1. Introduction

1 Protecting and improving health and mitigation of climate change have a shared agenda. Public policies intended to respond to climate change can help reduce health problems. Various policies can reduce greenhouse gas emissions and produce important health co-benefits (Zivin & Neidell [2013]) notably through retrofitting of housing (WHO Regional Office for Europe [2004]). In northern countries, many households live in homes that have poor thermal conditions and are therefore expensive to heat, accounting for a substantial share of emissions which in turn contribute to climate change. Inadequate thermal performance and ventilation of existing buildings are two of the biggest challenges for the European housing sector with respect to the sustainability and health of the built environment. Moreover, recent research has demonstrated that it is possible to improve indoor environmental quality, health and well-being of building occupants along with energy efficiency (Pampuri et al. [2018]; Prasauskas et al. [2016]; Rashid & Zimring [2008]; Stabile et al. [2019]).

2 This paper makes three contributions to the literature. Our first objective is to assess the link between fuel poverty and health over a recent period (2008-2016). Identifying a precise, direct and mid-term link between fuel poverty and health seems crucial in the design of relevant public policies. Households affected by fuel poverty are not always the same as those affected by monetary problems, even if the two phenomena are inextricably linked, representing an aspect of multidimensional poverty (Legendre & Ricci [2015]). Income poverty is now much studied in the literature, but its multiple dimensions sometimes require more in-depth study of certain aspects of deprivation, such as fuel poverty or the notion of health poverty (Clarke & Erreygers [2020]). Fuel poverty has generally been treated as a monetary poverty problem (Charlier & Legendre [2019]). Thus, like income poverty, energy poverty can be defined by the minimum energy consumption needed to sustain life. But unlike income poverty, which is based on the concept of a poverty line defined by the minimum consumption of food and non-food goods and services necessary to sustain a livelihood, energy poverty lacks a well-established energy poverty line to determine the minimum amount of energy needed for living. Moreover, a fuel-poor household may not be poor in monetary terms, and face difficulty in sustaining its energy needs. In that sense, energy poverty differs from monetary poverty. Low-income households may have difficulty paying for energy due to the poor thermal conditions of the buildings they occupy combined with rising energy prices, and therefore restrict their energy consumption. These households are considered to be fuel-poor (Boardman [1991], [2010]; Hills [2011], [2012]) and are subject to increased health problems due to the inferior energy performance of the buildings they live in.

3 Second, we want to provide a better understanding of the profile of households in poor health to help policy makers prevent other households from falling into this state and reduce future health expenditures. Implementing an effective policy implies anticipating both intermediate and final objectives: tackling fuel poverty as the intermediate objective and improving public health as the final objective. The literature shows that poor material conditions can harm health (Adena & Myck [2014]). Studies also demonstrate that improving housing insulation and decreasing carbon emissions are particularly important for public health: energy efficiency has been linked to an extensive range of health benefits and also to improvements in psychosocial functioning (Næss et al. [2007]; Wilkinson et al. [2007]). Haines et al. [2009] examined the effect of hypothetical strategies to improve energy efficiency in UK housing stock, assuming that low energy efficiency is also correlated with higher air pollution. They concluded that such interventions were generally beneficial for health: a strategy combining insulation, ventilation, fuel switching, and behavioral changes implied 850 fewer disability-adjusted life-years (DALYs), and a savings of 0.6 megatons of carbon dioxide, per million population in one year. This suggests that household energy interventions could yield important benefits in pursuit of both health and climate goals. Therefore, studying fuel poverty, including issues of housing quality, income and consumption patterns, will make it possible to focus on one aspect of the material difficulties households may experience.

4 Third, we explore the impact of being fuel poor on health empirically using a panel dataset controlling for endogeneity, state dependency of health and initial conditions, which is to the best of our knowledge relatively new. Indeed, the literature on the relationship between fuel poverty and health is quite limited, while research into the determinants of health is extensive. We focus on the French situation where the direct medical cost related to poor-quality housing is estimated to be 930 million euros (Euro-found [2016]). The indirect annual costs are estimated to be 20.3 billion euros. Using panel data (EU-SILC) enables us to consider individual health and household fuel poverty trajectories, but also to avoid obtaining results affected by 1-year climate variables. The advantages of a panel dataset are numerous, such as a greater capacity for modeling the complexity of human behavior than a single cross-section (Ben-Porath [1973]) or improved inference accuracy of model parameters (Hsaio et al. [1995]). Based on the example of Ben-Porath [1973] where a woman’s employment status is a perfect predictor of her future employment status, we consider that an individual’s self-assessed health status is a good predictor of their future health status. While cross-sectional data cannot distinguish between these two options, panel data can because the sequential observations contain information about individuals’ health problems at different subintervals of their life cycle. Using dynamic probit models, we can test the influence of fuel poverty on health. Moreover, health-state dependence can also serve as an explanation for observed spending phenomena, such as decreasing consumption in old age (Börsch-Supan & Stahl [1991]). Thus, we controlled for health-state dependence as we consider health status to be closely related to health status in the previous year. We also controlled for initial conditions: those found to be fuel poor in the base year may be a non-random sample (Cappellari & Jenkins [2004]; Heckman [1981]). Considering that unobserved heterogeneity might have an impact on health status and fuel poverty simultaneously, we corrected for endogeneity bias which could affect our results. Both short- and long-term policies are needed to improve the living conditions of households in fuel poverty. Supplementing household income and reducing energy costs will help in the short term but improving the energy efficiency of housing (e.g., modernizing heating systems, installing thermal insulation) is a long-term solution to address fuel poverty and its health consequences. Energy-efficient housing not only has benefits for the health of occupants, but also for society as a whole. Thus, governments could see a double benefit in energy retrofitting: an environmental gain and a reduction in health care costs.

5 The next section presents the existing literature on the relationship between energy efficiency and health. The theoretical background is then introduced. Data are presented in Section 4, and the empirical strategy is introduced in Section 5. The results are presented in Section 6, along with a key policy recommendation. We conclude in Section 7.

2. Methods

2.2. Data

10 In this study, we shed light on the potential causal effect of energy precariousness on health. Econometric analysis was conducted using the French portion of the European Union – Community Statistics on Income and Living Conditions database (EU-SILC) [4]. We chose data from 2008 to 2016, which is the last survey to date. This time frame was selected since the year 2008 marked significant changes, such as improving the accuracy of household resources using data from public institutions, and specifying the type of energy used inside dwellings instead of a global indicator, which is fundamental to the construction of relevant fuel poverty indicators (Hills [2011, 2012]). The database includes some health indicators, such as self-assessed health condition, the presence of chronic diseases and disabilities, unmet medical needs and the reasons for this, and other health-related variables, representing a good set of tools for describing and controlling the effect of fuel poverty on health. It also contains many variables characterizing socio-economic status, demographic status and living conditions related to housing and energy use of individuals, thus providing a solid basis for assessing health and fuel poverty under the many aspects and definitions reviewed in the literature. The entire sample consisted of 239,477 observations over nine years.

2.2.1. Fuel poverty indicators

11 Our health indicator is the self-assessed health status of the respondents; this will be the binary dependent variable, with 1 for poor health. This self-assessed health indicator encompasses more than just an assessment of an individual’s physical health (Bassett & Lumsdaine [2001]). For example, cultural differences may explain some heterogeneity between self-reported measures of health (Meijer et al. [2011]). Sometimes used with more objective measures, self-assessed health has been shown to improve both inference and predictability of outcomes in a variety of situations. However, questions that rely on self-assessment are now commonplace in health surveys where researchers have demonstrated their usefulness as a proxy for objective health measures when the latter are not readily observed (Damian et al. [1999]; Pinquart [2001]). According to Simon et al. [2005], 80% of respondents referred to one or more physical aspects (chronic illness, physical problems, medical treatment, age-related complaints, prognosis, body mechanics, and resilience) when answering the question “How is your health in general?” They also include aspects that go beyond the physical dimension of health. Thus, self-assessed health allows us to consider several consequences of fuel poverty on health in general, but also mental illness. Finally, the authors also demonstrate that health behavior or lifestyle factors (behavioral dimension) are relatively unimportant in health self-assessments.

12 In our study, some 8.3% of individuals reported a status of poor health over the period studied. While a categorical self-assessed health variable is a wide measure of one’s mental and physical condition, its reliability is questionable (Crossley & Kennedy [2002]). However, it remains one of the best tools available to measure health as it predicts future mortality and morbidity adequately (Idler & Kasl [1995]; Lundberg & Manderbacka [1996]). We also use a stated binary variable for chronic disease as a control (robustness check); 37.3% of individuals in the sample report a chronic disease. A chi square test of independence shows that having a chronic disease is not independent from reporting a status of poor health. Among people who reported a status of poor health, 94% had a chronic disease. In this study, we are interested in the possible continuity of health status. In our sample, the 95.9% of individuals who reported a status of good health, reported this over the entire period. Only 4.1% of individuals saw their health deteriorate, whereas 41% reported a status of poor health at the beginning of the period and saw their health improve.

13 To measure fuel poverty, we chose to apply the most common indicators of fuel poverty defined and used in the literature (which will be our main predictors): the 10% ratio [5] (Boardman [1991]) and Low Income High Costs (LIHC) [6] (Hills [2011], [2012]). Both of these measures are calculated based on the standard of living [7] of the households observed rather than the disposable income of the household. When standard indicators incorporate income to measure fuel-poverty, it is necessary to specify what definition of income is used: standard of living (standard of living is the material well-being of the average person in a given population) or disposable income (i.e., income minus taxes and social contributions). The 10% ratio permits comparisons and serves as a reference. We built it by aggregating electricity, natural gas and other heating expenditures whether or not included in the rent, and calculating its share of the standard of living which gives us the energy effort rate of individuals, and compared it to the 10% threshold, resulting in a binary variable (1=fuel poor, 0=not fuel poor). There is some criticism in the literature about this measure of fuel poverty, mainly due to the fact that high income individuals who spend a lot on energy can be considered to be fuel poor while being nowhere near fuel poverty socially speaking (Healy & Clinch [2002]), so we have also compared our results with the LIHC indicator [8]. We did this by comparing an individual’s standard of living to a poverty threshold [9] on the one hand and then comparing the individual’s fuel expenditures to an expenditure threshold [10]. If one has a standard of living below the poverty threshold, and spends more on energy than the expenditure threshold, the individual is then considered to be fuel poor (1=fuel poor, 0=not fuel poor). In our sample 4.78% are fuel poor according to the 10% ratio and 5.74% according to the LIHC indicator.

14 We tested the independence of fuel poverty (10% definition) and health status. The chi square test allows us to reject the hypothesis of independence between fuel poverty and poor health at the 1% level. Similar results are obtained if we consider another definition of fuel poverty or another measure of health i.e., having a chronic disease. Although the evolution over the period does not show the dependence between fuel poverty and health status (Figure 1) at first glance, more fuel-poor households are in poor health (Figure 2).

Figure 1. Evolution of the proportion of observations in good health and in fuel poverty (10%)

Figure 1. Evolution of the proportion of observations in good health and in fuel poverty (10%)

Figure 2. Health status according to fuel poverty definition

Figure 2. Health status according to fuel poverty definition

2.2.2. Independent variables

15 Finally, we introduced variables for controlling for health status: socio-demographic characteristics and local conditions (weather and pollution). Indeed, the economic and epidemiologic literature on the nexus between air pollution and health is vast (Contoyannis & Jones [2004]; Neidell [2004]). Health status might be linked to climate (WHO [2013]) and local air quality exposure is a significant driver of health problems (Currie et al. [2009]; Jans et al. [2018]). By reducing air pollution, countries can reduce the burden of disease from stroke, heart disease, lung cancer, and both chronic and acute respiratory diseases, including asthma (WHO [2013], [2018]). The relationship between socioeconomic characteristics and health is well-established. Negative health behaviors and psychosocial characteristics are clustered in low socioeconomic status groups (Lynch et al. [1997]). The notion that low socioeconomic status, mainly due to low income, causes poor health is widely supported by empirical research (Contoyannis et al. [2004]): poverty is associated with poor health even in advanced industrial societies (Benzeval & Judge [2001]). Education has also been found to have a positive association with physical and mental health (the so-called education-health gradient) because individuals invest in their health more effectively and allocate their resources better (Cutler & Lleras-Muney [2006]). Older people are often in poor health also (Ohrnberger et al. [2017]). Our dataset gave us access to information about the climate zone where each household is located. This was matched with meteorological data of Meteo France (unified degree days [11]) to provide a proxy for the actual meteorological conditions and to obtain unified degree days. Finally, we approximated the income variable with level of education to avoid multicollinearity with fuel poverty (based on income). The descriptive statistics are presented in Table 1, and all variables are defined in Appendix 1.

Table 1. Descriptive statistics

Table 1. Descriptive statistics

Note: * The variable has been decomposed into a between and within (, the global mean being added back in to make results comparable). N is the number of observations in the sample, n is the number of individuals

2.3. Empirical model

16 This paper is based on the traditional microeconomic approach: the consumer maximizes his utility under budget constraints. We estimate the impact of fuel poverty on the probability of being in poor health controlling for other characteristics. We use panel probit models and propose four configurations, first to control for state-dependence of health (Contoyannis et al. [2004]; Kools & Knoef [2019]), and second to control for endogeneity of fuel poverty (Awaworyi Churchill & Smyth [2018]).

17 So, we have as the first model specification (Model 1):

18 With:

19 h_it represents the health status of the individual i in t, equal to 1 for good health at time t, X_it is the vector of observed variables (including age, level of education, etc.), W_it is a vector of living conditions including local air pollution and climate, FP_it is the fuel poverty regressor, equal to 1 for a fuel-poor individual, according to the 10% threshold, u_i the individual-specific effect assumed to be unrelated to control variables, and v_it the idiosyncratic error term.

20 However, health status has a certain degree of persistence over time (Contoyannis et al. [2004]; Kools & Knoef [2019]), which leads to the first configuration being biased. To account for this state-dependence (Carro & Traferri [2014]; Halliday [2008]), we include a lagged health status variable representing the health of individuals in the previous year.

21 So, we propose a dynamic random effect model:

22 h_{it − 1} representing the lagged health status.

23 The use of a lagged variable presents the initial conditions problem (Heckman [1981]). The historic stochastic process of health status is not observed at the beginning. The initial values of this variable cannot be considered exogenous (Akay [2009]). Health during childhood, or other exogenous parameters affect the entire health trajectory, for example (Contoyannis et al. [2004]; Halliday [2008]). So, we can express health at the first period in the panel as following:

24 With Z_i0 including exogenous attributes affecting health status in the first period. The initial conditions problem leads h_i0 then to be correlated with the unobserved heterogeneity. Consequently, in equation 2 we have E (u_i | h_i0) ≠ 0 when θ ≠ 0.

25 In order to treat the initial conditions problem, we use the inverse of the Mills ratio [12] within Orme’s two-step method (Arulampalam & Stewart [2007]; Orme [1997]). We estimate a probit of the probability of being in good health in the first year in the panel. The medical density i.e., the ratio of physicians (practitioners or specialists) to the population in a geographic area, is used as an exogenous instrument (Z_i0) to explain the health status at time t₀ (Chaix et al. [2005]; Macinko et al. [2003]).

26 Our second specification, correcting the initial conditions problem, might be expressed as follows (Model 2):

27 in which the inverse Mills ratio, E [u_i | y_i0], has been estimated in the first step as:

28 The inverse Mills ratio allows us to control for initial health conditions.

29 The third specification of the model accounts for the potential endogeneity of fuel poverty. Endogeneity occurs when a variable, observed or unobserved, that is not included in our models, is related to a variable we incorporated in our model. In this paper, we confront one potential problem of endogeneity related to unobserved heterogeneity that simultaneously affects health and fuel poverty (Awaworyi Churchill & Smyth [2018]). We adopt a two-step instrumental method (Heckman [1979]) by first estimating fuel poverty.

30 To confirm the presence of endogeneity, we considered instruments that can explain fuel poverty and potential health problems not directly measured by the variables introduced in the model. A good instrument should explain fuel poverty, but not health status. Thus, energy prices appear to meet this requirement. Indeed, electricity and natural gas prices are used as instruments as they are the most commonly used sources of energy for heating (Ambrosio et al. [2015]), but intuitively, there is no reason why they should explain the state of health of the individual. Energy prices are often recognized as one of the most important determinants of residential energy demand (see Labandeira [2017], Labandeira et al. [2017] for a literature review) and as a consequence of energy expenditure. Spending a large share of one’s income on energy is one of the main components of fuel poverty (Charlier & Kahouli [2019]; ONPE [2014]). We extrapolate the fuel price rate of each household using its fuel bills and the surface area of the dwelling.

31 In order to estimate fuel poverty more precisely, we also introduce another variable as an instrument to consider energy efficiency from the solar exposure of the dwelling (Charlier & Kahouli [2019]), a binary variable if the dwelling is dark. Exposure is a criterion that is taken into consideration to measure the theoretical energy efficiency of a dwelling (and therefore the energy requirement). How a house is oriented in relation to the sun can have an impact on heating and cooling expenditures, which are the largest energy contributor in most homes. In a building heated by passive solar energy, glass areas are oriented and arranged so as to optimize the capture of solar light and heat. When buildings are highly insulated and energy efficient, passive solar energy can meet a substantial share of the heating demand, even in cold climates (Laustsen [2008]). However, because a building’s exposure can vary over the year and the day, it is necessary to store and balance solar energy. If too much solar energy is captured, it could require cooling which offsets the efficiency gains. Thus, exposure appears to be a good explanatory factor for energy expenditure. However, window exposure should not have a direct impact on health.

32 The predicted value of fuel poverty is thus obtained using equation 5:

33 With p_it = 1 if and 0 otherwise.

34 Gas_p_it, Elec_p_it are the gas price and the electricity price, GasElec_p_it the interaction parameter between both energy prices to account for potential multicollinearity between evolution of energy prices [13] and D_it, a dummy variable indicating if the dwelling is dark.

35 The third specification of the model is a dynamic probit model, with correction of the endogeneity bias caused by the fuel poverty variable (Model 3):

36 Finally, the fourth specification controls the sensitivity to the choice of health variable. We estimate the third model and replace the health variable with a binary variable for chronic disease (Model 4). Then, we apply bootstrap techniques in each step to avoid bias (Efron & Tibshirani [1993]; Horowitz [2003]).

3. Results and discussion

37 The main results are reported in Table 2. For the sake of clarity, the results of the initial conditions equation and the endogeneity treatment of fuel poverty equation are reported in the Appendix (See Appendix 2).

38 The estimates confirm the negative impact of fuel poverty on health status, as expected in the theoretical model, and as previously established in different cross-sectional case studies. Using a dynamic probit panel model, we found that being fuel poor significantly increases the probability of being in bad health (Table 2, Model 1). The marginal effect of fuel poverty on poor health reaches 1.9% [14] (Table 3, Model 1) for an individual. These results persist even after controlling for the state dependency of health and the initial conditions (Table 2, Model 2), which demonstrates the importance of health trajectories: being fuel poor increases the probability of being in bad health by 1.7% (Table 3, Model 2) for individuals in good health in the previous period, and 6.5% for those already in poor health.

39 However, treating the endogeneity of fuel poverty modifies the marginal effect substantially. The impact is thus strongly reinforced (Table 2, Model 3): being fuel poor increases the risk of bad health by more than a factor of 7 (Table 3, Model 3) for those already in poor health, and by a factor of 1.82 for those who were in good health in the previous year. Finally, to test the robustness of the results established by Models 1 to 3, we use an objective measure of health, the existence of chronic diseases (Model 4). This last model confirms that being fuel poor multiplies the risk of chronic disease by more than a factor of 6.14 (Table 3, Model 4) for those in good health (no chronic disease) in the previous year, and by 7.88 for those who already reported at least one chronic disease. If we account for the unobserved heterogeneity affecting exposure to fuel poverty and deteriorating health simultaneously, we can conclude there is a very strong causal relationship between both phenomena.

40 The empirical strategy adopted first shows that the nature of the health variable has to be accounted for, and second that neglecting endogeneity of fuel poverty would lead to a significant underestimation of the health risk. The entire health trajectory of each individual has to be controlled: the intermediate estimate consisting of the control of initial conditions shows that medical density is a good instrument to explain the health of individuals before their entry in the panel (Table 6, Appendix 2A). Medical density significantly decreases the probability of being in poor health and of having a chronic disease. The first step in Models 3 and 4 which provides the estimated fuel poverty probability also allows us to conclude that energy prices are good instruments and strongly affect exposure to fuel poverty (Table 7 in Appendix 2B). When electricity and gas prices increase, the risk of fuel poverty increases significantly for each individual. Having a dark dwelling also significantly increases the probability of being poor. This step enables us to provide a good prediction of fuel poverty: the correct prediction rate is about 87%.

41 To test the robustness of the results established in Models 3 and 4, we introduced an alternative indicator of fuel poverty: the low-income high costs indicator. According to this definition, households are fuel-poor if they have an income after fuel costs that is below the poverty threshold [15] as well as fuel costs above the median population fuel cost. These estimates are reported in the Appendix (Table 8 in Appendix 2C). We have validated the results and particularly the size of the effect: being fuel poor leads to an increase in the risk of bad health by more than a factor of 2 for a person previously in good health and by more than 8 for those already in poor health (Table 9 Model 3). The risk of reporting a chronic disease increases by a factor of 4.38 for a healthy person (Table 9 Model 4).

Table 2. Results – probability of being in poor health

Table 2. Results – probability of being in poor health

Robust standard errors in parentheses
*** p < 0.01, ** p < 0.05, * p < 0.1

Table 3. Marginal effect of fuel poverty (10%) on poor health

Model 1	0.019***
	Good health in t − 1	Poor health in t − 1
Model 2	0.017***	0.065***
Model 3	1.82***	7.05***
Model 4	6.14***	7.88***

Table 3. Marginal effect of fuel poverty (10%) on poor health

*** p < 0.01, ** p < 0.05, * p < 0.1

42 Finally, using the same methodology, we wanted to identify the threshold where the effort rate significantly increases the probability of being in poor health. Results for predictive values are presented in Figure 3 below.

Figure 3. Predictive margins for effort rate – linear prediction

Figure 3. Predictive margins for effort rate – linear prediction

43 The elements revealed in Figure 2 are confirmed when replaced in Model 3 (the most complete, the variable of fuel poverty multiplied by the effort rate). The results are shown in Table 4: when the effort rate increases marginally, the risk of reporting a state of poor health increases by 5.5.

Table 4. Results – probability of being in poor health

Poor health (lag)	1.613***
	(0.0329)
Predicted effort rate	5.505***
	(1.522)
Number of children	0.0158*
	(0.00936)
Unified Degree Days (log)	– 0.0245
	(0.0408)
Age	0.00172
	(0.00370)
Pollution problem	0.0203
	(0.0334)
Undergraduate degree	0.0514
	(0.0482)
Homeowner	-0.0560
	(0.0537)
Mills ratio	– 0.853***
	(0.165)
Panel-level variance (log)	– 1.727***
	(0.132)
Constant	– 0.433
	(0.526)
Observations	122,345
Number of individuals	37,855

Table 4. Results – probability of being in poor health

Robust standard errors in parentheses – Bootstrap 5000 replications
*** p < 0.01, ** p < 0.05, * p < 0.1

44 Our results reinforce the estimates provided by the National Housing Federation in 2010: the cost for treating ill health resulting from poor housing conditions has been estimated at GBP 2.5 billion per year in the UK. Eurofound [2016] models the cost of housing inadequacies and estimated that in France, 930 million euros in medical expenses would be saved by repairing inadequate housing. The previous reports combined with our results lead us to conclude that spillover effects exist in public policy: implementing policies which promote energy efficiency, such as encouraging investment in retrofitting are useful tools for tackling fuel poverty, and also reduce medical spending.

45 Retrofitting policies, in this sense, go beyond the simple monetary benefits: a household could occupy an energy-inefficient dwelling and cope with it by means of a high income. Retrofitting measures would still be highly beneficial to them due to improvement of the environment. In the context of continuous increases in energy prices in Europe as shown by Eurostat [2019], investment opportunities created by the need for energy efficiency are vast and highly profitable (Amstalden et al. [2007]). Retrofitting plans and energy efficiency measures have an impact on fuel poverty through different channels. First, they directly and robustly improve living conditions (reduce cold, dampness, mold). Second, they lead indirectly to a decrease in energy costs, particularly because the housing becomes less energy intensive. Finally, reducing residential energy consumption/demand reduces carbon dioxide emissions, which is an important component of the preservation of climate and natural resources, and the environment as a whole. “For example, cleaner energy systems could reduce carbon emissions, and cut the burden of household air pollution, which causes some 4.3 million deaths per year, and ambient air pollution, which causes about 3 million deaths every year” (WHO [2018]).

46 More broadly, our results beg the question whether it is possible to achieve energy transition without considering the impact of fuel costs on fuel poverty and the indirect impact on public health. Achieving energy transition without incurring excessive economic and social costs requires anticipating long chain reactions.

4. Conclusion

47 Fuel poverty has been treated extensively in recent literature. But very few studies target the causal link between fuel poverty and poor health. Numerous papers focus on specific experiments assessing the link between some timely retrofitting policies and the health of occupants (Chapman et al. [2009]; Ezratty et al. [2009]; Howden-Chapman [2015]; Howden-Chapman et al. [2007]; Lloyd et al. [2008]; Thomson & Snell [2013]). To the best of our knowledge only two papers (Lacroix & Chaton [2015]; Liddell & Morris [2010]) describe non-experimental and cross-sectional research on the relationship between fuel poverty and health. Our study aims to go beyond the limits of this previous work, namely the overly specific aspects of experimental studies and the non-generalizable results of cross-sectional research that neglect climate hazards and the path dependency of health.

48 In the present paper, we estimate the causal effect of fuel poverty on health status. To do this, we used panel data from the EU-SILC over the period 2008 to 2016 and estimated probit panel models in four steps. We begin with a simple probit model without controlling for state dependency of health and endogeneity of fuel poverty, and finish with a dynamic probit model controlling for both biases.

49 We conclude that fuel poverty has a significant impact on health. Being fuel poor increases the risk of bad health by slightly more than a factor of 7 for those already in poor health (1.82 for those in good health) and by 7.88 the risk of having a chronic disease for those who already have a chronic disease (6.14 for those previously in good health). In terms of policy recommendations, we stress the importance of energy-efficiency retrofits. Our results allow us to make the link between experimental studies and medical literature demonstrating the effects of humidity, cold and indoor pollution on health. Even more than underscoring the importance of supporting renovations, our results show what kind of renovations could yield a double dividend, both in terms of well-being for the inhabitants and public health expenditures. It thus appears that public policies should help households to improve their insulation, whether roof or wall insulation. At the same time, ensuring households have efficient and safe heating sources can also improve public health.

50 Lessons learned from our empirical strategy also enable us to conclude that neglecting path dependency of health and endogeneity of fuel poverty leads to a substantial underestimation of the impact on health. By controlling both biases, we first show that medical density must be accounted for to achieve a good understanding of the health history of each individual. Second, as energy prices are good instruments for controlling endogeneity, our results question the inextricably linked issues of energy transition, combatting fuel poverty and safeguarding public health. Is it possible to achieve energy transition without excessive economic and social costs, i.e., without increasing fuel poverty and damaging public health?