive ESG performance, with rigidly regulated sectors exhibiting the most robust effects from environmental sub-scores. These findings establish a prior expectation that high-ESG firms should exhibit superior accounting profitability. Critically, however, this expectation is conditioned on the economic environment: whether ESG confers a performance premium specifically during crisis periods, when liquidity constraints, demand destruction, and investor risk aversion are acute, demands separate analytical scrutiny.

Because theory admits opposing directional predictions for the crisis period, the first hypothesis is stated as a formal pair of competing hypotheses rather than a single directional claim:

H1a (insurance value): Firms with higher pre-crisis composite ESG scores exhibit a more favorable change in ROA during the COVID-19 crisis than firms with lower pre-crisis ESG scores.

H1b (short-run cost): Firms with higher pre-crisis composite ESG scores exhibit a less favorable change in ROA during the COVID-19 crisis than firms with lower pre-crisis ESG scores.

2.3 ESG as a corporate resilience mechanism during COVID-19

Since 2020, considerable empirical research has examined the idea that ESG performance acts as a resilience buffer during systemic shocks, rather than simply as a driver of performance during stable periods. The seminal paper is Lins et al. (2017), which finds that US firms with greater CSR engagement outperformed their peers by 4–7 percentage points during the 2008–2009 financial crisis. Their theoretical explanation focuses on stakeholder trust: accumulated social capital lowers the likelihood of opportunistic defection by employees, suppliers, and creditors at exactly the time such defection would be most costly. This insurance-value mechanism is theoretically compelling, but its prediction is highly contingent on whether stakeholder relationships translate into operational performance and not just reputational shields.

Broadstock et al. (2021) extend this methodology to COVID-19 using CSI 300 data and observe that high-ESG portfolios outperformed during the February–March 2020 crash, driven primarily by the environmental component. Albuquerque et al. (2020), exploiting the exogenous shock of the COVID-19 crash in the US, show that firms with higher environmental and social scores experienced lower equity volatility and higher abnormal returns, with their instrumental-variable analysis providing more credible causal identification than most observational studies on this topic. Ding et al. (2021) take this logic further, conducting a cross-country analysis across 56 countries and more than 6,000 firms and showing that pre-crisis ESG engagement is a strong positive predictor of relative equity market performance during the pandemic, a result that holds independently of firm size, leverage, and cash holdings. As documented by Pastor and Vorsatz (2020), ESG-screened mutual funds had significantly lower outflows and superior relative performance during the COVID-19 market turbulence, consistent with an investor flight-to-quality. Complementarily, Garel and Petit-Romec (2021) provide firm-level evidence that investors rewarded environmental responsibility during the COVID-19 crisis, with the benefit concentrated among firms stronger on employee relations and customer orientation. Similarly, Cheema-Fox et al. (2021) report that better ESG performance was associated with greater operational resilience in firms’ corporate response strategies. For ESG-rated exchange-traded funds (ETFs), Folger-Laronde et al. (2022) find that the net-asset-value drawdown of more highly rated ETFs was lower during the initial stage of the pandemic; Cardillo et al. (2023) document in a multi-country sample that ESG performance reduced firm-level financial distress stemming from the pandemic. Takahashi and Yamada (2021) find that ESG engagement had a cushioning effect during the first period of the market collapse that faded in the second period, implying that ESG resilience is time-varying; this motivates the explicit two-period DiD design adopted here. Cross-country evidence reinforces but also conditions this picture: El Khoury et al. (2022) report that responsible G20 companies performed comparatively well during the outbreak; Al Amosh and Khatib (2023a) and Al Amosh and Khatib (2023b) document positive associations between ESG and financial performance across G20 and broader cross-country samples during the pandemic; Aksoy and Yılmaz (2025) show that sustainability performance moderated, rather than uniformly improved, the crisis-period financial performance of emerging-market firms; and Binesh et al. (2025) find that the ESG–financial-distress relationship during COVID-19 was conditioned by innovation and capital intensity. Together, these studies indicate that crisis-period ESG effects are heterogeneous and moderated by firm and country characteristics, rather than uniform. The most recent evidence extends the picture: Song et al. (2026) find that ESG lowers financial-distress risk during economic downturns, with the effect strengthening after 2020 and concentrated among firms with stronger board monitoring, and Yang and Wang (2026) show that ESG performance improves supply-chain resilience through investment efficiency. Both identify risk-side and operational-continuity benefits that are, notably, distinct from short-run accounting margins, which is precisely the decoupling interrogated in this study.

However, the relevant question for this study is whether these market-based effects, reflected in equity returns and volatility, carry over to accounting-based profitability. The ESG stock-market premium during crises may be a product of investor sentiment or short-term signaling effects that are dissociated from operational fundamentals. Using a difference-in-differences identification strategy, the current analysis, which examines ROA as the outcome variable, directly asks whether ESG constitutes an actual operational advantage or just a valuation premium. Bae et al. (2021) go further: they find little evidence of pre-crisis CSR serving as a buffer for shareholder value during COVID-19 in a US sample and attribute their null findings to decoupling between reported scores and on-the-ground arrangements. This tension is what drives the disaggregated, pillar-level analysis undertaken in the following sections.

A final consideration shapes both the design and the interpretation of any negative high-ESG × COVID interaction. High-ESG firms differ systematically from low-ESG firms in ways other than ESG orientation: they tend to be larger (in this sample, the correlation between firm size and pre-crisis ESG is 0.69), and they may differ in industry composition, capital intensity, international exposure, workforce composition, supply-chain dependence, and the feasibility of remote operation (Ding et al., 2021; Binesh et al., 2025). Any of these characteristics could generate differential pandemic exposure that loads onto the ESG indicator. The firm fixed effects absorb all time-invariant differences, and firm size is controlled throughout, but industry-specific pandemic shocks and other time-varying confounds cannot be fully excluded. The interaction estimates reported below are therefore interpreted as differential-exposure associations, consistent with the cost hypothesis, and the competing explanations are evaluated as far as the data allow through the triple-difference and industry-by-year specifications in Section 4.7 and are addressed directly in Sections 5.1 and 5.4.

2.4 Environmental pillar: emissions management and operational efficiency

The environmental dimension of ESG is theorized to affect financial performance through two primary channels: operational cost reduction from energy efficiency and resource optimization, and risk reduction through lower exposure to carbon regulation and environmental litigation (Li and Xu, 2024). Firms with lower carbon intensity have already internalized the cost discipline required to minimize energy waste, translating directly into margin advantages that become proportionally more valuable when revenues contract during downturns. Chi et al. (2024) confirm that environmental sub-scores exhibit the most statistically reliable associations with ROA in cross-national panels, consistent with the efficiency-channel interpretation.

Broadstock et al. (2021) find that environmental performance is the strongest single predictor of cumulative abnormal returns during the early COVID-19 crisis, interpreting this as investor recognition that carbon efficiency signals organizational lean management and long-run risk preparedness. However, this evidence is market-side; whether environmental performance translates into superior accounting profitability measured through ROA during a demand-side shock such as COVID-19 is a more demanding empirical test. A pandemic-induced revenue collapse may overwhelm the cost-efficiency advantages of environmental performance, particularly in capital-intensive sectors where environmental investments represent sunk costs rather than variable savings. The present study examines this directly through the environmental pillar’s interaction with the COVID indicator in a two-way fixed-effects (TWFE) model.

H2 Firms with higher pre-crisis environmental scores will exhibit significantly greater ROA during the COVID-19 crisis period, reflecting the operational efficiency advantages of environmental performance.

2.5 Social pillar: human capital and organizational resilience

The social dimension of ESG encompasses workforce management practices, employee safety, diversity, and labor relations. The theoretical case for a social-resilience linkage during crises rests on the insurance-value hypothesis (Lins et al., 2017): firms that maintain strong employee relations, low turnover, and safe working conditions accumulate workforce loyalty that reduces absenteeism, productivity losses, and costly labor disruptions precisely when operational continuity is most critical.

Ding et al. (2021) find that stronger pre-crisis labor practices predict more modest revenue contractions during COVID-19 globally, while Engelhardt et al. (2021) document that ESG-rated firms outperformed non-rated peers during the pandemic downturn, with the social component explaining a disproportionate share of the return differential in labor-intensive industries. For ROA specifically, the transmission mechanism operates through the asset-utilization channel: firms with lower employee turnover and fewer labor disruptions deploy their asset base more productively during crisis periods, sustaining revenue generation relative to the asset denominator. Conversely, firms with poor labor relations face heightened absenteeism, recruitment costs, and potential strike activity precisely when financial slack is most constrained. The heterogeneity of these social effects across sectors, however, complicates generalization, and null findings are not uncommon when social scores are assessed against accounting rather than market outcomes (Bae et al., 2021).

H3 Firms with higher pre-crisis social scores will exhibit significantly greater ROA during the COVID-19 crisis, reflecting workforce stability advantages.

2.6 Governance pillar: board quality, accountability, and financial risk

Corporate governance is theoretically the most direct ESG mechanism for financial outcomes because governance structures shape the quality of strategic decision-making, capital-allocation discipline, and disclosure integrity, all of which bear directly on accounting profitability. Ahmad et al. (2023) identify governance as the single most important ESG pillar in moderating the adverse effects of financial crises on firm profitability in a UK context, arguing that well-governed firms exercise greater financial discipline during downturns and are less prone to the earnings manipulation and asset misallocation that generate accounting losses. Manita et al. (2018) document a positive relationship between board gender diversity and ESG disclosure quality, while Kara et al. (2022) show that gender-diverse boards maintained stronger risk-management practices during the pandemic in the banking sector. Related governance evidence points in the same direction: Ozdemir et al. (2022) show that board diversity shapes firm risk-taking conditional on board independence, CEO duality, and free cash flows, and Yılmaz et al. (2023) link board gender and cultural diversity to stronger corporate governance and social performance in emerging markets.

The mechanism linking governance to crisis-period ROA operates primarily through leverage discipline and risk-management quality. Well-governed firms enter crises with more conservative capital structures because boards exercise stronger oversight of management’s debt appetite. The controls in the present study are consistent with this: financial leverage is the strongest and most consistently significant predictor of ROA in the specifications in which it is included (β = −0.052, p < 0.001 in Model 1). Whether the governance ESG score itself, distinct from leverage, which is controlled separately, independently predicts ROA during COVID-19 is the specific question addressed by the pillar-level DiD interaction.

H4 Firms with higher pre-crisis governance scores will exhibit significantly greater ROA during the COVID-19 crisis, reflecting superior risk-management discipline and capital-structure conservatism.

The resulting conceptual framework can be summarized verbally as follows. Pre-crisis ESG orientation operates on crisis-period ROA through two competing channels: a stakeholder-trust channel that supports revenue and operational continuity (H1a) and a cost-rigidity channel that increases the share of inflexible expenditures and compresses margins under a demand shock (H1b). The three pillars (H2–H4) represent channel-specific variants of the same opposition, and capital structure (financial leverage) enters as a parallel, non-ESG resilience mechanism against which the ESG channels are benchmarked. All hypotheses are numbered consistently on this framework: H1a/H1b for the composite, H2 (environmental), H3 (social), and H4 (governance), and each is developed from the literature in Sections 2.2–2.6.

3 Data and methodology

3.1 Sample and data sources

The empirical study is conducted on a panel of 70 publicly traded US industrial companies observed across 2012–2021, equaling 690 firm-year observations. The initial sample consisted of all US companies with a non-missing composite ESG score and ROA in the Refinitiv Eikon database during the 2012–2021 period, limited to Standard Industrial Classification (SIC) codes 2000–3999 (manufacturing) and 4000–4999 (transportation and related services). Companies in the financial sector were excluded because their leverage ratios and asset mix are not structurally comparable with those of non-financial companies. Firms were excluded if their ESG and financial data were missing for the previous 5 years as this would make the pre-crisis average unreliable. Following these filters, the final estimation sample for Models 1–2 consists of 70 firms and 690 firm-year observations; the panel is unbalanced, with 6–10 observations per firm. Model 3 is estimated on the sub-panel of firms with available pillar-level (E, S, and G) pre-crisis sub-scores (67 firms; 661 firm-year observations). Table 1 reports the complete verified sample cascade, from the raw export through the merged analysis file to each model’s estimation sample, together with the high/low-ESG split and the panel imbalance; because the surviving firms must have long, continuous ESG histories, the sample is tilted toward larger, disclosure-intensive firms, and the resulting selection and survivorship risks are acknowledged explicitly in Section 5.4.

TABLE 1. Sample construction.

Screening criteria applied within the data platforms prior to export are as follows: (i) US-listed firms in SIC codes 2000–3999 and 4000–4999, excluding the financial sector; (ii) non-missing composite ESG score and ROA over 2012–2021; (iii) at least five consecutive prior years of ESG and financial data. All counts below the first row are verified by direct re-estimation from the final dataset. The estimation panel is unbalanced (690 of a possible 700 firm-years for the 70 firms, with 6–10 observations per firm); per-year observation counts are available from the authors on request. The long-history requirement tilts the sample toward larger, disclosure-intensive firms; the associated selection and survivorship risks are discussed in Section 5.4.

The Refinitiv Eikon database provided ESG data, comprising the composite ESG score and its environmental (E), social (S), and governance (G) pillar components, standardized on a 0–100 scale (Refinitiv, 2022). Financial variables, namely, ROA, financial leverage, quick ratio, total assets, and asset growth, were obtained from Bloomberg. The study window is split into a pre-crisis period (2012–2019) and a crisis period (2020–2021) covering the COVID-19 pandemic. Treating 2020–2021 as a single crisis window is justified on macroeconomic grounds: both years fall within the same sustained global disruption, with the output collapse in 2020 followed by an incomplete and uneven recovery in 2021, which remained dominated by pandemic conditions (IMF, 2021). The pooling is, nonetheless, a simplifying assumption: some industries began recovering in late 2021, while others remained depressed, so the estimated interaction is an average over heterogeneous within-window dynamics. Estimating the effects for 2020 and 2021 separately is a natural refinement of the design, and this heterogeneity is acknowledged in the limitations. This structure allows a DiD design in which the pandemic serves as a common, externally originating macroeconomic shock (Ding et al., 2021).

3.2 Variable construction

The dependent variable is firm performance measured as ROA. The control variables are firm size (natural logarithm of total assets), financial leverage, quick ratio (liquidity), and asset growth. The control set follows the crisis-resilience determinants identified by Ding et al. (2021), namely, capital structure, liquidity, size, and growth, which are the structural financial characteristics most consistently associated with pandemic-period performance. R&D and advertising intensity are plausible additional confounders; R&D intensity, in particular, co-moves with ESG practices (E-Vahdati and Binesh, 2022), and pre-crisis profitability trends are addressed through the event-study test in Section 4.4. Both are, therefore, added as controls in robustness specification R6 (Supplementary Table SB1), each measured as annual expense scaled by total assets; the ESG × COVID estimate remains unchanged (Section 4.7). Continuous financial variables are winsorized at the 1st and 99th percentiles, meaning that, for each variable, values below the 1st percentile or above the 99th percentile of the full panel distribution are replaced with the respective percentile value. This treatment is motivated by the extreme right skew of financial leverage (skewness = 6.47) and asset growth (5.73); it limits the influence of a small number of extreme firm-years on the estimates without deleting observations.

ESG measures are built as pre-crisis averages (2012–2019) to mitigate endogeneity and guarantee temporal predeterminedness. Specifically, ESG_pre is the equal-weighted arithmetic mean of the firm’s available annual composite scores across 2012–2019; no year receives extra weight. Averaging over the full pre-crisis window rather than using a single-year value is motivated by measurement reliability: ESG scores from a single provider exhibit year-to-year noise driven by changes in the firm’s disclosure practices and minor methodology updates by the rater. Averaging over 8 years reduces this measurement error and captures a firm’s structural ESG orientation rather than a snapshot influenced by transitory reporting decisions. Using the 2019 value alone or a shorter (for example, 2017–2019) pre-pandemic average as the assignment variable represents a natural alternative construction for the robustness program. Precisely, firm-level ESG indicators are defined as follows: ESG_pre denotes the composite average ESG score; Env_pre, Soc_pre, and Gov_pre denote the means of the environmental, social, and governance pillar scores, respectively. Because dichotomization discards information and depends on an arbitrary threshold, the continuous, mean-centered ESG_pre score interacted with the crisis indicator serves as the primary specification for inference (reported as specification R1 in Supplementary Table SB1), and the median split is retained as an interpretable categorical benchmark and robustness device, rather than as the principal design. For the categorical form, treatment groups are determined by a median split of ESG_pre: firms whose ESG_pre score exceeds the sample median are classified as high-ESG (High_ESG = 1), and the remaining firms as low-ESG (High_ESG = 0). A binary indicator (COVID = 1 in 2020–2021 and 0 otherwise) captures the crisis period.

3.3 ESG measurement properties

The environmental, social, and governance pillars are conceptually distinct dimensions of corporate conduct rather than interchangeable indicators of a single reflective latent construct; a composite ESG score is therefore best understood as a formative index. For this reason, no claim of construct validation is made for the composite. Cronbach’s alpha (Nunnally, 1978) is reported only as a descriptive summary of the degree of inter-pillar association in this sample, which is relevant for interpreting the collinearity of the pillar-level interactions in Model 3 (see Supplementary Appendix A), and not as evidence that the three pillars measure one underlying construct.

3.4 Econometric specification

To examine the relationship between ESG practices and corporate resilience, this study employs three complementary panel-data models, estimated in both continuous and categorical treatment forms as described in Section 3.2.

Model 1 DiD specification.

The primary model is a TWFE difference-in-differences specification. The TWFE DiD estimator is the standard panel approach for identifying differential responses to a shock (Callaway and Sant'Anna, 2021; Goodman-Bacon, 2021). The design has two features that limit the potential for TWFE bias identified in the recent methodological literature: (1) treatment assignment (High_ESG or, in the continuous form, centered ESG_pre) depends only on pre-crisis information and not on post-treatment years; and (2) the shock timing is common to all firms, so there is no staggered adoption, and the negative-weighting problems associated with staggered TWFE designs do not arise (Roth et al., 2023). The categorical DiD specification is as follows:where αᵢ represents firm fixed effects and γ_t denotes year fixed effects. For expositional transparency the equation displays the High_ESG and COVID main effects, but readers should note that both are absorbed by the fixed effects: High_ESGᵢ is time-invariant and is absorbed by αᵢ, and COVID_t is common to all firms in a year and is absorbed by γ_t. Only the interaction term is separately estimated, which is why Table 2 reports no main-effect rows. In the continuous primary form, High_ESGᵢ is replaced by the mean-centered continuous score ESG_preᵢ^c, so that β₃ measures the differential crisis-period change in ROA per unit of pre-crisis ESG. The coefficient of primary interest is β₃, which captures the differential change in ROA during the COVID-19 crisis for firms with stronger pre-crisis ESG engagement relative to weaker-ESG firms. Because the pandemic affected all firms and ESG standing is a pre-existing characteristic rather than an assigned treatment, β₃ is interpreted as a common-shock differential-exposure estimate: it identifies whether firms with different pre-crisis ESG profiles experienced different changes in ROA, and it provides evidence on, but does not by itself establish, an ESG-specific causal mechanism.

ROA ᵢ t = β 0 + β 1 High_ESG i + β 2 COVID t + β 3 High_ESG i × COVID t + β 4 ⁡ ln TotalAssets ᵢ t + β 5 Leverage ᵢ t + β 6 QuickRatio ᵢ t + β 7 AssetGrowth ᵢ t + α ᵢ + γ t + ε ᵢ t ,

TABLE 2. Panel regression results: ESG and ROA across three model specifications.

Dependent variable: ROA (%). Model 1 = two-way fixed-effects DiD; Model 2 = random-effects crisis-period; Model 3 = two-way fixed-effects pillar interaction. HC3 heteroskedasticity-robust standard errors in parentheses; Models 1 and 3 additionally clustered by firm group. By design, and not by specification search, financial leverage and quick ratio are not included in Model 3 as these controls were not part of the original pillar specification; this control-set difference is acknowledged in Section 5.4, and Supplementary Appendix A confirms via variance inflation factor (VIF) diagnostics that the pillar interaction terms are not collinear with financial leverage and the quick ratio. R² denotes the within R² for the fixed-effects Models 1 and 3, which excludes the variance absorbed by the firm and year fixed effects, and the R² of the GLS-transformed model for the random-effects Model 2. Models 1–2 are estimated on the 690-observation panel (Model 2 on its 140-observation crisis subsample); Model 3 uses the 661-observation pillar sub-panel. *p < 0.05; p < 0.01; *p < 0.001; †p < 0.10.

Identification and endogeneity. Identification rests on the parallel-trends assumption that, absent the pandemic, high- and low-ESG firms would have followed similar ROA paths; Section 4.4 reports a dynamic event-study specification (Table 3; Figure 1), and a placebo-year test is reported in Supplementary Table SB1. Three design features mitigate endogeneity: the treatment variable is a predetermined eight-year pre-crisis average, firm fixed effects absorb all time-invariant heterogeneity, and year fixed effects absorb common shocks. Reverse causality from crisis-period ROA to pre-crisis ESG is ruled out by construction. Remaining threats are time-varying confounds correlated with pre-crisis ESG, most importantly industry-specific pandemic exposure, which is why the TWFE estimator provides the most robust identification strategy for this natural experiment. Four features make the two-way fixed-effects DiD the most credible strategy available in this setting. (i) The shock is common to all firms, and its timing is exogenous to any individual firm, so there is no staggered adoption, and none of the negative-weighting pathologies of staggered TWFE designs (de Chaisemartin and D’Haultfoeuille, 2020; Goodman-Bacon, 2021; Roth et al., 2023). (ii) The treatment variable is predetermined (an eight-year pre-crisis average), which rules out reverse causality by construction and eliminates the anticipation problems that arise when treatment status responds to the outcome. (iii) Firm fixed effects absorb every time-invariant source of selection into ESG status (industry, business model, culture, and historical governance), and year fixed effects absorb every common shock, so identification requires only that unobserved determinants of ROA did not change differentially with pre-crisis ESG status during the pandemic. That assumption is testable in the pre-period, and it is assessed through the dynamic event-study specification (Section 4.4) and a placebo shock (Supplementary Table SB1). (iv) The design is extended rather than merely defended: a triple-difference specification exploiting industry pandemic exposure and an industry-by-year fixed-effects specification (Section 4.7) absorb the industry-specific shocks that constitute the leading residual threat. Instrumental-variable and system-GMM estimators were considered and rejected on substantive grounds. A valid instrument would have to shift a firm’s pre-crisis ESG orientation while affecting crisis-period ROA only through ESG, and no candidate in this setting (industry-peer ESG averages, rating-methodology changes, and state-level environmental regulation) plausibly satisfies that exclusion restriction since each also affects profitability directly through competition, disclosure, or compliance costs; internally instrumented system GMM would rely on lag exclusion restrictions that are unlikely to hold for a slow-moving firm characteristic, such as ESG orientation. Weakly or invalidly instrumented estimates would, therefore, be less credible than a transparent fixed-effects design whose identifying assumption is stated, tested, and probed. As a complementary check on selection, a one-to-one propensity-score-matched DiD is reported in Section 5.2. The absence of an external instrument and the resulting associational reading of the estimates are stated as limitations.

TABLE 3. Dynamic event-study estimates: high_ESG × year coefficients (reference year 2019).

Dependent variable: ROA (%). Two-way fixed-effects estimation on the full 2012–2021 panel (70 firms; 690 firm-year observations) with the Model 1 control set; HC3 standard errors clustered by firm. Coefficients are the differential ROA of high-ESG firms in each year relative to 2019, the last pre-crisis year. The joint Wald test uses the clustered covariance matrix. Figure 1 plots these estimates with 95% confidence intervals.

Model 2 Crisis-period random-effects model.

To further assess the role of ESG during the crisis period, a random-effects (RE) model is estimated on the crisis subsample (2020–2021):

ROA ᵢ t = β 0 + β 1 ESG_pre i + β 2 ⁡ ln TotalAssets ᵢ t + β 3 Leverage ᵢ t + β 4 QuickRatio ᵢ t + β 5 AssetGrowth ᵢ t + ε ᵢ t .

This specification evaluates whether pre-crisis ESG engagement is associated with firm performance during the crisis. A random-effects estimator is employed because the key explanatory variable (ESG_pre) is time-invariant and would be absorbed in a fixed-effects framework. The RE estimator requires the unobserved firm effect to be uncorrelated with the regressors, an assumption unlikely to hold exactly for ESG; Model 2 is therefore interpreted as a descriptive within-crisis association rather than a causal estimate, and a Hausman test rejects the orthogonality assumption empirically [χ² (2) = 8.90, p = 0.012], reinforcing the descriptive reading; a correlated-random-effects (Mundlak) specification, which relaxes this assumption, is identified as a refinement for the robustness program.

Model 3 ESG pillar disaggregation.

To identify the relative contribution of individual ESG dimensions, the composite ESG measure is decomposed into its environmental, social, and governance components:

ROA ᵢ t = β 0 + β 1 Env _ pre ᵢ × COVID t + β 2 Soc _ pre ᵢ × COVID t + β 3 Gov _ pre ᵢ × COVID t + β 4 ⁡ ln TotalAssets ᵢ t + β 5 AssetGrowth ᵢ t + α ᵢ + γ t + ε ᵢ t .

This model allows a more granular analysis of ESG effects by examining whether specific pillars drive the relationship between ESG and firm performance during crisis periods. All models are estimated using HC3 heteroskedasticity-robust standard errors; for fixed-effects specifications, standard errors are additionally clustered at the firm level to account for within-firm serial correlation (Petersen, 2009). The analysis is conducted using R with the plm package (Croissant and Millo, 2008).

4 Findings

4.1 Inter-pillar association

As described in Section 3.3, no construct-validation claim is attached to the composite ESG score. As a descriptive summary of inter-pillar association, Cronbach’s alpha across the environmental, social, and governance sub-scores is 0.72 (95% CI = 0.68–0.76; standardized alpha = 0.73; average inter-item r = 0.47), indicating moderate shared variance among conceptually distinct pillars. This value is reported because it foreshadows the collinearity structure of the pillar-level interactions examined in Model 3 and Supplementary Appendix A, not as evidence of a single latent ESG construct.

4.2 Descriptive statistics

Descriptive statistics for the key variables are reported in Table 4. ROA has a mean of 8.24% (SD = 5.42), in line with profitability levels in multi-country studies of large-cap firms (Ding et al., 2021). Pre-crisis ESG scores average 52.09 (SD = 16.75, range 14.21–87.27), reflecting substantial firm-level variation in sustainability performance. Financial leverage is highly right‐skewed (skewness = 6.47), consistent with results for diverse industry samples, and asset growth is also highly right‐skewed (5.73), motivating the winsorization at the 1st and 99th percentiles described in Section 3.2.

TABLE 4. Descriptive statistics for key study variables.

All continuous variables winsorized at 1st and 99th percentiles. ESG_pre, mean pre-crisis (2012–2019) ESG score; ROA, return on assets (%); SD, standard deviation. Table 4 is regenerated directly from the final dataset and computed on the merged analysis file (70 firms; 691 firm-years; financial leverage available for 690). Models 1–2 are estimated on the 690-observation complete-case panel, Model 2 on its 140-observation crisis subsample, and Model 3 on the 661-observation pillar sub-panel (67 firms); see Table 1.

4.3 Correlation analysis

Pearson correlations show a negative bivariate correlation between pre-crisis ESG and ROA (r = −0.30), visible in the correlation heatmap in Figure 2 and in the slopes of the ordinary least squares (OLS) regressions in Figure 3, Panel B, where both the pre-crisis and crisis trend lines are consistently negative. This pattern, higher ESG associated with slightly smaller accounting returns, is in line with the cost-of-investment explanation by Alareeni and Hamdan (2020), whereby initial sustainability investment produces transitional profitability drag. The bivariate correlation between financial leverage and ROA is −0.17. Bivariate magnitudes are, however, not informative about the conditional importance of predictors, which is assessed within the regression framework below. The largest pairwise predictor correlation is between firm size and pre-crisis ESG (r = 0.69), an expected pattern given that larger firms disclose more and score higher, which is addressed by including firm size in every specification and by the collinearity diagnostics reported in Supplementary Appendix A; no other predictor pair exceeds |r| = 0.50.

4.4 Main difference-in-differences results

Table 2 reports estimates for the three panel specifications: the two-way fixed-effects DiD model (Model 1), the random-effects crisis-period model (Model 2), and the two-way fixed-effects pillar-interaction model (Model 3). All standard errors are HC3-robust; Models 1 and 3 are additionally clustered by firm.

Model 1 evaluates the competing directional hypotheses H1a (insurance value: positive interaction) and H1b (short-run cost: negative interaction). The DiD interaction High_ESG × COVID is negative and statistically significant (b = −1.967, SE = 0.833, t = −2.36, p = 0.019, 95% CI = −3.60 to −0.33), a result inconsistent with the insurance-value prediction (H1a) and consistent with the short-run cost hypothesis (H1b). Rather than outperforming, high-ESG companies experienced an ROA approximately 2 percentage points lower than low-ESG counterparts during the crisis period. The primary continuous specification points the same way with greater precision: the centered ESG_pre × COVID interaction (Supplementary Table SB1, specification R1) is −0.063 ROA percentage points per unit of pre-crisis ESG (SE = 0.023, p = 0.006, 95% CI = −0.108 to −0.018), which corresponds to approximately −1.06 percentage points of ROA per one-standard-deviation (16.75-point) increase in pre-crisis ESG. In economic terms, the categorical estimate of −1.97 percentage points equals roughly 24% of the sample-mean ROA of 8.24% (N = 70 firms, 690 firm-year observations), and the one-SD continuous estimate equals roughly 13% of mean ROA; these are economically consequential, rather than merely statistically detectable, magnitudes. The pattern is visible in Figure 3, Panel A: the high-ESG group maintains a broadly parallel pre-crisis ROA trajectory to the low-ESG group but falls more sharply after 2020, consistent with the negative interaction coefficient. Figure 4, Panel B, further shows that the low-ESG group has a higher median ROA than the high-ESG group throughout the sample period.

To probe the parallel-trends assumption and to display the dynamics of the effect, a full event-study specification was estimated on the 2012–2021 panel: ROA is regressed on the interactions of High_ESG with indicators for every year, except 2019, the last pre-crisis year, which serves as the reference, together with the full control set and firm and year fixed effects, with HC3 standard errors clustered by firm. Table 3 reports the year-by-year coefficients, and Figure 1 plots them with 95% confidence intervals. All seven pre-crisis coefficients (2012–2018) are individually insignificant and small relative to the crisis coefficients; the largest, for 2018 (−1.34, p = 0.13), is imprecise, and the joint Wald test does not reject the null that all pre-crisis coefficients are zero (χ² (7) = 3.60, p = 0.82). The equivalent pooled pre-period test previously reported, F (7, 530) = 0.214, p = 0.982, is retained for comparability. In the crisis years, the high-ESG differential is negative and significant in both 2020 (−2.61, SE = 1.03, p = 0.011) and 2021 (−2.27, SE = 0.99, p = 0.023), indicating that the penalty was not confined to the initial shock year but persisted through the incomplete 2021 recovery; the pooled DiD estimate (−1.97) corresponds to the difference between the average crisis-year coefficient and the average pre-crisis coefficient. This event-study evidence supports, although, as with any pre-trends test, cannot definitively confirm, the identifying assumption (Roth et al., 2023), and Figure 3, Panel A, displays the underlying group trajectories.

An additional test is the crisis-period random-effects model (Model 2), which estimates the contemporaneous association between pre-crisis ESG scores and ROA during the pandemic period. The ESG main effect (b = 0.027, SE = 0.044, p = 0.545) is statistically indistinguishable from zero, indicating that no within-crisis profitability premium accrued to firms with higher ESG. Models 1 and 2 reach a convergent conclusion: pre-crisis composite ESG orientation did not produce an ROA resilience benefit during COVID-19 and, in the DiD framework, was associated with a statistically significant accounting penalty for high-ESG firms.

4.5 Pillar-level analysis

Model 3, estimated on the pillar sub-panel (67 firms; N = 661), splits ESG into its three pillars and evaluates each pillar–COVID interaction against its directional hypothesis. H2 predicted that higher pre-crisis environmental scores would be associated with significantly higher ROA during the crisis, reflecting operational-efficiency benefits. The Env_pre × COVID interaction (b = −0.043, p = 0.110) is negative and not statistically significant; H2 is not supported. H3 predicted that greater pre-crisis social scores would deliver ROA benefits through workforce stability. The Soc_pre × COVID interaction (b = −0.003, p = 0.923) is negligible and insignificant; H3 is not supported. H4 predicted that higher pre-crisis governance scores would be associated with better ROA through risk-management discipline. The Gov_pre × COVID interaction (b = −0.006, p = 0.847) is also insignificant; H4 is not supported. To make the informational content of these null results explicit, minimum detectable effects (MDEs) were computed for each interaction at 80% power and a 5% two-sided test, using the conventional approximation MDE ≈ 2.8 × SE applied to the reported robust standard errors. The MDEs are approximately 0.075 (environmental), 0.093 (social), and 0.092 (governance) ROA percentage points per pillar point. All three point estimates fall below their respective MDEs, so the design cannot distinguish protective effects smaller than these bounds from true nulls. The pillar-level results are, therefore, characterized as precision-limited nulls: the data neither support a protective pillar-level effect nor rule out effects below the stated bounds, and no claim of a latent or suppressed effect is made.

4.6 Control variables

Firm size (ln total assets) is negative and significant in all three models (−3.17 to −3.51, p < 0.001), indicating diminishing marginal returns to scale in fixed-effects estimation, consistent with evidence that large firms exhibit proportionally higher ESG compliance burdens (Linnenluecke, 2022). Financial leverage has a consistently negative coefficient in Models 1 and 2 (b = −0.052 and −0.099, respectively; both p < 0.001), the most consistent result in the analysis. Asset growth is marginally significant in Model 1 (p = 0.099). Liquidity (quick ratio) is not significant in any specification. Two caveats attach to the control set. First, leverage, liquidity, size, and asset growth are entered contemporaneously and may themselves have responded to the pandemic; they are, therefore, not strictly exogenous conditioning variables, and leverage, in particular, is a candidate resilience mechanism in its own right (Section 5.3). Re-estimating the models with lagged or pre-pandemic (2019) values of the financial controls, and a baseline specification without conditioning variables, are identified for the robustness program. Second, for comparative statements about ESG versus leverage, unstandardized coefficients on differently scaled variables are not directly comparable; per-standard-deviation associations are used in Section 5.3.

4.7 Robustness: additional controls, triple-difference, and industry-by-year specifications

Three further specifications address omitted-variable and industry-exposure concerns directly; full estimates are reported in Supplementary Appendix B. First, adding R&D intensity and advertising intensity, each measured as annual expense scaled by total assets, to Model 1 leaves the High_ESG × COVID interaction essentially unchanged (b = −1.933, SE = 0.832, p = 0.020; Supplementary Table SB1, R6, versus −1.967 in the baseline), and neither intensity measure predicts ROA within firm (R&D intensity: b = 15.97, SE = 17.69, p = 0.367; advertising intensity: b = 21.70, SE = 36.90, p = 0.557). The crisis-period penalty for high-ESG firms is, therefore, not an artefact of differences in innovation or marketing intensity, and the omitted-control concern raised in Section 3.2 is resolved, rather than merely acknowledged.

Second, a DDD specification interacts the DiD term with an indicator for the industries most exposed to the pandemic’s demand collapse, namely, airlines, air freight and logistics, and road and rail (14 firms; 140 firm-year observations), so that the third difference compares the high-ESG crisis differential across high- and low-exposure industries; the exposure main effect is absorbed by the firm fixed effects, and its interaction with COVID is estimated. Supplementary Table SB2 reports the full estimates. In lower-exposure industries, the High_ESG × COVID coefficient is −1.08 (SE = 0.67, p = 0.107); the triple interaction is −2.59 (SE = 3.35, p = 0.440), implying an ESG × COVID differential of −3.67 (SE = 3.28) in high-exposure industries. The point estimates are thus ordered as predicted by the sustainability trade-off mechanism (the penalty is larger where revenue collapsed most), but the DDD cannot reject equality across exposure groups at this sample size, and the low-exposure coefficient falls short of conventional significance. Two readings are consistent with this pattern and cannot be distinguished here: that rigid sustainability commitments were most costly where demand fell furthest, and that part of the average penalty reflects the concentration of high-ESG firms in exposed industries.

Third, replacing year fixed effects with industry-by-year fixed effects (12 industries × 10 years = 120 cells), which absorb all industry-specific pandemic shocks, attenuates the High_ESG × COVID coefficient to −1.14 (SE = 0.70, p = 0.104; Supplementary Table SB1, R8), roughly 42% smaller than the baseline and no longer significant at conventional levels. The within-industry-year differential remains negative and of similar magnitude to the one-standard-deviation continuous estimate (−1.06), so the pattern is not eliminated by industry adjustment, but this specification indicates that a material share of the raw differential reflects industry-level exposure correlated with ESG status. This result is explicitly reflected in the interpretation (Section 5.1) and the limitations (Section 5.4).

5 Discussion

5.1 The short-run cost mechanism of environmental and social commitments

The research question that inspired this study is whether firms more engaged in ESG are more resilient in terms of financial performance, in particular ROA, when facing systemic economic shocks such as the COVID-19 pandemic. The statistically significant negative DiD interaction (b = −1.967, p = 0.019; continuous form b = −0.063, p = 0.006) is the main substantive finding of the study and is inconsistent with the canonical ESG-resilience hypothesis, under which accumulated social capital is theorized to offer an insurance premium during times of crisis by keeping stakeholders loyal and minimizing opportunistic defection (Lins et al., 2017). Recent evidence likewise shows that the ESG–performance link is not always robust when financial performance is measured by accounting outcomes such as ROA, especially during crisis periods (Bae et al., 2021; Broadstock et al., 2021). An emerging literature suggests that sustainability benefits are more apparent in market‐based indicators than in operating profitability, which can explain why ESG might not generate an ROA buffer during a crisis (Albuquerque et al., 2020; Siwiec et al., 2024). The positive effects of ESG during the COVID-19 crash reported by Broadstock et al. (2021) and Albuquerque et al. (2020) rely on equity returns rather than accounting-based profitability. The current research examines ROA, a direct metric of operational continuity and stakeholder value creation. Read through the sustainability trade-off mechanism developed in Section 2.1, this is the pattern expected when environmental and social commitments behave like fixed capacity investments: their relational value is priced by markets, but their operating leverage is borne by accounting margins when demand collapses. Consistent with the same decoupling, Song et al. (2026) find that ESG lowers financial-distress risk during downturns, a risk-side benefit that can coexist with the accounting-margin penalty documented here.

The cost-based reading of this finding must be stated with care. One plausible interpretation is that high-ESG companies sustain relatively rigid spending on employee welfare, environmental compliance, and governance infrastructure, obligations that cannot be reversed quickly without damaging the social capital on which they rest (Lins et al., 2017; Fatemi et al., 2018), so that when revenues collapsed in 2020–2021, these structural commitments produced a disproportionate profitability drag. This mechanism is consistent with the estimates, but it is not evaluated directly: the study does not measure ESG-related expenditures, cost rigidity, or the financing of sustainability programs, and the negative interaction is therefore evidence of a differential profitability pattern, rather than a demonstration of its cause. Competing explanations deserve equal statement. High- and low-ESG firms differ on characteristics other than ESG orientation, including size (r = 0.69 with pre-crisis ESG in this sample), industry composition within the industrial sector, capital intensity, international exposure, workforce composition, supply-chain dependence, and the feasibility of remote operation (Ding et al., 2021; Binesh et al., 2025). Firm fixed effects absorb the time-invariant portion of these differences, and firm size is controlled directly, but industry-specific pandemic shocks and other time-varying confounds correlated with ESG standing could contribute to the estimated interaction. The triple-difference and industry-by-year specifications in Section 4.7 take these alternatives as far as the data allow: the penalty is imprecisely larger in the most exposed industries, consistent with the trade-off mechanism, but industry-by-year adjustment attenuates the average estimate by approximately 42% and renders it marginal (p = 0.10), indicating that part of the raw differential reflects industry exposure correlated with ESG status. Fully separating the ESG-cost mechanism from exposure therefore still requires mechanism data (for example, ESG-related expenditure and cost-structure disclosures), which are identified as the priority for future work. This reading aligns with Bae et al. (2021) and Demers et al. (2021), who similarly find no significant ESG protection for US firms during the COVID-19 crisis, attributing null results to decoupling between rated ESG scores and actual operating activities. The present study extends these findings in a multi-sector industrial panel.

5.2 Pillar-level evidence: capacity-like commitments and precision-limited nulls

The trade-off mechanism predicts that the crisis penalty should be concentrated in the environmental and social pillars, whose commitments are most capacity-like, rather than in governance. The pillar estimates are directionally consistent for the environmental pillar and uninformative for the social pillar, but they cannot confirm or reject the prediction: as the minimum-detectable-effect analysis in Section 4.5 shows, the design lacks the power to resolve pillar-specific effects of the size implied by dividing the composite differential across three correlated pillars. The pillar-level results are therefore reported as evidence on the limits of disaggregation in samples of this size, and the recent finding that pillar-level financial effects vary sharply across regions and industries (Soltanisehat and Sibdari, 2026) suggests that larger and more heterogeneous samples will be needed to assess the second implication of the mechanism decisively.

H2 held that increased pre-crisis environmental scores would produce higher ROA in the crisis, reflecting the operational-efficiency benefits that energy savings, waste minimization, and leaner supply chains are hypothesized to generate. The environmental–COVID interaction (b = −0.043, p = 0.110) is not statistically significant, and, according to the MDE analysis in Section 4.5, the design cannot detect effects smaller than approximately 0.075 ROA points per environmental point; the estimate is therefore an upper bound on any protective environmental effect in this sample, rather than evidence that no effect exists, and equally, it provides no evidence of a hidden effect. Broadstock et al. (2021) identify environmental performance as the most active ESG predictor of abnormal returns in the 2020 crash, interpreting it as a signal of organizational lean management. The accounting counterpart of this signal, real cost savings from energy efficiency and emissions reduction flowing through to ROA, may not materialize quickly enough to register within a 2-year crisis window, during which demand shocks dominate the profit-and-loss account. Previous research finds that environmental performance can correlate positively with crisis-period profitability, although the effect does not generalize across sectors or outcome measures, suggesting that the ROA channel is weaker than the stock-return channel (Broadstock et al., 2021; Siwiec et al., 2024). Studies on COVID-19 ESG also indicate that resilience may operate via financing constraints and green innovation, such that the environmental pillar may be indirectly relevant without being immediately visible in short-run profitability (Gao and Geng, 2024; Li and Xu, 2024). H2 is not supported in the present data.

H3 hypothesized that greater pre-crisis social scores would provide ROA benefits through workforce stability: retaining talented workers, sustaining morale, and ensuring operational continuity during crisis-induced disruption. The negligible Soc_pre × COVID interaction (b = −0.003, p = 0.923) does not support H3. Social-pillar effects are theorized to operate via employee retention and stakeholder loyalty, yet crisis-period evidence is inconsistent and frequently non-significant when accounting profitability rather than market value is the outcome (Bae et al., 2021; Lins et al., 2017). A reading consistent with the broader literature is that social investments build long-term relational buffers whose accounting payoff, if any, accrues over multi-year horizons rather than within an acute demand shock (Albuquerque et al., 2020; Broadstock et al., 2021); on the present evidence, this remains an interpretation, rather than an established finding.

H4 hypothesized that governance quality would convert into higher ROA through risk-management discipline and capital-structure conservatism. The Gov_pre × COVID interaction (b = −0.006, p = 0.847) is insignificant, and H4 is not supported. This contrasts with Ahmad et al. (2023), who identify governance as the most significant ESG pillar for safeguarding profitability in UK financial crises, and Kara et al. (2022), who report that board gender diversity was positively related to responsible banking during COVID-19. Candidate explanations for the discrepancy include the reliability of the governance sub-score, whose within-scale association is the weakest of the three pillars in this sample (item-rest correlation = 0.37), and the possibility that governance operates on crisis profitability indirectly through leverage discipline, which is controlled separately; both remain untested here. It is widely believed that governance is accompanied by more effective crisis management and reduced downside risk, but the empirical findings are not consistent across countries and samples (Siwiec et al., 2024; Gao and Geng, 2024), and the operationalization of the governance construct can affect the estimated governance–profitability relationship (Broadstock et al., 2021; Li and Xu, 2024).

Collinearity diagnostics for the pillar-level specification, including variance inflation factors for the main effects and interactions and single-pillar re-estimations, are reported in Supplementary Appendix A (Fox and Weisberg, 2019); they indicate that the elevated interaction VIFs are the mechanical consequence of entering interactions alongside their components, rather than evidence of collinearity among the underlying pillar scores, and the single-pillar models reproduce the non-significant interactions of the joint specification.

Supplementary Table SB1 reports the High_ESG × COVID interaction across robustness specifications. The effect is statistically significant (all p ≤ 0.043) and negative in all non-placebo variants (Baseline, R1, and R2). Critically, the placebo test (R3), assigning a pseudo-shock in 2017 using pre-COVID data only, generates a non-significant coefficient (b = −0.889, p = 0.192), supporting the conclusion that the baseline result is not a statistical artefact of the sample or threshold choice. The per-unit continuous interaction coefficient is −0.063 (p = 0.006); evaluated at the high/low group difference in mean ESG, this is consistent with the median-split estimate. The extreme-tercile specification (R2) yields a larger coefficient of −2.069 (p = 0.043), in line with the cost hypothesis. An extreme-quartile split (R4) yields −2.135 (p = 0.098), marginally significant with wider uncertainty but consistent in sign. Finally, as a check responsive to selection concerns, a one-to-one propensity-score-matched DiD was estimated (R5; 35 treated and 35 control firms, no observations discarded): the matched estimate is identical to the baseline (b = −1.967, p = 0.019). Post-matching covariate balance, nonetheless, remains imperfect (absolute standardized mean differences up to 2.0), so matching does not by itself resolve non-random ESG status, and this estimate is reported for transparency, rather than as an independent identification strategy. The additional-control (R6) and industry-by-year (R8) specifications and the triple-difference estimates are discussed in Section 4.7. Overall, the robustness battery indicates that the baseline result is not driven by the median-split treatment assignment, and, as set out in Section 3.2, the continuous specification (R1) serves as the primary basis for inference, with the categorical models as interpretable benchmarks.

5.3 The ESG–capital structure tension: debt-financed sustainability and amplified fixed costs

The most consistent predictor across the specifications in which it appears is financial leverage (b = −0.052 to −0.099, p < 0.001). Comparing the importance of leverage and ESG requires scale-free terms because the two variables are measured on different scales, and leverage does not enter Model 3. In per-standard-deviation terms, a one-SD (13.89-point) increase in leverage is associated with roughly −0.72 percentage points of ROA in Model 1 (and −1.38 in the crisis-period Model 2), while a one-SD (16.75-point) increase in pre-crisis ESG is associated with roughly −1.06 percentage points of crisis-period differential ROA via the continuous interaction. These per-SD associations are of comparable order but are not directly commensurable since the leverage estimate is an unconditional main effect and the ESG estimate is a crisis-specific differential; the comparison is suggestive rather than a formal test of relative importance, for which standardized estimates, partial explanatory power, or model-fit comparisons would be required. What the estimates do establish is that capital-structure conservatism is a robust correlate of crisis-period profitability in this sample. Ding et al. (2021) reached a similar conclusion, finding that pre-pandemic cash holdings, undrawn credit, and low leverage were among the most effective predictors of COVID-19 stock resilience across 61 economies. The accounting-profitability analog is direct: firms entering the crisis with lower debt-to-asset ratios were better placed to service interest, and their net margins were not forced to contract around fixed debt service as with high-leverage firms. One policy-relevant possibility consistent with, but not established by, these estimates is that ESG investment and balance-sheet conservatism can come into tension when sustainability programs are debt-financed: the fixed cost of ESG commitments added to the fixed cost of leverage may be amplified in a crisis, a channel made plausible by the growth of green bonds and environment-linked credit (Zhang et al., 2020). To disentangle these mechanisms, future research should directly examine the joint predictability of ESG and leverage as determinants of crisis-period profitability. Shakil (2021) provides initial evidence that ESG controversies and board gender diversity moderate the relationship between ESG performance and financial risk, suggesting that governance quality may be a key conditioning variable in this ESG–leverage interaction.

This tension can be theorized more precisely. Sustainability programs are increasingly financed with debt: green bonds, sustainability-linked loans, and environment-linked credit lines have increased rapidly since the mid-2010s, and large industrial firms have been prominent issuers (Zhang et al., 2020). Debt-financed sustainability converts a discretionary investment into two layers of fixed charge: the fixed operating cost of running the program, which is the operating-leverage channel of Section 2.1, and the fixed financing cost of servicing the associated debt, a financial-leverage channel. Under normal demand, both layers are covered by revenue, and the relational value of the program accrues; under a revenue collapse, both layers persist, while the revenue that covered them disappears, so the accounting penalty is amplified multiplicatively rather than additively. High-ESG firms that financed sustainability with debt therefore enter a crisis with higher combined operating and financial leverage than either their ESG score or leverage ratio alone would indicate. The mechanism yields three predictions: (i) a negative ESG × crisis interaction; (ii) a negative main effect of leverage on crisis-period profitability; and (iii) a negative ESG × leverage × crisis interaction concentrated among firms whose sustainability programs are debt-financed. The present estimates are consistent with (i) and (ii); (iii) is not identifiable in this sample without data on how sustainability programs are financed and is stated as the central testable proposition for future work. Two boundary conditions qualify the argument. First, well-governed firms may deliberately pair high sustainability commitments with conservative capital structures, in which case the tension is managed rather than realized; the null governance-pillar result is consistent with governance operating through this capital-structure choice, which is controlled separately, rather than through a direct profitability channel. Second, the tension is a short-run phenomenon: over horizons long enough for the relational and efficiency value of sustainability investment to accrue, sustainability programs may lower financing costs and increase resilience (Song et al., 2026; Yang and Wang, 2026), so the mechanism describes the timing of returns to sustainability investment, not their sign.

5.4 Conclusion

The aim was to understand how ESG relates to corporate performance during crisis conditions. In contrast to prevailing assumptions in the literature, the findings do not support the proposition that ESG enhances short-term financial resilience. The difference-in-differences estimates indicate that companies with higher pre-crisis ESG engagement experienced a larger relative decline in profitability during the crisis, and the crisis-subsample analysis shows no statistically significant within-crisis ESG–performance association. The event-study estimates locate the penalty in both crisis years with no differential pre-trends, and it survives controls for R&D and advertising intensity, while industry-by-year adjustment attenuates it, so part of the average differential reflects industry exposure. The scope of these conclusions must be stated precisely: within a relatively small sample of US industrial firms, over the 2020–2021 window, and under the reported models, high pre-crisis ESG status was associated with a less favorable relative change in ROA. The study does not disprove the general resilience value of ESG, does not measure the cost mechanism directly, and does not license generalization to other sectors, countries, crises, or outcome measures. The visual and econometric evidence consistently shows that high-ESG firms did not outperform low-ESG firms in the run-up to the crisis or during it, with the performance gap widening under economic stress. These outcomes suggest that ESG investment may impose a short-term financial constraint that limits operational flexibility during acute uncertainty and that ESG’s risk-reducing effectiveness is not uniform across conditions. Together, the results point to a temporal decoupling between ESG activity and financial performance: although ESG can contribute to long-term value creation, its short-term accounting outcomes, particularly during times of crisis, are less favorable. The study therefore calls for a sharper definition of the ESG value proposition that distinguishes long-term strategic benefits from short-run financial trade-offs.

The findings have implications for practitioners and scholars. At the managerial level, they indicate that ESG investments cannot be regarded as short-term shock absorbers in an economic crisis. High ESG engagement may reduce financial flexibility during crises because resources are committed to compliance and stakeholder obligations. The managerial implication is emphatically not that firms should reduce ESG investment; it is that ESG ratings should not be treated as guarantees of short-term accounting resilience and that ESG strategy should be integrated with financial risk management, in particular capital-structure policy, so that sustainability commitments remain sustainable under adverse conditions. There is a common assumption among many corporate managers and investors that high-ESG companies are safer during economic crises. The present evidence suggests that ESG ratings are not reliable indicators of short-term financial stability under systemic stress, which calls for a more discriminating use of ESG ratings in portfolio construction and risk management.

For the environmental strategy and sustainability policy audience of this journal, the findings carry a specific message. A crisis-window accounting penalty for high-ESG firms speaks to the timing profile of returns to environmental and social investment, not to its long-run value: environmental commitments behave like capacity investments whose costs are rigid in the short run and whose payoffs accrue over longer horizons. Policymakers and standard-setters should, therefore, avoid reading crisis-period profitability gaps as evidence against environmental investment and should instead consider instruments that smooth the short-run burden of sustainability commitments in downturns, for example, transition finance, countercyclical compliance flexibility, and disclosure frameworks that separate financially material environmental actions from composite scores. Investors using ESG ratings for resilience screening should complement composite scores with balance-sheet indicators, which this study finds to be the more reliable crisis-period signal.

Theoretically, the research contributes to the ESG–performance debate by incorporating time as a conditioning factor. Although the existing literature emphasizes the long-term benefits of ESG, the present results point to potential short-run costs. This supports a trade-off perspective in which ESG engagement underwrites long-run sustainability while constraining short-run profitability, especially during crises, and promotes a reconsideration of ESG within dynamic, context-specific frameworks rather than as a universally protective construct.

Although this represents a detailed investigation, several limitations should be acknowledged. First, the sample is limited to 70 US industrial firms; beyond constraining generalizability across sectors and geographies, the requirement of long continuous ESG histories tilts the sample toward large, disclosure-intensive survivors, so selection and survivorship bias cannot be excluded. Second, the modest sample size limits statistical power, especially for pillar-level effects, as quantified by the minimum-detectable-effect analysis in Section 4.5. Third, ESG standing is not randomly assigned: high- and low-ESG firms differ on size, industry mix, capital intensity, international exposure, and other characteristics, and, although the triple-difference and industry-by-year specifications now address industry-specific pandemic shocks directly, the industry-adjusted estimate is attenuated and imprecise, so the estimates remain differential-exposure associations rather than demonstrated ESG effects, and no direct mechanism test (of ESG expenditures or cost rigidity) is performed. Fourth, reliance on a single rating provider (Refinitiv) introduces measurement dependence since providers diverge substantially (Berg et al., 2022; Hahn et al., 2021), and results may differ under MSCI, Bloomberg, or Sustainalytics data. Fifth, the median split used in the categorical models discards information and imposes an arbitrary threshold, which is why the continuous specification is primary and both categorical forms are reported as robustness. Sixth, the contemporaneous financial controls may themselves respond to the pandemic and thus act partly as mediators; lagged-control specifications are flagged for future estimation. Seventh, the parallel-trends assumption is supported by the event-study test but cannot be definitively validated, and the pooled 2020–2021 window averages over heterogeneous within-crisis dynamics, including late-2021 recoveries in some industries. Finally, no external instrument for ESG is available, so omitted time-varying heterogeneity cannot be fully excluded.

Future research should investigate the dynamic nature of the ESG–performance relationship across periods and economic conditions. Concrete extensions follow directly from this study’s limitations: estimating the effects separately for 2020 and 2021 and tracing crisis-recovery horizons; re-estimating with lagged or pre-pandemic financial controls and a correlated-random-effects (Mundlak) form of the crisis model; replicating the design with alternative rating providers; collecting mechanism data such as ESG-related expenditure, cost-structure, and sustainability-financing disclosures to evaluate the cost channel and the debt-financing amplification proposed in Section 5.3 directly; and extending the outcome set to alternative resilience indicators such as employment stability, survival and exit rates, or credit-rating changes. Sector-specific analyses could identify heterogeneity in ESG performance effects, and institutional and regulatory contexts may help explain cross-country differences in when and how ESG creates value.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary Material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fenvs.2026.1909351/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Fortune Ganda, Walter Sisulu University, South Africa

Reviewed by: Nasim Binesh, University of Florida, United States

Diana Alhajjeah, Applied Science Private University, Jordan