Heterogeneous impacts of fear and policy on building energy use during COVID-19 in South Korea

Summary

The COVID-19 pandemic created a unique natural experiment in energy consumption, fundamentally reshaping demand. We analyzed electricity use in over one million South Korean buildings across six functional categories (May 2020 to April 2022). Using a panel regression with an instrumental variable strategy, we disentangle fear-driven behavioral responses (proxied by local mortality) from policy-driven mandates. Our empirical findings reveal heterogeneous impacts: fear reduced industrial and civic/business electricity use, yet counterintuitively increased residential and commercial/retail demand, often driven by safety signaling (e.g., enhanced ventilation). Stricter anti-contagion policies increased energy intensity in most nonresidential buildings due to compliance with heightened ventilation standards, but unexpectedly reduced residential consumption, likely reflecting restricted social activities and efficiency gains from appliance upgrades. These effects generally diminish over time, with distinct adaptation speeds across building types. Our study highlights that future crisis-era energy policies must be sector-specific, time-sensitive, and regionally adaptive to effectively manage demand shocks and sectoral vulnerabilities.

Graphical abstract

Highlights

• •Fear increased residential and retail use but reduced industrial and civic demand
• •Policy constraints raised non-residential energy intensity yet cut residential use
• •Impacts on energy use diminished over time with distinct adaptation across sectors
• •Future energy policy must be sector specific, time sensitive, and regionally adaptive

Health sciences; Environmental science; Energy resources; Social sciences; Interdisciplinary application studies

Introduction

The COVID-19 pandemic introduced a unique natural experiment in global energy consumption, triggering major shifts in demand patterns across all sectors.^1,2 In 2020, worldwide energy use fell by approximately 6% relative to 2019—seven times the decline seen during the 2008–2009 financial crisis.³ Multiple factors, including government-imposed restrictions,^4,5 voluntary risk avoidance,^6,7 and disruptions to global supply chains,² reshaped how energy was consumed, often in ways unanticipated by existing energy systems. Although aggregate energy demand recovered in many developed countries by 2021 (Figure 1), the complex variability in energy use driven by diverse behavioral changes and policy interventions is yet to be fully understood over time. Even as the global emergency phase has passed, lasting behavioral shifts, such as remote work and risk avoidance, point to the pandemic’s enduring influence on energy demand, raising crucial questions about how societies adapt to energy consumption under conditions of chronic disruption.

Figure 1.

Electricity consumption between 2010 and 2022 in OECD, USA, South Korea, and EU

Note: The red shading marks the 2020–2022 pandemic period. Source: IEA (2024).⁸

One of the key challenges in understanding energy consumption dynamics during the pandemic is disentangling the effects of policy-driven behaviors and fear-based or risk-driven behavioral responses. Understanding how the relative importance of these channels has evolved over time, and how their impacts differ across functional uses of electricity, is essential for designing crisis-resilient energy systems. Previous research has documented short-term load changes during early lockdowns, but the evidence remains fragmented, often limited to single buildings, short windows, or national aggregates that blur granular detail. This leaves two unresolved questions: (i) how do the policy and fear effects interact during an extended crisis? and (ii) do these effects attenuate or persist as societies adapt?

To address the limitations of previous research, this study leveraged a unique dataset of electricity use from over one million buildings across South Korea, covering May 2020 to April 2022. The dataset spans six functional categories—residential, commercial/retail, civic/business, industrial, religious/cultural, and community amenities—and is matched to regional COVID-19 mortality and a daily policy-stringency index. By combining building-level fixed-effects regressions with an instrumental-variables (IV) strategy, we separate the elasticities of electricity demand to policy stringency and local mortality (our proxy for fear).

Our empirical analysis revealed the strikingly divergent impacts of policy-driven and fear-based behavioral responses on electricity demand across different building types. We observed how fear-induced behaviors increased electricity use in residential and commercial/retail facilities, while reducing electricity use in the industrial and civic/business sectors. Furthermore, stricter anti-contagion policies counterintuitively raised energy intensity in most non-residential buildings owing to compliance with ventilation standards, while unexpectedly reducing residential consumption. These heterogeneous impacts generally diminish over time with distinct adaptation speeds across building types. These findings underscore the need for future crisis-era energy policies to be sector-specific, time-sensitive, and regionally adaptive. Thus, our study provides crucial empirical insights for designing more resilient energy systems capable of managing demand shocks and sectoral vulnerabilities during future public health emergencies, environmental disasters, and geopolitical shocks.

Theory and contextual background

Policy-driven and fear-based behaviors in energy consumption

Policy-driven behavior refers to changes in energy use resulting from official mandates, including mobility restrictions, lockdowns, social distancing requirements, business closures, and ventilation requirements imposed to contain viral spread. In contrast, fear-induced behavior encompasses voluntary adaptations to perceived health risks even without formal restrictions, such as avoiding public spaces or increasing time spent at home.⁷ As Goolsbee and Syverson⁷ highlighted, both channels shape energy use responses during public health crises, but their relative influence and sectoral impacts are highly heterogeneous. Understanding the interplay between these behavioral channels is critical for designing resilient energy systems.

The existing literature provides empirical evidence from diverse geographic contexts, illustrating the heterogeneous and dynamic nature of these impacts. In Dalian, China, Su et al.⁹ observed a 55%–63% monthly reduction in electricity consumption in a large commercial green building during lockdowns, along with improved indoor air quality, owing to reduced occupancy. In Great Britain, national-level data showed only modest aggregate changes during the lockdowns, but building-level analysis revealed that residential water heating and appliance use increased, while space heating decreased, highlighting sectoral heterogeneity and behavioral adaptation to extended home occupancy.¹⁰ Similarly, research from European municipalities revealed substantial variations in building functions: Energy demand in schools and universities declined sharply owing to closures, whereas administrative and office buildings experienced smaller changes. A European study further indicated that building energy consumption does not always directly correspond to occupancy levels.¹¹ These examples collectively demonstrate the importance of context, timing, and functional diversity in shaping energy trends in the pandemic era.

Although a growing body of literature has examined the energy implications of COVID-19 across residential and commercial sectors, many studies rely on short observation windows, highly localized data, or aggregated national-level indicators.^12,13,14 Few attempts have been made to empirically separate the effects of policy- and fear-driven behaviors or examine the persistence of these effects as societies adapt to ongoing or resurgent public health risks.

South Korea as a compelling context

By disentangling the effects of fear- and policy-based behavioral shifts, our study contributes not only to a retrospective understanding, but also to forward-looking preparedness for future systemic shocks. South Korea provides a compelling context for analysis because of its transparent and consistent policy trajectory, high-quality energy data infrastructure, and regional variations in pandemic severity. The country has one of the world’s highest per capita electricity consumption rates (Figure 1) and has experienced a sharp interruption due to the pandemic in its long-term growth trajectory.

As Figure 2 suggests, South Korea’s pandemic response evolved from strict early containment to more relaxed measures amid rising infections, providing a unique setting to distinguish between the impacts of policy interventions and public risk perceptions. Furthermore, the regional and temporal heterogeneity of outbreaks, coupled with detailed building-level energy data, enables a granular analysis of how different sectors and regions respond to both policy and fear-based drivers.

Figure 2.

Trends in electricity consumption per building area across six building types and national COVID-19 positive cases

Note: The stacked bars show the monthly electricity use (kWh/m²) for six building types: civic or business facilities, community amenities, commercial or retail facilities, cultural or religious facilities, industrial services, and residential buildings. The blue and red lines represent monthly COVID-19 deaths and aggregated daily anti-contagion policy measures, respectively. Source: MOLIT (2023)¹⁵ and KDCPA (2024).¹⁶

The dataset covered six major building categories: residential, commercial/retail, civic/business, industrial, religious/cultural, and community. Figure 2 shows this complexity; it tracks the monthly electricity consumption for six building types in South Korea, alongside COVID-19 deaths and anti-contagion policy measures. Remarkably, industrial facilities began to recover by August 2021, whereas other categories followed more varied trajectories. Moreover, the timing of the changes suggests that stringent policies drove initial reductions, while later shifts appeared to be more fear driven. These patterns highlight the importance of examining the independent effects of each channel over time.

A further dimension of heterogeneity emerges geographically as outbreaks occur unevenly across time and regions. Figure 3 illustrates how COVID-19 mortality distribution changed over time across South Korea’s 16 regions. The figure shows that initially low mortality rates were clustered in a few hotspots, while mortality eventually spread more uniformly nationwide.

Figure 3.

Regional distribution and time series of COVID-19 death case ratios in South Korea

Note: The top panel shows the ratios of deaths per population in May 2020, May 2021, and April 2022. The bottom panel depicts the time series of COVID-19 death cases on a logarithmic scale. Source: KDCPA (2024).¹⁶

We examined four key aspects of the heterogeneous responses to the pandemic: building type, temporal evolution, regional variation, and the distinction between policy-driven and fear-based behaviors. Using a high-resolution longitudinal dataset, this study explores how policy interventions (captured by government policy stringency indices) and public anxiety (proxied by local COVID-19 mortality rates) influence energy use in a disaggregated, temporally dynamic manner. Through panel regression modeling, this study goes beyond aggregate trends to highlight pronounced differences across building types, distinct temporal patterns, regional variations in pandemic intensity, and striking contrasts between policy mandates and voluntary behavioral responses.

Results

Our empirical analysis revealed substantial heterogeneity in how the pandemic reshaped energy consumption patterns across building types, driven by the complex interplay between fear-driven behavioral responses and policy-driven adjustments (Table 1). By leveraging our instrumental variables strategy, we can separate these two channels and examine their distinct impacts on electricity demand.

Table 1.

Main empirical result

Dependent variable: ELECi,j,t	Commercial or retail facilities	Community amenities	Civic/business facilities	Industrial services	Religious/cultural facilities	Residential buildings
COVID-19 deaths ($\widehat{{\alpha }_1})$	0.0311 (0.0109)∗∗∗	−0.0572 (0.0036)∗∗∗	−0.0949 (0.0088)∗∗∗	−0.1181 (0.0068)∗∗∗	−0.0805 (0.0108)∗∗∗	0.0583 (0.0053)∗∗∗
Deaths interacted with time ($\widehat{{\alpha }_2})$	−0.0019 (0.0003)∗∗∗	0.0001 (0.0001)∗	0.0006 (0.0002)∗∗∗	0.0021 (0.0002)∗∗∗	0.0003 (0.0002)	−0.0006 (0.0003)∗
Anti-contagion policy ($\widehat{{\alpha }_3})$	−0.0002 (0.0006)	0.0038 (0.0002)∗∗∗	0.0070 (0.0006)∗∗∗	0.0072 (0.0004)∗∗∗	0.0049 (0.0006)∗∗∗	−0.0042 (0.0005)∗∗∗
Policy interacted with time ($\widehat{{\alpha }_4})$	−0.0000 (0.0000)	−0.0001 (0.0000)∗∗∗	−0.0003 (0.0000)∗∗∗	−0.0003 (0.0000)∗∗∗	−0.0002 (0.0000)∗∗∗	0.0002 (0.0000)∗∗∗
Unemployment rate ($\widehat{{\alpha }_5}$)	0.0118 (0.0029)∗∗∗	0.0081 (0.0009)∗∗∗	0.0054 (0.0023)∗∗	−0.0036 (0.0020)∗	−0.0012 (0.0042)	0.0106 (0.0020)∗∗∗
Time ($\widehat{{\alpha }_6})$	−0.0071 (0.0052)	−0.0387 (0.0016)∗∗∗	−0.0616 (0.0045)∗∗∗	−0.0560 (0.0036)∗∗∗	−0.0455 (0.0057)∗∗∗	0.0300 (0.0039)∗∗∗
Time2 ($\widehat{{\alpha }_7}$)	0.0005 (0.0003)	0.0022 (0.0001)∗∗∗	0.0034 (0.0003)∗∗∗	0.0032 (0.0002) ∗∗∗	0.0027 (0.0003)∗∗∗	−0.0017 (0.0003)∗∗∗
Temperature ($\widehat{{\beta }_1})$	−0.0118 (0.0030)∗∗∗	−0.0305 (0.0010)∗∗∗	−0.0396 (0.0022)∗∗∗	−0.0424 (0.0018) ∗∗∗	−0.0456 (0.0029)∗∗∗	−0.0040 (0.0012)∗∗∗
Temperature2 ($\widehat{{\beta }_2})$	0.0004 (0.0001)∗∗∗	0.0007 (0.0000)∗∗∗	0.0009 (0.0000)∗∗∗	0.0008 (0.0000)∗∗∗	0.0011 (0.0001)∗∗∗	0.0002 (0.0000)∗∗∗
Precipitation ($\widehat{{\beta }_3})$	−0.0091 (0.0017)∗∗∗	0.0040 (0.0007)∗∗∗	0.0092 (0.0019)∗∗∗	0.0170 (0.0013)∗∗∗	0.0052 (0.0022)∗∗	−0.0106 (0.0010)∗∗∗
Precipitation2 ($\widehat{{\beta }_4})$	0.0002 (0.0001)∗∗∗	−0.0003 (0.0000)∗∗∗	−0.0005 (0.0001)∗∗∗	−0.0007 (0.0001)∗∗∗	−0.0002 (0.0001)∗∗	0.0002 (0.0000)∗∗∗
Observations	716,113	14,327,839	1,509,352	3,578,302	717,376	2,166,747
R2	0.91	0.87	0.91	0.92	0.95	0.95

Note: Robust standard errors clustered at the district level are in parentheses. Asterisks denote the level of significance: ∗p < 0.10, ∗∗p < 0.05, ∗∗∗p < 0.01.

The coefficient on COVID-19 deaths (α₁) captures fear-driven behavioral changes as local mortality rates signal heightened perceived risk. This fear channel produced divergent responses across the building types. For community amenities, civic/business facilities, industrial services, and religious/cultural facilities, the doubling of local COVID-19 death counts led to substantial reductions in electricity consumption, ranging from −5.7% to −11.8%, with these sectors experiencing the steepest decline. These reductions reflect operational retrenchment: facilities scale back operations, implement remote work arrangements, and minimize on-site activities aimed at limiting contagion risks. Here, the reduction in the core operational energy demand far exceeded any incremental consumption from the enhanced safety protocols for essential staff.

Conversely, for commercial/retail facilities and residential buildings, heightened risk perception drove electricity consumption by 3.1% and 5.8%, respectively. The mechanisms underlying this counterintuitive increase differ by sector. Unlike industrial or business facilities that are converted to work-from-home protocols, commercial/retail facilities do not bind to protocols and thus remain operational to mitigate revenue losses. In response to declining visitor numbers as COVID-19 severity intensified, commercial establishments engaged in what we term “safety signaling” a rational economic response in which businesses invest in visible safety measures, particularly enhanced ventilation and air purification systems, to maintain customer confidence. As documented extensively in the Korean business media during this period, creating an environment perceived as safe became essential for competitive survival.¹⁷

To provide robust, model-free evidence for this “intensity over occupancy” mechanism, we conducted a supplementary analysis for the Seoul metropolitan area comparing our electricity data with Google’s Community Mobility Reports (see Figure S3). The analysis revealed a clear paradox: during the most severe COVID-19 waves, as public mobility to retail and recreational venues declined, electricity consumption in the same buildings significantly increased. This supports our interpretation that facilities ramped up energy-intensive safety measures such as ventilation and climate control to a degree that overwhelmed the effect of reduced foot traffic.

For residential buildings, the increase reflects both extended occupancy as households adhere to confinement measures, driving higher usage of lighting and electronic devices, and prioritized health behaviors, such as increased ventilation despite the increased utility costs. These upward results remain robust across alternative specifications and sample data (see Tables S2–S5).

The temporal dynamics captured by the interaction term (α₂) reveal how these fear-driven responses evolve as the crisis extends. Across all building types, the elasticity of electricity consumption to local mortality diminished over time, which is consistent with the behavioral economics literature on risk habituation. As vaccination rates increased and risk management practices evolved, the heightened responses observed early during the pandemic gradually subsided. This adaptation pattern aligns with recent literature on risk perception and energy consumption behavior, where initial sharp responses to environmental threats gradually attenuate as individuals become desensitized to risk¹⁸ and return to more routine consumption patterns.^19,20 The adaptive responses in Industrial Services and Residential Buildings—the two extremes—are distinct. Industrial Services exhibited rapid adaptation, with the initial negative effect on electricity consumption diminishing by approximately 0.21% monthly, suggesting swift operational normalization despite persistent COVID-19 mortality. In contrast, Residential Buildings showed more persistent changes, with the positive effect declining by only 0.06% monthly, indicating sustained shifts in residential occupancy patterns. This asymmetric adaptation helps explain the observed recovery in aggregate electricity consumption during 2021–2022 following the initial shock in 2020.

Turning to the policy channel, the coefficient on anti-contagion policy (α₃) isolates the impact of government mandates from the confounding effects of fear-driven behaviors. Under stricter policies, community amenities, civic/business facilities, industrial services, and religious/cultural facilities exhibited increased electricity consumption. This result can be explained by compliance with COVID-19 mitigation requirements, including enhanced ventilation, hybrid service delivery systems, and intensified climate control, which often increase the baseline energy intensity. For example, the Korean government mandated specific ventilation frequencies and maximized fresh-air intake, imposing binding constraints on building operations that manifest differently across sectors. These buildings, often already operating with reduced occupancy owing to fear-driven behavioral changes, must nonetheless maintain or even increase Heating, Ventilation, and Air Conditioning (HVAC) operations to achieve regulatory compliance. The resulting increase in electricity consumption, ranging from 0.35% to 0.75%, represents a pure policy effect and the energy cost of maintaining the mandated indoor air quality standards, regardless of actual occupancy levels. This increase in building energy intensity has occurred globally³ as well as in South Korea.²¹ Our sensitivity analyses see Tables S2–S5) confirm that stricter policies can paradoxically raise baseline energy consumption.

Commercial and retail facilities present an intriguing case in which the policy effect is statistically indistinguishable from zero. This null result likely reflects offsetting forces; while ventilation mandates should increase energy consumption, concurrent restrictions on operating hours and customer capacity should reduce the baseline demand. Moreover, many commercial establishments have already adopted intensive ventilation practices as part of their fear-driven safety signaling strategy, suggesting that the policy mandate may have been non-binding for this sector.

The most surprising result emerged for residential buildings, where stricter policies were associated with reduced electricity consumption. As previous research has shown that various types of interventions can influence energy consumption behaviors in residential buildings,²² we suspect that this trend can be attributed to two key factors: restricted social gatherings reducing electricity-intensive activities, and government disaster relief funding encouraging energy-efficient appliance upgrades. Our supplementary analyses supported these two complementary mechanisms.

First, we document an energy efficiency channel: government stimulus payments distributed during high-stringency periods are associated with subsequent reductions in household electricity consumption of 7.0%–8.6%, with larger effects for larger homes (Table S12). This pattern is consistent with households using transfers to invest in energy-efficient appliances—a rational response when facing extended periods of elevated home occupancy, as documented by International Energy Agency (IEA).³ Second, we identify an activity restriction channel through triple-interaction analysis: policy mandates significantly dampen temperature-sensitive electricity consumption during extreme weather events (Table S13). This suggests that gathering restrictions successfully curtail the most energy-intensive social activities that typically occur during temperature extremes when the marginal energy cost of hosting is the highest. This counterintuitive reduction in residential electricity use persists across various empirical specifications, including the alternative treatment of outliers and different fixed-effects structures (see Tables S2–S5).

The policy adaptation dynamics (α₄) reveal differential patterns of adjustment across sectors. While most building types show diminishing policy effects over time as institutions and individuals develop workarounds and optimize their responses to regulations, commercial/retail facilities exhibit no such adaptation. This persistence suggests that policy-driven changes in retail operations, such as reduced hours, capacity constraints, and altered consumer shopping patterns, may represent structural shifts rather than temporary adjustments. The limited recovery in foot traffic following South Korea’s transition to the “With-COVID” strategy in late 2021 reinforces this interpretation, pointing to potentially permanent changes in retail energy demand patterns induced by the pandemic policy response.

Unemployment effects, a proxy for macroeconomic conditions, reveal another layer of heterogeneity in energy consumption patterns. A 1% increase in the unemployment rate reduces electricity consumption in industrial services by 0.004%, which is consistent with reduced industrial activity during economic downturns. However, we observed positive associations between unemployment and electricity consumption in commercial and retail facilities (0.01%), community amenities (0.009%), and residential buildings (0.01%). This pattern suggests a spatial reallocation of energy consumption during economic contractions, possibly reflecting greater reliance on public spaces and community services by displaced workers, along with higher residential occupancy. The positive effect in commercial/retail facilities may indicate countercyclical consumption behaviors, consistent with the “lipstick effect” documented in retail spending during recessions.

The long-term trajectories, captured by the $\widehat{{\alpha }_6}$ and $\widehat{{\alpha }_7}$, reveal divergent secular trends across building categories. Most nonresidential buildings show a U-shaped pattern in electricity consumption over time, with initial declines followed by gradual recovery. This nonlinear pattern is particularly pronounced in civil and business facilities, and industrial services. By contrast, residential buildings exhibit an inverted U-shape, reflecting improvements in energy efficiency and evolving occupancy patterns.

Weather effects demonstrate robust and significant impacts across all building types. The quadratic specification of temperature captures the well-documented U-shaped relationship between energy consumption and temperature^23,24 precipitation plays a significant role, with slightly different responses depending on the building type. These stable determinants of electricity usage underscore the importance of integrating structural and environmental considerations into energy modeling and policy design.

Discussion

The empirical results highlight a complex behavioral response shaped by policy mandates and public fear. Fear-based reactions increased residential and retail energy use, while reducing consumption in the industrial and civic sectors, with regional variation tied to pandemic severity. Policy-driven measures raised energy intensity in nonresidential buildings but unexpectedly reduced residential demand. Temporal dimensions are critical. Our significant interaction terms (COVID_j,t × t and POLICY_t-1 × t) show that the influence of both fear and policy diminished over time, reflecting societal adaptation. This dynamic is particularly relevant to the “post-policy” phase, when formal restrictions were relaxed but viral transmission persisted. The attenuation of consumption elasticity to mortality risk suggests that while voluntary behavioral responses persist, their intensity fades relative to the initial shock. These findings highlight that crisis response policies should be calibrated to sectors, regions, and evolving stages of prolonged disruption.

First, our results underscore the need to address the trade-offs between public health and energy efficiency. While stricter anti-contagion policies likely reduced virus transmission, they also raised the energy intensity in many non-residential buildings. This increase appears to stem from mandated ventilation and intensified climate control, even when occupancy is reduced. Policymakers should consider implementing health measures in a more energy-efficient manner, for example, through technological upgrades or operational strategies that maintain air quality without unnecessarily increasing consumption.

Second, in commercial and retail buildings, electricity use increased owing to fear-driven safety signaling, such as voluntary investments in enhanced ventilation and air purification. Although the direct policy effects were minimal in this sector, many businesses have already adopted these energy-intensive measures. Future policies can support such initiatives through incentive-based programs aimed at improving air quality while managing energy use.

Third, residential buildings exhibited a counterintuitive decline in electricity consumption under stricter policies. This appears to be linked to reduced social activity, disaster relief funding, and increased household attention toward energy efficiency, including appliance upgrades. Policymakers could reinforce these trends through targeted programs, such as appliance-replacement vouchers or weatherization support, particularly once basic safety standards are established.

Finally, the distributional dimension of energy behavior deserves closer attention. Our findings suggest that reductions in residential energy use are more evident in larger homes, which may reflect a greater ability to adopt energy-efficient technologies. This pattern raises the possibility that lower-income households may face greater obstacles to adopting energy-efficient technologies. To foster more inclusive resilience, governments should consider tiered subsidies, targeted vouchers, or other forms of assistance to help lower-income households manage their energy needs during future crises.

As COVID-19 continues to resurface in various parts of the world, understanding how people and institutions adapt to such disruptions remains highly relevant. Our results provide empirical insight into how policy and fear shape energy demand during crises and inform efforts to manage energy behavior during future disruptions, such as public health emergencies, environmental disasters, or geopolitical shocks. This study also highlighted the value of high-resolution data in supporting responsive and effective energy governance.

Limitations of the study

Despite the strengths of this study, including its large-scale and sectorally disaggregated dataset covering more than one million buildings, several limitations should be acknowledged. First, our analysis relied on the COVID-19 mortality rate as a proxy for public fear. While this indicator is grounded in the behavioral research literature, it may not fully capture subjective variations in risk perceptions. Future research could combine such indicators with longitudinal survey data to measure perceived risk and voluntary behavioral change more directly.

Second, although our study successfully identified and quantified the high-level behavioral drivers of energy shifts, our dataset did not include submetered end-use information. Consequently, the observed consumption shifts cannot be precisely attributed to specific systems, such as HVAC, lighting, or digital equipment. Incorporating facility-level energy audits in future studies would help decompose aggregate changes and improve the understanding of how energy behavior evolves within buildings during crises.

Finally, while South Korea provides a compelling context owing to its centralized policy regime, consistent implementation of anti-contagion measures, and high-quality data infrastructure, the generalizability of our findings to countries with more decentralized or fragmented governance remains uncertain. Comparative cross-country analyses would, therefore, be valuable to test the extent to which fear- and policy-driven responses are context-dependent. Addressing these limitations will advance our understanding of how energy systems and human behavior adapt to future large-scale disruptions.

Resource availability

Lead contact

Request for further information and resources should be directed to and fulfilled by the lead contact, Minjung Kim (minjungkim@inha.ac.kr).

Materials availability

This study did not generate new unique physical materials.

Data and code availability

• •The de-identified analysis dataset has been deposited in a public repository (https://github.com/jonghyunyoo/COVID19-in-South-Korea).
• •Analysis-generation code has been deposited in the same repository at publication (https://github.com/jonghyunyoo/COVID19-in-South-Korea).
• •All other information is available in the paper and supplemental information. Access requests are answered within a reasonable time frame.

Acknowledgments

This work was supported by the Korea Environment Industry & Technology Institute (KEITI) through the Climate Change R&D Project for New Climate Regime funded by the Korea Ministry of Environment (MOE) (RS-2023-00218794).

Author contributions

J.Y., writing – review and editing, writing – original draft, visualization, formal analysis, methodology, conceptualization, and funding acquisition; D.K., writing – original draft, data curation, and investigation; M.K., writing – review and editing, writing – original draft, conceptualization, and project administration.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
De-identified analysis dataset supporting results	This paper	GitHub: https://github.com/jonghyunyoo/COVID19-in-South-Korea
Building-level electricity consumption administrative microdata (monthly)	Ministry of Land, Infrastructure and Transport (MOLIT), Open Architectural Data System (EAIS)	https://open.eais.go.kr/
National Building Registry (attributes)	Ministry of Land, Infrastructure and Transport (MOLIT), Open Architectural Data System (EAIS)	https://open.eais.go.kr/
National/regional COVID-19 cases and deaths (daily → monthly aggregates)	KDCA (Korea Disease Control and Prevention Agency)	https://www.kdca.go.kr/index.es?sid=a3
Anti-contagion policy intensity/index (daily → monthly aggregation)	Ha et al.25	Article DOI/URL (above)
Regional unemployment rate (monthly)	Statistics Korea (Economically Active Population Survey/KOSIS)	https://kosis.kr/statHtml/statHtml.do?orgId=101&tblId=DT_1DA7104S&conn_path=I3
Global COVID-19 deaths (daily → monthly)	Our World in Data (WHO source)	https://ourworldindata.org/grapher/total-daily-covid-deaths
Weather: monthly regional temperature and precipitation	Korea Meteorological Administration (KMA) Weather Data Service	https://data.kma.go.kr/stcs/grnd/grndRnList.do
Software and algorithms
R (v4.3.2)	The R Foundation/CRAN	https://www.r-project.org
Analysis code for data cleaning and econometric models	This paper	GitHub: https://github.com/jonghyunyoo/COVID19-in-South-Korea (same as above)

Method details

Overview and study design

We quantify how building-level electricity consumption in South Korea responded to COVID-19 and national anti-contagion policies. The unit of observation is building–month. We estimate linear panel models with fixed effects and instrument regional COVID-19 deaths using contemporaneous global deaths. To capture purposeful heterogeneity by energy-use type, all models are estimated separately for six building-use categories.

Data sources and preprocessing

Building electricity. We analyze monthly electricity consumption from May 2020–April 2022 for roughly one million parcels/buildings across 16 regions (Si–Do), excluding Jeju, yielding >23 million building–month observations. Electricity is normalized by floor area (kWh per m²). The source provides full addresses, floor area, and building-use codes; the original 54 codes are aggregated to six categories for analysis (mapping in supplemental information).

Quality control. We trim outliers below the 1st and above the 99th percentile of electricity intensity and remove observations with missing values or duplicates. Robustness checks show results are not sensitive to this choice using 5th–95th trims and an IQR rule (Tables S4 and S5).

Pandemic measures

Regional mortality (endogenous regressor). Daily COVID-19 deaths (and cases) by Si–Do are drawn from official sources and aggregated to monthly counts. We use deaths—not cases—to proxy public risk perception because deaths are recorded more reliably and are less susceptible to underreporting; and we use monthly (not cumulative) counts to reflect the contemporaneous severity that drives behavior.

Policy intensity (exogenous control). We compile a national anti-contagion policy index from official directives (e.g., distancing levels, limits on private gatherings). Daily levels are normalized to a 0–3 scale and summed to monthly values; we use a one-month lag (POLICY_{t−1}) to reflect adjustment lags in energy use. Compared with the Oxford Stringency Index, this bespoke index better targets Korea-specific policy levers affecting social/economic activity; the two measures are moderately correlated (Table S11).

Socioeconomic and climate covariates

• •Unemployment: Monthly regional unemployment rates from Statistics Korea’s Economically Active Population Survey (Si–Do level).
• •Weather: Monthly average temperature and precipitation from the Korea Meteorological Administration. Each enters with a quadratic term to capture non-linear demand responses (e.g., cooling/heating loads; humidity effects).

Empirical specification

For building i in region j and month t (where t=0,1,…,23), the main specification is:

$${ELEC}_{i,j,t}={\alpha }_1\widehat{{COVID}_{j,t}}+{\alpha }_2\left(\widehat{{COVID}_{j,t}}\times t\right)+{\alpha }_{3}POLIC{Y}_{t-1}+{\alpha }_4\left(POLIC{Y}_{t-1}\times t\right)+{\alpha }_{5}U{R}_{j,t}+{\alpha }_{6}t+{\alpha }_7{t}^2+\beta f\left({WTHR}_{j,t}\right)+Bld{g}_i+MONT{H}_m+{\epsilon }_{i,j,t}$$ (Equation 1)

where the dependent variable ELEC_i,j,t represents the logarithmic value of electricity consumption per square meter for building i located in region j over month t. We normalize electricity consumption by the floor area of each building to account for differences in building size, assuming that electricity use increases linearly with building area, all else being equal.

We include two key variables to capture the pandemic’s effects: the logarithmic value of regional COVID-19 death numbers COVID_j,t and the national anti-contagion policy POLICY_t. Like Goolsbee & Syverson,⁷ we distinguish between policy-induced shifts (e.g., mandatory ventilation) and fear-based behavioral changes (e.g., voluntary distancing). These channels cannot be perfectly separated in practice but capturing them in separate variables clarifies how each uniquely shapes building energy consumption.

We use monthly COVID-19 death counts (deaths within each specific month) rather than cumulative totals, as monthly mortality provides a more reliable measure of current pandemic severity influencing public behavior and energy consumption decisions. COVID-19 death counts are chosen over positive case numbers because deaths are more likely to influence public fear, which directly impacts behavioral changes such as energy consumption. Furthermore, case data is prone to underreporting and measurement errors, while death data is more reliable as it is recorded directly by hospitals and reported to the government. Because local COVID-19 fatality counts may be correlated with unobserved regional factors, we adopt a two-stage least squares (2SLS) procedure. The first stage regresses regional COVID-19 deaths COVID_j,t on global deaths and other controls, generating a predicted value that serves as an instrument in the second stage $\widehat{{COVID}_{j,t}}$. This approach helps address potential endogeneity in our core explanatory variable (described below). Because we use the log of COVID-19 deaths, the coefficient α₁ represents the elasticity of electricity consumption with respect to COVID-19 deaths in the initial period.

Policy interventions intended to reduce virus transmission—such as social distancing mandates, quarantine protocols, and remote-working policies—also shape energy usage patterns. To maintain temporal consistency, we normalized South Korea’s official anti-contagion policy levels to a standardized 0-3 scale, which we aggregate to a monthly level by summing daily policy scales. Table S10 details this conversion for full reproducibility. We use a time-lagged policy variable POLICY_t-1 to account for the time required for adjustments in energy consumption behavior in response to policy changes as found in many studies.¹⁶ Although the severity of policies may correlate with COVID-19 deaths, this relationship is not necessarily linear, especially during the “With-COVID-19” phase, when deaths surged while policies were relaxed. We used the national policy measure rather than regional ones because anti-contagion policies in South Korea were largely centralized, with little to no regional variation.

Over time, public responses to both COVID-19 and the associated policies likely evolved. Early in the pandemic, heightened uncertainty and precaution reduced social activities and influenced electricity consumption. As society adapted, electricity usage patterns partially recovered even amid high infection rates. Similarly, strict compliance with newly introduced social distancing mandates relaxed over time. To account for this temporal heterogeneity, we include interaction terms, $\widehat{{COVID}_{j,t}}\times t$ and POLICY_t-1×t. Time t is assigned continuous values from 0 to 23, covering a total analysis period of 24 months. The instantaneous semi-elasticities at month t are (α₁+α₁t) for log deaths and (α₃+α₄t) for policy. A negative α₂ (or α₄) implies attenuation over time, while a positive value implies amplification.

We include the regional unemployment rate UR_jt to control for local economic conditions, as these can significantly influence electricity demand (for example, in industrial facilities. Among various economic indicators, we rely on regional unemployment rates as our primary economic indicator because they are officially reported at the “Si-Do” and monthly levels. Alternative measures (e.g., industrial production) either lack similar spatiotemporal resolution or proved insignificant in our tests, reinforcing unemployment as the most viable macro control. We also incorporate a quadratic time trend, α₆t+α₇t², to capture shifts in electricity usage unrelated to the pandemic. Such trends might stem from changes in the prevalence of electrical appliances, improvements in building energy efficiency, or evolving societal energy consumption behavior.

Building electricity use is also affected by weather.^21,23 To consider these factors, we included two weather variables—temperature and precipitation—each with a quadratic term: $\beta f\left({WTHR}_{j,t}\right)={\beta }_{1}TEM{P}_{j,t}+{\beta }_{2}TEM{P}_{j,t}^2+{\beta }_3{PRCP}_{j,t}+{\beta }_4{PRCP}_{j,t}^2$. The inverted-U relationship between temperature and electricity use is well-documented, and precipitation levels influence air-conditioning usage due to humidity.

We include two sets of fixed effects. Building-level fixed effects Bldg_i control for time-invariant building characteristics (e.g., construction materials, building height, or age), while MONTH_m month-of-year fixed effects (12 dummies) capturing seasonality (such as increased energy use in summer and winter). A robustness test in the Supplementary Information considers regional fixed effects instead of the building-level one and finds robust insights.

To capture the heterogeneous impact of energy use purpose, we run the regression model separately for each of the six building types. This means we split the data based on building types and perform the regression for each category individually. With sufficient observations for each building type, it is possible to estimate the effect for each one. Thus, the empirical results are provided for each type.

Instrument-variable strategy

To address the potential endogeneity of local COVID-19 death counts (COVID_j,t)—which may be influenced by local behaviors and policies that also affect energy use—we employ a two-stage least squares (2SLS) estimator. We instrument for COVID_j,t using the contemporaneous logarithmic global COVID-19 death count (GCOVID_t). The validity of an instrumental variable rests on two conditions: relevance and exogeneity. Our instrument is highly relevant, as shown by the first-stage instrument F-statistics comfortably exceed conventional weak-IV thresholds in all building types; exact values are reported in the replication logs, reflecting the predictable epidemiological pathway where global pandemic waves predict same-month mortality trends in South Korea. The second condition, the exclusion restriction, requires that GCOVID_t affects local electricity consumption only through its effect on local deaths. We argue this condition likely holds due to a crucial timing mismatch between fast-moving epidemiological channels and slower-moving macroeconomic channels. Potential confounding factors, such as disruptions to global supply chains, typically impact domestic economic activity with a time lag of several weeks or months. Our use of a same-month instrument leverages this mismatch, as it captures the immediate epidemiological shock which is plausibly exogenous to unobserved, lagged economic factors. This assumption is further supported by South Korea’s institutional context of stable, government-regulated electricity prices. To empirically test the validity of this timing assumption, we conducted two key robustness checks (see Tables S8 and S9).

Thus, our first-stage regression is specified as:

$${COVID}_{j,t}={\gamma }_{1}GCOVI{D}_t+{\gamma }_{2}POLIC{Y}_{t-1}+{\gamma }_3\left(POLIC{Y}_{t-1}\times t\right)+{\gamma }_{4}U{R}_{j,t}+{\gamma }_{5}t+{\gamma }_6{t}^2+\delta f\left({WTHR}_{j,t}\right)+MONT{H}_m+RE{G}_j+u_{jt}$$ (Equation 2)

where we include a regional fixed effect, REG_j, rather than the building-level fixed effect used in our main specification because the endogenous regressor varies only at the regional level and building fixed effects are not defined for a region-level outcome. In the second stage, we regress building-month electricity use on the fitted regional fatalities (and their interaction with time), including building fixed effects and month-of-year fixed effects to absorb time-invariant building heterogeneity and common seasonal patterns, with standard errors clustered by “Dong.”

The estimated coefficient of global fatalities $\widehat{{\gamma }_1}$ is statistically significant and ranges between 0.8 and 1.0 across all building types, indicating that global COVID-19 deaths are strong predictors of local COVID-19 deaths in South Korea. This elasticity close to unity suggests that global pandemic trends substantially influenced local mortality patterns. The first-stage F-statistics are large when clustering at the granular level (“Dong”), which we use to maintain consistency between the first and second stages. For robustness, we also estimate the model with clustering at the regional level (“Si-Do”). Due to computational constraints with our large sample size, we are unable to conduct other weak instrument tests including Stock-Yogo, though our F-statistics exceed conventional thresholds for instrument strength.

Robustness and sensitivity

Results are robust to alternative fixed-effects structures (e.g., replacing building FE with region FE plus richer time × type effects) and alternative weather constructions (degree-days), and remain stable under outlier handling using 5–95% trims and the IQR rule; using cases (instead of deaths) as a descriptive comparator does not change our conclusions. IV diagnostics indicate strong relevance and support exogeneity via a future-dated placebo instrument and region-specific trend checks. Our policy index tracks the Oxford Stringency Index on a month-by-month basis while retaining Korea-specific fidelity. Mechanism-oriented checks show post-stimulus residential reductions and policy-moderated temperature sensitivity. Leave-one-region-out confirms no single region drives results. Full details and estimates appear in the SI.

Software and reproducibility

Analyses were conducted in R 4.3.2 (Windows 10). Scripts are organized with a README. See the public repository for code, data, and a step-by-step reproduction guide: https://github.com/jonghyunyoo/COVID19-in-South-Korea.

Quantification and statistical analysis

• •Estimation: Linear panel regressions with building and month fixed effects; 2SLS for IV models; category-wise estimation by building-use type.
• •Uncertainty: Cluster-robust standard errors at the Dong level in the second stage. Coefficients are reported with 95% CIs and p-values.
• •Sample size: The unit is building–month; exact n for each category/specification is reported with the corresponding results (main text/SI).
• •Model selection / multiplicity: Baseline and category-wise models were pre-specified; robustness checks are reported transparently in the SI (no post-hoc pruning).
• •Diagnostics & assumptions: Instrument relevance; residual diagnostics; leverage checks; leave-one-region-out sensitivity; exogeneity probes via timing and placebo tests.
• •Significance: Two-sided tests with α=0.05 unless noted; all computations use double precision.

Published: December 18, 2025