Food-related energy consumption can help reveal poverty in rural Chinese households Authors: Ruohan Zhong, Jinjun Xue, Xinye Zheng, Chu Wei, Qian Sun Published: 2026 ABSTRACT Energy consumption for cooking is important for satisfying basic needs, complementing traditional poverty measurements. Here, we propose an energy consumption-based indicator (Energy Engel Coefficient) that measures the proportion of household energy dedicated to food preparation activities relative to the total. Using a nation-wide household survey in China, we find that this indicator exhibits consistent patterns with regional and household characteristics. In addition, it more accurately reflects Engel’s Law than traditional Engel Coefficient, which holds that households with lower incomes spend a larger proportion of their income on basic needs. Our proposed measure also shows stronger performance than traditional coefficient in comparing with non-monetary metrics like subjective experience and asset ownership. Our analysis reveals that the Energy Engel Coefficient identifies 13.9% of households experiencing poverty, with 85.8% of these households not identified by the traditional coefficient. These households are predominantly low-income, small-sized rural ones. Therefore, this indicator complements poverty identification and contributes to improving the distribution of welfare. Introduction A persistent challenge in achieving global sustainable development is addressing disparities in access to basic human needs such as food, energy, and essential services1,2. While international efforts have strengthened frameworks aimed at improving livelihoods, the measurement of progress toward equitable resource distribution remains a central concern3–5. Among basic needs, food and energy, particularly for cooking, are universally recognized as prerequisites for survival and well-being6,7. These necessities support critical dimensions of human welfare, ranging from nutritional security to the capacity to transform raw materials into meals8,9. However, quantifying access to these resources remains complex, as traditional monetary metrics often fail to account for non-market transactions or the physical realities of household energy consumption. This gap underscores the necessity for innovative indicators that complement existing economic measures by focusing on observable resource flows, such as energy use for food preparation. Poverty measurement has historically relied on indicators assessing access to basic needs, which are strongly linked to food10–16. Early methodologies emphasized monetary thresholds for acquiring life-sustaining goods, such as food bundles required to meet caloric needs17. However, because these metrics rely heavily on purchasing power parity (PPP), the periodic adjustments to the poverty line can make it less accurate in identifying those truly in need18,19. In addition, the satisfaction of basic needs is not only about expenditure, but also encompasses the ability to store, cook, and transform ingredients into meals. These challenges reveal the limitations of income- or expenditure-based metrics, which often fail to capture households reliant on subsistence farming, non-monetary exchanges, conspicuous consumption, or public energy subsidies20–22. To address these limitations, researchers have increasingly turned to physical indicators of resource use23–25. Among these, energy consumption, particularly for cooking, provides a practical perspective for assessing the fulfillment of basic needs26,27. Food-related energy use not only reflects dietary patterns but also reveals infrastructural constraints: households in resource-scarce areas often invest a lot of time and effort in fuel procurement, food preparation, and cooking. This aligns with Engel’s observation that resource-constrained households prioritize essential activities such as food preparation28–31, though traditional Engel Coefficient (EC) calculations, based on monetary expenditures, struggle to quantify these realities. In some low-income countries, subsistence agriculture and bartering practices may reduce the reliance on monetary exchange32, particularly in rural areas where non-market production and exchange are essential for sustaining household livelihoods33. As shown in Fig. 1, the negative correlation between the traditional EC and gross national income (GNI) per capita is much less pronounced in low-, lower middle- and upper middle-income countries than in high-income countries. Since expenditures are a combination of the quantity and price of goods, price differences across regions and countries can also influence poverty measurement. This is especially true for housing costs, where rental prices in lower-income rural areas differ widely from those in urban areas. Moreover, housing costs, being more volatile than the implied prices of other services34, can disproportionately reduce expenditures on basic needs, leading to a biased measure of poverty when using the traditional EC.Fig. 1The per capita income and traditional Engel Coefficient (EC).The x-axis (logarithmic form) represents the GNI per capita in USD (Atlas methodology), using data from the World Bank. Box plots display the distribution of the data, with the median (50th percentile) at the center, the interquartile range (25th to 75th percentiles) as the box, and whiskers extending to the maximum and minimum values within 1.5 times the interquartile range. The y-axis represents the EC from the U.S. Department of Agriculture’s Economic Research Service. Countries are classified into three income groups based on World Bank (GNI per capita = 4255 USD and 13205 USD): low and lower middle (N = 30), upper middle (N = 28), and high (N = 41)68. Circle color indicates the group of each country. Circle size represents the population size of each country. Ordinary least squares linear regression models weighted by population size are conducted with two-sided hypothesis tests and no multiple comparison adjustments. The shaded areas denote the 95% confidence intervals (CI) from the linear regression models, and the central lines represent the predicted values. R-squared (R2), significance levels (p), and 95% CI are reported for each fit. The per capita income and traditional Engel Coefficient (EC). The x-axis (logarithmic form) represents the GNI per capita in USD (Atlas methodology), using data from the World Bank. Box plots display the distribution of the data, with the median (50th percentile) at the center, the interquartile range (25th to 75th percentiles) as the box, and whiskers extending to the maximum and minimum values within 1.5 times the interquartile range. The y-axis represents the EC from the U.S. Department of Agriculture’s Economic Research Service. Countries are classified into three income groups based on World Bank (GNI per capita = 4255 USD and 13205 USD): low and lower middle (N = 30), upper middle (N = 28), and high (N = 41)68. Circle color indicates the group of each country. Circle size represents the population size of each country. Ordinary least squares linear regression models weighted by population size are conducted with two-sided hypothesis tests and no multiple comparison adjustments. The shaded areas denote the 95% confidence intervals (CI) from the linear regression models, and the central lines represent the predicted values. R-squared (R2), significance levels (p), and 95% CI are reported for each fit. Here, we follow the framework of basic needs35 and propose the Energy Engel Coefficient (EEC) as a complementary poverty indicator, which shifts the focus from monetary expenditure to household energy consumption. The EEC quantifies the proportion of total energy use dedicated to food preparation, a critical but often insufficiently considered dimension of basic needs. This indicator addresses some limitations of the expenditure-based traditional EC. It avoids distortions arising from price volatility, non-market energy sources, and liquidity-constrained spending behaviors. While the EEC also has some limitations in urban contexts with diversified energy use, it serves as a valuable complement to existing poverty metrics. This indicator focuses on cooking-related energy consumption, highlighting the physical burdens of food preparation in resource-constrained environments. It provides policymakers with valuable guidance for designing targeted interventions, ranging from clean stove programs to electrification projects, that can simultaneously alleviate poverty and accelerate the energy transition. Combining the EEC with expenditure-based measures enables a multidimensional assessment of poverty, which is essential for advancing Sustainable Development Goals (SDGs) in low-income countries. Using household data from the China Residential Energy Consumption Survey (CRECS), we test the performance of the proposed EEC indicator in a Chinese context. In our study sample, we demonstrate that the EEC values are consistent with regional variations and the patterns of basic household characteristics. More importantly, the EEC reflects household economic living standards more effectively than the EC. On one hand, when compared with monetary-based metrics such as total household income and expenditure, the EEC shows a stronger consistency with Engel’s Law. On the other hand, the EEC demonstrates clearer differentiation across levels of subjective experience and household asset ownership. After determining the “threshold” for EEC, we reveal that 25.3% of households experienced misclassification in poverty identification. Among these, 47.1% of households experiencing poverty are not identified by the EC, with a large proportion being rural households. By analyzing the characteristics of these households, we find that rural location, low-income level, and small household size are important factors contributing to this mismatch. Results Reconstructing poverty measurement with food-related energy consumption The EEC is defined as the proportion of household energy dedicated to food preparation activities relative to total energy consumption for all end-use activities (see “Methods”). This indicator provides a service-based perspective to assess household welfare through energy distribution patterns, particularly the critical role of cooking in meeting dietary needs. Figure 2 presents the framework for measuring poverty through different perspectives, all grounded in basic needs.Fig. 2Framework for measuring poverty based on basic needs.The framework presents both absolute and relative indicators for poverty assessment from the perspectives of caloric intake, commodity expenditure, and household end-use energy activities. Absolute indicators include caloric intake standards used in medical contexts, as well as monetary poverty lines applied in economics. Among relative indicators, the traditional Engel Coefficient focuses on food expenditure as a share of total expenditure, while the proposed Energy Engel Coefficient complements the traditional indicator by focusing on energy consumption for cooking, a food-related end-use activity that reflects basic needs more directly. Framework for measuring poverty based on basic needs. The framework presents both absolute and relative indicators for poverty assessment from the perspectives of caloric intake, commodity expenditure, and household end-use energy activities. Absolute indicators include caloric intake standards used in medical contexts, as well as monetary poverty lines applied in economics. Among relative indicators, the traditional Engel Coefficient focuses on food expenditure as a share of total expenditure, while the proposed Energy Engel Coefficient complements the traditional indicator by focusing on energy consumption for cooking, a food-related end-use activity that reflects basic needs more directly. Food-related indicators are important performers in measuring poverty (see Supplementary Note 1, Supplementary Table 1 for review of indicators to measure poverty). For instance, medical standards define nutritional requirements through calorie thresholds (1800–2340 kcal/day for adults)36, whereas economists monetize living costs through poverty lines like the $3.00 IPL for low-income countries37. Different from these absolute indicators, traditional EC quantifies poverty through commodity expenditure, measuring the proportion of food spending relative to total expenditure. In the traditional framework, this relative indicator reflects the trade-off between household expenditure on food representing basic needs and expenditure on other items, including education, transportation, clothing, and residence, thereby revealing the level of poverty from a monetary perspective38. Following this approach, the EEC differentiates from monetary-based metrics by tracking energy consumed in food preparation and comparing it to energy inputs for other purposes. This perspective circumvents the effects of market price distortions, proving particularly valuable in subsistence economies where monetary metrics prove inadequate. Within the proposed framework, from an energy-based perspective, basic needs and developmental needs are distinguished by the type and purpose of energy use. Basic needs, such as cooking, represent essential energy consumption required for survival and minimum well-being. In contrast, development needs, reflected in energy consumption for household appliances, enhanced thermal comfort, or entertainment equipment, correspond to higher living standards and improved quality of life. Therefore, a higher EEC indicates that households prioritize basic needs more heavily and face a higher likelihood of poverty through three mechanisms. First, households facing economic constraints prioritize cooking energy over discretionary uses such as space cooling or entertainment, and they have limited resilience to economic fluctuations39,40. Second, poverty is often associated with labor-intensive livelihoods, where food preparation constitutes a more central and time-intensive household activity41. Third, households experiencing poverty face both housing constraints that limit access to diverse energy appliances and inefficient cooking technologies, concentrating energy use in cooking and increasing the input required for food preparation42. As with any metric of poverty, the EEC we proposed is also influenced by various factors, such as household characteristics and geo-locational (see Supplementary Note 2). When using the framework to define a poverty indicator, we expected it to fulfill Hypothesis 1, “Consumption can bring timely utility,” and Hypothesis 2, “Consumption data are comparable.” As shown in Table 1, while the EC has some limitations in real-world applications, the EEC overcomes many of these, providing a complement for the accurate identification of poverty (see Supplementary Note 3 for review of comparison between EC and EEC).Table 1Theoretical comparison of EC and EECECEECDefinitionThe proportion of food expenditures in total expendituresThe proportion of energy consumption for cooking in the total energy consumption of all end-use activitiesPerspectivestock/flowservice/utilityHypothesis 1There are cases where goods are purchased but cannot be utilized: Food storage behavior in cold regions and conspicuous consumption.The performance of end-use activities is a direct reflection of utility.There are cases where utility is obtained without the purchase of goods: bartering and household production, such as farming.Hypothesis 2The range of total expenditures is uncertain as to whether the government bears expenditure through public services.The scope of household energy end-use activities is certain: cooking, space heating, space cooling, water heating, and home appliances.Monetary expenditure data is a value quantityEnergy consumption data is a physical quantityMore details about the calculation of the EEC can be seen in “Methods”. Table 1Theoretical comparison of EC and EECECEECDefinitionThe proportion of food expenditures in total expendituresThe proportion of energy consumption for cooking in the total energy consumption of all end-use activitiesPerspectivestock/flowservice/utilityHypothesis 1There are cases where goods are purchased but cannot be utilized: Food storage behavior in cold regions and conspicuous consumption.The performance of end-use activities is a direct reflection of utility.There are cases where utility is obtained without the purchase of goods: bartering and household production, such as farming.Hypothesis 2The range of total expenditures is uncertain as to whether the government bears expenditure through public services.The scope of household energy end-use activities is certain: cooking, space heating, space cooling, water heating, and home appliances.Monetary expenditure data is a value quantityEnergy consumption data is a physical quantityMore details about the calculation of the EEC can be seen in “Methods”. Theoretical comparison of EC and EEC More details about the calculation of the EEC can be seen in “Methods”. In response to the conceptual limitations of the EC, EEC provides a more direct measure of household utility by linking energy consumption to the services received. Compared to purchasing commodity expenditures, household end-use activities are more effective at capturing the various needs that are met daily and accurately portraying the flow of services24,43. These activities focus on essential services like food preparation, thermal comfort, and household maintenance, providing an accurate portrayal of daily needs and welfare44. As a reflection of individual behaviors within the household, these activities can solve the problems caused by the ambiguity of the scope of goods in various categories. Cooking, in particular, is a reliable indicator of food demand, as most raw ingredients require preparation to become edible. While some foods, like snacks and beverages, do not require cooking, these are often less essential nutritionally and more common in higher-income households. Including them in food expenditures could inflate the EC for these households, thus underestimating the poverty gap. By focusing on cooking, the bias introduced by the consumption of non-essential food, as it affects the measurement of basic needs, can be appropriately circumvented. By focusing on household-level energy consumption, the EEC aligns with the public’s common expectations and provides complementary analysis for poverty identification45. The proportion of energy consumption for different end uses can provide a more accurate measure of actual behavior than monetary expenditures alone46. In cases where subsistence farming or non-market exchanges dominate, energy consumption still reflects essential activities like cooking. In addition, measuring energy consumed by household end-use activities avoids problems related to flows and stocks. The calculation of energy consumption is time-bound and measurable through the duration and intensity of activities, along with the effectiveness and quantity of household equipment45. This makes it more comparable and reliable than expenditure data, which can suffer from underreporting. Quantifying the stylized poverty patterns using EEC In our study sample, we obtain a weight-adjusted national average EEC of 0.418 (see Supplementary Note 4 for calculation method). The EEC highlights the poverty gap between urban (0.327 ± 0.006) and rural (0.472 ± 0.006) areas, with the urban results exhibiting a more concentrated and right-skewed distribution compared to the broader distribution in rural regions (Fig. 3a), a difference confirmed by a t-test. Whether focusing on all end-use activities or only on cooking, the proportion of rural households using biomass energy is much higher than that of urban households (see Supplementary Fig. 1). Constrained by infrastructure limitations and facing higher affordability risks, rural households have fewer lifestyle upgrades and use less clean energy than urban households47. This gap may be driven by “spatial poverty traps,” where reliance on natural resources in rural areas perpetuates poverty48.Fig. 3Stylized facts of poverty as measured by EEC.a EEC across urban (N = 1743), rural (N = 2090), eastern (N = 1558), middle (N = 1371), western (N = 904), northern (N = 1644) and southern (N = 2189) regions. b Energy consumption (standard kg coal equivalent, kgce) by end-use activities among climate zones: severe cold (COLDH, N = 581), cold (COLDL, N = 1067), mild (MILD, N = 176), hot summer and cold winter (HSCW, N = 1697), hot summer and warm winter (HSWW, N = 312). c EEC across total household income and expenditure quintiles. Income quintile sample sizes from lowest to highest: Q1 (N = 708), Q2 (N = 729), Q3 (N = 719), Q4 (N = 658), Q5 (N = 702). Expenditure quintiles: Q1 (N = 759), Q2 (N = 541), Q3 (N = 755), Q4 (N = 635), Q5 (N = 533). d EEC across age groups of the household head. Age is divided into four groups: ≤ 24 years (N = 1172), 25–34 years (N = 793), 35–54 years (N = 774), and ≥ 55 years (N = 1094). e EEC across education groups of the household head. Year of education is divided into four groups: ≤ 6 years (N = 1372), 7–9 years (N = 1,079), 10–12 years (N = 716), and ≥ 13 years (N = 653). f EEC across life cycle groups based on the presence of children (under 12 years old) and older adults (over 60 years old). The four groups are households with neither children nor older adults (None, N = 1549), households with children but no older adults (Children, N = 631), households with older adults but no children (Older adults, N = 1271), and households with both children and older adults (Both, N = 382). g EEC across household size groups. Household size is divided into two 1-4 individuals (Small, N = 2990) and > 4 individuals (Large, N = 843). In (a) and (d–g), box plots display the distribution of the data, with the median (50th percentile) at the center, the interquartile range (25th to 75th percentiles) as the box, and the circular dots represent the mean values. Statistical significance is evaluated using the two-sided Welch’s t-test. In (b, c), data are presented as mean values ± SEM. Stylized facts of poverty as measured by EEC. a EEC across urban (N = 1743), rural (N = 2090), eastern (N = 1558), middle (N = 1371), western (N = 904), northern (N = 1644) and southern (N = 2189) regions. b Energy consumption (standard kg coal equivalent, kgce) by end-use activities among climate zones: severe cold (COLDH, N = 581), cold (COLDL, N = 1067), mild (MILD, N = 176), hot summer and cold winter (HSCW, N = 1697), hot summer and warm winter (HSWW, N = 312). c EEC across total household income and expenditure quintiles. Income quintile sample sizes from lowest to highest: Q1 (N = 708), Q2 (N = 729), Q3 (N = 719), Q4 (N = 658), Q5 (N = 702). Expenditure quintiles: Q1 (N = 759), Q2 (N = 541), Q3 (N = 755), Q4 (N = 635), Q5 (N = 533). d EEC across age groups of the household head. Age is divided into four groups: ≤ 24 years (N = 1172), 25–34 years (N = 793), 35–54 years (N = 774), and ≥ 55 years (N = 1094). e EEC across education groups of the household head. Year of education is divided into four groups: ≤ 6 years (N = 1372), 7–9 years (N = 1,079), 10–12 years (N = 716), and ≥ 13 years (N = 653). f EEC across life cycle groups based on the presence of children (under 12 years old) and older adults (over 60 years old). The four groups are households with neither children nor older adults (None, N = 1549), households with children but no older adults (Children, N = 631), households with older adults but no children (Older adults, N = 1271), and households with both children and older adults (Both, N = 382). g EEC across household size groups. Household size is divided into two 1-4 individuals (Small, N = 2990) and > 4 individuals (Large, N = 843). In (a) and (d–g), box plots display the distribution of the data, with the median (50th percentile) at the center, the interquartile range (25th to 75th percentiles) as the box, and the circular dots represent the mean values. Statistical significance is evaluated using the two-sided Welch’s t-test. In (b, c), data are presented as mean values ± SEM. Figure 3a further shows regional patterns (see Supplementary Note 5 for regional classification). The eastern (0.334 ± 0.006) regions have EEC values below the national average, while the middle (0.444 ± 0.008) and western (0.473 ± 0.009) regions exceed it, reflecting China’s large regional poverty disparities. The EECs for the middle and western regions show large interquartile spacing and more dispersed results, suggesting larger disparities within these regions. By comparison, the EEC for the eastern areas shows smaller interquartile spacing and smaller values, with more centralized results. However, we also find that the EEC in the southern regions (0.496 ± 0.006) is much higher than that in the northern regions (0.287 ± 0.006), which is slightly different from the actual economic situation (see Supplementary Fig. 2, Supplementary Note 6 for regional analysis of EC). In fact, due to the large temperature difference between the north and south of China during the winter period, the northern region has a higher demand for heating. Next, analysis of different climatic zones reveals that EEC patterns are consistent with regional climatic characteristics (Fig. 3b). Taken together, the cold northern regions have relatively low EECs (COLDH: 0.212 ± 0.009, COLDL: 0.314 ± 0.008), which could be attributed to cold weather. Specifically, the effective energy for heating in both zones is over 550 kgce per household per year, which is comparable to the energy used for cooking49. This suggests that space heating should be recognized as a basic need in cold regions50. Therefore, cold regions need to balance heating and basic energy distribution to ensure low-income households have access to cooking energy. It also shows that our proposed EEC has some limitations in application in terms of comparability among regions. Household characteristics may also affect the results of EEC (Fig. 3c–g). According to monetary-based metrics, Fig. 3c reveals that higher-income households generally have lower EEC values, with the lowest income group (Q1) at 0.500 ( ± 0.012) and the highest (Q5) at 0.324 ( ± 0.009), reflecting a significant difference. This finding is in line with the traditional Engel’s Law, and thereby validates the rationality of EEC. On the one hand, low-income households tend to have higher energy use intensity51, while on the other hand, higher-income households have made more progress in energy consumption transition52, allowing the EEC indicator to be sensitive to capturing various types of demand for different income levels. The results based on expenditure also show the same pattern (Q1: 0.511 ± 0.011; Q5: 0.303 ± 0.010). Figure 3d, e show the variation in EEC by age and education level of household heads. Specifically, households with heads aged 55 years or older have an EEC of 0.437 ( ± 0.009), while those with heads aged 24 years or younger show 0.360 ( ± 0.008). Households with heads who have 0-6 years of education have the highest EEC of 0.492 ( ± 0.008), contrasting with 0.271 ( ± 0.009) in households with heads who have 13 years or more of education. These patterns indicate that EEC values are higher among households with older heads and heads with fewer years of education. As shown in Fig. 3f, households comprising both children (under 12 years old) and older adults (over 60 years old) exhibit the highest of 0.468 ( ± 0.015), while those with neither cohort show the lowest (0.377 ± 0.007). The EECs for households with only children or only older adults are at the middle range (Children: 0.411 ± 0.011; Older adults: 0.420 ± 0.008). Furthermore, Fig. 3g reveals that smaller (1–4 person) households have a lower EEC value (0.395 ± 0.005) than larger (5+ person) households (0.444 ± 0.010), reflecting a significant difference. Multi-generational and large households often face disproportionately high living costs, and the EEC pattern serves as a validation of this point. Comparing the EEC with the traditional indicator While EEC has shown its application in quantifying poverty, we next compare its main characteristics with those of the traditional EC. With a sample of 3089 households in 2014, the value for EEC is slightly higher than the traditional indicator EC (0.340 ± 0.004) (see Supplementary Note 7, Supplementary Fig. 3 and Supplementary Fig. 4 for other indicators’ results). In order to compare the differences in poverty levels as measured by these two indicators, Figs. 4, 5 present some differential characteristics.Fig. 4Comparison of the EEC and EC.a, b Comparison of Engel’s Law based on total household income at (a) cohort level (N = 64) and (b) city level (N = 85). c, d Comparison of Engel’s Law based on total household expenditure at (c) cohort level (N = 64) and (d) city level (N = 85). In (a–d), the x‑axis is logarithmic. Circle size represents the number of households in each observation. Ordinary least squares linear regression models weighted by the number of households are conducted with two-sided hypothesis tests and no multiple comparison adjustments. Dashed lines indicate the fitted curves. R-squared (R2), significance levels (p), and 95% confidence intervals (CI) are reported for each fit. e Comparison of subjective experience across three levels, which are “Below”, “Average”, and “Above”. EC sample sizes: Below (N = 1133), Average (N = 1682), Above (N = 280). EEC sample sizes: Below (N = 1376), Average (N = 2091), Above (N = 344). f Comparison of “Asset-poverty” and “Non-asset-poverty” groups identified using asset index (see Supplementary Note 9) and three poverty lines (10th, 15th, and 20th percentiles). EC sample sizes: 10th (N = 310, N = 2798), 15th (N = 608, N = 2500), 20th (N = 642, N = 2466). EEC sample sizes: 10th (N = 375, N = 3458), 15th (N = 729, N = 3104), 20th (N = 767, N = 3066). In (e, f), data are presented as mean values ± SEM. Statistical significance is evaluated using the two-sided Welch’s t-test.Fig. 5Characterization of groups after categorization using the EEC and EC.a Poverty status identified by the EEC and EC. Both indicators classify households into three groups, which are “low”, “medium”, and “high” (see Supplementary Note 11). The nine possible combinations are shown (Cat.1: both low; Cat.2: EEC medium, EC low; Cat.3: EEC high, EC low; Cat.4: EEC low, EC medium; Cat.5: both medium; Cat.6: EEC high, EC medium; Cat.7: EEC low, EC high; Cat.8: EEC medium, EC high; Cat.9: both high). Circle size represents the number of households per category. Circle color indicates the matching pattern between indicators. Urban and rural resident distributions are presented. b Decomposition of energy consumption by end-use activities. c Decomposition of monetary expenditure by goods and services. d, e Decomposition of energy consumption by fuel types (d) across all end-use activities and (e) within cooking activity. The categories are arranged based on the EEC identification results. Comparison of the EEC and EC. a, b Comparison of Engel’s Law based on total household income at (a) cohort level (N = 64) and (b) city level (N = 85). c, d Comparison of Engel’s Law based on total household expenditure at (c) cohort level (N = 64) and (d) city level (N = 85). In (a–d), the x‑axis is logarithmic. Circle size represents the number of households in each observation. Ordinary least squares linear regression models weighted by the number of households are conducted with two-sided hypothesis tests and no multiple comparison adjustments. Dashed lines indicate the fitted curves. R-squared (R2), significance levels (p), and 95% confidence intervals (CI) are reported for each fit. e Comparison of subjective experience across three levels, which are “Below”, “Average”, and “Above”. EC sample sizes: Below (N = 1133), Average (N = 1682), Above (N = 280). EEC sample sizes: Below (N = 1376), Average (N = 2091), Above (N = 344). f Comparison of “Asset-poverty” and “Non-asset-poverty” groups identified using asset index (see Supplementary Note 9) and three poverty lines (10th, 15th, and 20th percentiles). EC sample sizes: 10th (N = 310, N = 2798), 15th (N = 608, N = 2500), 20th (N = 642, N = 2466). EEC sample sizes: 10th (N = 375, N = 3458), 15th (N = 729, N = 3104), 20th (N = 767, N = 3066). In (e, f), data are presented as mean values ± SEM. Statistical significance is evaluated using the two-sided Welch’s t-test. Characterization of groups after categorization using the EEC and EC. a Poverty status identified by the EEC and EC. Both indicators classify households into three groups, which are “low”, “medium”, and “high” (see Supplementary Note 11). The nine possible combinations are shown (Cat.1: both low; Cat.2: EEC medium, EC low; Cat.3: EEC high, EC low; Cat.4: EEC low, EC medium; Cat.5: both medium; Cat.6: EEC high, EC medium; Cat.7: EEC low, EC high; Cat.8: EEC medium, EC high; Cat.9: both high). Circle size represents the number of households per category. Circle color indicates the matching pattern between indicators. Urban and rural resident distributions are presented. b Decomposition of energy consumption by end-use activities. c Decomposition of monetary expenditure by goods and services. d, e Decomposition of energy consumption by fuel types (d) across all end-use activities and (e) within cooking activity. The categories are arranged based on the EEC identification results. According to monetary-based metrics, we present the fitted curves for the EEC and EC in Fig. 4a, b. The results indicate that the EEC aligns with Engel’s Law, where the value is negatively correlated with income (Cohort level: R2 = 0.711, p < 0.0001, 95% CI = [−0.215, −0.154]; City level: R2 = 0.118, p = 0.0013, 95% CI = [−0.226, −0.057]). See Supplementary Note 8 for the cohort classification. The optimal fit, using a logarithmic function, reveals that as income increases, the EEC declines rapidly for low-income households, who can then afford other necessities like heating, cooling, and lighting, thereby increasing energy consumption for these activities. For higher-income households, the impact of rising income on EEC diminishes, as their needs are already met. As shown in cohort analyses in Fig. 4a, while EEC is negatively correlated with the standard of living, the EC is less responsive to changes in household income when fitted logarithmically (R2 = 0.005, p = 0.5837, 95% CI = [−0.034, 0.019]). However, city-level analyses in Fig. 4b show that EC is positively correlated with household income (R2 = 0.104, p = 0.0027, 95% CI = [0.019, 0.087]). Therefore, the EEC is more sensitive to shifts in living standards, complementing the traditional EC in identifying poverty levels. In addition, we conduct the same analysis based on total household expenditure in Fig. 4c, d. Similarly, we find that at both the cohort level and city level, EEC exhibits a significant negative correlation with total household expenditure (Cohort level: R2 = 0.553, p < 0.0001, 95% CI = [−0.240, −0.150]; City level: R2 = 0.243, p < 0.0001, 95% CI = [−0.314, −0.139]), while EC shows relatively little sensitivity to changes in total household expenditure (Cohort level: R2 = 0.029, p = 0.1851, 95% CI = [−0.056, 0.011]; City level: R2 = 0.113, p = 0.0016, 95% CI = [0.025, 0.102]). According to non-monetary metrics, we first compare the differences at different levels of subjective experience in Fig. 4e. Respondents are asked to rate their household economic status relative to their local area. Answers are grouped into three levels: below average (Below), average (Average), and above average (Above). We hypothesize that individuals who subjectively assess their household economic situation as being lower are more likely to experience poverty. For the EEC, as expected, the “Below” level has the highest value (0.443 ± 0.008), significantly exceeding that of the “Average” level (0.392 ± 0.006), which in turn is significantly higher than the “Above” level (0.343 ± 0.014), forming a significant difference. In contrast, the same analysis for the EC reveals a non-significant trend in the same characteristic. We then treat asset ownership as another non-monetary perspective for a robustness test. We hypothesize that more asset ownership will be associated with a higher economic standard of living. The comparisons based on asset indices constructed through the MCA approach53 are shown in Fig. 4f (see Supplementary Note 9 for construction). Across all three poverty lines, households identified as “Asset-poverty” exhibit significantly higher EEC values than those identified as “Non-asset-poverty”, with average differences ranging from 0.176 to 0.193 (p < 0.0001). In contrast, the EC shows no consistent or statistically significant differences across any lines. Other results based on asset ownership have also concluded the same (see Supplementary Note 10, Supplementary Fig. 5). To compare the poverty status identified by the two indicators, Fig. 5a presents the matching results for households across three groups: low, medium, and high (see Supplementary Note 11, Supplementary Fig. 6). Under this classification, households in the high group are considered to be experiencing poverty. The EC indicator shows a concentrated distribution, with 2018 households classified as low, 596 as medium, and 475 as high. In contrast, the EEC indicator is more dispersed, classifying 2195 households as low, 465 as medium, and 429 as high. In addition, 51.0% of households are perfectly matched by both indicators (Cat.1, Cat.5, and Cat.9). While the EEC classifies 13.9% of households as high – marginally lower than EC’s 15.4% – the composition of these groups differs radically. Among unmatched households, 414 households classified as high are not identified by the EEC (Cat.7 and Cat.8), while 368 households classified as high are not identified by the EC (Cat.3 and Cat.6). It is worth noting that among households not identified by the EC, rural households account for as much as 80.2%, indicating that if analysis is based only on monetary indicators, vulnerable rural households may be difficult to identify directly. Figure 5b, c characterize energy consumption by end-use activities and monetary expenditures by goods and services across the nine categories (see Supplementary Note 12, Supplementary Fig. 7 for characteristics of different categories). Among categories identified as experiencing poverty by the EEC only, Cat.3 (EEC high, EC low) averages 369 kgce per year in effective energy consumption during the sample period of 2014, with cooking (86.1%) and home appliances (11.9%) being the primary end-use activities. Their annual total household expenditures are 32,445 yuan (USD 4581), with the largest portion spent on the residence (40.8%). This category is under heavy pressure from housing costs. Such expenditure could reduce the resources for food, resulting in a biased EC that fails to identify this category. Our EEC can help to identify these households. The Cat.6 (EEC high, EC medium) consumes an average of 298 kgce, mainly for cooking (86.7%) and home appliances (12.9%). Their total expenditures amount to 16,961 yuan (USD 2395), with the highest spending on food (50.1%). Among categories identified as experiencing poverty by the EC only, the Cat.7 (EEC low, EC high) averages 876 kgce in effective energy consumption, with the largest energy uses being space heating (41.1%), water heating (22.8%), and cooking (21.9%). This may be partly explained by heating demand limiting the otherwise basic demand for cooking under the influence of climatic characteristics. As previously noted, our proposed EEC may exhibit some limitations due to cold temperatures. The total household expenditures of Cat.7 amount to 28,775 yuan (USD 4063), predominantly for food (69.7%). The traditional EC can identify this category of people based on their expenditures. The Cat.8 (EEC medium, EC high) consumes an average of 326 kgce, with cooking (68.3%) and home appliances (26.1%) as the primary end-use activities. Their total expenditure is 19,584 yuan (USD 2765), with the majority spent on food (70.6%). We further analyze the fuel types across all end-use activities and specifically within the cooking activity in Fig. 5d, e, respectively. For categories identified as experiencing poverty by the EEC, biomass energy has become their second-large fuel source (Cat.3: 28.3%; Cat.6: 25.8%; Cat.9: 36.8%), second only to electricity. Focusing only on the end-use activity of cooking reveals that biomass energy is the most important fuel for Cat.3 (32.8%) and Cat.9 (42.3%), accounting for a higher proportion than any other fuel source. This category of households relies heavily on biomass energy, indicating that traditional fuels still hold an important role. Combined with the “Energy Ladder” theory54, the high proportion of inferior fuel sources suggests these households are of a lower socioeconomic status and may face increased poverty risks. For categories identified as experiencing poverty by the EC, except for electricity, Cat.7 relies primarily on thermal energy (31.8%) as its main fuel source due to unavoidable heating needs. Cat.8 has a more diverse demand for other fuels. These include liquefied petroleum gas (LPG) at 21.4%, natural gas at 16.0%, and biomass at 11.7%. Within cooking, the primary fuel sources for Cat.7 and Cat.8 are natural gas (Cat.7: 43.4%; Cat.8: 21.1%), electricity (Cat.7: 26.9%; Cat.8: 30.6%), and LPG (Cat.7: 20.6%; Cat.8: 31.5%), collectively accounting for over 80% of the total share. These households have relatively more abundant fuel sources, with clean energy making a large contribution. Unveiling household characteristics associated with poverty identification mismatches We categorize all samples into “poverty” and “non-poverty” groups by the EEC and EC, then construct Logit and Probit models to identify which household characteristics influence the mismatch of poverty identification (see “Methods”). Figure 6 shows the main results (see Supplementary Table 2 for descriptive statistics of all variables, Supplementary Tables 3, 4 for regression results and Supplementary Tables 5, 6 for robustness test results).Fig. 6Regression coefficients for the relationship between poverty identification mismatches and household characteristics.a–f Relationships between poverty identification mismatches and household characteristics. The six panels show results for indicators of a Urban, b LnIncome, c Age, d Education, e Household size, and f the constant term from the regression models. Data are presented as coefficient estimates from empirical models, and the error bars represent 95% confidence intervals. Logit and Probit regression models are conducted with two-sided hypothesis tests and no multiple comparison adjustments. Standard errors are clustered at the province level. The dependent variable equals 1 for households experiencing poverty that are not identified by the relevant indicator and 0 otherwise. “EEC only” means the EEC identifies poverty, but the EC does not; “EC only” means the EC identifies poverty, but the EEC does not; “All mismatch” combines the two sets. Regression sample sizes are 2658 for “EEC only”, 2902 for “EC only” and “All mismatch”. Regression coefficients for the relationship between poverty identification mismatches and household characteristics. a–f Relationships between poverty identification mismatches and household characteristics. The six panels show results for indicators of a Urban, b LnIncome, c Age, d Education, e Household size, and f the constant term from the regression models. Data are presented as coefficient estimates from empirical models, and the error bars represent 95% confidence intervals. Logit and Probit regression models are conducted with two-sided hypothesis tests and no multiple comparison adjustments. Standard errors are clustered at the province level. The dependent variable equals 1 for households experiencing poverty that are not identified by the relevant indicator and 0 otherwise. “EEC only” means the EEC identifies poverty, but the EC does not; “EC only” means the EC identifies poverty, but the EEC does not; “All mismatch” combines the two sets. Regression sample sizes are 2658 for “EEC only”, 2902 for “EC only” and “All mismatch”. When analyzing “All mismatch” as the dependent variable, we found that small households with lower incomes, older heads of households, and heads with fewer years of education are easily mismatched by these two indicators, but at this point, the coefficient estimates from empirical models of urban-rural status are not significant. All these characteristics suggest that vulnerable households are likely to be inconsistently identified when poverty is assessed from different perspectives. For the “EEC only” group, small, low-income rural households are less likely to be identified by EC. Constrained by the level of rural development and their own experience, most members of these households tend to be engaged in agriculture. Thus, a portion of their basic food needs, such as vegetables, eggs, rice, and noodles, often can be met through self-sufficiency. In some cases, their expenditure on food is close to zero, resulting in a very small EC and a rather large difference between the two indicators. However, the poverty of this group is very much worth tracking, and they may determine the progress of poverty alleviation efforts. The EEC serves as a valuable complement in identifying these households. In contrast, for the “EC only” group, low-income urban households with older heads of household and heads with fewer years of education are less likely to be identified by EEC. For older adults with limited ability on cooking, community support, such as community kitchens, can provide basic service regards to energy supply, particularly in the context of population aging. Consequently, their energy consumption needs for cooking are no longer within the household, meaning EEC remains underestimated. In addition, older adults’ energy use behaviors are influenced by traditional lifestyles. They show lower adoption rates of new energy-using devices and maintain more fixed energy use patterns. This unexpected result highlights the importance of using multiple measurements of poverty. In addition, empirical findings from the eastern, middle, and western regions indicate that low-income households are most vulnerable to being mismatched. Among these, rural households in all three regions tend not to be identified by EC, while households with older heads in the eastern region and urban households in the middle region are less likely to be identified by EEC (see Supplementary Table 7–9). To conclude, both EEC and EC have their own scope of application in measuring poverty, and both have limitations. The EEC may be especially valuable in identifying low-income households in rural areas, as it can identify poverty that is hidden by the use of monetary-based indicators alone. But it may fail in urban areas. Therefore, it is best to use poverty indicators in combination. Discussion Tracking household energy flows within households through food systems offers a different perspective for welfare assessment, particularly as traditional measurements struggle with regional heterogeneity in consumption patterns. Traditional poverty indicators, such as the poverty line, capture basic needs by quantifying the minimum income required to meet them. While income- and expenditure-based thresholds quantify monetary deprivation, they often fail to capture non-monetary forms of poverty, especially inadequate cooking infrastructure that threatens food security55,56. This study fills this gap by mapping how households allocate limited energy resources between essential food preparation and other needs, providing a behavior-based complement to existing indicators. This framework focuses poverty measurement on cooking, specifically the thermodynamic process of transforming raw materials into edible meals. By quantifying energy invested in stove operation, fuel processing, and meal preparation relative to total household consumption, it avoids distortions arising from reliance on subsidized pricing or informal food economies. More importantly, our proposed EEC addresses two main shortcomings in EC. First, avoid counting non-nutritive processed foods in the diets of higher-income populations; second, avoid undercounting subsistence farming inputs due to monetary valuation. Our analysis shows that the EEC not only matches established socioeconomic patterns such as regional disparities (see Supplementary Fig. 8 for robustness test) and basic household characteristics, but also complements EC in assessing household economic living standards. On one hand, compared to direct monetary indicators such as income and expenditure, the EEC’s stronger compliance with Engel’s Law indicates that it more effectively captures the burden of basic needs when disposable income changes. On the other hand, its enhanced correlation with subjective experiences and asset holdings demonstrates that the EEC combines the multidimensional characteristics of economic well-being, reflecting both perceived deprivation and actual asset possessions. In addition, we find that low-income small rural households, potentially living in poverty, who may rely on subsistence agriculture and non-monetary transactions, are particularly not identified by EC. This underscores the important value of EEC in the rural scenarios of low-income countries. Moreover, a combined analysis of indicators from different perspectives, including EEC, allows for a more comprehensive assessment of poverty levels, facilitating more precisely targeted poverty reduction policies. The EEC demonstrates broad applicability across three important socioeconomic contexts, aligning with global efforts to refine poverty measurement frameworks. First, in low-income countries and rural regions characterized by non-monetary transactions and subsistence agriculture, this indicator effectively captures poverty conditions by analyzing energy consumption, overcoming the limitations of monetary-based metrics. Second, for regions with well-developed public services but large consumption mobility disparities, this indicator addresses expenditure distortions caused by subsidy policies and credit constraints. Third, the global proliferation of energy infrastructure has established cross-regional comparability, enabling standardized poverty assessments from Sub-Saharan Africa to South Asia. This multi-scale applicability makes the EEC a complementary perspective for measuring SDGs. Although the EEC offers advantages in complementing traditional indicators, its application still faces some limitations. First, the EEC focuses on household energy use and may be an incorrect measure for the standard of living for those who primarily eat out or rely on takeout. In this case, a traditional monetary-based indicator, like the EC, can serve as a useful complement. This limitation also applies to other groups with unusual energy consumption patterns. For example, in an increasingly aging society, elderly adults with limited cooking abilities may rely on community support systems such as public canteens or meal delivery services. In such cases, the household-centered EEC framework may not accurately reflect their standard of living, indicating the need for complementary metrics that capture access to public services. Second, the EEC may not apply to households with little or no access to energy57. However, as energy access continues to improve globally, particularly in low-income countries58,59, the EEC remains relevant for tracking poverty. It is worth noting that measurement challenges persist even in areas with accessible energy sources60. The possible bias from unrecorded energy sources, such as free straws and wood collected by rural residents, could affect the accuracy of EEC calculations. These non-market energy streams, while economically important to vulnerable households, are frequently underreported in standard surveys. Third, when we calculate the energy consumption of each activity, we assume in advance that the output power of all devices and the conversion coefficient are fixed. However, the efficacy of various device types is constantly changing as they are updated. This assumption may introduce additional bias, particularly given the substantial variations in actual energy use patterns. For instance, in the case of cooking, energy consumption varies widely among different cooking methods61. This heterogeneity in equipment efficiency and usage behaviors, especially pronounced in low-income countries where device types are highly diverse, prompts us to evaluate household energy use more carefully in the future. We believe that EEC can be combined with traditional indicators to serve as a pre-assessment tool and the other as a validation mechanism, thus reducing the risk of missing true poverty. First, we suggest using established monetary-based measures to identify candidate regions and households for anti-poverty programs, maintaining compatibility with current targeting systems. Second, we propose applying the EEC as a complementary indicator within these candidate populations to identify households exhibiting high expenditure consumption shares for basic cooking needs, which may indicate poverty not fully captured by monetary metrics alone. Third, we recommend implementing targeted interventions based on this combined assessment. For households identified as experiencing poverty by both metrics, comprehensive support, including livelihood assistance, could be appropriate; for those with adequate income but high EEC values, specifically targeted energy assistance, such as efficient cookstove programs or utility bill support, could be most relevant. Methods Data source and representativeness The data are retrieved from the China Residential Energy Consumption Survey (CRECS), which allows us to analyze household-level energy consumption and other important characteristics in China. The survey aims to collect a set of high-quality microdata on household energy consumption representing all regions of China, in order to understand the differences in energy consumption patterns between urban and rural areas, analyze the dynamic impact of the urbanization process on energy consumption, and assess the effectiveness of corresponding energy policies. Twelve rounds of the Household Energy Consumption Questionnaire Survey have been successfully implemented, creating a relatively stable questionnaire structure and module, and constructing a large-sample household energy consumption microdata base (containing a sample of 32,743 households). The CRECS protocol is approved by the National Survey Research Center (NSRC), an interdisciplinary research institute at Renmin University of China. In this study, we use data from 2014 (CRECS 2014). It is important to note that the CRECS 2014 was embedded within the China General Social Survey (CGSS) framework. The CGSS main sample was constructed using a stratified, multi-stage probability sampling, aiming to represent all urban and rural households in mainland China. This ensures the CRECS household sampling frame proportionally covers all regions. The CRECS subsample itself was generated through a random selection: one-third of respondents were randomly selected from the full CGSS sample to complete the CRECS module questionnaire electronically, ensuring each eligible household had an equal probability of being selected. Thus, the CRECS sample inherits the representative characteristics of the CGSS main sample, and its geographic distribution accurately reflects the national population distribution pattern under the probability proportional to size (PPS) sampling scheme. Data collection took place between July and October 2015, capturing household information for the 2014 calendar year. The CRECS 2014 sample includes 28 provinces (or municipalities), 85 cities, and 127 counties (or districts) in mainland China, with a total of 3863 households, of which 54.75% are rural residents. The data modules we use include the following two broad categories. The first category comprises basic household characteristics, including head of household details such as age, education level, and household size, as well as income and expenditure information data such as total household income, total household expenditure, and various household expenditure categories. The second category comprises energy consumption, including fuel types used in household end-use activities, along with the frequency and duration of use for various devices. Measurement of energy consumption The survey covers eight categories of fuel types: electricity, coal, natural gas, LPG, solar power, biomass, thermal and others. These fuels are used in five end-use activities. Cooking involves food preparation such as boiling, frying, steaming, and baking. Space heating maintains thermal comfort in cold seasons. Space cooling provides cooling and ventilation in warm seasons. Water heating supplies domestic hot water. Home appliances include lighting, refrigeration, entertainment, and other household electronics. These five activities comprehensively cover the scope of household energy consumption24,43. The consumption of cooking devices and household appliances is primarily determined by the unit energy consumption of the device, its usage frequency, and operating time. The energy consumption for home heating is influenced by the heating method, categorized into centralized heating systems and individual household heating systems. We estimate effective energy consumption using a bottom-up device-based accounting approach. This approach is widely used in accounting for energy consumption within households. It can disaggregate household energy consumption by fuel type and end-use activities. It has proven particularly valuable for quantifying non-commercial energy sources such as biomass, which typically do not require traditional metering systems. The general equation for estimating the energy consumption of household i, measured in kgce, is as follows:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{Energy}}_{i}={\sum}_{n}^{N}\left({P}_{i,n}\times {f}_{i,n}\times {t}_{i,n}\times {\eta }_{i,n}\times {D}_{i,n}\right)\times {\phi }_{n}$$\end{document}Energyi=∑nNPi,n×fi,n×ti,n×ηi,n×Di,n×ϕnwhere Pi,n represents the output power of appliance that using fuel n, fi,n is the daily usage frequency, ti,n is the average operating duration per use, ηi,n is the energy efficiency coefficient, Di,n is the number of operational days in the survey period, and ϕn is the conversion coefficient. Our bottom-up energy accounting approach relies on standardized energy conversion coefficients from official sources. While this ensures consistency and comparability with national energy statistics, it assumes uniform energy quality and conversion efficiency across all households. In reality, variations in fuel quality, appliance efficiency, and combustion conditions may introduce measurement errors. Regardless of the fuel type or appliance, only a portion of the energy can be converted into useful work during operation, meaning only part of the energy is utilized. To capture the actual delivered energy service obtained by households more precisely, we further measure the effective energy consumption as shown in Eq. (2), which adjusts final energy consumption by applying thermal efficiency weights for different fuel types.2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{Effective}{energy}}_{i}={\sum}_{n}^{N}{{Energy}}_{i,n}\times {\tau }_{n}$$\end{document}Effectiveenergyi=∑nNEnergyi,n×τnwhere τn is respective thermal efficiency coefficient of fuel n62. Calculation of the EEC The traditional Engel Coefficient is calculated as the share of food expenditures in total expenditures, as shown in Eq. (3).3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{\sum}_{j\in {J}_{{Food}}}{P}_{{Food},j}\times {Q}_{{Food},j}}{{\sum}_{i\in I,j\in {J}_{i}}{P}_{i,j}\times {Q}_{i,j}}$$\end{document}∑j∈JFoodPFood,j×QFood,j∑i∈I,j∈JiPi,j×Qi,jwhere Qi,j represents quantity of the jth (j∈Ji) consumer good and service in the ith (i∈I) consumption category, and Pi,j represents the price of this good and service. The total consumption category I = {Food, Clothing, Residence, Transportation, Recreation, Education, Health care, and Others}. Considering that each consumption category contains different consumer goods and services, the total set of consumer goods and services in each consumption category (Ji) depends on category i. It is clear that Ji1 ∩ Ji2 = ∅ for categories i1 and i2. Consumer goods and services in the numerator are taken to be i = Food, which includes only consumer goods in the food category, e.g., rice, vegetables, meat, belonging to the set JFood. Consumer goods and services in the denominator are aggregated across all categories. The Energy Engel Coefficient (EEC) is calculated based on energy consumption data. It can be divided into two main steps to compensate for the shortcomings of Qi,j and Pi,j in EC. First, we focus our perspective on the goods and services to the end-use activities within the household. Then, we replace the original monetary measures with the physical quantity of energy consumed by household end-use activities. Therefore, the EEC calculates the total amount of energy used for food preparation activities, known as cooking, as a proportion of the total energy consumption for all household end-use activities, as shown in Eq. (4).4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{{\sum}_{n\in N}{{Energy}}_{{Cooking},n}}{{\sum}_{m\in M}{\sum}_{n\in N}{{Energy}}_{m,n}}$$\end{document}∑n∈NEnergyCooking,n∑m∈M∑n∈NEnergym,nwhere Energym,n represents the effective energy consumption quantity of the nth fuel type used in the mth category of household end-use activity. The end-use activities are categorized into five types, M = {Cooking, Space heating, Space cooling, Water heating, and Home appliances}. And there are eight fuel types considered, N = {Electricity, Coal, Natural gas, LPG, Solar power, Biomass, Thermal, and Others}. Empirical strategy Our empirical model aims to examine the determinants of classification differences between EEC-based and EC-based poverty indicators. We specify a linear probability model with fixed effects. The selection of explanatory variables is theoretically motivated (see Supplementary Note 13). Income is regarded as the fundamental determinant of economic capacity, serving as the baseline for calibrating the two poverty indicators47,63. The urban/rural status reflects structural differences in cost of living, energy access, and infrastructure development, which influence how poverty is expressed across different geographic contexts64,65. Household size determines economies of scale in consumption and dependency burdens, both critical factors in poverty assessment66. The level of age and education represents human capital dimensions that affect consumption preferences67. Our dependent variable reflects disparities in poverty identification. By classifying households as either identified as experiencing poverty or not based on the EEC and EC, we identify three specific cases of divergence. The first type, “EEC only”, refers to households identified as experiencing poverty by the EEC but not by the EC. The second type, “EC only”, refers to households identified as experiencing poverty by the EC but not by the EEC. The third type, “All Mismatch”, combines the above two cases. Based on the above analysis, the empirical model is organized as Eq. (5).5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\rm{Prob}}}}\left({Mismatc}{h}_{i}=1\right)={\beta }_{0}+{{{{\boldsymbol{\beta }}}}}_{1}^{{\prime} }{{{{\boldsymbol{X}}}}}_{i}+{\sigma }_{c}+{\varepsilon }_{i}$$\end{document}ProbMismatchi=1=β0+β1′Xi+σc+εiwhere Mismatchi is a binary variable divided into three cases. For “EEC only”, it equals 1 if a household is identified as experiencing poverty by the EEC but not by the EC, and 0 otherwise. For “EC only”, it equals 1 if a household is identified as experiencing poverty by the EC but not by the EEC, and 0 otherwise. For “All Mismatch”, it equals 1 if a household falls into either of the above two cases, and 0 otherwise. Xi are key variables that may affect the consistency of poverty identification across indicators, including urban-rural status (Urbani), household annual income per capita in logarithmic form (LnIncomei), household size (Sizei), age of the head of household (Agei), and years of education of the household head (Edui). Urbani is a binary variable, taking 1 for urban and 0 for rural households. The other variables are continuous. σc represents the province-level fixed effects, and εi represents the error term. This empirical model helps us to find which household attributes lead to differences between monetary and energy-based poverty assessments, thereby showing the unique but complementary information captured by each indicator. The inclusion of province fixed effects controls for regional heterogeneity in economic development, energy infrastructure, and climate conditions that might affect both poverty incidence and measurement. This ensures that our estimates primarily reflect household-level characteristics rather than unobserved regional factors. Descriptive statistics for the variables are given in Supplementary Table 2. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Reporting Summary Transparent Peer Review file Supplementary Information Reporting Summary Transparent Peer Review file Source data Source Data Source Data