Energy poverty and health: Micro-level evidence from Germany

Buchner, Martin; Rehm, Miriam Working Paper Energy poverty and health: Micro-level evidence from Germany ifso working paper, No. 48 Provided in Cooperation with: University of Duisburg-Essen, Institute for Socioeconomics (ifso) Suggested Citation: Buchner, Martin; Rehm, Miriam (2025) : Energy poverty and health: Micro-level evidence from Germany, ifso working paper, No. 48, University of Duisburg-Essen, Institute for Socio-Economics (ifso), Duisburg This Version is available at: https://hdl.handle.net/10419/313638 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by/4.0 uni-due.de/soziooekonomie/wp ifso working paper Martin Buchner Miriam Rehm Energy Poverty and Health: Micro-Level Evidence from Germany 2025 no.48 Energy Poverty and Health: Micro-Level Evidence from Germany∗ Martin Buchner†Miriam Rehm‡ October 2024 ∗We thank Bruno Brinkmeier for excellent assistance in the preparation of this paper, and participants of the IfSO research seminar, the “Political Economy of Inequality” summer school, and the EuHEA PhD Conference for comments and feedback. Naturally, all remaining errors are our own. Martin Buchner gratefully acknowledges PhD funding from the Hans-Böckler Foundation. †RWI – Leibniz Institute for Economic Research & University of Duisburg-Essen, Institute for SocioEconomics, martin.buc[email protected] ‡University of Duisburg-Essen, Institute for Socio-Economics, [email protected] Abstract This paper aims to understand the health effects of energy poverty in Germany using SOEP panel data from 2010 to 2020. Linear probability models and fixed effects ordered logit models reveal a consistently negative relationship of three expendituresbased energy poverty indicators with general health. The association is stronger for the subjective energy poverty metric: members of households unable to keep the home comfortably warm due to financial reasons have an about 3.23 p.p. lower probability of being in at least satisfactory health. Investigating potential channels shows that mental health is consistently negatively linked to our energy poverty metrics, while physical health is weakly associated with energy poverty in Germany, with the exception of doctor visits. Finally, by instrumenting energy poverty with data on energy price indices and matching energy costs to the heating systems used by households, we show that living in a household that experiences a transition to energy poverty due to rising energy prices is also linked to a lower likelihood of being in good health. Keywords: energy poverty, health, fixed effects ordered logit models, Germany JEL Codes: I10, I32, Q41 1 1 Introduction Energy poverty of households can make their members sick. When households struggle to attain an adequate level of energy services, leading to energy or fuel poverty (Boardman, 1991; Bouzarovski, 2014), it can result in financial strain on the one hand, and to insufficient heating (and cooling) of living quarters on the other hand (Davillas et al., 2022). Such circumstances can precipitate various physical health problems, including increased risk of hypertension, inflammation, cardiovascular diseases, thrombosis, and respiratory illnesses in children, as well as mental health issues (Gallerani et al., 2004; Fares, 2013; BallesterosArjona et al., 2022). Energy poverty is thus a policy concern both nationally, highlighted by advocacy groups in the United Kingdom since the 1970s, and at the European level, where the European Commission has issued a recommendation on addressing energy poverty (European Commission, 2020). In Germany, recent estimates suggest a prevalence of energy poverty of about 6.6% (Destatis, 2023). With short-run spikes in energy prices and longterm trends stemming from the German energy transition, the financial burden of energy costs on households is likely to increase. This paper aims to investigate whether a robust link exists between energy poverty and mental and physical health in Germany. While there is ample research on lowand middle-income countries (Banerjee et al., 2021; Jayasinghe et al., 2021; Nawaz, 2021; Nie et al., 2021; Pan et al., 2021), recent studies have uncovered robust relationships between energy poverty and health in several highincome countries. Notably, four key studies using survey panel data for France (Kahouli, 2020; Baudu et al., 2020, EU-SILC), Australia (Churchill and Smyth, 2021, HILDA), and the UK (Davillas et al., 2022, UKHLS) have developed a framework for estimating the causal effect of various energy poverty measures on objective and self-rated health outcomes. Self-reported health serves as a common outcome variable across these studies, with Churchill and Smyth (2021) employing a composite indicator of subjective health and Davillas et al. (2022) in addition accessing objective health data from blood samples. The explanatory variable, energy poverty, is measured through both objective and subjective dimensions, with all three studies employing the low income-high cost measure1(Hills, 2012) as an objective measure, along with subjective assessments regarding the ability to adequately heat the home. Churchill and Smyth (2021) and Davillas et al. (2022) also incorporate a composite of objective and subjective measures for energy poverty. To address endogeneity concerns, energy poverty is instrumented with regional energy prices in all three studies, given their correlation with the potentially endogenous explanatory variable, energy poverty, and their presumed lack of direct linkage to health outcomes. 1This measure indicates that the share of energy expenditures exceeds 10% of equivalised disposable household income. 2 Additionally, standard controls include individual and dwelling characteristics, as well as climate factors. This research yields valuable insights into the relationship between energy poverty and health. In addition to the robust headline correlation of energy poverty and adverse health outcomes (Champagne et al., 2023), it shows consistent findings even when going into more detail. For instance, subjectively measured energy poverty exhibits a stronger negative effect on health compared to measures assessing the financial burden of energy costs, such as the "ten percent rule" (energy costs amounting to more than ten percent of the household budget) or the "low income-high cost" indicator (a composite indicator based on its two namesake concepts). However, some substantial variations persist. Notably, instrumental variable (IV) effect sizes vary considerably across studies (e.g., Baudu et al., 2020; Kahouli, 2020; Davillas et al., 2022; Churchill and Smyth, 2021), emphasizing the significance of the concrete IV specification but also of national institutional characteristics. Germany is characterised by a more decentralised institutional framework, including the social policy regime, than France or the United Kingdom, and by a less temperate climate than Australia or France. Thus, this study seeks to ascertain whether a robust link between energy poverty and health outcomes exists in the German context.2Leveraging representative panel data from the German Socio-Economic Panel (SOEP) spanning from 2010 to 2020, our outcome variable health is measured subjectively using a self-assessed indicator. Self-rated health metrics are recognized as reliable proxies for actual health status, combining both mental and physical health considerations (Schnittker and Bacak, 2014), and are commonly employed in empirical health economics (e.g., Kuehnle and Wunder, 2017) including studies on energy poverty (Llorca et al., 2020). For energy poverty measurement, we employ three expenditure-based metrics that consider the ratio of household income to energy expenditures alongside a subjective indicator, which constitute the most widely used metrics in energy poverty research (Brabo-Catala et al., 2024). The latter may better capture underconsumption of energy due to financial constraints and it avoids labelling high-income, high-energy use households as energy-poor (Thema and Vondung, 2020; Drescher and Janzen, 2021). Our estimation strategy aims to establish a robust link between energy poverty and health outcomes. Initially, we fit a linear probability model before turning to panel-data ordered logit models with fixed effects using the blow-up and cluster (BUC) estimator proposed by Baetschmann et al. (2020) and Baetschmann et al. (2015). This approach 2Reibling and Jutz (2017) present a first quantitative indication for a negative correlation between energy poverty and mental health. Due to data restrictions, however, this study focuses exclusively on heating expenditures (without taking electricity costs into account), uses a single expenditure-based measure of energy poverty, and cannot assess causality. 3 accommodates the ordered scaling of the dependent variable (self-rated health) and exploits the panel structure of the data by controlling for potential unobserved individual timeinvariant confounders. Additionally, we address possible biases in two-way fixed effects models with staggered and intermittent treatment by implementing the innovative Fixed Effects counterfactual estimator proposed by Liu et al. (2024). Across all three approaches, we consistently observe a statistically significant negative association between expenditurebased energy poverty indicators and overall health. Notably, both the linear probability model and the fixed effects ordered logit model show that the link between subjective energy poverty and health is especially strong. Subsequently, we invesigate potential channels through which energy poverty impacts physical and mental health outcomes. Our findings suggest that the negative association primarily stems from deteriorations in mental rather than physical health in Germany, as evidenced by composite indicators and individual variables. Furthermore, we adopt an IV approach to address the potential endogeneity between health outcomes and energy poverty. To this end, we instrument energy poverty using data on price indices for oil, gas, district heat, solid fuels, and electricity. Our data permit us to identify the primary energy source households use for heating, which is crucial for instrumenting energy poverty since leveraging the variation in energy prices is the leading approach in instrumenting energy poverty. They are generally assumed to be relevant, since price increases are likely to be correlated with energy poverty. The exclusion restriction is harder to satisfy (Kahouli, 2020; Churchill and Smyth, 2021): An increase in energy prices can affect health status beyond inducing energy poverty. For instance, it may prompt households to reduce direct expenditures on health-related products and services, such as gym memberships or healthy diets, in response to the price surge. However, this substitution effect is less of a concern if energy expenditure changes are small relative to total expenditures or if health expenditures are price inelastic (Davillas et al., 2022). The results of the IV point to a strong negative link between energy poverty and health, and are thus in line both with our multivariate findings and the literature. Finally, we extensively test the robustness of these results by checking that our findings are not influenced by features of the German welfare state, which covers heating expenditures with some social transfers. Additionally, we restrict the sample to direct survey respondents and non-COVID years, and exclude over-sampled migrants. The rest of this paper is structured as follows: Section 2 outlines our methodology and the empirical strategy, while Section 3 describes the data and variables. Section 4 presents our findings, commencing with the linear probability models in Section 4.1, followed by fixed-effects ordered logit models in Section 4.2, and two-way fixed effects counterfactual 4 models in Section 4.3. Section 4.4 explores potential channels, and Section 4.5 addresses endogeneity with the IV approach. Section 5 examines the robustness of our results, and Section 6 concludes. 2 Empirical Strategy To assess the relationship between energy poverty and an individual’s overall health status, we first utilize a Linear Probability Model (LPM) for consistency with the existing literature. LPMs offer two key advantages: first, their results afford a simple and intuitive interpretation; and second, they are well-established in the literature and thus facilitate comparability with other studies, such as Kahouli (2020). The estimated equation is as follows: good_healthit =β1EPit +X n βnXn,it +ϑi+γt+εit,(1) where good_healthit is the self-rated health of respondent iat time t, which is assigned a value of 1 if the respondent’s self-rated health status is “very good”, “good” or “satisfactory” and 0 if the response is “poor” or “bad”. EP is the respective indicator for energy poverty and Xis a vector of n individual time-varying observed control variables. Individual fixed effects are captured by ϑi, while γtdenotes wave dummies covering time fixed effects. We proceed by employing fixed-effects ordered logistic regression models, our preferred approach. These models are advantageous as they avoid the need to dichotomize the dependent variable, self-rated health, which is measured on an ordered categorical scale, while still leveraging the panel data structure to account for unobserved timeand individualinvariant factors. However, in order to estimate the fixed effects consistently despite the so-called incidental parameters problem (Lancaster, 2000), Baetschmann et al. (2015) introduce a method leveraging the Conditional Maximum Likelihood (CML) estimator. For ordered categorial dependent variables, the Blow-Up and Cluster estimator (BUC) combines the „information of the CML estimators obtained from dichotomizing samples at different cutoff points“ (Baetschmann et al., 2020) by replacing the sample with copies of itself and applying the CML estimator for the entire enlarged sample.3The BUC-approach not only allows for a consistent estimation of fixed effects ordered logit models, but it also does not rely on the assumption that the thresholds are constant across individuals, in contrast to standard ordered logit models for cross-sectional data. 3Since the copies of the same units are not independent of each other, standard errors are clustered at the individual level. 5 We operationalize the fixed-effects ordered logistic models by employing self-rated health as dependent variable, which is scaled from 1 (bad) to 5 (very good health status). The approach can handle the same independent variables, including energy poverty, individual and year fixed effects, and controls, as the linear probability model. Although the fixed effects ordered logit model is our preferred approach given the data structure, it has been shown that TWFE estimates may suffer from bias stemming from weighting issues, particularly when treatment effects exhibit temporal variation in staggered treatment designs (Chaisemartin and d’Haultfoeuille, 2023; Goodman-Bacon, 2021; Chaisemartin and d’Haultfoeuille, 2020). Our panel data is susceptible to this concern, given that individuals residing in households experience transitions in and out of energy poverty throughout our analysis period.4To address this issue, we employ the innovative Fixed Effects counterfactual estimator (FEct) proposed by Liu et al. (2024), which calculates the average treatment effect on the treated by directly imputing counterfactual outcomes for treated observations. This approach treats treated observations as missing during modelling and estimates the counterfactual outcome as a weighted sum of all untreated observations. This method compares the observed outcome of treated observations with the predicted counterfactual for each matched pair, thus removing biases caused by improper weighting that affect conventional TWFE models (Liu et al., 2024). We employ the FEct estimator for our expenditure-based energy poverty indicators, test for the parallel trends assumption, and perform placebo tests by removing pre-treatment observations used at the modelling stage. Despite controlling for time-invariant heterogeneity and numerous potential time-variant confounding variables in all discussed models, as well as conducting tests for pre-trends and placebo treatments in the FEct models, concerns regarding identification may still arise if energy poverty is correlated with the error term. There are at least three possible sources of endogeneity: First, unobserved time-varying confounding variables not accounted for in the model specifications (see Equation 1) could bias the estimation of the regression coefficients. For instance, individual expectations concerning job loss or income may influence both health and energy poverty (Kahouli, 2020). A second potential source of endogeneity is reverse causation. For instance, individuals in poor health might earn less, thereby increasing the likelihood of being in a state of energy poverty, which could bias the regression estimates upward. Finally, endogeneity may stem from measurement error. Survey respondents might not accurately recall their energy bills, leading to systematic overor underestimation of their annual energy expenditures. 4Figure A5 in the Appendix provides a visual representation of individuals’ treatment status in our sample over time. 6 from the Federal Statistical Office (Statistisches Bundesamt, 2024). However, it is important to note that due to a methodological change in the calculation of the price indices, data for the district heat price index begins in 2015. We incorporate one-year lagged energy price index values for our Instrumental Variable (IV) estimates, driven by two considerations: First, in the SOEP dataset, homeowners typically report energy expenditures from the preceding year, while renters, who report current expenditures, are influenced by utility bills based on prior energy prices. Second, given that SOEP interviews predominantly occur during the spring months14, energy prices from the previous year retain greater salience when respondents report their energy expenditure. Consequently, our instrument comprises 49 unique values, reflecting variations across 5 heating types and 11 time periods (see Figure 3). In assigning heating systems to households, we leverage a unique feature of the SOEP dataset. In 2015 and 2020, households were surveyed about their expenditure on heating per energy source15. When heating costs are reported for a single energy source, we assign that specific heating system to the household for the respective year. In instances where expenses are incurred for multiple energy sources, we prioritise based on the following hierarchy, reflecting the relative frequency in Germany: gas > oil > district heating > electricity > solid fuels > other (Destatis, 2022). We extend this assigned heating system to all years preceding and succeeding the observed year, provided the household does not report a move. We assume that as long as there is no move, the household’s heating system remains unchanged. However, if the heating systems differ between 2015 and 2020, individuals residing in those households are excluded from our instrumental variable (IV) analysis, as we cannot determine the timing of the switch. Our IV analysis thus only includes observations from households that consistently utilize the same energy source for heating over time. The summary statistics for the reduced subsample, consisting of individuals residing in households where we have successfully assigned a heating system and obtained the corresponding price index from the previous year, are presented in Table A2. Remarkably, no substantial deviations are observed between this reduced sample and our main sample concerning the dependent variable good_health and the energy poverty indicators. 14See Figure A4 in the Appendix, which shows that approximately 50% of interviews take place in February and March, and over 75% occur in the first half of the year. 15Expenditures can be reported across various categories, including oil, gas, district heating, electricity, environmental, pellets, coal, biomass, liquid gas, and solar. For the analysis presented in this paper, we aggregate these categories into gas, oil, district heating, electricity, solid fuels, and other, as price data are available only for the first five of these energy sources. 13 Figure 3: Energy price indices This figure shows the energy price indices for gas, oil, district heat, electricity, and solid fules, indexed to 2020 as 100. We employ the one-year lag of the price indices in our instrumental variable (IV) models. Data: Statistisches Bundesamt (2024). 4 Results This section first presents the results for energy poverty and health through standard linear probability models, ordered logistic fixed-effects, and two-way fixed effects counterfactual models. It then moves to examining the evidence for potential channels between the two and attempts to address their potential endogeneity using an IV approach. It concludes with robustness checks. 4.1 Linear Probability Models Table 2 shows linear probability models focusing on the threshold between very good to satisfactory health versus poor to bad health status. The results indicate a negative correlation between health status and energy poverty across all four energy poverty indicators (columns 1-4). Notably, as shown in column 4, the effect size for the subjective energy poverty indicator is substantially larger compared to the expenditure-based indicators. When energy poverty is subjectively measured, it is associated with a 3.2 percentage 14 points lower probability of reporting at least a satisfactory health status. All four specifications in Table 3 control for other factors potentially related to health as discussed in Section 3, including logarithmized and equivalized household income, income poverty, socio-economic characteristics such as age and age squared, education, labour force status (including being registered as unemployed), marital status, and the number of dependants living in the household. Most relevant for our research question are the covariates capturing income poverty and unemployment, both of which are statistically significantly associated with worse health. 4.2 Ordered Logistic Fixed-Effects Models Next, we examine whether the negative association between energy poverty and health persists when considering the ordered categorical nature of the dependent variable and the panel data structure. Fixed-effects ordered logistic estimates of the relationship between energy poverty and general health variables are presented in odds ratios in Table 3. Similar to the linear probability model, we incorporate the full set of controls. The exponentiated coefficients for all three expenditure-based energy poverty indicators (columns 1-3) are statistically significant at least at the 5%-level and smaller than one, suggesting that energy poverty is associated with a decrease in the odds ratio for being in better health categories. For instance, the odds ratio of 0.945 in column 1 indicates that living in a household classified as energy poor by the Ten Percent Rule (tpr) decreases the odds ratio of being in a better health category by about 5.5%. In other words, individuals in Ten Percent Rule energy-poor households have 0.945 times the odds of being in better health categories than those in non-energy-poor households. Similarly, the odds ratios for the Two Times Median Share of Income (mtwo) and the Low Income High Cost (lihc) indicators are 0.943 and 0.919, respectively. As shown in column 4, the odds ratio for the subjective energy poverty indicator is substantially larger than those of the expenditure-based indicators. Its point estimate of 0.818 indicates that living in a household unable to keep the home comfortably warm due to financial reasons decreases the odds ratio of being in a better health category by about 18,2%. Table 4 presents the marginal effects at the sample average, which may be more intuitive to interpret.16 These indicate that being energy poor according to our four metrics (columns 1-4) is associated with lower probabilities of being in the two best health categories (lines 4-5) and higher probabilities of being in the three lower health categories 16Marginal effects at the average use the relative frequencies of the corresponding categories of the outcome variable in the estimation sample to compute the sample average. Note that the marginal effects at the sample average differ from the marginal effects at the average of the regressors and also differ from the average marginal effect. For a more detailed discussion, see Baetschmann et al. (2020). 15 Table 2: Energy poverty and health – linear probability model (1) (2) (3) (4) Good health Good health Good health Good health Ten percent rule (tpr) -0.00828∗∗∗ [0.00268] Two times median share (mtwo) -0.00879∗∗ [0.00370] Low income high costs (lihc) -0.00892∗∗ [0.00396] subjective -0.0323∗∗ [0.0134] Income poverty -0.0108∗∗∗ -0.0112∗∗∗ -0.0114∗∗∗ -0.0119 [0.00375] [0.00373] [0.00372] [0.00723] Log of eq. HH income 0.00270 0.00366 0.00448 0.00443 [0.00364] [0.00361] [0.00357] [0.00681] Age 0.00215∗0.00218∗0.00213∗0.0179∗∗∗ [0.00111] [0.00111] [0.00111] [0.00346] Age squared -0.0000812∗∗∗ -0.0000810∗∗∗ -0.0000808∗∗∗ -0.000209∗∗∗ [0.0000107] [0.0000107] [0.0000107] [0.0000335] Registered Unemployed -0.0551∗∗∗ -0.0554∗∗∗ -0.0554∗∗∗ -0.0542∗∗∗ [0.00497] [0.00497] [0.00497] [0.0100] Not employed -0.0202∗∗∗ -0.0203∗∗∗ -0.0202∗∗∗ -0.00912 [0.00399] [0.00399] [0.00399] [0.00836] In Education / Apprentice / Community Service -0.00177 -0.00188 -0.00181 -0.0116 [0.00469] [0.00469] [0.00469] [0.00925] Pensioner 0.0125∗∗ 0.0124∗∗ 0.0125∗∗ 0.0129 [0.00569] [0.00569] [0.00569] [0.0111] High School 0.00338 0.00362 0.00364 -0.0299 [0.00868] [0.00867] [0.00867] [0.0182] More than High School 0.00380 0.00404 0.00393 -0.0470∗ [0.0116] [0.0116] [0.0116] [0.0245] Married 0.00182 0.00180 0.00180 0.00165 [0.00550] [0.00550] [0.00550] [0.0137] Separated or divorced 0.0118 0.0115 0.0115 0.0166 [0.00832] [0.00832] [0.00832] [0.0196] Widowed -0.00288 -0.00294 -0.00360 0.00641 [0.0143] [0.0144] [0.0144] [0.0342] Number of dependants in hh 0.000885 0.000942 0.00101 -0.00453 [0.00185] [0.00185] [0.00185] [0.00441] Constant 0.926∗∗∗ 0.916∗∗∗ 0.912∗∗∗ 0.512∗∗∗ [0.0341] [0.0340] [0.0338] [0.0945] Individual Fixed Effects Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes n (total) 255684 255684 255684 86211 This table shows the estimates of a linear probability model for the four energy poverty indicators on the probability of being in good health (dichotomized as 1 for very good, good or satisfactory, and 0 for poor or bad). The reference category for labour force status is (self-)employed, for education is less than high school, and for marital status is single. Cluster robust standard errors in brackets. ∗p < 0.1,∗∗ p < 0.05,∗∗∗ p < 0.01 16 Table 3: Ordered logistic regression (1) (2) (3) (4) Self rated Self rated Self rated Self rated health health health health Ten percent rule (tpr) 0.945∗∗∗ [0.0176] Two times median share (mtwo) 0.943∗∗ [0.0230] Low income high costs (lihc) 0.919∗∗∗ [0.0235] subjective 0.818∗∗ [0.0744] Socio-econ. controls Yes Yes Yes Yes Control for income poverty Yes Yes Yes Yes Individual Fixed Effects Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes Observations (total) 255684 255684 255684 86211 Observations (w. variation in outcome) 206662 206662 206662 52671 Panel units (w. variation in outcome) 31797 31797 31797 15581 This table shows the exponentiated coefficients of a fixed-effects ordered logistic regression for the four energy poverty indicators on self-rated health measured on a 5-point scale with 1 = bad and 5 = very good. Cluster robust standard errors in brackets. Data: SOEP (2022). ∗p < 0.1,∗∗ p < 0.05,∗∗∗ p < 0.01 (lines 1-3). The results suggest that if a household falls into Ten Percent Rule energy poverty, there is a decrease of 1.4 percentage points in the probability of being in good or very good health for its members. These results are consistent across all four indicators; for subjective energy poverty, the association again appears to be the strongest, with a reduction of approximately 5 percentage points in the probability of being in good or very good health. 4.3 Two-way Fixed Effects Counterfactual Models As a third modeling approach, we employ fixed effects counterfactual models (FEct) to counteract potential weighting issues stemming from our staggered and intermittent treatment. The dependent variable is a dummy variable that captures whether individuals report satisfactory, good, or very good health, as opposed to poor or bad health. The main explanatory variables are the expenditure-based indicators, as the subjective indicator for energy poverty is available for too few periods to adequately test for parallel trends. Figures A6 to A8 in the Appendix provide a visual assessment of the parallel trends assumption, complemented by an F-test for zero residual averages in the pretreatment periods, where a higher F-test p-value indicates a better fit for pre-trend analysis. The ATT plots and the p-values of the FTest collectively indicate no substantial presence 17 Table 4: Marginal effects (1) (2) (3) (4) tpr mtwo lihc subjective self-rated health 1 0.00192∗∗∗ 0.00200∗∗ 0.00285∗∗∗ 0.00712∗∗ [0.000632] [0.000827] [0.000868] [0.00323] 2 0.00660∗∗∗ 0.00691∗∗ 0.00984∗∗∗ 0.0250∗∗ [0.00218] [0.00285] [0.00299] [0.0113] 3 0.00556∗∗∗ 0.00582∗∗ 0.00829∗∗∗ 0.0176∗∗ [0.00184] [0.00240] [0.00252] [0.00798] 4 -0.00910∗∗∗ -0.00953∗∗ -0.0136∗∗∗ -0.0317∗∗ [0.00300] [0.00393] [0.00413] [0.0144] 5 -0.00498∗∗∗ -0.00521∗∗ -0.00742∗∗∗ -0.0180∗∗ [0.00164] [0.00215] [0.00226] [0.00815] This table shows the marginal effects of the four energy poverty indicators in a fixed-effect ordered logistic regression, calculated at the sample average of the dependent variable (self-rated health on a 5-point scale with 1 = bad and 5 = very good). The reported standard errors are for effects at the sample average and not for effects at the population average. Data: SOEP (2022). ∗p < 0.1,∗∗ p < 0.05,∗∗∗ p < 0.01 of pretreatment differential trends for most models while the p-value of the t-test for the placebo test in the Low Income High Cost model is borderline statistically significant at the 10% level.17 Furthermore, we conduct a placebo test wherein all observations from periods -2 to 0 relative to the treatment timing are excluded. Instead, the untreated outcomes of these omitted periods are predicted using a model trained with the remaining untreated observations (Liu et al., 2024). The outcomes are depicted in Figures A9-A11 in the Appendix: In the placebo tests, we cannot reject the null hypothesis of an ATT = 0 for the Two Times Median Share of Income and Low Income High Cost models, with p-values of 0.967 and 0.825, respectively. This outcome supports the validity of the underlying assumptions for these models. 17This is noteworthy, especially considering that the F-Test is particularly sensitive to deviations from zero: „[W]hen the sample size is large, a small confounder (or a few outliers) that only contributes to a neglectable amount of bias in the causal estimates will almost always cause rejection of the null hypothesis of joint zero means using the F test“. (Liu et al., 2024). 18 Table 5: Fixed effects counterfactual models EP Indicator ATT N sd lower CI upper CI p-value tpr -0.01008 28502 0.003204 -0.01529 -0.005 0.002 mtwo -0.01268 13842 0.004397 -0.0197 -0.0051 0.004 lihc -0.01024 11966 0.004931 -0.01844 -0.00223 0.038 This table shows fixed effects counterfactual estimations. Standard errors are obtained through nonparametric bootstrap procedures (500 bootstrap runs). Units must have a minimum of 1 observed period under control to be considered. Data: SOEP (2022) The average treatment effects of the treated (ATT) are presented in Table 5. We observe similar but somewhat larger negative ATTs compared to the coefficients obtained from the linear probability model. Specifically, being in energy poverty is associated with a decrease in the probability of being in good or very good health by approximately 1 percentage point for the Ten Percent Rule and the Low Income High Cost indicator, and about 1.3 percentages points for the Two Times Median indicator. Our results thus far demonstrate a robust negative association between energy poverty and health in Germany, aligning with findings from the literature on other high-income countries. As discussed in Section 1, Kahouli (2020), Baudu et al. (2020), and Churchill and Smyth (2021), and Davillas et al. (2022) establish this relationship for France, Australia, and the UK. Moreover, Kahouli (2020) and Churchill and Smyth (2021) also observe stronger effects of consensual-based (i.e., subjective) indicators compared to expenditurebased metrics. In the next section, we explore potential channels through which energy poverty may affect health in Germany. 4.4 Potential Channels We now turn to exploring the potential channels through which energy poverty may impact health. Table 6 presents the results of fixed-effects regressions for our full model, utilizing mental health (left-hand side panel) and physical health (right-hand side panel) summary scales as dependent variables.18 The associations with our energy poverty metrics exhibit stark variations: Mental health is negatively associated with energy poverty, statistically significant at least at the 5% level for all our expenditure-based metrics. However, there appears to be no such association for physical health. While the magnitude of the coefficients for the mental health summary scale associ18Recall that the dependent variable is available biennially, leading to a reduced number of observations of 121,227. Consequently, we cannot utilize the subjective energy poverty metric due to its limited overlap with the interval censored physical and mental health summary scales. 19 ated with the three energy poverty metrics may not be directly interpretable due to the composite nature of the outcome variable, the coefficients exhibit consistent sizes. This association persists even after extensive controls for potential confounding factors such as income poverty and unemployment, mirroring the findings from the previous Section. Table 6: Mental and physical health summary scales – SF12-questionaire Mental Summary Scale Physical Summary Scale Ten percent rule (tpr) -0.354∗∗∗ -0.0438 [0.108] [0.0868] Two times median share (mtwo) -0.401∗∗∗ 0.0460 [0.146] [0.118] Low income high costs (lihc) -0.322∗∗ 0.241∗ [0.162] [0.130] Socio-econ. controls Yes Yes Yes Yes Yes Yes Control for income poverty Yes Yes Yes Yes Yes Yes Individual fixed effects Yes Yes Yes Yes Yes Yes Year fixed effects Yes Yes Yes Yes Yes Yes n 121227 121227 121227 121227 121227 121227 This table shows fixed-effects regressions for the three objective energy poverty indicators on mental and on physical health summary scales (ranging from 0 to 100). Due to data availability, this regression covers every other year (2010, 2012, ..., 2020). Individual clustered robust standard errors in parentheses. Data: SOEP (2022). ∗p < 0.1,∗∗ p < 0.05,∗∗∗ p < 0.01 Next, we explore these two channels further by employing individual items related to mental and physical health. Our data permits us to substitute the dependent variable with three “hard” indicators for physical health and five “soft” indicators for mental health: the logarithmized number of visits to a doctor’s office in the previous year, the number of stays at a hospital in the previous year, and the logarithmized number of days off work sick for physical health; and life satisfaction, the frequency of being happy or sad in the last four weeks, health satisfaction, and concerns about health for mental health. The results are reported in the Appendix. They reinforce the conclusions drawn from the mental and physical health summary scales: Physical health demonstrates a weak, if any, connection to energy poverty after adjusting for income poverty and unemployment. While the number of doctor visits (Table A3) displays some correlation with energy poverty, coefficients for hospital stays (Table A4) and the number of days off work (Table A5) are statistically insignificant. Conversely, indicators related to mental health consistently substantiate a link with energy poverty even after controlling for other factors. Life satisfaction (Table A6), health satisfaction (to a lesser extent, Table A7), and the frequency of being happy (Table A8) are negatively associated with all four energy poverty indicators, while the frequency of being sad (Table A9) and, less consistently, concerns about one’s health (Table A10) tend 20 to increase with energy poverty. This robust relationship thus supports the hypothesis that, at least within the German context, the negative health impacts of energy poverty may primarily originate from its effects on mental well-being. However, as discussed in Section 2, various potential sources of endogeneity of energy poverty to health necessitate caution in interpreting these findings. Consequently, the next section investigates whether these findings remain robust in an IV analysis. 4.5 Addressing Endogeneity To assess whether the negative association between energy poverty and health persists while addressing potential endogeneity, we employ 2SLS regressions, instrumenting for energy poverty with household-specific energy prices using energy price data. The results of the IV fixed-effects regression are presented in Table 7. We incorporate the same covariates as in our main specification, including a dummy for income poverty. The KleibergenPaap Wald F-statistic of the first stage suggests that the instrument is not weak for the Ten Percent Rule and Two Times Median Share of Income indicators (Stock and Yogo, 2005). However, for the Low Income High Cost indicator, the F-statistic is marginally larger than 10, indicating only a weak correlation between household-specific energy prices and this form of energy poverty. Notably, only the F-statistic for the Ten Percent Rule indicator exceeds the threshold of 104.7, as proposed by Lee et al. (2022). One potential explanation is that the Ten Percent Rule indicator does not rely on a reference population. Consequently, when energy prices rise, the median energy expenditure in the population is likely to rise too, somewhat offsetting the probability of becoming Two Times Median Share of Income or Low Income High Cost energy poor. Therefore, we primarily focus on the results derived from the Ten Percent Rule indicator. In the second stage of the 2SLS regression, we find a negative and statistically significant coefficient for both the Ten Percent Rule and the Two Times Median Share of Income metric. Living in a household that experiences a transition to Ten Percent Rule energy poverty due to rising energy prices is linked to a decline of being in good health of about 15.5 percentage points. For Two Times Median Share of Income energy poverty, this association is even stronger at about 37.2 percentage points. However, given the strong correlation in the first stage of the 2SLS estimates, we consider the Ten Percent Rule indicator our preferred measure of energy poverty here. 21 Table 7: Energy poverty and health (IV estimates) tpr mtwo lihc Second stage Energy poverty -0.1546∗-0.3717∗-0.7364 [0.0876] [0.2162] [0.4656)] Observations 146,331 146,331 146,331 First stage Energy prices (CPI) 0.0008∗∗∗ 0.0004∗∗∗ 0.0002∗∗∗ [0.0000771] [0.0000561] [0.0000533] Kleibergen-Paap Wald F statistic 120.49 39.376 11.114 This table shows the instrumental variable estimates for the three objective energy poverty indicators on the probability of being in good health (dichotomized as 1 for very good, good or satisfactory, and 0 for poor or bad). All regressions include the usual covariates. Individual clustered robust standard errors in parentheses. Data: SOEP (2022). *p < 0.1,** p < 0.05,*** p < 0.01 The endogeneity bias-corrected estimates derived from the 2SLS model appear substantial, a pattern consistent with much of the literature utilizing energy prices as instruments (Churchill and Smyth (2021); Churchill and Smyth (2020) Kahouli (2020); Prakash and Munyanyi (2021)). However, this may indicate lingering concerns regarding the exclusion restriction, urging caution against an overly confident interpretation of causality. Nonetheless, the sign of the coefficients associated with energy poverty in the 2SLS estimates remains consistent with the baseline specification, reinforcing our overall conclusion that energy poverty significantly and detrimentally impacts individuals’ health. 5 Robustness Checks We test the robustness of our findings through several strategies. First, we address a specificity of the German welfare system by excluding households potentially receiving transfers covering energy costs. Second, we restrict the analysis to household heads, since they respond to the SOEP household questionnaire which contains the energy expenditurerelated questions. Third, we omit the year 2020 from our analysis to mitigate potential distortions attributable to the COVID-19 pandemic. Finally, we eliminate respondents who were part of the migration samples in the SOEP data to ensure the consistency and reliability of our results. First, certain households in Germany receive transfers covering housing costs, including heating (“Arbeitslosengeld” or “Wohngeld”). Although our data do not allow for the 22 A Appendix A.1 Data Description Table A1: Between and within variation for energy poverty metrics Variable Mean Std. dev. Observations tpr overall .2177453 .4127142 N = 255684 between .3631767 n = 56263 within .25968 T-bar = 4.54444 mtwo overall .0852263 .2792187 N = 255684 between .2509484 n = 56263 within .1881695 T-bar = 4.54444 lihc overall .0713967 .2574869 N = 255684 between .2253253 n = 56263 within .1758772 T-bar = 4.54444 subjective overall .0158796 .1250106 N = 86211 between .1039507 n = 31941 within .0797766 T-bar = 2.69907 This table shows the variation for the four energy poverty measures between and within individuals. Data: SOEP (2022). Figure A1: Venn diagrams Objective indicator (2010-2020) Objective and subjective indicator (20162019) Venn diagrams illustrating the intersection of energy poverty indicators within our pooled sample (left panel: 2010-2020; right panel: 2016-2019.). Reading guide for left panel: 13,187 observations (individual-year combinations) in our pooled sample are identified as energy poor by both the tpr and the mtwo indicators. Additionally, 199,921 observations are not classified as energy poor according to any of our objective energy poverty metrics. Data: SOEP (2022). 29 Figure A4: Month of SOEP interview Distribution of person interviews in our sample by month and year. 141 observations are excluded due to missing information on the month the interview took place. Data: SOEP (2022). 30 Table A2: Summary Statistics - IV Sample Mean SD Min Max N self-rated health 3.39 0.95 1.00 5.00 146331 good health 0.83 0.38 0.00 1.00 146331 tpr 0.20 0.40 0.00 1.00 146331 mtwo 0.08 0.27 0.00 1.00 146331 lihc 0.07 0.25 0.00 1.00 146331 subjective 0.02 0.12 0.00 1.00 60417 gas 0.56 0.50 0.00 1.00 146331 oil 0.28 0.45 0.00 1.00 146331 district heat 0.08 0.27 0.00 1.00 146331 electricity 0.04 0.20 0.00 1.00 146331 solid fuels 0.05 0.21 0.00 1.00 146331 Summary statistics for the reduced IV-Sample. Data: SOEP (2022). 31 A.2 Two-way Fixed Effects Counterfactual (FEct) Figure A5: Treatment Status Ten Percent Rule Energy Poverty Visualization of the treatment status for energy poverty (according to the tpr, mtwo, lihc, and subjective indicators, respectively) of all individuals in the sample. Units are sorted based on the timing of receiving the treatment for visualization purposes. Data: SOEP (2022). 32 Figure A6: FEct – tpr Figure A7: FEct – mtwo Figure A8: FEct – lihc Fixed Effects Counterfactual estimations. Standard errors are obtained through non-parametric bootstrap procedures (500 bootstrap runs). Units must have a minimum of 1 observed period under control to be considered. Plot is limited to periods where the number of treated observations exceeds 20 percent of the largest number of treated observations in a period (default is 30 percent; we had to lower this threshold in order to have at least 3 pre-treatment periods). We test for the presence of pretreatment differential trends by using a variant of the F-Test built-in the fect-package that tests for zero residual averages in the pretreatment periods (larger F-test p-values suggest better pre-trend fitting). Data: SOEP (2022). 33 Figure A9: FEct placebo test – tpr Figure A10: FEct placebo test – mtwo Figure A11: FEct placebo test – lihc The placebo test is performed by removing all observations from the periods -2 to 0 relative to treatment timing for model fitting. The p-value indicates whether the estimated ATT in this range is significantly different from zero. Standard errors are obtained through non-parametric bootstrap procedures (500 runs). Data: SOEP (2022). 34 A.3 Robustness Checks Table A3: Fixed-effects regression - log number of doctor visits (1) (2) (3) (4) Log doctor Log doctor Log doctor Log doctor visits visits visits visits Ten percent rule (tpr) 0.0124∗∗ [0.00515] Two times median share (mtwo) 0.0116∗ [0.00700] Low income high costs (lihc) 0.00540 [0.00740] subjective 0.0510∗∗ [0.0256] Socio-econ. controls Yes Yes Yes Yes Control for income poverty Yes Yes Yes Yes Individual Fixed Effects Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes n (total) 248713 248713 248713 85812 Fixed-effects regression. Dependent variable: Log of number of visits to a doctor’s office in the previous three months. Cluster robust standard errors in brackets. All regressions include the usual covariates. ∗p < 0.1,∗∗ p < 0.05,∗∗∗ p < 0.01 Table A4: Fixed-effects regression - hospital stays (1) (2) (3) (4) hospital stays hospital stays hospital stays hospital stays Ten percent rule (tpr) -0.00291 [0.00309] Two times median share (mtwo) 0.00333 [0.00446] Low income high costs (lihc) -0.0000771 [0.00445] subjective 0.00159 [0.0148] Socio-econ. controls Yes Yes Yes Yes Control for income poverty Yes Yes Yes Yes Individual Fixed Effects Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes n (total) 196707 196707 196707 73713 Fixed-effects regression. Dependent variable: Hospital stays in survey year (0=no stay, 1=at least one hospital stay). Cluster robust standard errors in brackets. All regressions include the usual covariates. Data: SOEP (2022). ∗p < 0.1,∗∗ p < 0.05,∗∗∗ p < 0.01 Table A5: Fixed-effects regression - number of days off work (1) (2) (3) (4) Log days Log days Log days Log days off work off work off work off work Ten percent rule (tpr) -0.0245 [0.0156] Two times median share (mtwo) 0.0259 [0.0247] Low income high costs (lihc) 0.0145 [0.0248] subjective 0.0320 [0.0682] Socio-econ. controls Yes Yes Yes Yes Control for income poverty Yes Yes Yes Yes Individual Fixed Effects Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes n (total) 123568 123568 123568 48916 Fixed-effects regression. Dependent variable: Log of number of days off work sick in the respective survey year. Cluster robust standard errors in brackets. All regressions include the usual covariates. Data: SOEP (2022). ∗p < 0.1,∗∗ p < 0.05,∗∗∗ p < 0.01 36 Table A6: Fixed-effects regression - life satisfaction (1) (2) (3) (4) Life satisf. Life satis. Life satisf. Life satisf. Ten percent rule (tpr) -0.0655∗∗∗ [0.0114] Two times median share (mtwo) -0.0593∗∗∗ [0.0163] Low income high costs (lihc) -0.101∗∗∗ [0.0176] subjective -0.159∗∗∗ [0.0607] Socio-econ. controls Yes Yes Yes Yes Control for income poverty Yes Yes Yes Yes Individual Fixed Effects Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes n (total) 252040 252040 252040 86037 Fixed-effects regression. Dependent variable: Current life satisfaction (0=Low, 10=High). Cluster robust standard errors in brackets. All regressions include the usual covariates. Data: SOEP (2022). ∗p < 0.1,∗∗ p < 0.05,∗∗∗ p < 0.01 Table A7: Fixed-effects Regression - health satisfaction (1) (2) (3) (4) Health satisf. Health satisf. Health satisf. Health satisf. Ten percent rule (tpr) -0.0153 [0.0134] Two times median share (mtwo) -0.0471∗∗ [0.0188] Low income high costs (lihc) -0.0335∗ [0.0200] subjective -0.192∗∗∗ [0.0651] Socio-econ. controls Yes Yes Yes Yes Control for income poverty Yes Yes Yes Yes Individual Fixed Effects Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes n (total) 250495 250495 250495 85928 Fixed-effects regression. Dependent variable: Current health satisfaction (0= completely dissatisfied, 10=completely satisfied). Cluster robust standard errors in brackets. All regressions include the usual covariates. Data: SOEP (2022). ∗p < 0.1,∗∗ p < 0.05,∗∗∗ p < 0.01 37 Table A8: Ordered logistic regression - frequency of being happy (1) (2) (3) (4) Freq. happy Freq. happy Freq. happy Freq. happy Ten percent rule (tpr) 0.908∗∗∗ [0.0185] Two times median share (mtwo) 0.882∗∗∗ [0.0231] Low income high costs (lihc) 0.906∗∗∗ [0.0255] subjective 0.837∗∗ [0.0736] Socio-econ. controls Yes Yes Yes Yes Control for income poverty Yes Yes Yes Yes Individual Fixed Effects Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes Observations (total) 227573 227573 227573 85873 Observations (w. variation in outcome) 180294 180294 180294 50033 Panel units (w. variation in outcome) 28639 28639 28639 14803 Fixed-effects ordered logit regression. Dependent variable: Frequency of being happy in the last 4 weeks (1=very seldom, 5=very often). Exponentiated coefficients (odds ratios); Cluster robust standard errors in brackets. All regressions include the usual covariates. Data: SOEP (2022). ∗p < 0.1,∗∗ p < 0.05,∗∗∗ p < 0.01 Table A9: Ordered logistic regression - frequency of being sad (1) (2) (3) (4) Freq. sad Freq. sad Freq. sad Freq. sad Ten percent rule (tpr) 1.114∗∗∗ [0.0208] Two times median share (mtwo) 1.123∗∗∗ [0.0283] Low income high costs (lihc) 1.098∗∗∗ [0.0291] subjective 1.397∗∗∗ [0.122] Socio-econ. controls Yes Yes Yes Yes Control for income poverty Yes Yes Yes Yes Individual Fixed Effects Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes Observations (total) 227580 227580 227580 85873 Observations (w. variation in outcome) 198536 198536 198536 61117 Panel units (w. variation in outcome) 32306 32306 32306 18162 Fixed-effects ordered logit regression. Dependent variable: Frequency of being sad in the last 4 weeks (1=very seldom, 5=very often). Exponentiated coefficients (odds ratios); Cluster robust standard errors in brackets. All regressions include the usual covariates. Data: SOEP (2022). ∗p < 0.1,∗∗ p < 0.05,∗∗∗ p < 0.01 38