Energy Burden in the United States: An Analysis Using Decision Trees

Abstract

The concept of energy burden (EB) continues to gain prominence in energy and associated policy research as energy prices rise and electricity and heating options diversify. This research offers a deeper understanding of EB dynamics and how EB can be addressed more effectively by discerning the interplay between regional environmental, social, and economic factors. Using decision trees (DTs), a powerful machine learning technique, we explore the multifaceted dynamics that shape EB across the United States (U.S.) by examining how factors like housing quality, demographic variations, access to energy sources, and regional economic conditions interact, creating distinct EB profiles across communities. Following a comprehensive review of existing literature and DT analysis, we map the results to identify the most significant factors influencing EB. We find that no single variable has a determinant effect on EB levels. While there is no uniform regional pattern, regions with higher population density exhibit a stronger correlation between EB and socioeconomic and other demographic factors such as educational attainment levels and racial segregation. Our findings underscore the significance of regional ecologies in shaping EB, revealing how localized environmental and economic contexts amplify or mitigate systemic inequities. Specifically, our analysis reveals significant regional disparities, highlighting the need for localized policies and interventions. We find that a one-size-fits-all approach is insufficient and that targeted, place-based strategies are necessary to address the specific needs of different communities. Policy interventions should prioritize energy democracy, address systemic inequities, and ensure universal energy access through participatory planning, financial assistance, and targeted initiatives such as housing rehabilitation, energy efficiency improvements, and incentives for underrepresented communities.

1. Introduction

The concept of energy burden refers to the proportion of household income spent on energy costs, encompassing both electricity and heating [1]. It is a critical metric in assessing energy affordability and economic vulnerability among different populations.

Energy burden is a concept that gained prominence in energy research. It encapsulates the economic strain experienced by households in meeting their energy needs. This article delves into the significance of different variables in predicting areas of energy burden. The article explores the multifaceted dynamics that shape energy burden and examines how factors like housing quality, income levels, demographic variations, access to energy sources, and regional economic conditions interact, creating distinct energy profiles across communities.

We examine these phenomena through the concept of regional ecologies—or the idea that a unique blend of environmental, social, and economic conditions that characterize different areas plays a significant role in influencing the degree and nature of energy burdens. By exploring regional ecologies, we shed light on localized factors that contribute to variations in energy affordability and help identify strategies that can address these disparities. Traditional, broad-based policies often fail to capture the intricate dynamics of energy burden and can generate interventions ill-suited to realities on the ground. Localized, data-driven interventions better consider the specific needs of communities. For example, urban and rural areas frequently face distinct challenges in energy access and affordability. Factors such as infrastructure development, energy sources, housing conditions, economic opportunities, and the demographic makeup of communities influence these aspects. By acknowledging these differences, we can develop nuanced policies that more effectively target the root causes of energy burden. Addressing energy burden disparities is vital for alleviating energy poverty, promoting energy equity, and improving the health, economic stability, and overall well-being of disadvantaged communities.

Regional ecologies—the unique blend of environmental, social, and economic conditions that characterize different areas—play a significant role in influencing the degree and nature of energy burdens. By exploring these regional ecologies, we gain a deeper understanding of how localized factors contribute to variations in energy affordability and identify strategies that can be more finely tuned to address these disparities. To better understand these dynamics, this study employs decision trees (DTs), a powerful machine learning technique, as a primary analytical tool. By using DTs, the article seeks to uncover not only the direct contributors to energy burdens but also the complex web of interrelated factors that shape the energy landscape. Through this approach, we aim to bridge the gap in the existing literature, which often overlooks the localized and context-specific factors that influence energy burdens. Filling the gap in understanding the role of energy justice within the scope of granular local conditions will offer a comprehensive view of energy burdens and contribute to the development of more targeted and equitable solutions. By leveraging a detailed, geographically specific dataset that integrates multidisciplinary factors, this study enhances the precision of local energy burden analysis.

A key motivation for this research is the recognition that energy burdens are deeply influenced by context-specific factors, which are often overlooked in favor of broader, more generalized approaches. Our goal is to advance a more comprehensive understanding of energy burdens. In doing so, we aim to inform policy decisions that promote equitable energy solutions and ultimately contribute to reducing the disparities that define the energy experiences of diverse communities. Studying the complex interplay of factors, including housing quality, socioeconomic status, demographics, and geographic location, shows that no single variable fully accounts for energy burden levels. Furthermore, regional ecologies indicate that the combination of factors involved varies significantly across different regions and between rural and urban areas. The DT analysis reveals specific profiles where variables like older housing stock and the percentage of minority populations play a crucial role in determining energy burdens. As such, broad, one-size-fits-all policies are insufficient to address the diverse challenges faced by different communities. No prior study has identified energy burden profiles and mapped the geographical distribution of energy burdens and their primary influencing factors.

The paper is organized into several key sections that systematically address the complexities of energy burden and its determinants. We begin with a comprehensive review of the literature, highlighting existing research on energy burden, its impact on households, and the gaps in current policy frameworks. The methodology section then details the use of DT analysis as a tool to identify the most influential factors contributing to energy burden. In the results section, we present our findings and map the results. This section focuses on how localized variables shape energy burden profiles in different regions. This is followed by a discussion that interprets the results, emphasizing the importance of regional ecologies and the need for tailored policy responses. Finally, we conclude with a set of policy recommendations based on our analysis, suggesting targeted interventions at both the local and state levels to reduce energy burden and promote a more equitable energy landscape. Our findings underscore the need for targeted, place-based interventions that consider the unique conditions of each area, focusing on tailored solutions such as subsidies for energy-efficient upgrades in older homes or financial assistance in communities with higher economic vulnerability. We conclude with recommendations for how policymakers can localize analysis and design more effective strategies that not only reduce energy burden but also promote equity in energy access and affordability.

2. Literature Review

2.1. Energy Burden as an Emerging Concept

The problem of energy costs has long been a concern in the United States (US), especially for low-income families and marginalized groups. The term “energy burden” rose to prominence in the 20th century as studies began looking into how energy expenses were not evenly distributed among different social classes. Research conducted by Hirst and Brown (1990) and Low (2016) shed light on the disproportionate burden low-income households face, attributing it to inefficient housing, lack of access to energy-saving technologies, and stagnant incomes [2,3]. Initial attempts at measuring the energy burden primarily concentrated on calculations like the proportion of household income allocated toward energy expenses, which offered valuable initial perspectives on the extent of energy-related hardship endured by these households. These measurements were not advanced enough and did not fully grasp the complex nature of energy challenges [3]. While Blumstein et al.’s (1980) work began to highlight institutional influences on energy challenges, such as conflicting motivations and regulatory issues, through detailed research and interviews, they did not adopt a structured method to pinpoint the specific groups facing energy challenges [4]. Higgins and Lutzenheiser (1995), in their research study, acknowledged the shortcomings in distribution efficiency for aiding low-income households under the Low-Income Weatherization Program and Low-Income Home Energy Assistance Program to effectively reduce their energy cost burden [5]. Jones and Reyes (2023) highlight how these complex challenges impede energy projects’ ability to holistically understand and measure energy burdens, which limits more sustainable, equitable, just, and regionally appropriate energy policies to effectively meet the needs of communities [6]. However, due to the growing recognition of energy burden as a salient issue in both Europe and the US, there has been a concerted effort to research past normalized statistical metrics and study a variety of related factors, such as low-income household energy behaviors, for a more complete analysis of energy burden across communities [6]. Much of the previous literature surrounding this energy burden research can be organized into thematic categories, including energy justice, energy efficiency policy, energy assistance programming, renewable energy transitions, energy usage, and energy poverty evaluations [6]. Over the last several decades in the US, various factors have influenced the challenges people face regarding energy costs. The increasing expenses linked to energy usage, worsened by changes in fuel prices and the effects of climate change, have led to an energy burden, especially affecting disadvantaged communities [7]. The transition from fossil fuels to embracing low-carbon energy sources also brings concerns regarding the distribution of advantages and disadvantages across different communities, as evidenced by challenges such as job cuts and economic downturns in regions prioritizing justice initiatives. Additionally, some low-income households may face challenges in accessing opportunities related to clean energy transitions, like vehicles [8]. Between 2006 and 2012 in Germany, energy transitions resulted in a twofold rise in electricity costs for households. This trend is projected to persist as sustainable energy solutions are embraced, putting additional strain, particularly on low-income families [9].

Additionally, the aging infrastructure and inadequate housing stock in many urban areas have led to higher energy consumption and an increased burden on residents [10]. From a social justice standpoint, the energy burden reflects broader disparities in access to resources and opportunities, disproportionately affecting communities of color and Indigenous populations [11].

Environmental justice research has emphasized the interconnections between energy poverty and other aspects of socioeconomic marginalization, stressing the importance of holistic solutions that tackle systemic inequalities [12]. This further strengthens the argument that sustainable energy development hinges on more inclusive energy policy regimes [13].

2.2. Measurement Challenges and Advancements

Measuring the energy burden on households can be challenging due to the ways people use energy and fluctuations in energy prices based on socioeconomic factors, as well as the availability of data in different regions. The lack of an agreed-upon definition and approach to assessing energy burden can lead to inconsistencies among researchers, making it difficult to compare findings. Additionally, data-limited areas, such as low-income neighborhoods, pose obstacles to accurately measuring and analyzing energy usage [14]. Evaluating energy burden using a single or primary metric like income level risks overlooking more nuanced variables, such as people’s actual behavior patterns [15]. Furthermore, variations in types of available housing, weather patterns, and social support systems across areas mean there is no effective one-size-fits-all method. This necessitates tailored measurement strategies that are site-specific. These obstacles highlight the importance of techniques and cooperation across fields to advance research on energy burden.

2.2.1. Metrics

To address these challenges, researchers have created more sophisticated metrics to capture the subtleties of energy strain and potential solutions more effectively. These metrics take into account aspects such as household income levels, costs related to energy, housing attributes, and regional geographic differences. As an example, the Multidimensional Energy Poverty Index (MEPI), introduced by Sokolowski et al. (2020), encompasses aspects of energy poverty, including earnings, housing standards, and availability of energy facilities [16]. The energy equity gap introduced by Cong et al. (2022) is a measurement that captures complementary energy-limiting behavior to identify inequities [15]. New research from Yang et al. (2024) used a Bidirectional Flow (BDF)—New Energy Microgrid (NEM) framework—to promote equitable energy dispatch optimization in rural areas [17]. Additionally, the European Commission has established approaches for assessing energy poverty by considering income levels and expenses related to energy usage alongside factors affecting energy efficiency [18].

2.2.2. Data Advancements

Advancements in technology and improved data collection methods have strengthened the measurement of energy burden. The more widespread use of smart meters and increased energy usage data availability has allowed researchers to analyze energy consumption patterns more granularly [19]. Geographic Information Systems (GISs) also provide detailed spatial analysis of energy burden across regions, which is essential to highlight disparities and contribute to more informed policy interventions [20].

As a result of the multifaceted factors that contribute to the energy burden highlighted, researchers have also developed indexes for more comprehensive measurement, such as the Multidimensional Energy Poverty Index (MEPI). This index provides an approach to evaluating energy scarcity through factors such as income status, living conditions, and accessibility to energy amenities [16]. Likewise, the European Commission has set up methods for assessing energy scarcity, facilitating cross-national comparisons and energy policy evaluation. Machine learning techniques like logistic regression have demonstrated an opportunity to potentially better handle unbalanced data and uncover the impacts of multiple socioeconomic factors on energy burden [21]. To further improve this opportunity and more cohesively capture all factors impacting a specific community or region’s energy burden, using predictive machine learning tools such as DTs that can utilize multiple regression models has promise to help better forecast energy burden for effective policy interventions.

2.3. Energy Burden and Associated Factors

Studies conducted recently have uncovered factors to take into account when evaluating energy burden levels. For example, variables such as lifestyle, household traits, and climate change impacts influence home energy characteristics [22]. This stresses gaps in capacity and the need for more adaptive strategies that mitigate vulnerabilities.

2.3.1. Weather Conditions

The speed of the wind plays a role in determining energy burdens as it affects the generation of renewable energy, specifically wind power, and also impacts the heating and cooling requirements of buildings in different areas. Therefore, regions with high average wind speeds can harness wind power to potentially reduce energy costs and lower energy scarcity for residents. Additionally, wind power is a cost-effective alternative energy option that can reduce reliance on fossil fuels and, depending on availability, stabilize and reduce electricity prices [23]. Regions with wind resources usually attract investments in wind power facilities, which offer a sustainable energy option and bring about economic advantages for nearby towns, like job opportunities and income from locally operated wind farms. The presence of wind power can also help maintain energy costs and act as a safeguard against the unpredictability of fossil fuel markets by assisting in reducing pressure on energy supply [24]. Increased wind speeds can also impact the thermal comfort and energy usage of buildings [25]. During colder months, elevated wind speeds can result in increased heat loss from buildings, requiring more energy for heating purposes [25]. Meanwhile, during warmer months, wind can help cool down buildings, which may lessen the need for air conditioning [26,27]. Therefore, wind speed can indirectly affect the energy burden by changing the demands for heating and cooling energy consumption. Understanding these factors is essential for identifying opportunities to develop targeted approaches to reduce energy burden, particularly in regions with higher-than-average wind speeds [28,29].

2.3.2. Built and Digital Infrastructure

Other factors that influence the energy burden are the age of the housing stock and infrastructure, as well as technological access. The relationship between incomplete plumbing systems in the United States and energy burden shows how outdated systems (leaky faucets, malfunctioning toilets, aging pipes, etc.) lead to wasted water, often heated for domestic use. Minor leaks in plumbing systems can significantly raise the amount of water used [30]. For low-income homes, this may result in increased energy requirements for hot water and create financial difficulties. Moreover, the 2024 American Housing Survey suggests a connection between income levels and the higher occurrence of inefficient plumbing problems [31].

Climate change is increasing the frequency and severity of extreme weather events, revealing systemic inequities in energy distribution systems [32]. Weather catastrophes disparately impact low-income, racially marginalized, and rural communities with increased power outages because they have power lines with increased failures or are far away from grid power sources [33]. To address the threat of climate change, the US and countries around the world are focusing on the significant challenge of how to decarbonize their economies [34]. There will be massive electrification initiatives in buildings and transportation, such as the promotion of the broad use of electric vehicles (EVs), which, without sufficient EV infrastructure, areas cannot access their economic and environmental benefits [33]. Such efforts are predicted to increase energy consumption and demand across the US [35]. Catalyzing decarbonization efforts will rely not only on widely adopting renewable energy but also on updating electric grid infrastructure [35]. New smart grid (SG) technologies may worsen inequities if they only focus on optimization and do not incorporate social dimensions into their energy systems [33]. Furthermore, if modernizing power restoration practices after outages and grid infrastructure do not consider factors connected to equity, vulnerable communities risk taking on a greater energy burden [36].

Internet burden measures the percentage of income a household spends on its internet costs. The divide between those who have access to the internet and technology and those who do not is often described as a division between rural and urban locations, but it is also emerging as a division along socioeconomic lines. As many as 42 million Americans may lack broadband access, and rural and tribal communities experience the lowest rates of broadband access; 97.8 percent of urban areas have access to 100/10 Mbps internet, compared to only 66.8 percent of rural areas and 63.7 percent of tribal lands [37]. Rural energy burden is also disproportionately high; the share of annual income spent on energy, or energy burden, is 4.4 percent for rural areas, compared to 3.3 percent for urban areas. This disparity is higher for rural low-income households, where the energy burden is three times greater than that for higher-income rural households. Reasons for higher rural energy burdens can include old homes, energy-inefficient heating systems, economic hardship, and a lack of financial assistance and energy efficiency programs in rural areas. Smart energy devices could alleviate energy costs through the management of energy use, but homes without broadband cannot access such technologies [37].

The interconnection between energy and internet infrastructures is becoming more entwined with the rise in smart grid technologies, smart buildings, smart transportation, and home energy management systems. While many households in the US have access to internet and energy services, vulnerable households, such as low-income households, senior citizens, people of color, single women, and renters, are more likely to suffer from issues of energy burden and energy insecurity. Similarly, given the impact of income on internet and computer access, low-income households are more likely to experience internet burdens, and the inequalities have been exacerbated by inadequate energy and internet infrastructure to support shifts brought on by COVID-19 [37].

2.3.3. Socioeconomic Conditions

Other socioeconomic factors also contribute to energy burden, and recent studies have shown significant findings in energy burden trends, drivers, and household impacts. For example, when examining the intersection of energy poverty and gender inequality, an emphasis on women’s disproportionate burden in accessing clean and affordable energy services is prevalent [38]. New methods are needed to include environmental factors in energy hardship measurements for a more comprehensive view of energy poverty and its causes [39]. Additionally, when empirically analyzing energy insecurity in the US, the research reveals the role of race and housing conditions in exacerbating the energy burden [40]. Furthermore, there is a link between education level and energy burden, where having less than a college degree is a significant predictor of energy burden due to a higher likelihood of financial hardship and social vulnerability [41]. There is a statistically significant relationship between having less than a college degree and energy burden [42]. There is also a disparity in energy burden between those with a college degree or higher and those with a high school degree or less, and residents with a high school degree or less have a higher overall energy burden [43]. Furthermore, as decarbonization efforts grow and more efficient technologies like EVs and decentralized Home Energy Management Systems (HEMSs) are adopted, already vulnerable communities such as minorities and renters lack the financial resources and flexibility to engage in them, amplifying their energy burden [44].

2.3.4. Racial Demographics

Race is another variable deeply embedded in energy burdens through complex interactions of socioeconomic status, housing quality, and systemic inequalities, affecting households of different types in distinct ways. We hypothesize that such disparities are evident in differential access to energy-efficient technologies, housing conditions, and economic resources. Socioeconomic conditions magnify these issues for vulnerable households. For example, lower-income families from underrepresented groups often live in homes that lack proper energy efficiency features [45]. These further burden them with higher energy costs compared to their income [45]. These economic limitations hinder their capacity to adopt energy-saving behaviors and perpetuate the cycle of energy instability. Research in 2016 and a subsequent 2017 study of energy burden demonstrates the correlation between race and socioeconomic status concerning heating usage and efficiency within urban settings such as Detroit [46]. The results from these two studies reveal how underserved groups in the US encounter disproportionately higher energy costs due to disparities in environmental conditions and infrastructure, thereby elevating the spatial dimensions of energy justice.

Public funding in energy research and development innovations is critical to promoting a just low-carbon energy transition [47]. Increased investment in resilient and inclusive energy systems holds promise to advance more equitable clean energy access, which can improve social well-being [47]. Despite recent historic US federal funding policies such as the Inflation Reduction Act and Bipartisan Infrastructure Act seeking to address energy affordability in tandem with equity efforts like the Justice 40 initiative that aim to ensure significant federal benefit investments in disadvantaged communities alongside existing energy assistance programs, these combined efforts are still insufficient in meaningfully addressing energy burdens for every vulnerable household [6]. Therefore, we aim to explore energy burden patterns in the US with households as our unit of analysis, drawing out additional implications for a more targeted approach to address energy burden disparities across multiple regions. Improving the energy efficiency of households is an effective way to lower utility bills and decrease energy burden; however, many residents who experience energy burden also rent, and landlords are often not incentivized to make upgrades to housing stock that would reduce costs for renters. This paper seeks to highlight the disparities in energy burden across rural and urban regions, including subregional and regional patterns, as well as to investigate the intersection of multiple demographic, socioeconomic, and geographic variables to evaluate energy burden using DTs. We expect to contribute to an understanding of some of the factors that predict energy burden, allowing us to propose targeted policy interventions to alleviate energy burden in the most vulnerable communities.

3. Materials and Methods—Decision Tree (DT) Concepts and Principles

DTs serve as a methodology for constructing classification and regression models based on tree structures. As predictive and data mining tools, DTs enable the forecasting of future events and the identification of trends and profiles by traversing the created tree paths. The versatility of DTs extends from handling simple to intricate multi-stage problem structuring [48].

The DT consists of nodes that progress downward, creating a rooted tree structure. These nodes originate at the root node, which is characterized by having no incoming edges. Nodes may or may not have outgoing edges; if they extend the growth of the tree, they are referred to as internal or test nodes. Conversely, if a node lacks an outgoing edge, it is termed a leaf node, also known as a terminal or decision node [48].

3.1. Data Collection

The dataset employed to build the DT was self-compiled, consisting of 19,244 observations. Each observation corresponds to a distinct location across the USA, encompassing its own dependent and independent variables. The dataset is a compilation of variables from diverse sources, carefully integrated to create a unified dataset that spans multiple fields of knowledge. The compilation of variables follows deliberate selection criteria, focusing on factors that characterize either the household or the place, thereby providing a rich foundation for conducting in-depth decision tree analysis. By combining variables from various domains, this dataset enables a nuanced understanding of the complex relationships between energy burden, socioeconomic factors, demographic characteristics, and other relevant variables. The dataset provides uniform coverage across the United States, with no evidence of geographical bias. All observations utilized in this analysis pertain to the year 2019. The variables have diverse sources, and a directory detailing them is provided in Appendix A.

The dataset of our study and the LEAD dataset (LEAD) share some of the variables available at the American Community Survey (Census), including the Building Age, Building Type (by number of units), Rent/Own, Heating Fuel Type, Group Education, and Demographics. The dataset used in the study extended the areas of knowledge to topics like Climate Conditions, Housing and Labor conditions, and State Energy Profiles. The data integration allowed for the creation of a complete and detailed dataset that provides a broader view of the complex relationships impacting energy burden.

The observations pertain to the year 2019. Assessing pre-pandemic conditions is crucial as it provides a baseline unaffected by the economic and social disruptions caused by COVID-19. This allows for a clearer understanding of energy burden dynamics under ’normal’ conditions and helps isolate the structural factors influencing energy vulnerability before external shocks introduce additional variability.

Data preprocessing is a crucial step in any data analysis or machine learning project, as it ensures the quality and comparability of the data before any analysis or modeling is conducted. In the context of an initial dataset comprising almost 20,000 samples related to various places, with inherent differences in size, population, and other characteristics, data preprocessing becomes especially important for making the variables and their information comparable. This process involves cleaning and transforming the data, as well as handling missing data. After preprocessing, the dataset is composed of a total of 18,786 places across the USA. The steps followed to prepare the data included:

• The places that did not have energy burden level information were deleted.
• Standardization of categories: Variables that can be broken down under the same category were grouped.
• Normalization: To compare the places with significant differences in size and population, the absolute numbers were transformed into percentages within their category.

An example of the process described in step 3 is presented in

Table 1. Data preprocessing example; normalization.

Reference Variable (Inhabitants)	Raw Variable (Inhabitants)	Transformed Variable (Percentage)
Total Population	Population over 25 y/o	Population over 25 y/o
2573	1870	72.68%

:

3.2. Data Analysis

3.2.1. DT Model Development

Different algorithms are available for building a decision tree (DT); the choice depends on the data type (categorical vs. numerical) and the learning task (classification vs. regression). For this study, the Classification and Regression Tree (CRT) algorithm has been applied, as it is a foundational method used for both classification and regression tasks. The core idea behind the CRT algorithm is to recursively split the data into subsets, making the data within each subset more homogeneous with respect to the target variable [49]. The CRT algorithm was chosen for its capability to handle both categorical and numerical data.

Dependent Variable: Energy Burden

• The energy burden values extracted from the Low-Income Energy Affordability Data (LEAD) Tool were classified into three distinct categories, as shown in Table 2: Class A—Low Energy Burden: Households with less than 4% energy burden. These households allocate a relatively small portion of their income to energy costs, suggesting better energy affordability.
• Class B—Medium Energy Burden: Households that spend more than 4% but less than 6%. This range indicates a moderate level of energy spending, raising the need for cautious energy management.
• Class C—High Energy Burden: Households with an energy burden greater than 6%. This group faces significant energy affordability challenges, dedicating a substantial portion of their income to energy costs, which can impact other areas of household spending and financial stability.

The classification of energy burden levels is closely aligned with the thresholds defined in the report “The High Cost of Energy in Rural America: Household Energy Burdens and Opportunities for Energy Efficiency” by the American Council for an Energy-Efficient Economy [50]. In this report, a nuanced understanding of the energy burden across the U.S. is presented, highlighting that an energy burden of 6% serves as a critical benchmark. This average indicates the typical proportion of household income spent on energy costs across the nation. Furthermore, the report identifies a 10% energy burden as a threshold for an extremely high level of energy burden, emphasizing the severe financial stress experienced by households exceeding this threshold.

Table 2. Dataset conformation by class type.

Class	Count (Places)	Percentage
A—Low Energy Burden (less than 4%)	3038	16.2%
B—Medium Energy Burden (between 4 and 6%)	8765	46.7%
C—High Energy Burden (more than 6%)	6983	37.2%
Total	18,786	100%

Table 2. Dataset conformation by class type.

Class	Count (Places)	Percentage
A—Low Energy Burden (less than 4%)	3038	16.2%
B—Medium Energy Burden (between 4 and 6%)	8765	46.7%
C—High Energy Burden (more than 6%)	6983	37.2%
Total	18,786	100%

Independent Variables: A total of 185 variables, organized into eight key areas, were analyzed to gain a comprehensive understanding of the factors influencing household energy burden. The criteria for categorizing these variables involved grouping individual variables under broader thematic areas. The categorization aims to structure the variables and does not have any impact on the analysis.

• Demographics: captures characteristics of the household occupants.
• Labor Conditions: focuses on factors related to employment status and industry.
• Geography: considers the physical characteristics of the locations.
• Climate: local climate characteristics are captured in this category.
• Housing: explores characteristics of the dwelling unit itself.
• Housing Utilities: focuses on the specific utilities used within the dwelling.
• Other Place Characteristics: captures additional factors specific to the places.
• State Energy Profile: considers the state’s energy-related characteristics.

By examining these eight categories, as described in

Table 3. Categorical division of the independent variables.

1. Demographics	2. Labor Conditions	3. Geography	4. Climate
Total Population	Total Workers	Population Density	Median Precipitation
Race	Military Population	Area	Median Wind Speed
Age	Civilian Population	Land Cover Median	Median Temperature
Gender	Labor Force	Land Cover Minimum
Population over 25	Work industry	Land Cover Maximum
Educational Attainment
5. Housing	6. Housing Utilities	7. Other Place Characteristics	8. State Energy Profile
Structure Type	Internet type	Housing Age Distribution	Average Electricity Retail Price
Number of Rooms	Heating Fuel Type	Commuters	Net Summer Capacity
People per Room	Plumbing Status	Renters by Commute Method	Net Generation
People per Household	Kitchen Status	Housing Units	Natural Gas Price
Vehicles per Renting Household	Telephone Service	Renter Households
Renters per Room	Computer availability	Total Households
Head of Household Education Level		Renter Occupied Housing Units

, a comprehensive understanding of the multifaceted factors contributing to household energy burden was sought.

3.2.2. DT Model Features

For the DT construction, the splitting criterion used was the Gini impurity measure [49]. The Gini impurity at a node reflects the level of homogeneity (similarity) within that node’s data points concerning Energy Burden (as the dependent variable). The DT aims to find splits that minimize the overall Gini impurity across the entire tree, effectively creating a series of increasingly homogeneous subgroups based on the chosen variables.

The model was built with a minimum node size of 30, which signifies that a node will only be split if it contains at least 30 data points. This hyperparameter setting helps prevent overfitting by avoiding overly granular splits with very small subgroups and potentially high mini-impurity.

3.2.3. Evaluation Metrics

The model’s performance was evaluated using k-fold cross-validation. This methodology divides the data into ‘k’ parts or folds, iteratively using one-fold for validation and the remaining folds for training. This process is repeated k times, ensuring all data points are used for training and testing [51]. A 10-fold was selected.

In the confusion matrix (see

Table 4. Confusion matrix.

Actual Class—Predicted Class (Training)	Count	A	B	C	Percentage Correct
A—Low Energy Burden (less than 4%)	3038	2568	395	75	84.5%
B—Medium Energy Burden (between 4 and 6%)	8765	1446	5084	2235	58.0%
C—High Energy Burden (more than 6%)	6983	123	1551	5309	76.0%
All	18,786	4137	7030	7619	69.0%
Actual Class—Predicted Class (Test)	Count	A	B	C	Percentage Correct
A—Low Energy Burden (less than 4%)	3038	2485	467	86	81.8%
B—Medium Energy Burden (between 4 and 6%)	8765	1318	4964	2483	56.6%
C—High Energy Burden (more than 6%)	6983	98	1592	5293	75.8%
All	18,786	3901	7023	7862	67.8%

), each row represents the instances of every class, while each column represents the instances in a predicted class. The diagonal cells from the top left to the bottom right show the number of correct predictions for each class (true positives for that class). In contrast, the off-diagonal cells show the misclassifications (how many instances of one class were predicted as another class).

The DT model achieved an overall accuracy of 69.0%. While category A demonstrated the highest accuracy at 84.5%, category B showed the lowest performance with 58.0% accuracy. Category C achieved a moderate accuracy of 76.0%. These results indicate that the model is more effective in classifying higher and lower energy burdens.

4. Results

4.1. DT Analysis Outcomes

The outcome variables of the DT have been tabulated, providing details about each variable’s general statistical information for a broader comprehension, as shown in

Table 5. Summary statistics: DT outcome variables.

Variable	Units	Minimum	Maximum	Mean	Std. Deviation
NO_INTERNET	%	0.00	100.00	20.58	12.08
HAS_COMPUTER	%	0.00	100.00	72.08	15.09
BACHELORS EDU	%	0.00	100.00	13.69	9.13
TOTAL POP2	inhabitants	0	8,419,316	10,478.18	84,505.28
HEATING_UTILITY_BOTTLED	%	0.00	100.00	10.57	19.06
PERSON_HOUSEHOLD	%	0.00	100.00	29.95	10.42
HOUSEHOLDS	numbers	0	3,167,034	3925.61	31,573.02
BLACK	%	0.00	100.00	7.97	16.99
ASIAN	%	0.00	100.00	1.45	04.06
LABOR FORCE	%	0.00	396.94	70.17	15.71
HIGHSCHOOL EDU	%	0.00	100.00	35.50	12.16
BUILT_1939_EARLIER	%	0.00	100.00	21.38	18.18
INTERNET_IN_HOUSEHOLD	%	0.00	100.00	84.15	13.03

.

The relevant variables represent a broader set that includes the variables shown in the tree and those significantly influencing the energy burden levels, even if they are not explicitly used to build the tree. It is determined that the tree pulls three variables from the top five in Figure 1 below, including households with no internet, populations in the municipality aged 25 and over with a high school degree as their final educational attainment, and households that have a computer.

The above bar graph highlights the importance of the top 20% of variables used in the Decision Tree (DT) model. The most influential factors affecting energy burden levels include households without internet access, availability of computers and broadband, bachelor’s degree education, and internet subscription; these top five variables play a significant role in shaping the model’s predictions.

The tree diagram serves as the visual depiction of the model. The “rollback method” is the strategy employed to identify the optimal pathway in a DT [52]. This method involves analyzing the tree from the bottom to the top (or right to left, if applicable) and prioritizing later decisions.

By analyzing the DT structure, the variables selected by the model, and the splitting criteria used at each step, we can gain insights into how the selected variables interact with energy burden outcomes. In this example, computer availability is not just a variable but a crucial starting point for the model, highlighting its importance. However, it is important to note that other factors also significantly contribute to a more nuanced understanding of energy burden.

The DT analysis reveals that households reporting computer ownership below 82.72 percent are more likely to be severely energy-burdened, with most falling into Classes B and C, as indicated in the lower half of the full DT in Figure 2. Among the households that belong to this classification, those that have a computer are less than or equal to 82.72 percent; if the number of households that do not have access to the internet is less than or equal to 22.71 percent, then they are also more likely to be energy burdened. In this category, while Classes B and C make up the majority, the share of Class B is higher than that of Class C, which suggests that households’ access to a computer appears to be more closely associated with the severe energy-burdened classification compared to having a computer but without internet access. On the flip side, if the percentage of households with a computer exceeds 82.72 percent, they are less likely to be severely energy-burdened, with no Class C representation in nodes 19–23 at the upper part of the DT diagram (Figure 2). For detailed classification rules, please see

Table A2. Classification Rules.

Final Node	Classification Rules
Node 1.	IF (HasComputer <= 82.72) AND (NoInternet <= 22.71) AND (BachelorsEdu <= 19.60) AND (TotalPop =< 1977.5) AND (HasComputer <= 70.78) AND (OnePersonHousehold <= 33.3) AND (BachelorsEdu <= 11.13) THEN A = 2.40%, B = 41.30%, C = 56.30%.
Node 2.	IF (HasComputer <= 82.72) AND (NoInternet <= 22.71) AND (BachelorsEdu <= 19.60) AND (TotalPop =< 1977.5) AND (HasComputer <= 70.78) AND (OnePersonHousehold <= 33.3) AND (BachelorsEdu => 11.13) THEN A = 3%, B = 60%, C = 37%.
Node 3.	IF (HasComputer <= 82.72) AND (NoInternet <= 22.71) AND (BachelorsEdu <= 19.60) AND (TotalPop =< 1977.5) AND (HasComputer <= 70.78) AND (OnePersonHousehold => 33.3) THEN A = 2.1%, B = 29%, C = 68.9%.
Node 4.	IF (HasComputer <= 82.72) AND (NoInternet <= 22.71) AND (BachelorsEdu <= 19.60) AND (TotalPop =< 1977.5) AND (HasComputer => 70.78) AND (BachelorsEdu <= 9.23) AND (HeatingUtilityBottled <= 27.57) AND (OnePersonHousehold <= 29.24) THEN A = 1.40%, B = 62.40%, C = 36.2%.
Node 5.	IF (HasComputer <= 82.72) AND (NoInternet <= 22.71) AND (BachelorsEdu <= 19.60) AND (TotalPop =< 1977.5) AND (HasComputer => 70.78) AND (BachelorsEdu <= 9.23) AND (HeatingUtilityBottled <= 27.57) AND (OnePersonHousehold =>29.24) THEN A = 1.60%, B = 44.2%, C = 54.2%.
Node 6.	IF (HasComputer <= 82.72) AND (NoInternet <= 22.71) AND (BachelorsEdu <= 19.60) AND (TotalPop =< 1977.5) AND (HasComputer => 70.78) AND (BachelorsEdu <= 9.23) AND (HeatingUtilityBottled =>27.57) THEN A = 1.80%, B = 31.6%, C = 66.6%.
Node 7.	IF (HasComputer <= 82.72) AND (NoInternet <= 22.71) AND (BachelorsEdu <= 19.60) AND (TotalPop =< 1977.5) AND (HasComputer => 70.78) AND (BachelorsEdu => 9.23) THEN A = 3.60%, B = 68.2%, C = 28.2%.
Node 8.	IF (HasComputer <= 82.72) AND (NoInternet <= 22.71) AND (BachelorsEdu <= 19.60) AND (TotalPop => 1977.5) THEN A = 6%, B = 78.8%, C = 15.2%.
Node 9.	IF (HasComputer <= 82.72) AND (NoInternet <= 22.71) AND (BachelorsEdu => 19.60) AND (NoInternet <= 11.91) AND (HasComputer <= 39.50) THEN A = 8.50%, B = 59.4%, C = 32.1%.
Node 10.	IF (HasComputer <= 82.72) AND (NoInternet <= 22.71) AND (BachelorsEdu => 19.60) AND (NoInternet <= 11.91) AND (HasComputer => 39.50) THEN A = 44.8%, B = 47.8%, C = 7.4%.
Node 11.	IF (HasComputer <= 82.72) AND (NoInternet <= 22.71) AND (BachelorsEdu => 19.60) AND (NoInternet <= 11.91) AND (Asian <= 5.85) THEN A = 11.6%, B = 69.2%, C = 19.2%.
Node 12.	IF (HasComputer <= 82.72) AND (NoInternet <= 22.71) AND (BachelorsEdu => 19.60) AND (NoInternet <= 11.91) AND (Asian => 5.85) THEN A = 58.5%, B = 36.5%, C = 5%.
Node 13.	IF (HasComputer <= 82.72) AND (NoInternet => 22.71) AND (HasComputer <= 65.76) THEN A = 0.7%, B = 20.40%, C = 78.9%.
Node 14.	IF (HasComputer <= 82.72) AND (NoInternet => 22.71) AND (HasComputer <= 65.76) THEN A = 0.7%, B = 20.40%, C = 78.9%.
Node 15.	IF (HasComputer <= 82.72) AND (NoInternet => 22.71) AND (HasComputer => 65.76) AND (Households <= 535) THEN A = 0.9%, B = 38.9%, C = 60.2%.
Node 16.	IF (HasComputer <= 82.72) AND (NoInternet => 22.71) AND (HasComputer => 65.76) AND (Households => 535) AND (Black <= 31.83) AND (LaborForce <= 67.98) AND (Households <= 1735.5) THEN A = 0%, B = 45%, C = 55%.
Node 17.	IF (HasComputer <= 82.72) AND (NoInternet => 22.71) AND (HasComputer => 65.76) AND (Households => 535) AND (Black <= 31.83) AND (LaborForce => 67.98) THEN A = 2.7%, B = 74.6%, C = 22.7%.
Node 18.	IF (HasComputer <= 82.72) AND (NoInternet => 22.71) AND (HasComputer => 65.76) AND (Households => 535) AND (Black => 31.83) THEN A = 0%, B = 33%, C = 67%.
Node 19.	IF (HasComputer => 82.72) AND (Highschool Edu <= 30.75) THEN A = 70.6%, B = 27.7%, C = 1.7%.
Node 20.	IF (HasComputer => 82.72) AND (Highschool Edu => 30.75) AND (Built1939earlier <= 20.89) AND (NaturalGasPrice <= 8.75) THEN A = 36.2%, B = 57.40%, C = 6.4%.
Node 21.	IF (HasComputer => 82.72) AND (Highschool Edu => 30.75) AND (Built1939earlier <= 20.89) AND (NaturalGasPrice => 8.75) AND (HasComputer <= 87.16) THEN A = 11.4%, B = 80.7%, C = 7.9%.
Node 22.	IF (HasComputer => 82.72) AND (Highschool Edu => 30.75) AND (Built1939earlier <= 20.89) AND (NaturalGasPrice => 8.75) AND (HasComputer => 87.16) THEN A = 39.6%, B = 58.3%, C = 2.1%.
Node 23	IF (HasComputer => 82.72) AND (Highschool Edu => 30.75) AND (Built1939earlier => 20.89) AND THEN A = 6.6%, B = 71.2%, C = 22.2%.

in Appendix B.

Among the households that belong to the above classification of households that have a computer (more than 82.72 percent); if the total population over 25 with a high school education as their final educational attainment is less than or equal to 30.75 percent, then the number of households is more likely to be less severely energy-burdened (roughly 70 percent belonging to Class A). This reinforces the finding that households with a computer are an important explanatory factor associated with moderate levels of energy burden. This may also suggest that a higher share of the population over 25 belongs to the higher categories of final educational attainment, equivalent to a bachelor’s degree and beyond. Contrastingly, in the same category, if the total population over 25 with a high school education as their final educational attainment is greater than 30.75 percent, the households tend to be more severely energy-burdened, with the majority in Classes B and C. This suggests that given a specific profile of households with a computer and a specific educational attainment profile, the factor of when the houses were built is closely associated with higher energy burdens in these households (see node 23 in Figure 3).

According to the DT analysis, the most severely energy-burdened household characteristics are classified by households lacking access to the internet, those without a computer, those living in larger communities with a greater number of households, and those with a higher proportion of Black populations. Specifically, if the number of households that lack access to the internet exceeds 22.71 percent; if this particular group of households with a computer exceeds 65 percent; if the number of households exceeds 535; and finally, if the percentage of the population that is Black is greater than 31.83 percent, then these households appear to be highly energy burdened, with no households in Class A, 33 percent in Class B, and 67 percent in Class C (see node 18 in Figure 4). Following the same branch, given that the number of households in the municipality is less than or equal to 535, the households tend to be severely energy burdened (see node 14). Similarly, following along in “branch 2” of the DT, given that the number of households with a computer is less than or equal to 82.72 percent; if the number of households that lack access to the internet exceeds 22.71 percent; if this particular group of households with a computer exceeds 65 percent; if the number of households exceeds 535; if the percentage of the population that is Black is less than or equal to 31.83 percent; if the labor force, meaning the total number of workers in a municipality, including both military and civilian, is less than or equal to 67.98; and when the total number of households in the municipality is less than or equal to 1735.5, the households tend to be severely energy burdened with 45 percent in Class B and 55 percent in Class C (see node 15 in Figure 4). This finding clearly indicates a strong association between severe energy burden and access to computers and the internet, combined with household numbers in a given locality and household profiles of racial representation.

“Branch 3” also represents another observation that ties access to the internet, final education attainment, and racial and ethnic factors as variables selected to classify excessively energy-burdened and less severely energy-burdened households. For example, given that the number of households that have a computer is less than or equal to 82.72 percent; if the number of households that have no access to the internet is less than or equal to 22.71 percent; if the total population over 25 with the education attainment of a bachelor’s degree is higher than 19.6 percent; and if the share of the population that is greater than 5.85 percent is Asian, then the households tend to be less severely energy burdened, with the majority belonging to Class A at 58.5 percent (see node 12 in Figure 5).

“Branch 4” in Figure 6 suggests a specific profile of households closely associated with variables such as access to a computer, internet access, educational attainment, total population of the municipality, and type of heating used. Generally, access to critical infrastructure, education, and population density seems closely linked to energy burden based on the variables the DT pulls. The type of heating households connect to may vary by location, particularly in rural areas where utility heating may not be accessible. This suggests that households in smaller municipalities, distant from metropolitan areas, tend to be more severely energy-burdened (also associated with the lack of infrastructure, characterized by access to the internet).

In general, the DT results indicate that variables such as internet access, educational attainment, housing built prior to 1939, racial representation, and population density are highly correlated with varying levels of energy burden. The DT model identified computer access and internet connectivity as key classification variables that led to various context-specific energy burden outcomes. For instance, in certain municipalities, if the percentage of households with computer access is 82% or less and the percentage of households without internet access is 23% or less, and in these areas, if more than 20% of individuals over the age of 25 have a bachelor’s degree, then these households tend to exhibit a low energy burden.

Conversely, the same computer access and internet connectivity indicators suggest a high energy burden profile when the DT model incorporates the household density variable. The maps in the following section will further illustrate these context-specific findings and highlight the regional variations identified by the DT analysis. Each node in the DT represents a specific profile of energy-burdened households within a given municipality, providing a unique opportunity to map the results and identify specific geographical implications and patterns.

4.2. Key Variable Mapping and Analysis

This section provides a spatial mapping of energy burden levels determined by the Decision Tree (DT) model across six selected regions in the United States: Philadelphia, Chicago, Los Angeles, Seattle, Dallas, and Atlanta. Alongside the energy burden data, key influencing variables—such as high school education levels, computer ownership, the percentage of homes built in 1939 or earlier, and the proportion of Black and Asian residents—are also mapped for each region.

As observed in Figure 7, the distribution of energy burden varies within the study areas. There are noticeable clusters of high (class C), medium (class B), and low (class A) energy burden areas, particularly across California and the Southern Chicago Metropolitan Area when compared to the Mid-Atlantic Region and the Southeast block, where the three classes of energy burden are scattered heterogeneously across the territories.

Class A seems to be predominant in major cities like Chicago, Seattle, and Los Angeles. Other metropolitan areas like Philadelphia, Dallas, and Atlanta appear to be less homogeneous, although class B households in the central part of the city primarily represent Philadelphia. A clear north–south division is visible in Dallas and Atlanta. The suburban areas surrounding Philadelphia, Dallas, and Atlanta represent a mix of all three classes, with B and C more predominant in the latter two metropolitan areas. In general, rural areas tend to display a higher prevalence of higher energy burden (classes B and C).

Figure 8 illustrates that in Philadelphia, school education and the percentage of Asian residents are homogenous across the region. Computer ownership is difficult to correlate with energy burden. Still, houses built earlier than 1939 in more rural areas surrounding Philadelphia seem consistent with a higher energy burden, as does a higher percentage of Black residents.

Figure 9 illustrates the key variables and energy burden levels in Chicago. It is observed that a lack of computer ownership and a higher percentage of residents with only a high school education correlate with a higher energy burden. The percentage of Black residents correlates with a higher energy burden, and the percentage of Asian residents correlates with a lower energy burden.

Figure 10 illustrates the key variables and energy burden levels in the Los Angeles-San Diego area. It is observed that the percentage of residents with only a high school education and a lack of computer ownership correlates with a higher energy burden. The percentage of Black and Asian residents does not correlate with the energy burden. Energy burden types are concentrated in zones compared to the other variables, which are more mixed throughout the region.

Figure 11 shows that the energy burden levels in Seattle and its suburban area are generally homogeneous and healthy. Houses built in 1939 or earlier are weakly correlated with class B energy burden, as is the percentage of residents with only a high school education and a lack of computer ownership, particularly toward the coast.

Figure 12 illustrates the key variables and energy burden levels in the Dallas metropolitan region. There is a correlation between higher energy burden and the percentage of residents with only a high school education, as well as between the percentage of Black residents and the lack of computer ownership. Asian populations seem to be concentrated in areas with mostly lower energy burdens. All of the variables have a clear zone differentiation aligned with energy burden levels within the Dallas metropolitan region.

Figure 13 illustrates the key variables and energy burden levels in the Atlanta region. Atlanta has the most apparent correlation among all variables except for houses built in 1939 or earlier. The percentage of residents with only a high school education, a lack of computer ownership, and higher percentages of Black residents correlates with a higher energy burden. Higher percentages of Asian residents correlate with a lower energy burden.

The five variables examined—houses built earlier than 1939, high school education, percentage of Black population, percentage of Asian population, and computer ownership—all have some degree of correlation with energy burden. Houses built earlier than 1939 have a weak correlation with higher energy burden, if any. The percentage of residents whose highest level of education attained is high school is somewhat correlated with higher levels of energy burden, except in the Philadelphia metropolitan area. Higher percentages of Black residents also tend to be relatively correlated with higher levels of energy burden, except in the Los Angeles and Seattle metropolitan areas examined. Higher percentages of Asian residents are weakly correlated with lower levels of energy burden when correlation is noted. Finally, a lack of computer ownership correlates with higher energy burden levels, except in the Philadelphia metropolitan area. The age of housing stock is correlated with a higher energy burden in and around Philadelphia (class C energy burden) and Seattle (class B energy burden). However, there is no correlation between the other regions examined, and Dallas and Atlanta, in particular, are mainly composed of newer housing. It may be inferred that in Philadelphia specifically, the colder weather combined with inefficient old housing and a reliance on natural gas for heating contributes to the energy burden. For those whose highest level of education attained is high school, there was some correlation with higher energy burdens in California, Seattle, and Atlanta. In comparison, Chicago and Dallas showed a stronger correlation with higher energy burden, and Philadelphia appeared more homogeneous. Dallas, Atlanta, and Chicago had higher levels of segregation among Black residents, correlating with a higher energy burden where the percentages of Black residents were greater; Philadelphia had a weaker correlation between Black populations and energy burden, and California and Seattle showed little correlation with energy burden. In areas where the percentage of Asian residents is higher, Chicago and Dallas both showed some correlation with lower energy burden, but Philadelphia, California, Seattle, and Atlanta did not show a correlation between the percentage of Asian residents and energy burden; however, it is worth noting that the percentage of Asian residents is not very high in any part of the Philadelphia metropolitan area. Computer ownership showed a higher correlation between lower access to computers and higher energy burden in Chicago (class C energy burden), California, and Atlanta; Seattle and Dallas had some correlation with higher energy burden, but Philadelphia did not show any correlation between these variables. Overall, the energy burden levels in Philadelphia and Seattle seem to be more influenced by housing stock and less by the socioeconomic demographics of residents, unlike the other regions (Atlanta, Dallas, and Chicago), where the correlation between energy burden and socioeconomic demographics is stronger, and where racial and class stratification plays a larger role in the geographic and socioeconomic distribution of energy burden.

5. Discussion

5.1. Cross-Sectoral Relationships and Impact

Our multidisciplinary analysis reveals that energy burden is a complex issue with multiple interrelated determinants. No single factor can fully explain the energy burden, and the relative importance of different factors may vary across regions and communities. Either through the DT or the maps, each place presents its mix of characteristics, resulting in certain energy burden levels. For example, the DT selected the variables of houses older than 1939 and the percentage of the Black population as influential for specific profiles; the maps clearly show a disparity between urban and rural areas for these variables. Philadelphia has a higher percentage of the Black population than its suburban and rural counties. Similarly, older houses are more concentrated in historical urban centers like Atlanta and Philadelphia than in their surroundings. As evident from our DT analysis, the proportion of multiracial households in many municipalities was identified as a determining factor (see nodes 12 and 18). It is important to note that these patterns are also evident from our efforts to map the distributions of Black and Asian households relative to energy burden levels at these various regional and subregional locations. Our findings corroborate other research related to households and communities of color: Black households in the U.S. often experience higher energy burdens compared to households of different ethnic backgrounds due to historical and ongoing socioeconomic disadvantages. In general, we found that places with higher population densities demonstrated a more significant correlation with socioeconomic demographics, such as educational attainment and racial representation. In less populated areas, particularly in some rural regions, housing stock seems more related to disparate levels of energy burden, as demonstrated in parts of Philadelphia and Seattle. Based on our analysis, no single variable has a determining effect on energy burden levels, and there is no single regional pattern that can be applied across the entire country. In some highly dense areas, there were indications of a correlation between Black populations and severe energy burdens shown in parts of Dallas, Atlanta, and Chicago. This was not the case in Los Angeles, for instance, where high Black populations and Class A representation were apparent. Computer ownership also shows a significant correlation across some parts of the country we examined, but not all. This suggests that every place requires a targeted investigation of community profiles to design policy measures or programs to help severely energy-burdened households and communities, taking into account the region-specific factors illustrated above.

5.2. Rural and Urban Dynamics

Identifying such variables contrasted in the maps highlights a rural-urban energy burden differentiation in some regions. Although some variables may be more evident than others in explaining the rural-urban gap, this analysis has shown possibly overlooked variables that intensify such differences.

In the rural regions surrounding Atlanta and Philadelphia, communities face a higher energy burden than their urban counterparts. Nevertheless, each region has different levels of correlation with each variable. This argument supports the interdependence of factors, which interact with each other to create a complex web of determinants and particular energy burden profiles. Classifying profiles reinforces the targeting of policies, aiming to prevent energy vulnerability with localized actions.

In addition, state-level averages of energy burden can hide significant disparities across regions and communities. These averages often do not reflect the local realities of energy affordability, leading to a biased understanding of where energy burden levels are highest. As a result, state-level policies may not adequately account for the unique energy needs of different regions within a state. Observing the energy burden at a more localized level is critical for effectively identifying and addressing inequities.

For example, about 5% of occupied housing units in the US use “bottled, tank, or LP gas” as a source of heating fuel. This is approximately ten times less than the occupied housing units that rely on utility gas as a source of heating fuel. The share of fuel or heating oil was slightly less, at just under 4%. In MA, for example, the price of LPG as of May 2024 was, on average, $3.58 per gallon, an increase of 3.96% compared to 2023 [53]. On average, the price of heating oil as of May 2024 was $3.79 per gallon, an increase of 7.86% from 2023 prices (Mass.gov, 2024). In Missouri, for example, the share of occupied housing units that rely on “bottled, tank, or LP gas” for heating is over 8%, almost double the national share (US Census). By contrast, in California, the share of bottled, tank, or LP gas use in heating is just under 3.5% [54]. This signals the geographic variation in LPG or propane use as a heating source. The variable results depend on several factors, but the higher proportion of alternative fuel use could be propelled by the limited interconnection to utility gas.

The concept of regional ecologies underscores the importance of crafting policies and regulations specifically tailored to local communities’ unique characteristics and needs. Localized data are crucial for creating policies that address energy inequities. The observation that energy burden levels differ from household or state characteristics and local ones (as illustrated by the selection of maps) underscores a notable gap in targeting local and community-level interventions that could have wider-reaching impacts, particularly for vulnerable places. Aligning energy strategies among government tiers and industries is essential for addressing specific issues encountered in urban and rural settings since these areas frequently experience varying energy landscapes and economic influences. A comprehensive energy policy may support creating customized solutions that address the requirements of diverse populations in different regions. Cities could consider improving energy efficiency in apartment buildings as an initiative. In contrast, rural areas might benefit more from investing in expanding infrastructure for better access to sustainable energy sources that effectively utilize local resources. Working together across levels allows stakeholders to exchange successful methods and allocate resources efficiently to develop inclusive energy plans that recognize the diverse needs of urban and rural communities.

5.3. Policy Implications

Energy burden analysis has a significant impact on intervention policies. Policy interventions should aim to democratize energy, address inequities, aid financial assistance programs, and spur cross-sector collaboration. When strategies across the energy, housing, and transportation sectors are aligned, these policy interventions will accelerate the adoption of sustainable energy and generate job opportunities that positively impact local economic development. Energy democracy may be promoted by fostering participatory approaches in energy planning that support historically underrepresented groups, particularly Black and Hispanic communities, which can lead to more equitable energy transitions and benefit distribution. To begin addressing and alleviating systemic inequities, policymakers must acknowledge and confront the historical and systemic injustices that contribute to energy burdens and develop targeted policies such as housing rehabilitation programs to improve energy efficiency, universal energy access initiatives to ensure reliable energy for all households and incentives for energy-efficient upgrades for low-income families. With investment in efficiency, energy costs decrease, and living conditions improve. Therefore, place-based policies focused on efficiency can address energy injustice and mitigate racial disparities [55]. This approach highlights the importance of targeted interventions to address and reduce energy poverty. Efficiency programming that embraces home retrofits or financial assistance is essential in mitigating energy burdens for households in low-income, underserved communities despite an abundance of energy resources in the US [14,56]. While Asian households tend to face less severe energy burden levels (primarily in Classes A and B), they encounter unique challenges related to energy burden, influenced by factors such as immigration status, language barriers, and varying economic circumstances. Furthermore, racial disparities in energy poverty emphasize the need for more culturally sensitive energy assistance programs and policies that consider the diverse needs of Asian communities [57]. Policymakers in energy governance may implement efforts that target green job workforce development, inclusionary zoning to maintain affordable housing and community land trusts that focus on energy equity and housing quality. By aligning strategies across the energy, housing, and transportation sectors, the widespread adoption of sustainable energy can be expedited, spurring job creation that ultimately bolsters the growth and development of local economies.

Community-driven energy initiatives that empower local residents to participate in decision-making processes should be nurtured, and stakeholders may be engaged to identify their needs and preferences in energy solutions. These initiatives not only foster a sense of ownership and engagement but also help ensure that the resulting projects align with local priorities and challenges. Some examples of community-based impacts include community solar projects, such as the community solar gardens in Minnesota, which are centrally located PV systems that provide electricity to subscribers and eliminate the need for individual installations; energy cooperatives like Co-op Power in New England and New York that facilitate communities in building their own community-owned clean energy businesses, organizing and investing in their local energy projects; and participatory budgeting, such as the New York City Council’s participatory budgeting process, where community members can vote on energy projects and allocate city funds for energy efficiency upgrades. These gardens make solar energy more accessible to renters, low-income households, and those without suitable roofs. Moreover, programs like Colorado’s Solar Rewards Community enhance these efforts by offering financial incentives for participation and making clean energy more affordable for underrepresented groups. Cook and Shah (2018) suggest that policies and programs designed to improve rooftop solar access for low-income households, such as expanding existing weatherization programs to include solar or encouraging utility investment in solar for low-income households, would alleviate the energy burden for these households [58]. For example, the Solar for All program in Washington, D.C., not only subsidizes rooftop solar installations but also provides free solar panels to qualifying low-income families, directly reducing their energy bills. Similarly, California’s Single-family Affordable Solar Homes (SASH) program targets economically disadvantaged households, pairing solar access with energy efficiency upgrades for maximum impact. Collaborative frameworks also enable the inclusion of social equity aspects in energy policy decisions to ensure that clean energy efforts benefit marginalized communities equitably.

Another localized recommendation could be to expand access to energy efficiency programs for low-income households. Hernandez and Bird suggest policy interventions in energy conservation, energy literacy, and utility rate affordability that would improve conditions for low-income families experiencing energy burdens [50]. Policymakers may consider financial assistance programs, such as those offered through LIHEAP, that address immediate energy needs by providing utility bill support to low-income households facing high energy costs, or through WAP, which could provide subsidies for energy-efficient upgrades and rebates for installing energy-efficient appliances or making home weatherization improvements [50]. These cooperatives not only focus on renewable energy generation but also promote workforce development by creating green jobs within the community. Co-op Power has, for example, helped establish biodiesel manufacturing facilities and energy-efficient housing projects that directly benefit local economies while reducing dependence on fossil fuels. Similar initiatives, such as the Black Hills Electric Cooperative in South Dakota, have furthered these goals by implementing comprehensive energy efficiency programs and supporting member education on energy conservation practices.

These programs should be informed by local energy burden assessments. Implementing direct cash assistance for families in areas with extreme temperature fluctuations or providing tailored financing options for renewable energy installations in underserved neighborhoods presents potential options for targeted policies. Other methods of improving access to energy efficiency programs include local initiatives like the Better Buildings Initiative, which partners the Department of Energy with private businesses, cities, and states to enhance building energy efficiency through workshops, novel technologies, and financial incentives [59]. Collective energy audits in neighborhoods and cities allow residents to identify energy-saving opportunities collaboratively, while community resilience hubs and microgrids, such as a new resilience hub funded by the Bipartisan Infrastructure Law in Louisiana, strengthen local emergency energy resources during extended outages [60]. Youth-led energy initiatives like The Climate Initiative provide community-based education and empowerment programs to engage youth and encourage them to become climate leaders [61]. Finally, local energy advocacy groups such as the Citizens Utility Board in Illinois advocate for community needs in energy policy by fighting utility rate hikes, promoting clean energy, and ensuring consumer protection in Illinois [62]. These frameworks guide policymakers in designing inclusive policies that address systemic inequities, ensuring that renewable energy solutions prioritize vulnerable populations.

To tackle these issues, social equity factors like gender, race, and socioeconomic status must be considered in energy policies [63]. Adopting more comprehensive and intersectional approaches that consider gender, race, and socioeconomic status will increase the development of interventions that foster access to affordable, sustainable, and just energy solutions that equitably distribute benefits across diverse and vulnerable U.S. communities. By addressing variables such as internet access, multi-factor impact analysis is significant for developing policy regimes to reduce energy burdens and effectively promote an equitable energy landscape for all communities. Access to reliable internet enables households to use energy management tools and platforms that can help monitor and reduce energy consumption. Additionally, internet access facilitates participation in demand response programs, where consumers can receive incentives for reducing their energy use during peak demand times. Ensuring that all communities, especially those in low-income and rural areas, have access to resources like high-speed internet enables residents to become self-determined in making informed decisions concerning their energy consumption and improves access to available support programs. Integrating internet access into energy governance can ameliorate communication strategies and outreach initiatives to better educate communities about relevant energy-efficient practices and available financial assistance programs. Policymakers can also utilize online platforms to gather local energy burden assessments, which will ensure the developed programs are tailored to each community’s specific needs.

5.4. Limitations and Future Direction

While this study comprehensively analyzes the factors influencing energy burden, several limitations should be acknowledged. First, 2019 data may only partially capture the current impacts of more recent events, such as the COVID-19 pandemic, which has significantly altered economic conditions, energy usage, and affordability patterns. Additionally, while helpful in handling complex relationships, the DT methodology may oversimplify some interactions or overlook variables, such as qualitative factors like behavioral responses to energy costs and cultural or institutional barriers. For future studies, offering a point of comparison for examining post-pandemic impacts and incorporating more recent data would allow for a more comprehensive understanding of how the pandemic and other recent events have reshaped energy consumption and affordability trends. A follow-up analysis focusing on the pandemic years of 2020 and 2021 would provide a compelling contrast to pre-pandemic conditions, shedding light on how these unprecedented disruptions influenced energy trends. This represents important further work to be undertaken.

Regarding the DT methodology, future studies can complement this approach with qualitative methods, such as interviews or case studies, to better account for non-quantifiable factors (behavioral responses, cultural factors, and institutional barriers). For example, Philadelphia and Seattle’s energy burden levels seem to be more influenced by housing stock and less by the socioeconomic status of residents, unlike the other regions (Atlanta, Dallas, and Chicago), where the correlation between energy burden and socioeconomic demographics is stronger. Additional investigations could be conducted using multiple methodologies and tools.

6. Conclusions

This article investigates the multifaceted issue of energy burden across the U.S. we examine the ways in which socioeconomic disparities and systemic injustices contribute to unequal energy experiences among different communities. The findings from this analysis contribute to the existing literature on energy burden by revealing the complex interplay of socioeconomic, geographic, and demographic factors—regional ecologies—that shape energy inequities in the U.S. However, while this study uncovers how these factors interact, it does not identify any single factor or regional pattern that can fully explain energy burden. Building upon prior work by researchers like Hernandez, Sovacool, and Wang, this study highlights that local factors, such as housing age, energy source, and educational attainment, intensify systemic inequities related to housing, race, and labor status. Taken together, these factors drive energy burden across various metropolitan and rural regions. These results align with findings by Sovacool et al. (2020) that underscore the importance of context-sensitive approaches in addressing energy burden across communities [56]. The interrelated factors influencing energy burden include socioeconomic status, housing characteristics, educational attainment, racial demographics, access to technology, and energy system characteristics.

By utilizing DT analysis of these factors, this article provides a detailed analysis of the regional ecologies at play. The analysis of energy burden across various regions reveals distinct patterns: in Philadelphia, moderate energy burden is often associated with historical housing stock and racial demographics. Similarly, Chicago exhibits a notable correlation between higher energy burden levels and neighborhoods with high percentages of Black populations, as well as neighborhoods with lower percentages of computer ownership. Los Angeles and San Diego demonstrate a correlation between high school educational attainment and a lack of computer ownership, and energy burden. Seattle showcases healthy energy burden levels, suggesting the city has effective energy policies. Dallas and Atlanta, which have more prevalent Class B and C energy burden scores, correspondingly, and higher levels of segregation of Black and Asian populations, reveal a clear correlation between energy burden and racial demographics, where higher percentages of Black and Asian residents correlate with higher and lower energy burden, respectively. These findings highlight the intersection of socioeconomic factors and urban development. The implications are significant for understanding how place-based factors amplify or mitigate energy burdens within high-burden areas. This underscores the critical importance of considering regional and local ecologies as essential determinants in understanding energy burden. They reveal how environmental, social, and economic factors operate together in distinct geographical settings.

Regional ecologies not only reflect the natural environment and urban infrastructure but also include the policies, institutions, and historical contexts that shape energy access and distribution. Regional ecologies thus influence how energy costs are perceived and experienced, and they help explain why certain communities experience disproportionate burdens. For example, regions with historical patterns of racial segregation, like those seen in Chicago and Atlanta, exhibit energy burden profiles that reflect both the social infrastructure of those areas and the inequitable access to energy resources. Similarly, areas with older housing stock or poor infrastructure are more likely to see higher energy burdens because the built environment does not support energy efficiency. By incorporating regional ecologies into energy burden analysis, we better understand that energy inequity is not a standalone issue but is intertwined with the broader social, economic, and environmental fabric of a given region. Energy burden is, therefore, not merely a function of economic hardship but is deeply tied to regional and local ecologies, which necessitates localized interventions. Our nuanced analysis and findings contribute to the growing recognition within energy studies scholarship that demographic variables alone are insufficient to capture the localized drivers of energy burden.

These insights further suggest a need for localized energy interventions rather than one-size-fits-all strategies. We offer recommendations aimed at promoting energy equity while simultaneously alleviating the challenges faced by historically marginalized and underserved populations. For example, targeting housing stock improvements in older urban areas may help alleviate the energy burden for specific communities. Similarly, expanding infrastructure in rural areas could improve access to sustainable energy sources. We recommend implementing targeted policy interventions that promote energy equity, such as financial assistance programs for low-income households, investments in energy efficiency retrofits, and community-driven energy initiatives. Additionally, we advocate for the integration of social equity considerations into energy policies to ensure that marginalized communities benefit from the clean energy transition and experience improved energy access. Policies should consider energy burden through the lens of energy democracy. Such policies might better foster community-based approaches that engage residents in energy planning and ensure equitable benefits. Collaborative frameworks that connect energy, housing, infrastructure, and transportation policy, such as those recommended by Johnson et al. (2020), could also ensure a more coordinated approach to energy equity [63].

The broader implications of this study point to the need for targeted assistance programs that consider the full spectrum of regional ecologies. Programs such as the Low-Income Home Energy Assistance Program (LIHEAP), which provides financial assistance to low-income households to help them manage their energy costs, and the Weatherization Assistance Program (WAP), which offers funding for energy-efficient upgrades to homes, improve energy efficiency, and reduce overall energy expenses for eligible families. However, LIHEAP and WAP could benefit from using localized data to better target areas with high energy burdens. Expanding initiatives such as community solar programs or resilience hubs to address specific regional needs will further foster both resilience and sustainability for communities. These approaches underscore the value of collaborative partnerships across federal, state, and local levels to facilitate equitable energy transitions and allow for resources to be allocated based on specific community profiles.

This article has laid the groundwork for understanding the role of regional ecologies in shaping energy burden. Future work should focus on integrating more recent data and exploring non-quantifiable factors, such as cultural or behavioral responses, into the analysis. Future work could extend the findings of this study by incorporating qualitative methods, such as surveys and interviews, to gain deeper insights into the determinants of energy burden and the interactions between various factors. For example, surveys could be used to target low-income households to explore their perceptions of energy affordability and the impact of local energy policies on their daily lives. Case studies involving interviews with community members in areas facing high energy burdens could provide data on the specific challenges they face and the strategies they use to cope with energy burden. These data would further enhance our understanding of the multifaceted nature of energy inequity. Finally, the impacts of the COVID-19 pandemic continue to reshape energy usage patterns and economic conditions. Future research could also examine post-pandemic trends to assess ongoing and emergent disparities and how these relate to regional ecologies.

In conclusion, this article underscores the critical need to move beyond a one-size-fits-all approach to addressing energy burden. By highlighting the significance of regional ecologies, we advance our understanding of the intersection between energy burden, socioeconomic disparities, and place-based factors. Regional ecologies enhance our understanding of the intersection between energy burden, socioeconomic disparities, and place-based factors. We advocate for further nuanced, localized solutions that meet the unique needs of diverse communities. This research lights a path toward energy equity—one that empowers communities and paves the way for a fairer, more resilient energy future.