Unraveling the Complexities of Energy
Poverty in Germany
A Comparative Analysis of Determinants, Dynamics, and Indicators
Supervisor: Student:
Prof. Thomas Sterner Freya Leverenz
Master’s Thesis in Economics
Spring 2023
Abstract
This master thesis investigates the determinants and dynamics of energy
poverty in Germany based on different calculation methods. Using a dy-
namic random-effects probit model, socioeconomic factors associated with a
household’s risk of experiencing energy poverty are identified. Factors such
as household composition, employment status, and living space are found
to be strongly correlated with the prevalence of energy poverty. Significant
state dependence highlights the persistence of energy poverty, whereby house-
holds that have experienced energy poverty in the past are at higher risk of
remaining energy poor in the future.
In a further step, the dynamics of energy poverty are investigated by
means of survival analysis. This analysis reveals that both transitory and
persistent aspects of energy poverty exist. A decreasing trend in the condi-
tional probability of exiting energy poverty illustrates the transitory element,
while the significant proportion of households unable to exit energy poverty
throughout the observation period as well as statistically significant state
dependence indicate the persistence of the issue.
This study is based on the capability approach and emphasizes that ad-
dressing energy poverty should be done not only by increasing income, but
also by improving individuals’ capabilities to achieve well-being. The conclu-
sions drawn from this study serve as a basis for informed policy decisions and
enable the development of targeted policies to address and alleviate energy
poverty.
Keywords: Energy poverty; Dynamic random effects probit; Survival analysis;
Capability approach.
Contents
1 Introduction 1
2 Literature Review and Theoretical Framework 2
2.1 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.1.1 Energy Poverty as an independent poverty measure . . . . 2
2.1.2 Evolution and Dynamics of Energy Poverty . . . . . . . . . 3
2.2 Theoretical Framework . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3 Hypotheses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3 Data and Empirical Strategy 8
3.1 Data and Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.2 Key Measurement Concepts of Energy Poverty . . . . . . . . . . . 12
3.2.1 10 percent Indicator . . . . . . . . . . . . . . . . . . . . . . 12
3.2.2 LIHC Indicator . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2.3 2M Indicator . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.3 Random Effects Estimator . . . . . . . . . . . . . . . . . . . . . . . 13
3.4 Survival Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4 Empirical Results 18
4.1 Extent and Development of Energy Poverty . . . . . . . . . . . . . 18
4.2 Energy Poverty Determinants . . . . . . . . . . . . . . . . . . . . . 20
4.3 Energy Poverty Dynamics . . . . . . . . . . . . . . . . . . . . . . . 23
4.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
5 Conclusion 30
I References 32
II Appendix 37
i. Supplementary Data . . . . . . . . . . . . . . . . . . . . . . . . . . 37
ii. Robustness Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
List of Tables
1 Defining categories of energy capabilities according to Sen’s and
Nussbaum’s capability approach . . . . . . . . . . . . . . . . . . . 6
2 Income profiles and average monthly expenditures on domestic en-
ergy services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3 Descriptive statistics . . . . . . . . . . . . . . . . . . . . . . . . . . 10
4 Income profiles and average monthly expenditures on domestic en-
ergy services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
5 Average marginal effects - energy poverty determinants . . . . . . 21
6 Distribution of periods spent in energy and income poverty . . . . 23
7 Average year-to-year movements . . . . . . . . . . . . . . . . . . . 24
8 Survival analysis - exits out of energy poverty . . . . . . . . . . . . 24
9 Policy recommendations . . . . . . . . . . . . . . . . . . . . . . . . 28
List of Figures
1 Percentage share of energy-poor households according to the dif-
ferent indicators . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1 Introduction
Energy poverty - an increasingly prominent issue in Europe that particularly
affects low-income households due to rising energy prices. It is a phenomenon that
in developed countries is closely linked to the affordability of a minimum package
of energy services, including lighting, cooking, heating, and cooling (Faiella and
Lavecchia, 2021). According to European Commission reports, in 2022, 38 million
Europeans - around 8% of the total EU population - were unable to adequately
heat their homes (European Commission, 2022). Unlike other markets, there
is no possibility of a general market exit in the energy sector, as energy is an
indispensable commodity. This underscores the critical role that energy plays in
daily life and highlights the need to develop policies that address energy poverty
by providing every household with access to affordable energy to meet basic needs
without undue burden.
However, despite the increased attention energy poverty has gained in Europe
as well as the development of various initiatives to address this issue, there are
significant discrepancies. The discrepancy stems from the absence of a uniform
calculation method to quantify energy poverty, which makes it difficult to draw
comparable conclusions about the extent of the problem and possible policy im-
plications (European Parliamentary Research Service, 2022). Against this back-
ground, this thesis aims to analyze the development of energy poverty in Germany
over the period from 2016 to 2020 and to investigate and compare determinants
and dynamics.
In the first step, a dynamic random effects probit estimation is conducted to
identify the determinants of energy poverty and answer the research question:
1. What socioeconomic factors are associated with the risk of a household ex-
periencing energy poverty in Germany, and how do these factors impact the
capabilities to access affordable, reliable, and safe energy services?
The model reveals that factors like household composition, employment, and living
area are strongly correlated with the prevalence of energy poverty. The study also
uncovered significant state-dependent effects, revealing that if households were
energy poor in a previous period, they were up to 6.6% more likely to face energy
poverty in the subsequent period.
In the second step, this master thesis introduces a novel approach to un-
derstanding the dynamics of energy poverty in Germany by employing survival
analysis. This step aims to answer the research question:
2. How do households transition in and out of energy poverty over time, and
how do patterns of persistence influence their ability to overcome energy
poverty?
1
The analysis of energy poverty dynamics highlights both transitory and persis-
tent aspects of the issue, with survival analysis illustrating the temporary nature
through a decreasing trend in the conditional probability of exiting energy poverty.
On the other hand, the influence of state-dependent effects found in the probit
estimation and the significant proportion of households unable to exit energy
poverty throughout the observation period underline its persistent element.
The study employs the capability approach as a theoretical framework and
emphasizes that combating energy poverty is not simply about increasing income,
but about improving the ability of individuals to achieve well-being. The conclu-
sions drawn therefore provide a basis for informed policy decisions and enable the
development of targeted policies to address and alleviate energy poverty based on
the capability approach.
2 Literature Review and Theoretical Framework
2.1 Literature Review
2.1.1 Energy Poverty as an independent poverty measure
Energy poverty is a complex and multifaceted phenomenon. Defining energy
poverty, along with addressing its relationship to income poverty is an important
challenge for energy poverty research. There are two largely independent debates
on energy poverty in developing and industrialized countries. In developing coun-
tries, energy poverty refers to an inadequate supply of modern forms of energy,
such as electricity as many people rely exclusively on traditional fuels (Sagar,
2005). In industrialized countries, on the other hand, energy poverty is related to
a lack of affordability and the associated restrictions on basic needs (Bouzarovski
and Petrova, 2015), which will be the focus of this paper. The term energy poverty
was first introduced in the academic literature by Isherwood and Hancock (1979)
in the course of the 1970s oil crisis. Building on this, modern energy poverty
research begins with the dissertation of Boardman (1991), where she articulates
energy poverty in response to the rising prices of fossil fuels. In it, she defines
energy poverty as follows:
”Energy poverty is the inability to afford adequate heat due to energy
inefficiency in the home”
Building on this definition, the United Kingdom established an official definition
of energy poverty in 2001 and introduced a strategy to address it (Liddell et al.,
2012). According to the national definition, a household is considered energy poor
if it cannot afford to provide adequate heat at a reasonable cost (DEFRA, 2001).
2
After the official adoption of the concept by the UK, energy poverty research has
expanded, predominantly in Europe.
A number of voices in the literature consider energy poverty to be already
sufficiently covered by general poverty research and alleviation (Hills, 2011; Bun-
destag, 2017; Kopatz, 2013). However, while there is a significant overlap between
households experiencing energy poverty and income poverty, the majority of re-
searchers argue that energy poverty is a unique form of poverty with distinct
causes and solutions (Hills, 2011; Strünck, 2017).
Heindl et al. (2017) assert that energy poverty links directly to the satisfaction
of basic needs, particularly the consumption of essential domestic energy services
such as heating, hot water, lighting, and operating major electrical appliances.
Barnes et al. (2008) and Churchill et al. (2020; 2021) see the urgency of energy
poverty as a distinct issue underscored by the disproportionate rise in energy prices
and growing inequality in Germany, in addition to severe repercussions such as
reduced physical and mental health, and limited educational opportunities for
children. The Ukraine war and its impact on energy prices in the recent past only
intensified the urgency of addressing energy poverty. Tews (2013) argues that
while both low income and high energy costs are constituent factors to energy
poverty, they are not structural causes. He sees the issue of energy poverty as
unique due to the specific impacts of energy efficiency on costs for households.
Mayer (2013) and Hills (2011) find that energy costs can vary substantially for
households with similar income levels due to the differing energy efficiency of their
homes. This point is echoed by Boardman (1991) who asserted that while general
poverty could be addressed through income support, mitigating energy poverty
requires capital investment in the energy efficiency of homes and heating systems.
This distinguishing feature necessitates specific policies targeting energy poverty.
2.1.2 Evolution and Dynamics of Energy Poverty
There are numerous country-level and cross-country studies on energy poverty,
with a focus on quantifying and explaining the prevalence of energy poverty at a
specific point in time. The majority of research on energy poverty has been con-
ducted in European countries (Meyer et al., 2018; Aristondo and Onaindia, 2018;
Karpinska and Śmiech, 2020). Meyer et al. (2018) developed an energy poverty
barometer using a set of complementary indicators to capture the multifaceted na-
ture of energy poverty in Belgium. Results from their barometer indicator show
an energy poverty rate of 21.3% in at least one of its forms between the years
2009 and 2015. The paper by Aristondo and Onaindia (2018) reveals that energy
poverty in Spain worsened between 2004 and 2015, with a greater than 25% in-
crease in individuals experiencing deprivation in at least two or three dimensions,
3
and identifies higher energy poverty levels in rural areas compared to densely pop-
ulated ones. Cross-sectional analyses in the European Union by Thomson et al.
(2017) have highlighted the lack of sufficient data at EU level and found that en-
ergy poverty is more prevalent in Southern and Eastern European Member States
(Bouzarovski and Simcock, 2017; Herrero, 2017).
Energy Poverty, however, is not solely researched in Europe. There are also
various studies on the determinants and trends of energy poverty in non-European
countries such as Japan and the United States (Okushima, 2016; Teller-Elsberg
et al., 2016). Studying the development in Vermont, Teller-Elsberg et al. (2016)
reveal that energy poverty in Vermont has increased by 76% between 2000 and
2012, affecting one in five residents.
Research on energy poverty in Germany is relatively limited, with Heindl and
Schuessler (2015) and Neuhoff et al. (2013) being the first to address the issue of
energy poverty in Germany. Heindl and Schuessler (2015) evaluate the dynamic
behavior of various energy poverty indicators, concluding that common indicators
such as the 2M and LIHC, fail to capture crucial changes in expenditure or income.
Instead, they introduce variations to the existing energy poverty indicators such
as combining the first condition of the LIHC indicator (see equation 3) with the
10 percent indicator. They argue that this so-called LIHCt indicator has a better
ability to capture changes in expenditures and income. Neuhoff et al. (2013)
examined the distributional impact of renewable energy costs (”EEG Umlage”)
on households. The study reveals that while German consumers are bearing the
financial burden through their electricity bills, the impact is more pronounced for
poorer households, necessitating interventions such as adjusted transfers, reduced
electricity taxes, and enhanced energy efficiency measures.
Various studies have examined the household-level determinants of energy
poverty to highlight the complex interplay between household-level factors and
energy poverty (Legendre and Ricci, 2015; Drescher and Janzen, 2021).Legendre
and Ricci (2015) employ logit, C log-log, and mixed-effect logit models to bet-
ter identify and understand fuel vulnerability in French households. Drescher
and Janzen (2021) use a dynamic random effects probit model to investigate the
determinants of the transition between different states of energy poverty in Ger-
many. These studies found significant correlations between factors like income,
energy expenditure, household composition, education, housing conditions, and
the likelihood of being in energy poverty.
Few studies have investigated the dynamics of energy poverty using longitu-
dinal data, with Phimister et al. (2015) for Spain, Chaton and Lacroix (2018) for
France, Karpinska and Śmiech (2020) for Poland, and Drescher and Janzen (2021)
for Germany being some notable examples. The studies use different analysis ap-
proaches to investigate the dynamics of energy poverty. Phimister et al. (2015)
4
utilize a survival analysis to identify the risk factors associated with entering and
exiting energy poverty, while Chaton and Lacroix (2018) employ a mover-stayer
model on French household data. Similarly, Karpinska and Śmiech (2020) used
cluster analysis to identify different profiles of energy-poor households in Poland.
Despite these methodological differences, all these studies suggest that en-
ergy poverty is a complex and dynamic phenomenon, with households moving in
and out of energy poverty over time. All studies conclude that energy poverty
is predominantly a transitory state, with the majority of households not being
persistently energy poor.
The Master thesis contributes to the existing literature by offering an updated
assessment of the level of energy poverty in Germany, exploring its determinants
based on the model employed by Drescher and Janzen (2021), and presenting a
novel examination of the energy poverty dynamics in Germany using panel data,
similar to the approach taken by Phimister et al. (2015).
What sets this thesis apart from previous research is the use of survival anal-
ysis to quantify the probabilities of exiting or staying in energy poverty. To the
best of my knowledge, this approach has not been applied before in the context
of measuring energy poverty in Germany, and thus constitutes a valuable contri-
bution to the field. By quantifying the dynamics of energy poverty in this way,
the thesis provides insights into the probabilities of transitioning into and out
of energy poverty over time. Additionally, this thesis contributes to the field by
integrating the capabilities approach, to be introduced in the subsequent section,
into the analysis of determinants and dynamics of energy poverty. Overall, the
research represents an important step forward in the understanding of energy
poverty in Germany and how to formulate appropriate policy measures to combat
this phenomenon.
2.2 Theoretical Framework
Energy poverty occurs in different forms and intensities and therefore affects
households differently. The complexity of energy poverty goes beyond a house-
hold’s energy expenditures including external factors such as socioeconomic house-
hold factors, climate variability, energy prices, and residential energy efficiency
(Lee et al., 2021). In this thesis, the capability approach is utilized as the theo-
retical framework for addressing energy poverty as an issue of social justice. The
capability approach, initially formulated by Nobel laureate Amartya Sen (1985;
2004; 2005), extended by Martha Nussbaum (2003) and other prominent social
scientists (Alkire, 2005; Robeyns, 2005), asserts that real development must be
gauged not merely by an increase in income, but the enhancement of people’s
capabilities to achieve well-being. In contrast to utilitarian or resource-based per-
5
spectives on social justice, the capability framework underlines a strong connection
between human freedom and well-being (Sen, 2009). This approach regards the
issue of energy poverty as a state of capability deprivation that prevents people
from realizing their life goals and values because they do not have sufficient access
to affordable, reliable, and secure energy sources. The capability approach invites
a shift in focus from ”who gets how much of what” to ”what needs to be done
in order to help everybody live a fully human life” (Berges, 2007). It suggests a
context-specific understanding of energy poverty issues and calls for differentiated
solutions that take into account the diversity of human needs. Sen (2004) empha-
sizes public discourse and reasoning in operationalizing the capability approach
and argues for policies tailored to specific cultural and climatic contexts.
To utilize the approach of a capability-based view on energy poverty, this paper
proposes an assessment framework with three categories of energy capabilities
derived from the approaches of Sen, extended by Nussbaum:
Table 1: Defining categories of energy capabilities according to Sen’s and Nuss-
baum’s capability approach
Secondary
Categories of
Categories of energy capabilities Energy
capabilities
Capabilities
Space heating and
Capabilities related Energy capabilities related to biological cooling, lighting,
to biological and and physical needs. Energy use has a cooking, access to
physical human direct impact on health by providing a water and food,
needs livable environment and food supply. personal hygiene,
medical care etc.
Energy capabilities related to Skills for accessing
Capabilities related
intellectual and emotional needs. information,
to fundamental education,
Energy use can contribute to the
interests of a transportation,
development of intellectual and
human agent social interaction
emotional skills. with others etc.
Energy capabilities related to social
Capabilities related and political needs. People should be Participation in
to fundamental empowered to make energy-related decision-making
processes, resistance
interests of a social decisions that affect their ability to
to unjust energy
being pursue valued goals and live with decisions
dignity.
Sources: Representation of the categories is borrowed from Lee et al. (2021); Categories of capabilities are borrowed
from (Nielsen and Axelsen, 2016); relevant capabilities are based on Nussbaum’s list of central human functional
capabilities (Nussbaum, 2003)
Energy services are essential for maintaining a fully human life and achieving
one’s goals and values (Lee et al., 2021). Empirical studies show that energy use
correlates with a wide range of essential capabilities, from maintaining health to
6
developing intellectual skills and social relationships (Barnes et al., 2008; Churchill
et al., 2020; Churchill and Smyth, 2021). Utilizing the capability approach, this
paper acknowledges the multitude of human needs and the corresponding array of
functions that capabilities can fulfill, thereby avoiding a deficit-oriented view that
confines energy poverty to income poverty. This avoids oversimplification of the
complex and diverse experience of energy poverty. The emphasis of the capability
approach on factors beyond household income underlines the multiple dimensions
that contribute to energy poverty. Based on this theoretical framework as well
as the analysis of determinants and dynamics, this paper aims to develop policies
that address the three categories of energy capabilities.
2.3 Hypotheses
In accordance with literature and the theoretical framework presented above, this
thesis aims to explore the relationship between energy poverty, household charac-
teristics, and the dynamic nature of the issue. This investigation is guided by the
formulation of three specific hypotheses, each designed to shed light on one of the
three capability categories of the underlying theoretical framework of this study.
The focus of hypotheses 1a and 1b is on various socio-economic factors and
their influence on a household’s energy poverty risk. It is proposed that these
factors interact with energy capabilities related to both biological/physical needs
and intellectual/emotional needs.
Hypothesis 1a: Lower household income significantly increases the risk of ex-
periencing energy poverty. This is due to the direct influence of financial resources
on the ability to access and utilize energy services.
Hypothesis 1b: Higher levels of education within a household decrease the risk
of energy poverty. More educated households have better access to information
about energy efficiency and affordable energy services, which helps them make
more informed energy consumption decisions.
Given the positive correlation between education and income (see table A.4 in
the appendix), hypothesis 1b seeks to identify the nuanced interplay between
education and the risk of experiencing energy poverty, raised in the capability
approach. This is based on the understanding that capabilities such as educa-
tion do not exist in isolation, but are interrelated and can be strongly influenced
by factors such as income. Education itself, in the context of its relationship to
income, can uniquely influence a household’s risk of energy poverty by influenc-
ing the household’s access to information about energy efficiency and affordable
energy services, enabling better decisions about energy use.
7
This paper does not only consider the static state of energy poverty but as-
sumes that energy poverty, much like the capabilities approach that guides this
paper, is inherently dynamic. This aligns with the aspect of ’Energy capabilities
related to social and political needs,’ emphasizing the need for individuals to be
empowered to make energy-related decisions, which can influence their energy
poverty status and their ability to pursue valued life goals with dignity.
Hypothesis 2: The majority of households in Germany experience energy poverty
as a transitory, rather than a persistent, condition. This hypothesis aligns with the
idea that changes in energy capabilities, particularly related to social and political
needs, can lead to shifts in energy poverty status.
The selection of these hypotheses aims to enhance our understanding of energy
poverty’s complexities and dynamics, leading to more informed policy measures.
The choice is informed by economic theory, particularly the capabilities approach,
and strives to uncover unique and significant associations that have not been
fully explored in previous studies as elaborated in section 2.1.2. The hypotheses
will be operationalized in Section 3.3 and Section 3.4. However, prior to this,
the subsequent chapter will provide an introduction to the data and the key
measurement concepts of energy poverty that are utilized in this paper.
3 Data and Empirical Strategy
3.1 Data and Variables
To analyze the determinants as well as the dynamics of energy poverty in Germany,
this paper draws on the dataset of the German Socio-Economic Panel (SOEP).
The SOEP is a national, representative household survey that is conducted annu-
ally since 1984 under the auspices of the German Institute for Economic Research
(DIW). The survey covers selected households and all persons living in them, gen-
erating longitudinal data on income, consumption expenditure and housing. As
representative microdata, the SOEP data is available to all researchers in the dis-
ciplines of social, behavioral, and economic sciences. The annual survey rhythm of
the SOEP allows for an analysis of energy poverty over time. This paper uses the
latest version (v37) of the SOEP dataset, which includes data up to and including
2020. For the analysis, data from 2013 and subsequent years (v30-v37) will be
considered.
The present dataset encountered numerous missing values in the SOEP dataset,
so a rigorous process of data cleaning and filtering was required to ensure the
quality and validity of the data. As a result, some variables, such as the primary
energy source were excluded from the analysis due to the scarcity of the available
8
data. It should be noted that the absence of such information could potentially
bias the present analysis, as primary energy sources can be highly correlated with
energy efficiency. In addition, the dataset did not include descriptive variables on
household living conditions, also potentially correlated with energy efficiency and
thereby total energy costs.
In order to mitigate the impact of panel attrition, this paper focuses exclusively
on households that exhibit consistent participation throughout the time frame
spanning 2016 to 2020. Confining the sample to households that were present
throughout the five-year period intentionally minimizes potential bias and in-
creases internal validity of the study’s results. The resulting balanced sub-sample
includes 24.230 individuals from 2.346 households spanning a five-year period. To
the best of my understanding, this sub-sample is free of inconsistencies, errors, or
missing values.
Table 2: Income profiles and average monthly expenditures on domestic energy
services
Q1 Q2 Q3 Q4 Q5 Sample Mean
Household Income 932 1 479 1 935 2 575 4 508 2 259
Electricity Costs 49 52 62 69 77 62
Heating Costs 67 74 94 100 108 88
Total Energy Costs 116 126 156 169 185 150
*all values in Euro (EUR)
Table 2 provides a detailed overview of household income profiles and average
monthly expenditure on energy services of the pooled sample from 2016 to 2020.
The data is divided into five income quintiles (Q1 to Q5). For each income
quintile, the table shows the average household income and total energy costs,
which is the sum of electricity costs and heating costs. A trend can be observed
in which higher income quintiles correlate with increased household income as
well as increased energy costs. This correlation suggests that households with
higher incomes tend to spend a greater amount on energy services. The sample
mean acts as a comparative measure, as it represents the average values of each
variable within the pooled sample. It is significant to note that the sample mean
household income is EUR 2259, while the de facto mean net income in Germany
is EUR 2005 (Statista, 2023). This discrepancy could limit the direct transfer-
ability of the results to the entire German population and thus poses a potential
threat to external validity.
Table 3 presents the descriptive statistics of the explanatory variables that,
in addition to household income and total energy costs, potentially influence the
probability of energy poverty in Germany. The selection of covariates is based
9
Table 3: Descriptive statistics
Description Frequency
Household Type
Couple, no Children 25.84%
Single Parent 35.41%
One-Person Household 16.14%
Couple with Children 22.61%
Employment =1 if employed 56,44%
House Type
Detached 13.06%
Semi-attached 11.19%
Apartment Building 75.74%
Migration Background =1 if migrational background 26.72%
Years of Education
No Degree 36.83%
Lower Secondary Degree 21.60%
Higher Secondary Degree 19.60%
Tertiary Degree 21.97%
Living Area
12-67 sqm 31.91%
68-80 sqm 21.42%
81-100 sqm 21.24%
101-270 sqm 8.27%
Owner =1 if owner 0.12%
10
on the existing literature on the determinants of energy poverty (Schreiner, 2015;
Cappellari and Jenkins, 2014). Relevant factors include aspects such as household
composition, employment status, house type, living area, and education level.
Education level and employment status provide information on human capital and
labor market participation. The precision of the analysis regarding the variable
employment is subject to limited status differentiation. The available data only
differentiates between employed and unemployed persons, which does not allow
for a more detailed categorization, as no distinction is made between pensioners,
unemployed persons, as well as part-time and full-time employees.
A dummy variable is integrated to indicate whether the household head has a
direct migration background. This variable is based on empirical findings sug-
gesting that race and ethnicity positively influence household energy demand
and make migrant households more vulnerable to energy poverty (Drescher and
Janzen, 2021). In addition, a migrant background is negatively correlated with
household income (see table A.4 in the appendix).
In accordance with OECD guidelines for statistics, household income, and en-
ergy expenditure variables were adjusted using a square root scale. This involved
dividing the variables by the square root of household size to produce energy
poverty indicators in a consistent and standardized manner. The application of
this scale results in a lower weighting for each additional household member, which
is considered plausible as it accounts for economies of scale in consumption and
assumes that additional household members do not generate the same energy con-
sumption as the first members (OECD, 2013). This deflation of variables allows
for more accurate and comparable measures of energy poverty in households of
different sizes.
The analysis also takes into account the distinction between owners and ten-
ants. In the surveyed population, a mere 0.12% of households were identified as
homeowners, primarily attributed to their exclusion from the analytical sample
due to their failure to provide complete data on overall energy expenditures. It
is significant to highlight that the low proportion of homeowners in our sample
is not representative of the overall population in Germany. In 2018, 42.1% of
households lived in their own home, while around 57.9% were renting (Statista,
2021). Despite the small number of homeowners in the subsample, the inclusion
of this variable can provide valuable insight into potential differences in the ex-
perience of energy poverty between homeowners and tenants, thus deepening the
overall understanding of the issue. Therefore, these few observations are retained
in the subsample. However, the reduced presence of homeowners in our sample
may limit the external validity of our results and may prevent a generalizable
transfer of findings to all German households. Finally, federal state fixed effects
are incorporated in the analysis to account for differences in electricity grid usage
11
across states and fixed effects for different waves to control for both observed and
unobserved wave-specific factors, such as temperature variations.
3.2 Key Measurement Concepts of Energy Poverty
In energy poverty research, different indicators are used to quantify energy poverty.
In Europe alone, Rademaekers et al. (2016) record 77 different operationalizations
of energy poverty. The best-known indicators of energy poverty fall into two cat-
egories: consensual- and expenditure-based indicators.
The consensual approach is a method of measuring poverty based on identi-
fying areas of deprivation in basic needs through a consensus of both objective
and subjective indicators. Consensual indicators are usually collected by ques-
tionnaires or interviews.
Expenditure-based energy poverty measures refer to methods of measuring
energy poverty based on a household’s ability to afford adequate energy services,
such as heating and electricity using income and expenditure data.
This master thesis focuses on the three most prominent expenditure-based in-
dicators for quantifying energy poverty in Germany, which are derived from var-
ious interpretations of the interplay between high energy costs and low income.
Due to the lack of a national definition for energy poverty in Germany, this paper
deliberately does not commit to a single definition of energy poverty but considers
and compares different definitions in order to gain a comprehensive understanding
of energy poverty. This approach is consistent with the literature, as other stud-
ies also often refrain from using a single definition of energy poverty and instead
only concretize an operational definition in the form of a measurement construct
(Drescher and Janzen, 2021; Heindl and Schuessler, 2015; Phimister et al., 2015).
This paper does not evaluate consensual indicators due to the limitations in avail-
able data. All indicators studied are based on the percentage of annual average
household income (y) spent on annual average total energy costs (e), which is
represented by σ in the first step of the quantification process.
e
σ = (1)
y
3.2.1 10 percent Indicator
The 10 percent indicator is the best-known measure of energy poverty. The 10
percent threshold originates from Boardman (1991). According to this concept, a
household is considered energy poor if the share of energy expenditure in income
(σ) is greater than 10 percent.
σi > 10% (2)
12
In principle, however, any other threshold of the form ”X” percent share of
energy expenditure in income is conceivable. This indicator represents an absolute
threshold of poverty, given that it is primarily concerned with the absolute level
of energy expenditure in relation to household income. This approach does not
consider the distribution of income within the population, making it independent
of relative poverty lines used in other poverty measures.
3.2.2 LIHC Indicator
The low-income-high-cost (LIHC) indicator originated from Hills (2011, 2012).
Since its elaboration, it has replaced the 10 percent measure as the official energy
poverty indicator in England. According to the LIHC indicator, a household
is identified as energy poor if two conditions are met: (1) energy expenditures
are higher than those of the median household (ē) and (2) income net of energy
expenditures is below an income poverty threshold of 60% of the populations
median income (ȳ). Both thresholds are thus relatively defined.
ei > ē
(3)
yi − ei < 0.6 ȳ
3.2.3 2M Indicator
The 2M indicator approach to energy poverty suggested by the EU Energy Poverty
Observatory (EPOV) is based on an income poverty line that captures the burden
that energy bills put on households relative to their disposable income, using
the national median of energy expenditures (σ̄) as a reference point (European
Commission, 2023). This indicator represents a relative poverty measure and
households are considered energy poor if the share of energy expenditure as a
percentage of income is more than twice as high as that of the median household.
σi > 2σ̄ (4)
The measure does neither capture low-income households that under-consume
relevant energy services nor high-income households with proportionally high en-
ergy expenditures that may be defined as energy poor. It is dependent on the
distributions of income and energy expenditures, and thresholds may change over
time.
3.3 Random Effects Estimator
Following the approach of Drescher and Janzen (2021) as well as Alem and Demeke
(2020), a dynamic random effects probit estimation is employed to identify the
13
determining factors of energy poverty. This approach is often used in research to
capture state dependence on income poverty, unemployment or social assistance
benefit receipt (Biewen, 2009; Stewart, 2007; Cappellari and Jenkins, 2014). In
the context of energy poverty, state dependence means that an energy-poor person
or household in the previous period is more likely to experience energy poverty in
the current period. This assumption could hold for a variety of reasons, including
inadequate infrastructure, low energy efficiency, lack of access to affordable energy,
or persistently low income. Static panel models assume that previous energy
poverty states do not influence the current probability of experiencing energy
poverty.
Unlike a fixed effects model, the dynamic random effects model incorporates
both within-household and between-household variation. This is particularly rele-
vant with a relatively short observation period of five years, during which socioeco-
nomic household characteristics are not likely to change significantly. By utilizing
both types of variation, the random effects model can provide more efficient esti-
mates and a more comprehensive understanding of the factors influencing energy
poverty, making it superior to the fixed effects model in this context. Additionally,
the random effects model also has the ability to capture unobserved heterogeneity
among households that could affect the energy poverty status. In the context of
energy poverty, these unobserved factors may include cultural practices, house-
hold preferences, or simply household energy management. These unobserved
household effects are assumed to be independent of the observed variables and
random shocks, i.e., they are not systematically correlated with the observed ex-
planatory variables. By controlling and appropriately accounting for unobserved
heterogeneity, biased and inconsistent estimates can be prevented.
In the context of the present Master thesis, the dependent variable of interest,
yit, denotes the observed binary outcome for household i at time t. The latent
index form of the model can be represented as follows, where y∗it represents a
latent dependent variable for unit i at time t:
y ∗it = 1 [yit ≥ 0] (5)
y∗it = αyit−1 + βX
′
it + ui + εit (6)
The variable yit−1 denotes the energy poverty status in the previous period,
′
while Xit is a vector of exogenous explanatory variables. The error term εit is
assumed to follow a normal distribution. The unobserved factors contributing
to energy poverty are captured by the random intercept ui for each individual
or household. This intercept represents the unique, time-invariant characteristics
of each household that are not directly accounted for by the observed explana-
tory variables. The individual-specific intercepts are assumed to be normally
14
distributed across the population. Furthermore, the error term εit and the ex-
′
planatory variables Xit are assumed to be independent of ui, implying that the
unobserved individual effects do not depend on the observed variables or random
shocks.
Utilizing the model, the probability of a household encountering energy poverty
at time t (yit = 1) can be assessed, given their poverty status in the previous period
′
(yit−1), the explanatory variables (Xit), and the unobserved household-specific ef-
fect (ui).
However, this approach can suffer from the initial conditions problem, which
refers to the potential correlation between the initial observation yi1 and the time-
invariant individual-specific effect ui. This issue may arise when the stochastic
process begins before the observation period. The initial observation may be
affected by both the random intercept and pre-sample responses, with the impact
becoming particularly pronounced when the observation period is relatively short.
Assuming the initial state of the dependent variable to be exogenous may yield
inconsistent estimates of the parameters in the dynamic random effects model.
To address the initial conditions problem, the Woodlridge conditional maxi-
mum likelihood (WCML) estimator proposed by Wooldridge (2005) is employed.
This method models the distribution of binary receipt indicators from year two
(2017) to the final year (2020), conditioning on the set of explanatory variables
and the binary receipt indicator for the initial year. By doing so, the method
accounts for the effect of individual-specific effects and pre-sample responses on
the initial binary receipt indicator. The WCML estimator is computationally
less intense than other methods such as modeling the initial dependent variable
jointly with the subsequent response (Heckman, 1987) and easier to implement
using standard statistical software (Drescher and Janzen, 2021).
The individual-specific intercept can be modeled as:
ui = η0 + η1yi1 + η X̄
′ + η X ′ + a a |X̄ ′, y ,X ′ ∼ N(0, σ22 i 3 i1 i i i i1 i1 a) (7)
where yi1 and Xi1 represent the initial value of the response variable and
time-varying explanatory variables, respectively, X̄ ′i represents the within-mean
of the time varying exogenous variables and ai is a unit-specific time-constant
error term, normally distributed with mean zero and variance σ2a.
Assuming that unobserved heterogeneity is captured by ui, the t − 1 lagged
value of the response variable can be interpreted as genuine state dependence,
or the causal effect of experiencing poverty in one wave on the experience of
poverty status in the subsequent time point. Accounting for initial conditions,
15
the response probability for yit becomes:
Pr(yit = 1|X ′ ′i1, X̄i, yi1) =
(8)
ϕ(αy ′it−1 + βxit + η0 + η1yi1 + η2X̄′i + η3X′i1 + ai + εit)
where Φ(.) denotes the normal cumulative distribution function. In this con-
text, the latent index model can be expressed as:
y∗it = αy
′
it−1 + βXit + η0 + η1yi1 + η2X̄′i + η3X′i1 + ai + εit (9)
The primary coefficient of interest, α, signifies the genuine state dependence of
energy poverty. In order to accommodate potential heteroskedasticity and auto-
correlation, standard errors are clustered at the household level. The model is
calculated both with and without initial conditions for each individual indicator,
yielding robust and dependable outcomes.
3.4 Survival Analysis
After quantifying the determinants of energy poverty, the dynamics of energy
poverty are addressed in a second step. Survival analysis is utilized in this paper
to examine the transitions into and out of energy poverty and to gain insight into
its underlying dynamics. This method has already been used by Phimister et al.
(2015) to analyze the dynamics of energy poverty in Spain.
Survival analysis is a statistical technique for analyzing time-to-event data,
where the event of interest in this context is the exit from energy poverty. It is par-
ticularly useful in situations where the outcome of interest is not only whether an
event occurs, but also when it occurs. Examples of the use of survival analysis in
research include analyzing the duration of unemployment (Ciuca and Maer Matei,
2010; Khoo et al., 2022) or the time to exit poverty (Callander and Schofield, 2016;
Chen et al., 2022).
The main objective is to examine the duration of energy poverty spells and the
likelihood of exiting energy poverty over time. A spell is a continuous period of
time during which a particular condition is maintained. In this paper, we focus on
two types of energy poverty spells: time spent in energy poverty and time spent
out of energy poverty. A spell begins when a household’s energy poverty status
changes (either entering or exiting energy poverty) and ends when the status
changes again. To ensure an unbiased analysis, left-censored spells are excluded
from the dataset, i.e., observations where we do not have information on when a
household originally entered (or exited) energy poverty. Only non-left-censored
spells where the beginning of each period is observed are considered in the analysis.
This results in a maximum available period of 4 years. The analysis also includes
16
right-censored spells, where the end of a period is unknown. Despite the short
duration of the periods, the analysis examines whether the probability of emerging
from energy poverty decreases as the duration of energy poverty increases.
Two fundamental concepts in survival analysis are the survivor function and
the hazard function. The survivor function represents the probability that a house-
hold will remain in energy poverty for at least t periods, under the condition that
the household has just entered energy poverty. It is calculated as the complement
of the cumulative distribution function (CDF) of the time-to-event variable. An
event in this context refers to the specific change in a household’s energy poverty
status. This could either be a household entering or exciting energy poverty at a
specific point in time t.
Mathematically, the survivor function is defined as:
S (t) = P (T ≥ t) = 1− F (t) (10)
where F (t) is the CDF of the time-to-event variable. The survivor function can
be estimated using the Kaplan-Meier estim(ator, gi)ven by:∏ di
Si (t) = 1− (11)
ni
i:ti≤t
where ti is the time of the i
th observed event (a household transitioning in or out
of energy poverty), di is the number of events at time ti, and ni is the number of
households at risk (households that have not yet experienced the event) at time
ti. The Kaplan-Meier estimator is a non-parametric method used to estimate the
survival function from censored data, taking into account both observed event
dates and censored observations, allowing for a more accurate estimate of the
survival function.
The hazard function, symbolized by h(t), quantifies the instantaneous tran-
sition rate between states (into or out of energy poverty) at time t under the
condition that the household has survived until that time. In other words, the
hazard function measures the conditional probability that a household will expe-
rience a change in energy poverty status at a given point in time, assuming that
it has maintained that status until that point in time.
The hazard function is related to the survivor function by:
−Sit − Sit−1hi(t) = (12)
Sit−1
In summary, survival analysis provides a valuable method for studying the
dynamics of energy poverty by examining the duration of energy poverty spells
and the factors influencing the likelihood of transitioning in and out of energy
17
poverty. The concepts of the hazard and survivor functions enable us to gain
insights into the underlying mechanisms.
4 Empirical Results
4.1 Extent and Development of Energy Poverty
Examining energy poverty rates in Germany between 2016 and 2020 using the
three indicators presented in Section 2.3 shows insightful results on the extent
and development of energy poverty. Table 4 illustrates the energy poverty rates
as a function of income quintiles. The results underline the great importance of
income level in relation to energy poverty. On the one hand, households with
lower incomes seem to be more at risk of energy poverty. On the other hand,
the results show that energy poverty and income poverty differ from each other,
especially when looking at the indicators that do not include household income
in the measurement (10 percent indicator and 2M indicator). Notably, the 2M
and the 10 percent indicator reveal that even households in the highest income
quintiles can be classified as energy poor.
Table 4: Income profiles and average monthly expenditures on domestic energy
services
Q1 Q2 Q3 Q4 Q5
2M Indicator 29.22% 7.59% 4.62% 1.54% 0.04%
LIHC Indicator 59.39% 22.08% 5.19% 0.00% 0.00%
10% Indicator 50.47% 21.17% 11.97% 3.23% 1.06%
Trends in energy poverty by different indicators are shown in Figure 1, where
panel (a) shows households in the entire sample and panel (b) shows households
from the lowest three income quintiles. Looking at the entire sample, the 10 per-
cent indicator indicates that energy poverty has declined since 2016. The percent-
age of households that spent more than 10% of their disposable income on energy
decreased from 21.34% in 2016 to 15.19% in 2020. In contrast, energy poverty has
increased to varying degrees under the other two indicators: The percentage of
households classified as energy poor under the 2M indicator increased from 7.57%
in 2016 to 10.99% in 2020, while the LIHC indicator remained relatively stable,
increasing from 17.17% in 2016 to 18.16% in 2020.
Looking at the sample of the lower income quintile in panel (b) a much higher
percentage of energy-poor households is presented. Trends in energy poverty here
are similar to those in the overall sample for the three indicators.
18
Figure 1: Percentage share of energy-poor households according to the different
indicators
The discrepancy between the trends in energy poverty for the 10 percent in-
dicator and the other two indicators (2M and LIHC) can be attributed to the
different methods used to calculate energy poverty. The 10 percent indicator is
an absolute threshold that considers only the share of household income spent on
energy. As energy costs on average increase slower than income increases in the
observation period, this measure shows a decrease in energy poverty.
The 2M indicator shows an increase in energy poverty when the gap between
households with higher energy expenditures relative to their income and the me-
dian household increases. The increase in energy poverty according to this in-
dicator shows that the relative position of households with high energy costs in
relation to their income in the total population is deteriorating.
The increase in energy poverty according to the LIHC indicator, in contrast
to the decrease according to the 10 percent indicator, could be due to several
factors. For example, energy costs may have increased for a significant portion of
the population, causing their energy expenditures to be higher than the median. It
also suggests growing income inequality, which affects net household income after
accounting for energy expenditures. The different trends between the indicators
highlight that these measures capture different dimensions of energy poverty.
19
4.2 Energy Poverty Determinants
Table 5 presents the average marginal effects of the dynamic probit random effects
models for various indicators of energy poverty. Columns (1), (3) and (5) present
the results based on the estimation of the model according to equation (6). Here,
the initial condition yit−1 is considered exogenous. Columns (2), (4) and (6)
show the results obtained using the conditional maximum likelihood estimator
proposed by Wooldridge according to equation (8). In this case, a distribution of
heterogeneity conditional on a household’s energy poverty status at the beginning
of the observation period is considered.
A strong statistically significant evidence of state dependence is found both
with and without controlling for initial conditions. When initial conditions are not
controlled for, the probability of being energy-poor increases by 40.0% (column
(1)), 45.6% (column (3)), and 17.5% (column (5)), respectively), if the household
was energy poor in the previous period. However, the coefficients are likely to be
biased upwards and do not reflect the true effects of state dependence (Heckman,
1987). Therefore, the interpretation below will focus on the results using the
WCML estimator.
The coefficients of interest remain positive, however, decrease when control-
ling for initial conditions. The chances of being classified as energy poor under
the 2M indicator increase by 3.5% if a household was classified energy poor in
the previous period (6.6% for the LIHC indicator and 3.6% for the 10 percent
indicator, respectively). These results closely align with the underlying analysis
of Drescher and Janzen (2021) as they also observed lower, however, statistically
significant state dependence when controlling for initial conditions.
The relationship between energy poverty and most other covariates is in line
with previous expectations. The analysis of the regression results with respect
to household type reveals that the statistical significance of a single parent is
eliminated when controlling for initial conditions for all indicators. For the 2M
indicator and the 10 percent indicator, a household consisting of a couple with
children loses statistical significance in the model using the WMCL estimator.
Being a one-person household is associated with a higher probability of experi-
encing energy poverty throughout all indicators, statistically significant at the 1%
level. The coefficients increase for all indicators when controlling for initial con-
ditions. This is in line with expectations as we assume that additional household
members do not generate the same energy consumption as the first members,
considering economies of scale (OECD, 2013). One possible explanation for the
loss of statistical significance in determining energy poverty for parents is the ex-
istence of family-oriented social policy measures in Germany, such as child and
family allowances. Such additional income could both reduce the likelihood that
20
Table 5: Average marginal effects - energy poverty determinants
10 Percent Indicator LIHC Indicator 2M Indicator
(1) (2) (3) (4) (5) (6)
Energy Poverty 0.400∗∗∗ 0.036∗∗∗ 0.465∗∗∗ 0.066∗∗∗ 0.175∗∗∗ 0.035∗∗∗t-1
(0.037) (0.005) (0.012) (0.010) (0.039) (0.008)
Household Type
Couple, no Children Ref Ref Ref Ref Ref Ref
Single Parent 0.049∗∗∗ -0.010 0.079∗∗∗ 0.038 0.030∗∗∗ -0.004
(0.007) (0.025) (0.006) (0.024) (0.005) (0.015)
One-person household 0.069∗∗∗ 0.162∗∗∗ 0.041∗∗∗ 0.048∗ 0.054∗∗∗ 0.091∗∗∗
(0.007) (0.036) (0.007) (0.025) (0.007) (0.029)
Couple, with Children 0.016∗∗ -0.016 0.044∗∗∗ 0.029∗ 0.003 -0.007
(0.007) (0.018) (0.006) (0.016) (0.004) (0.004)
Employed -0.073∗∗∗ -0.030∗∗∗ -0.095∗∗∗ -0.045∗∗∗ -0.037∗∗∗ -0.007∗∗∗
(0.006) (0.006) (0.005) (0.006) (0.004) (0.004)
House Type
Detached Ref Ref Ref Ref Ref Ref
Semi-attached -0.037∗∗∗ -0.026 -0.021∗∗ -0.026 -0.027∗∗∗ -0.022
(0.006) (0.025) (0.009) (0.026) (0.009) (0.015)
Apartment Building -0.032∗∗∗ -0.018 -0.038∗∗∗ -0.058∗ -0.034∗∗∗ -0.002
(0.008) (0.022) (0.007) (0.024) (0.007) (0.014)
Migration Background 0.032∗∗∗ -0.019∗∗ 0.036∗∗∗ 0.033∗∗∗ -0.005 -0.001
(0.008) (0.007) (0.005) (0.006) (0.005) (0.004)
Years of Education -0.016∗∗∗ -0.011∗∗∗ -0.014∗∗∗ -0.010∗∗∗ -0.012∗∗∗ -0.008∗∗∗
(0.001) (0.001) (0.001) (0.001) (0.001) (0.001)
Living Area
12-67 sqm. Ref Ref Ref Ref Ref Ref
68-80 sqm. -0.021∗∗∗ -0.026 -0.055∗∗∗ -0.013∗∗∗ -0.015∗∗∗ 0.019
(0.006) (0.025) (0.006) (0.015) (0.005) (0.012)
81-100 sqm. -0.016∗∗ 0.061∗∗ -0.086∗∗∗ -0.036∗ -0.015∗∗ 0.046∗∗
(0.007) (0.028) (0.007) (0.022) (0.006) (0.021)
101-270 sqm. -0.033∗∗∗ -0.018 -0.082∗∗∗ -0.049∗∗ -0.023∗∗∗ 0.074∗∗∗
(0.009) (0.027) (0.008) (0.020) (0.007) (0.029)
Owner 0.017 -0.026 0.020 -0.062 0.072 0.057
(0.097) (0.071) (0.078) (0.054) (0.101) (0.101)
State Fixed effects Yes Yes Yes Yes Yes Yes
Wave Fixed effects Yes Yes Yes Yes Yes Yes
Number of Observations 19,384 19,384 19,384 19,384 19,384 19,384
Number of Groups 4,846 4,846 4,846 4,846 4,846 4,846
Log Likelihood -6,016.92 -5,514.06 -5,148.68 -4,781.14 -4,260.52 -1,000.67
Notes: Columns (1), (3), and (5) present the average marginal effects of random effects probit estimators that do not
account for the initial conditions problem. Columns (2), (4), and (6) present the average marginal effects obtained with
the WCML estimator. The columns are for the entire sample (i.e., 2016-2020). Standard errors are clustered at the
individual level and are reported in parentheses. ’Ref’ stands for the reference category.
*** p<0.01, ** p<0.05, * p<0.10.
21
a household is experiencing energy poverty and correlate with certain household
types. If the model controls for initial conditions, the potential impact of these
family-oriented social policies on the likelihood of energy poverty may be better
accounted for. It is also possible that family-oriented social policies have changed
over the observation period, resulting in a differential impact on parental energy
poverty over time. As a result, the estimated effects of household types on energy
poverty may be closer to the true effects when controlling for initial conditions,
which may lead to the reduction of the statistical significance of parents (single
parents and couples with children).
Being employed significantly reduces the probability of experiencing energy
poverty in both models for all indicators. These findings indicate that unemployed
or retired households are more vulnerable to energy poverty. House type is not a
statistically significant variable for the 2M indicator and the 10 percent indicator
when controlling for initial conditions. For the LIHC indicator, however, living in
an apartment building compared to a detached building reduces the probability
of experiencing energy poverty by 5.8%. The LIHC indicator focuses primarily on
low-income households, with households in detached dwellings in particular facing
potentially higher energy costs. The coefficient migration background shows pos-
itive signs, as expected, indicating that households with a migration background
have a higher probability of experiencing energy poverty. Statistical significance is
shown for the LIHC indicator and the 2M indicator. The underlying cause of this
relationship could be related to lower incomes among households with a migrant
background.
For all indicators in both models increasing years of education is associated
with a lower probability of experiencing energy poverty. Again, this relationship is
most likely driven by income, as higher education is associated with higher income.
The signs of living area show different trends for the various indicators, which can
be attributed to their underlying calculation. For the 2M indicator, the coefficients
for larger dwellings between 81 and 270 m² are positive and statistically significant,
suggesting that households with larger living areas are more likely to experience
energy poverty. Also, for the 10 percent indicator, the likelihood of suffering from
energy poverty increases when a household lives in 81-100 sqm compared to a
living area of 12-67 sqm. This relationship highlights an important weakness in
these calculation methods, which will be reviewed further in the discussion. For
the LIHC indicator, the probability of suffering from energy poverty decreases
statistically significantly when households have living areas above 81 sqm.
As anticipated the variable owner shows no statistically significant coefficients,
likely due to the limited data availability and the resulting inability to establish
significance.
22
4.3 Energy Poverty Dynamics
Since significant state-dependent effects on energy poverty were found for all in-
dicators in both models, further investigation of the dynamics of energy poverty
is essential to gain a deep and integrative understanding of this poverty phe-
nomenon. In this chapter, the results of the survival analysis will be presented to
provide valuable insights into the dynamics of energy poverty, revealing patterns
of household entry and exit from energy poverty.
Table 6 gives an overview of the number of years a household spends in energy
poverty over the entire observation period. The majority of households in the
sample have never experienced energy poverty (65.02% for 10 percent, 78.89% for
2M and 69.13% for LIHC). The 2M indicator shows the lowest overall persistence
of energy poverty, with only 1.86% of households experiencing energy poverty in
all five periods. In contrast, the LIHC and 10 percent indicators have higher levels
of persistent energy poverty, at 7.00% and 5.65%, respectively.
Table 6: Distribution of periods spent in energy and income poverty
0 1 2 3 4 5
10 Percent Indicator 65.02% 12.15% 6.17% 6.54% 4.46% 5.65%
LIHC Indicator 69.13% 8.40% 5.61% 5.43% 4.44% 7.00%
2M Indicator 78.89% 9.39% 5.37% 2.79% 1.71% 1.86%
Table 7 shows the average rate of movement between each of the energy poverty
indicators over the five years period. The table is an estimate of the corresponding
Markov transition matrix and provides first insights into the dynamics of energy
poverty. Each row of the table represents the initial energy poverty state, while
each column denotes the classification for the subsequent state. The values in the
cells of the table represent the average proportion of individuals who transition
from one energy poverty state to another during the specified time periods.
For the 2M and LIHC indicators, the majority of non-energy poor house-
holds remain so in the following year, at 97.20% and 97.60%, respectively. The
10 percent indicator shows a slightly lower rate, with 96.90% of non-energy-poor
households remaining so in the subsequent year. Among energy-poor households,
the 2M indicator and 10 percent indicator display similar year-to-year transitions,
with 81.50% and 83.80% remaining energy poor, respectively. The LIHC indica-
tor, however, shows a slightly higher persistence of energy poverty, with 89.30%
of households remaining energy poor in the following year. The analysis reveals
that the choice of indicator impacts the observed dynamics of energy poverty.
While the Markov transition matrix in Table 5 provides valuable insight into
the overall persistence of energy poverty, it also has its limitations. In particu-
lar, it assumes that individuals in different poverty categories are identical and
23
Table 7: Average year-to-year movements
Year t + 1
10 percent Indicator
Not Energy Poor Energy Poor N
Not Energy poor 96.90% 3.10% 17 884
Year t
Energy poor 16.20% 83.80% 3 977
LIHC Indicator
Not Energy Poor Energy Poor
Not Energy poor 97.60% 2.40% 18 006
Year t
Energy poor 10.70% 89.30% 3 855
2M Indicator
Not Energy Poor Energy Poor
Not Energy poor 97.20% 2.20% 19 991
Year t
Energy poor 18.50% 81.50% 1 870
have the same likelihood of exiting and entering poverty. This assumption over-
simplifies the dynamics of energy poverty by ignoring the heterogeneity among
households and their unique characteristics that affect their propensity to escape
or remain in energy poverty. Thus, relying solely on Markov matrices may lead
to an incomplete understanding of the problem.
The following survival analysis can capture the subtle differences among house-
holds and improve the understanding of the underlying factors driving energy
poverty dynamics.
Table 8: Survival analysis - exits out of energy poverty
Time since 10 percent Indicator LIHC Indicator 2M Indicator
start of spell Survivor Hazard Survivor Hazard Survivor Hazard
1 year 0.708 0.292 0.789 0.211 0.740 0.260
2 years 0.493 0.202 0.614 0.222 0.468 0.367
3 years 0.370 0.250 0.467 0.238 0.349 0.255
4 years 0.143 0.613 0.225 0.519 0.100 0.714
Table 8 presents the survival analysis estimates of energy poverty. The survivor
function, which represents the probability that a spell that has just begun lasts
for t periods, shows a decreasing trend over time for all indicators, reflecting the
decreasing likelihood of households remaining in energy poverty with increasing
duration. In the first year, survivor functions range from 70.8% for the 10 percent
indicator to 78.9% for the LIHC indicator, suggesting that a substantial proportion
of households that just entered the state of energy poverty are still energy poor
after one year. In the second and third years, the probabilities of remaining in
energy poverty decline for all indicators. In the fourth year, survivor functions
24
drop dramatically, with values of 10% for the 2M indicator, 22.5% for the LIHC
indicator, and 14.3% for the 10 percent indicator. These results suggest that the
longer households are in energy poverty, the less likely they are to remain in this
condition.
For all three indicators, the probability of exit (hazard function) for those who
have just started a period of energy poverty is relatively low in the first period,
ranging from 21,1% for the LIHC indicator to 29.2% for the 10 percent indicator.
As time progresses notable variations in the hazard functions across the indicators
can be observed. The 2M indicator shows an increasing hazard function over
time, indicating that the probability of households exiting energy poverty rises
as the spell lengthens. In contrast, the LIHC and 10 percent indicators exhibit
fluctuating hazard functions, with the highest probabilities of exit occurring in
the fourth year, at 51.9% and 61.3%, respectively.
However, the results of the survival analysis should be interpreted with care
as the survivor and hazard functions are calculated without controlling for any
exploratory variables that might affect the probability of transitioning in and out
of energy poverty. The survival analysis provides an overview of the general pat-
terns of energy poverty spells. Informed predictions about the impact various
influential factors might have on the probability of escaping energy poverty can-
not be made. To gain a more comprehensive understanding of energy poverty
dynamics by investigating the possible relationship between the hazard rate and
the explanatory variables, the analysis could be extended using the Cox Propor-
tional Hazards (CPH) Model. The CPH model is a semi-parametric regression
procedure for survival analysis, developed in 1972 by the British statistician David
Cox. However, our sample size of households changing their energy poverty status
throughout the observed period is not sufficiently large to perform this subsequent
analysis step.
4.4 Discussion
Energy poverty is a complex and multi-layered phenomenon that is composed of
various aspects. A comprehensive view of energy poverty, incorporating different
calculation methods, can contribute to a better understanding of the phenomenon
and allow for targeted policy measures that meet both the individual needs of af-
fected households and societal and environmental demands. In order to assess
and understand the availability, accessibility, and usability of energy services for
households, this work has used the capabilities approach beyond purely economic
energy poverty indicators. The previous analysis has shown to what extent de-
terminants, as well as dynamics, are strongly dependent on the choice of the
indicator.
25
The LIHC approach demonstrates that a differentiated consideration of income
and energy aspects is important for the identification of energy-poor households
by attributing both elements in its calculation. Although income poverty and
energy poverty cannot be equated, there are overlaps between the two concepts of
poverty. When examining the intersection between the two concepts, the LIHC
indicator exhibits the lowest count of income-poor households that are not af-
fected by energy poverty (see table A.2 in the Appendix). Compared to other
indicators, the LIHC indicator also identifies the smallest number of non-income
poor households that are impacted by energy poverty. In contrast, the 10 percent
indicator and the 2M indicator suggest that high energy costs can cause energy
poverty, regardless of income level, since high-income households with high energy
expenditures can be classified as energy poor. The examination of the determi-
nants supports this observation by identifying that households with larger living
space, which is positively correlated with household income in our sample, have
an increased risk of energy poverty when the 2M indicator is applied.
Looking at the association between socioeconomic factors and the risk of en-
ergy poverty in light of our hypotheses, hypothesis 1a is supported. The analy-
sis shows that lower household income significantly increases the risk of energy
poverty. The presence of negative correlations between household income and
certain variables, such as migration background or smaller living spaces, is con-
vincingly demonstrated by the negative coefficients revealed in the random effects
probit analysis. In addition, the variable household income was included as a
covariate in the regression in the form of a robustness test (see table A.5 in the
Appendix) and a statistically significant negative correlation with energy poverty
status was observed. This finding confirms the direct influence of financial re-
sources on the ability to access and benefit from energy services, as emphasized in
the capabilities approach. In this sense, this paper highlights the need for policies
aimed at reducing income inequality to effectively address energy poverty.
Regarding hypothesis 1b, the results show that a higher level of education
within a household reduces the risk of experiencing energy poverty. This sup-
ports the capabilities approach by highlighting the positive impact of education
on expanding household capabilities with respect to energy services. Specifically,
better-educated households can access information on energy efficiency and afford-
able energy services and are thus better able to make informed decisions about
their energy use.
The results from the random effects regression analysis and the survival analy-
sis suggest that energy poverty has both temporary and persistent features. This
finding is consistent with previous studies on the topic (Drescher and Janzen,
2021; Phimister et al., 2015). Survival analysis highlights the temporary charac-
teristic of energy poverty. The conditional probability of exiting energy poverty
26
for individuals who have recently entered a spell of energy poverty demonstrates
a decreasing trend over time for all indicators. This implies that the majority of
households only experience energy poverty for limited periods of time and do not
remain in this state throughout the entire observation period.
These results support hypothesis 2 by showing that the majority of households
in Germany experience energy poverty as more of a transient condition. This
means that capabilities in terms of access to and use of energy services are dynamic
and can be influenced by social, economic, and political factors. This aspect
emphasizes the importance of the capabilities approach to identify transitional
states and potential intervention points to address energy poverty.
Simultaneously, the survival analysis also shows some degree of persistence.
For the LIHC indicator, the probability of remaining energy poor after four peri-
ods is still 22.5% (10% for the 2M indicator and 14.3% for the 10 percent indica-
tor). These findings emphasize the persistent nature of the issue for a significant
portion of the sample, highlighting the need to investigate and address the under-
lying factors contributing to this enduring facet of the issue. The chronic nature
of energy poverty is underlined by the statistically significant state dependence
identified for all indicators in the dynamic random effects estimation.
Such persistence may be related to a chronic deficit in certain capabilities, such
as low income, lack of access to energy efficiency information, or other structural
factors. This emphasizes the relevance of the capabilities approach for developing
effective strategies to reduce energy poverty by addressing both transient and
chronic aspects of the problem.
Table 9 provides potential policy implications for overcoming energy poverty
associated with the different capability categories. Each column suggests dif-
ferent policy strategies, including compensation-based, empowerment-based, and
sustainability-based policies that could address both chronic and transitory en-
ergy poverty. These strategies range from providing income support and legal
assistance to promoting energy-efficient infrastructure and public participation in
energy policy decisions. These strategies underscore the need for a multifaceted
approach to addressing energy poverty, as it is complex and dynamic. Compen-
satory measures, such as social assistance and direct subsidies, can be particularly
effective in helping both chronically and transitory affected households afford ba-
sic energy services that directly affect their biological and physical needs. In
addition, subsidizing energy-efficient appliances can help households reduce their
energy costs in the long run.
Empowerment measures, such as digital literacy programs, help strengthen
households’ ability to make informed decisions about their energy consumption.
In particular, such measures can help overcome chronic energy poverty. Public
participation in energy decision-making processes strengthens households’ ability
27
Table 9: Policy recommendations
Categories of Compensation- Empowerment- Sustainability-
capabilities based policy focused policy focused policy
Capabilities Implement Subsidize Invest in
related to income-support energy-efficient infrastructure for
measures to help appliances to energy-efficient
biological and
households afford reduce energy costs housing to reduce
physical human basic energy for cooking, the cost of heating
needs services. lighting, etc. and cooling.
Implement public
information
campaigns to raise
Capabilities Provide subsidies Promote digital awareness about
related to for educational literacy programs environmental
resources or digital to help households
fundamental impacts of
tools to access make informed
interests of a excessive energyenergy-related decisions about
human agent consumption andinformation. energy use.
benefits of
energy-saving
practices.
Capabilities Offer financial Invest inFacilitate public
related to support for communityengagement in
households to renewable energy
fundamental energy
legally challenge projects to enhance
interests of a policy-makingunjust energy local control over
social being processes.decisions. energy resources.
28
to resist unfair energy decisions. Sustainability-oriented policies such as invest-
ments in energy-efficient infrastructure and community renewable energy projects
can help address chronic energy poverty in the long term.
However, there are certain limitations to be considered in this paper. Firstly,
a comparison of consensus energy poverty indicators and expenditure-based indi-
cators would have been interesting, as this could have provided a more nuanced
picture of energy poverty. Another limiting factor is the lack of a variable measur-
ing energy efficiency. Energy efficiency could act as a significant influencing factor
on energy poverty, which affects the interpretation of the results. Additionally,
the study does not take into account household behavior in the form of energy-
saving measures, which could affect the elasticity of demand for energy services.
This could lead to under- or overestimation of energy poverty rates, as households
may adjust their energy consumption in response to changes in income or energy
prices.
Moreover, the time frame of the study, while extensive, may not fully capture
the long-term dynamics of energy poverty. A longer observation period would
likely provide richer insights into the persistence and transitions of energy poverty
over time. Additionally, the model does not account for household behavior in
the form of energy-saving measures, which are likely to affect the elasticity of
demand for energy services. Particularly during periods of high energy prices,
households may adjust their energy consumption in ways not captured by the
current model. The study does not factor in the potential influence of external
shocks that might affect energy poverty such as the Russian invasion of Ukraine,
leading to unexpectedly high energy prices and inflation. Therefore the ability
to extend the analysis to current periods of high energy prices is limited. Lastly,
the study was not able to apply the Cox Proportional Hazards Model due to
the insufficient sample size of households changing their energy poverty status
throughout the observed period. This model could provide more nuanced insights
into the determinants affecting the probability of transitioning in and out of energy
poverty, thus, its inclusion in future research would be highly beneficial.
In in my opinion, the 2M and 10 percent indicators do not adequately capture
the essence of energy poverty and could be misleading when referred to as poverty
indicators. High-income households who simply consume large amounts of energy
can be classified as energy poor under these indicators. Adopting such definitions
could dilute the significance of the issue and lend support to arguments in the
literature that suggest energy poverty is sufficiently addressed by general income
poverty research. Nonetheless, I believe that energy poverty should generally
be recognized as a distinct poverty measure, albeit with a notable overlap with
income poverty.
29
5 Conclusion
This thesis aimed to answer two main research questions: (1) What socioeconomic
factors are associated with the risk of a household experiencing energy poverty in
Germany, and how do these factors impact their capabilities to access affordable,
reliable, and safe energy services?, and (2) How do households transition in and
out of energy poverty over time, and how do patterns of persistence influence their
ability to overcome energy poverty?
These questions were answered using a combination of statistical methods to an-
alyze a dataset of German household data over a five-year span. The capabilities
approach was used as a theoretical framework to consider the dynamics and de-
terminants of energy poverty in a broader socioeconomic context.
Regarding the first question, the study found both similarities and statistically
significant differences in terms of socioeconomic factors associated with energy
poverty across the three chosen indicators. The findings show that a one-person
household has a higher risk of experiencing energy poverty for all indicators. Since
additional household members are generally expected to have lower energy con-
sumption than the first member, living alone may negatively impact their ability
to access affordable, reliable, and safe energy services. Statistical significance
of single-parent households or couples with children as determinants of energy
poverty diminished, possibly due to the influence of family-oriented social poli-
cies in Germany. Socioeconomic determinants in terms of house type, living area,
and migration background varied across the different indicators. The 2M and 10
percent indicators show an increased likelihood of energy poverty for households
with larger living spaces. In contrast, the LIHC indicator, which does not classify
high-income households as energy poor, suggests that households with larger living
spaces have a lower tendency to be affected by energy poverty. This discrepancy
could be explained by the positive correlation between larger living spaces and
higher household income. From the perspective of the capability approach, these
findings highlight how different socioeconomic factors and living conditions can
either facilitate or hinder a household’s ability to access sufficient and affordable
energy services.
The second question sought to understand the dynamics of energy poverty.
The findings highlight significant state dependence for all three indicators, as ev-
idenced by the random effects probit estimation. This state dependence implies
some degree of persistence of energy poverty over time, meaning that households
that experienced energy poverty in one period are more likely to experience it
in subsequent periods. The survival analysis provided more detailed information
about the dynamics of energy poverty over time. It showed that while many
households move in and out of energy poverty during the observation period -
30
underlining its temporary nature - there is a nontrivial probability of persistence.
In particular, for households identified as energy poor by the LIHC indicator, the
likelihood of remaining in energy poverty after four periods is still substantial
at 22.5% (compared to 10% for the 2M indicator and 14.3% for the 10 percent
indicator). This persistence in energy poverty may continue to limit their capa-
bilities and resources and significantly hinder their efforts to combat the issue.
These findings emphasize the need to consider a broad range of capabilities when
addressing energy poverty.
In summary, this paper highlights the complexity of energy poverty and the
importance of selecting appropriate indicators to capture its multifaceted nature.
Despite its limitations, the study provides valuable insights into the dynamics and
determinants of energy poverty by highlighting its transitory and persistent fea-
tures, as well as the influence of socioeconomic factors. The analysis underscores
the need for targeted policies that address both the transitory and persistent as-
pects of energy poverty. As we currently find ourselves in a period of high and
volatile energy prices, the need for accurate quantification of energy poverty be-
comes even more imperative. The establishment of a national energy poverty
indicator, coupled with targeted policy measures, is therefore urgently needed.
Additionally, the current research emphasizes the necessity of broadening the
perspective to consider not only the expenditure aspect but also socio-economic
viewpoints, as exemplified by the capabilities approach, to gain a more compre-
hensive understanding of energy poverty. This approach emphasizes the need not
only to ensure access to affordable energy services but also to strengthen the ca-
pacity of those affected to use these services in a sustainable manner. It implies
that there is a need to think beyond purely financial measures and adopt policies
that take into account the different capabilities and needs of households. In this
way, policymakers can not only alleviate immediate energy poverty through tar-
geted policy interventions in the form of direct subsidies but also contribute to a
more sustainable and equitable energy future by improving the energy efficiency
of affected households.
31
I References
References
Alem, Y. and Demeke, E. (2020). The persistence of energy poverty: A dynamic
probit analysis. Energy Economics, 88:104789.
Alkire, S. (2005). Why the capability approach? Journal of Human Development,
6(1):115–135.
Aristondo, O. and Onaindia, E. (2018). Counting energy poverty in spain between
2004 and 2015. Energy Policy, 113:420–429.
Barnes, M., Butt, S., Tomaszewski, W., et al. (2008). The dynamics of bad hous-
ing: The impact of bad housing on the living standards of children. National
Centre for Social Research London.
Berges, S. (2007). Why the capability approach is justified. Journal of Applied
Philosophy, 24(1):16–25.
Biewen, M. (2009). Measuring state dependence in individual poverty histories
when there is feedback to employment status and household composition. Jour-
nal of Applied Econometrics, 24(7):1095–1116.
Boardman, B. (1991). Fuel poverty: from cold homes to affordable warmth. Bel-
haven Press, London.
Bouzarovski, S. and Petrova, S. (2015). A global perspective on domestic en-
ergy deprivation: Overcoming the energy poverty–fuel poverty binary. Energy
Research & Social Science, 10:31–40.
Bouzarovski, S. and Simcock, N. (2017). Spatializing energy justice. Energy
Policy, 107:640–648.
Bundestag, D. (2017). Bericht der bundesregierung über ihre maßnahmen zur
bekämpfung von energiearmut. Technical report, Deutscher Bundestag.
Callander, E. J. and Schofield, D. J. (2016). The risk of falling into poverty after
developing heart disease: A survival analysis. BMC Public Health, 16(1):570.
Cappellari, L. and Jenkins, S. P. (2014). The dynamics of social assistance benefit
receipt in britain. Emerald Group Publishing Limited.
Chaton, C. and Lacroix, E. (2018). Does france have a fuel poverty trap? Energy
Policy, 113:258–268.
32
Chen, J., Wu, C., and Zheng, H. (2022). Who stays poor and who doesn’t?
an analysis based on joint assessment of income and assets. Journal of Social
Policy, pages 1–22.
Churchill, S. and Smyth, R. (2021). Energy poverty and health: panel data
evidence from australia. Energy Economics, 97:105219.
Churchill, S., Smyth, R., and Farrell, L. (2020). Fuel poverty and subjective
wellbeing. Energy Economics, 86:104650.
Ciuca, V. and Maer Matei, M. (2010). Survival analysis for the unemployment
duration.
DEFRA (2001). The uk fuel poverty strategy.
Drescher, K. and Janzen, B. (2021). Determinants, persistence, and dynamics of
energy poverty: An empirical assessment using german household survey data.
Energy Economics, 102:105433.
European Commission (2022). Communication from the commission to the euro-
pean parliament, the council, the european economic and social committee and
the committee of the regions: The european green deal.
European Commission (2023). Energy poverty advisory hub handbook: Diagnos-
ing energy poverty. Technical report, European Commission.
European Parliamentary Research Service (2022). Understanding energy poverty
in the european union. Technical report, European Parliament.
Faiella, I. and Lavecchia, L. (2021). Energy poverty. how can you fight it, if you
can’t measure it? Energy and Buildings, 233:110692.
Heckman, J. J. (1987). The incidental parameters problem and the problem of
initial conditions in estimating a discrete time-discrete data stochastic process
and some monte carlo evidence. University of Chicago Center for Mathematical
studies in Business and Economics.
Heindl, P., Kanschik, P., and Sch”ussler, R. (2017). Anforderungen an energiear-
mutsmaße. In Großmann, K., Schaffrin, A., and Smigiel, C., editors, Energie
und soziale Ungleichheit, pages 241–262. Springer Fachmedien Wiesbaden.
Heindl, P. and Schuessler, R. (2015). Dynamic properties of energy affordability
measures. Energy Policy, 86:123–132.
Herrero, S. T. (2017). Energy poverty indicators: A critical review of methods.
Indoor Built Environ, 26(7):1018–1031.
33
Hills, J. (2011). Fuel poverty: the problem and its measurement. (69).
Hills, J. (2012). Getting the measure of fuel poverty: final report of the fuel
poverty review. (72).
Isherwood, B. C. and Hancock, R. M. (1979). Household expenditure on fuel:
Distributional aspects. Technical report, Economic Adviser’s Office, DHSS,
London.
Karpinska, L. and Śmiech, S. (2020). Conceptualising housing costs: The hidden
face of energy poverty in poland. Energy Policy, 147:111819.
Khoo, W. C., Yeah, K. L., and Hong, S. Y. (2022). Modeling unemployment
duration, determinants and insurance premium pricing of malaysia: insights
from an upper middle-income developing country. SN Business Economics,
2:116.
Kopatz, M. (2013). Energiewende. aber fair! wie sich die energiezukunft sozial
tragfähig gestalten lässt.
Lee, J., Kim, H., and Byrne, J. (2021). Operationalising capability thinking in the
assessment of energy poverty relief policies: Moving from compensation-based
to empowerment-focused policy strategies. Energy Research & Social Science,
76:292–315.
Legendre, B. and Ricci, O. (2015). Measuring fuel poverty in france: Which
households are the most fuel vulnerable? Energy Economics, 49:620–628.
Liddell, C., Morris, C., McKenzie, S., and Rae, G. (2012). Measuring and moni-
toring fuel poverty in the uk: National and regional perspectives. Energy Policy,
49:27–32.
Mayer, I. (2013). Energiearmut: Der weiße fleck in der deutschen forschungsland-
schaft. Energiewirtschaftliche Tagesfragen, 63(6):61–63.
Meyer, S., Laurence, H., Bart, D., Middlemiss, L., and Mar’echal, K. (2018).
Capturing the multifaceted nature of energy poverty: Lessons from belgium.
Energy Research Social Science, 40:273–283.
Neuhoff, K., Bach, S., Diekmann, J., Beznoska, M., and El-Laboudy, T. (2013).
Distributional effects of energy transition: impacts of renewable electricity sup-
port in germany. Economics of Energy & Environmental Policy, 2(1):41–54.
Nielsen, L. and Axelsen, D. V. (2016). Capabilitarian sufficiency: Capabilities
and social justice. Journal of Human Development and Capabilities.
34
Nussbaum, M. (2003). Capabilities as fundamental entitlements: Sen and social
justice. Feminist Economics, 9:33–59.
OECD (2013). OECD Framework for Statistics on the Distribution of Household
Income, Consumption and Wealth. OECD Publishing, Paris.
Okushima, S. (2016). Measuring energy poverty in japan, 2004–2013. Energy
Policy, 98:557–564.
Phimister, E., Vera-Toscano, E., and Roberts, D. (2015). The dynamics of energy
poverty: evidence from spain. Econ. Energy Environ. Policy, 4(1):153–166.
Rademaekers, K., Yearwood, J., Ferreira, A., Pye, S., Hamilton, I., Agnolucci, P.,
Karasek, J., and Anisimova, N. (2016). Selecting indicators to measure energy
poverty: Final report. Technical report, European Commission.
Robeyns, I. (2005). The capability approach: a theoretical survey. Journal of
Human Development, 6(1):93–117.
Sagar, A. D. (2005). Alleviating energy poverty for the world’s poor. Energy
Policy, 33(11):1367–1372.
Schreiner, N. (2015). Auf der suche nach energiearmut: Eine potentialanalyse des
low-income-high-costs indikators f”ur deutschland. SOEPpapers 811, SOEPpa-
pers on Multidisciplinary Panel Data Research.
Sen, A. (1985). Commodities and Capabilities. North-Holland; Sole distributors
for the U.S.A. and Canada Elsevier Science Pub.
Sen, A. (2004). Capabilities, lists, and public reason: Continuing the conversation.
Feminist Economics, 10(3):77–80.
Sen, A. (2005). Human rights and capabilities. Journal of Human Development,
6(2):151–166.
Sen, A. (2009). The Idea of Justice. Harvard University Press, Cambridge, MA.
Statista (2021). Verteilung der Haushalte in Deutschland nach Miete und Eigen-
tum. Statista website. Accessed: May 11, 2023.
Statista (Accessed: 2023). Monatliche nettolöhne und gehälter je arbeitnehmer
in deutschland. Statista.
Stewart, M. B. (2007). The interrelated dynamics of unemployment and low-wage
employment. J. Appl. Econometrics, 22(3):511–531.
35
Strünck, C. (2017). Energiearmut bekämpfen – instrumente, maßnahmen und
erfolge in europa.
Teller-Elsberg, J., Sovacool, B., Smith, T., and Laine, E. (2016). Fuel poverty,
excess winter deaths, and energy costs in vermont: Burdensome for whom?
Energy Policy, 90:81–91.
Tews, K. (2013). Energiearmut definieren, identifizieren und bek”ampfen – eine
herausforderung der sozialvertr”aglichen gestaltung der energiewende. FFU-
Report 04-2013, Forschungszentrum f”ur Umweltpolitik, Berlin.
Thomson, H., Bouzarovski, S., and Snell, C. (2017). Rethinking the measurement
of energy poverty in europe: A critical analysis of indicators and data. Indoor
and Built Environment, 26(7):879–901.
Wooldridge, J. M. (2005). Simple solutions to the initial conditions problem in
dynamic, nonlinear panel data models with unobserved heterogeneity. Journal
of Applied Econometrics, 20(1):39–54.
36
II Appendix
i. Supplementary Data
Table A.1: Intersection between Income Poverty Energy Poverty
Not Income Poor Income Poor Total
Not Energy Poor 2,182 2,190 4 372
10% Indicator
Energy Poor 1 626 18 232 19 858
Not Energy Poor 2 890 1 405 4 295
LIHC Indicator
Energy Poor 918 19 017 19 935
Not Energy Poor 1 317 845 2 162
2M Indicator
Energy Poor 2 491 19 588 22 079
Table A.2: Household Characteristics of Energy Poor Households
Variable Description 2M Indicator LIHC Indicator 10% Indiactor Income Poverty
Household Type
Couple, no children 17,99% 15,95% 18,28% 19,30%
Single parent 40,75% 46,57% 42,34% 21,72%
One-person household 26,78% 20,77% 22,90% 22,53%
Couple with children 14,48% 16,72% 16,49% 36,45%
Employed .= 1 if employed 30,53% 25,94% 32,00% 30,75%
House Type
Detached 19,33% 12,41% 16,56% 8,30%
Semi-attached 9,94% 6,47% 8,81% 7,64%
Apartment Building 70,72% 81,12% 74,63% 84,06%
Migration Background .= 1 if migrational background 31,27% 40,49% 31,70% 39,47%
Years of Education No degree 60,08% 62,10% 58,90% 58,19%
Lower Secondary Degree 21,46% 18,30% 21,27% 19,30%
Higher Secondary Degree 11,84% 13,83% 12,58% 15,52%
Tertiary Degree 6,61% 5,77% 7,25% 6,99%
Living Area
12-67 sqm 49,86% 57,97% 47,16% 64,71%
68-80 sqm 24,24% 26,87% 27,56% 24,53%
81-100 sqm 18,96% 10,83% 18,53% 8,69%
101-270 sqm 6,94% 4,33% 6,75% 2,07%
Owner .= 1 if owner 0,05% 0,02% 0,02% 0,05%
*Notes: households with an income below 60% of the median income are classified as income-poor.
Table A.3: Mean Income of Energy Poor Households
2016 2017 2018 2019 2020
2M Indicator 1 452 1 504 1 525 1 592 1 616
LIHC Indicator 1 359 1 432 1 482 1 503 1 588
10% Indicator 1 682 1 701 1 694 1 779 1 719
37
Table A.4: Pearson correlation matrix for continuous variables
Household Employed Migration Living Years of
Income Back- Area Education
ground
Household Income 1
Employed 0.2901 1
Migration Background -0.0604 -0.0112 1
Living Area 0.4901 0.1075 -0.0657 1
Years of Education 0.2569 0.2127 -0.2034 0.1035 1
ii. Robustness Tests
The findings presented in this paper were checked by various robustness tests to
capture the behavior of the variables of interest under changing conditions. In a
first step, the dynamic model with random effects was applied using different time
spans to ensure that the estimation results do not depend on the choice of the
observation period. Two time spans, Panel A (2016 to 2018) and Panel B (2017 to
2029), were considered. The results are presented in the table A.5, with only the
values of the lagged variable energy poverty shown for clarity. Regardless of the
observation period, the sign and statistical significance of the state dependence
were preserved. However, for the 10 percent indicator and the 2M indicator, the
magnitude was lower in Panel B than in Panel A. All other variables of interest
showed consistent behavior throughout the observation period, both in terms of
sign and statistical significance.
In a second step, the model was estimated with the inclusion of log equiv-
alized household income as an additional covariate (Panel C). State dependence
maintained its sign and significance in the model specifications for both the 2M
indicator and the 10 percent indicator, but was slightly lower than in the specifi-
cation without household income. All other variables showed consistent behavior
in terms of sign, significance, and magnitude compared to the model discussed in
the paper. The only variable that lost statistical significance is education in the
case of the 2M indicator.
Regarding hypothesis 1b, which states that higher levels of education reduce
the risk of energy poverty, the results may suggest that the observed negative
correlation between education and energy poverty for the 2M indicator in models
without income is actually mediated by income. In other words, the impact of
education on energy poverty could be primarily due to the fact that individuals
with higher education tend to have higher incomes, making them less vulnerable
to energy poverty. Thus, by including log equivalized household income in the
model, hypothesis 1b is no longer independently supported of hypothesis 1a.
For the LIHC indicator, the dynamic random effects model failed to converge
38
when log equivalized household income was added to the model. The calculation of
the LIHC indicator includes both energy expenditures and income, so the inclusion
of household income in the model could potentially lead to multicollinearity. This
multicollinearity can lead to instabilities in the model and make it difficult for the
model to converge.
The decision not to use a fixed-effects model to test robustness was deliberate
and based on trade-offs between the model assumptions and the specific circum-
stances of the present analysis. In particular, the fixed-effects model requires the
assumption that the individual effects are time-invariant and correlate with the
explanatory variables. However, this assumption does not fit the structure and
conditions of the underlying paper. In the present analysis, the data structures
take into account both time-varying and time-invariant influencing factors, such
as changes in educational attainment, household income, and other socioeconomic
aspects over time. The fixed-effects model is primarily focused on time-invariant
effects. Moreover, a fixed-effects model that implies that the unobserved individ-
ual effects are correlated with the regressors would lead to estimator bias if this
assumption is not met. In this paper, the individual effects were assumed to be
uncorrelated with the explanatory variables, which is more consistent with the
assumptions of the random-effects model.
Overall, the findings of the robustness tests help to confirm and deepen the
conclusions and theories presented in the results section by showing both their
robustness under different conditions and the nuanced relationships among the
variables studied. They provide valuable insights into the relationships and inter-
actions between income, education, and energy poverty that can contribute to the
further development and refinement of future research approaches in this area.
39
Table A.5: Average Marginal Effects - Robustness Checks
10 percent Indicator LIHC Indicator 2M Indicator
Panel A (2016-2018)
0.110 0.119 0.043
Energy Poverty t-1 (0.001)*** (0.001)*** (0.001)***
Num. Obs. 9 692 9 692 9 692
Num. Groups 4846 4846 4846
Log likelihood -2 913.42 -2 444.95 -1 963.47
Panel B (2017-2019)
0.024 0.054 0.077
Energy Poverty t-1 (0.015)** (0.011)*** (0.008)***
Num. Obs. 9 692 9 692 9 692
Num. Groups 4 846 4 846 4 846
Log likelihood -2 751.04 -2 408.24 -1 985.73
Panel C
with Income (2016-2018)
0.013 0.019
Energy Poverty t-1 (0.014)** (0.000)***
-0.246 -0.381
Household Income
(0.000)*** (0.000)***
-0.002 -0.001
Education
(0,005)*** (0.287)
Num. Obs. 19 384 19 384
Num. Groups 4 846 4 846
Log likelihood -4 211,63 -2 943.49
Notes: Columns (1), (2), and (3) present the average marginal effects of random effects probit estimators for the
different panels.
*** p<0.01, ** p<0.05, * p<0.10.
40