Most Japanese households use a combination of four types of energy inside the home: electricity, city gas, LP gas, and kerosene⁴. In practice, however, because virtually all households use electricity, a household chooses among energy from city gas, LP gas, and kerosene. City and LP gases cannot be used simultaneously because they have different calorific values and require different equipment. Therefore, at the installation stage, households must decide what type of gas to use. Although LP gas is available in all areas, city gas is available only in urban areas connected to the city gas grid. Therefore, households in urban grid areas have six options for energy source combinations, whereas those in rural non-grid areas have only four options. To examine the availability of city gas in municipalities, we first determined whether there was a household using city gas in each municipality based on data from the Ministry of the Environment's Household CO₂ Survey. We classified the municipalities with households using city gas into urban grid areas, whereas the remaining municipalities were classified as rural non-grid areas.

Table 1 shows the number and percentage of households that selected a specific combination of energy sources. The nationwide percentage of fully electrified households is 13.8%, with more than two energy sources used by the remaining households. The table further shows that a fully electrified household is more popular in rural non-grid than in urban grid areas. Although households often install a solar panel when fully electrifying their houses, solar panels are more difficult to install in urban than in rural areas.

The most common combination in urban grid areas is Combination 2, electricity and city gas, with 34.4% of households choosing that combination. In addition, more than 39% of households use kerosene. In contrast, the most common combination in rural non-grid areas is Combination 6—electricity, LP gas, and kerosene—with 39.2% of households choosing that combination. In rural non-grid areas, 54.2% of households use kerosene. These statistical results suggest that it is not appropriate to discuss household energy consumption in winter without considering kerosene use.

Households use energy for various purposes, and several factors affect household energy consumption. In this study, we focus on CO₂ emissions generated by winter energy consumption (November to February) for items such as space heating and hot water. Specifically, we calculate the difference in CO₂ emissions between June, when CO₂ emissions are minimized for the year, and the winter months and then define that difference as winter CO₂ emissions. By investigating the change in CO₂ emissions between seasons for the same households, it becomes possible to more accurately evaluate the impact of weather changes on household CO₂ emissions.

Figures 1A,B shows the CO₂ emissions of the average household in each prefecture during June and January, respectively. The figures show a geographical variation in CO₂ emissions, namely, that CO₂ emissions in northern regions are higher than in southern regions. The figures further show that CO₂ emissions are substantially greater in January than in June. Indeed, CO₂ emissions triple in some prefectures during that time.

Weather conditions are expected to affect both energy source choices and energy consumption decisions. Households choose their energy choice combinations by referring to previous weather conditions. We obtained temperature data between 4/1/2009 and 3/31/2013 from the website of the Japan Meteorological Agency (2021)⁵. We then identified the nearest weather station for each household and counted the total number of heating required days (HRDs)⁶. It was assumed that heating was required if the average temperature of the corresponding day was below 14°C. Figure 2 presents the total number of HRDs at the prefecture level. The figure clearly shows that prefectures in the north require more heating. In later analysis, we examine whether the total number of HRDs determines households' energy source choices.

The weather conditions in the sampling year ultimately determine the intensity of use of space heating and water-heating appliances. We will now examine whether households using different combinations of energy sources increase CO₂ emissions at different speeds as the temperature drops.

To examine whether the choice of the energy source combination influences CO₂ emissions, we estimate a selection bias correction model. We define household i's CO₂ emissions due to winter energy use as the difference in CO₂ emissions between winter month m and the base month of June. We then assume that the increase in CO₂ emissions becomes ΔC_i,m,k when household i chooses the k^th energy source combination.

We next calculate the difference in the average temperature between the base month and month m in the area where household i lives and indicate this temperature change by ΔT_i,m > 0. To understand how households increase their CO₂ emissions with decreases in temperature, we consider the following empirical model:

where H_iΔT_i,m is the interaction term of household characteristics H_i and temperature change ΔT_i,m. Because households with different characteristics are likely to increase their energy consumption at differing intensities in response to reductions in temperature, we include the interaction terms in the estimation model. Specifically, we include annual household income, number of persons, floor area of the house, and housing ownership in H_i.

We assume that household i receives utility of ui,k* when it chooses the k^th energy source combination. We further assume that this utility is characterized by the following specification:

In addition to annual household income, floor area of the house, and housing ownership, we include the number of HRDs and the age of the house in Z_i. These two attributes influence the choice of energy source combination but do not determine energy consumption behavior directly. It is assumed that the probability that household i chooses the k^th energy source combination can be given by the multinominal logit model:

The estimated coefficients of Equation (1) (CO₂ emissions equation) may be biased because the disturbance term u_i,m,k can be correlated with ν_i,k. To avoid this potential bias, we employ the approach proposed by Dubin and McFadden (1984)⁷.

Table 2 presents the descriptive statistics of the variables used in the empirical analyses below. After removing the households lacking information about socioeconomic or housing characteristics, the number of households in the urban grid area decreased from 24,453 to 23,506, and the number of households in the rural non-grid area decreased from 6,690 to 6,381. Since the data are divided into urban grid and rural non-grid areas for analysis, we present the descriptive statistics separately for urban grid and rural non-grid areas. The number of sampling households is 23,506 for urban grid areas and 6,381 for rural non-grid areas. In the survey, households were asked to choose their annual income class from preprepared options, but we use the median income of the option in this study⁸. Similarly, households were asked to choose the age of the house from preprepared options, but we use the median year of the option in this study⁹. We obtained CO₂ emissions and temperature data from the four winter months: November, December, January, and February. Therefore, the number of these two types of data is quadrupled. The table shows that CO₂ emissions doubled, while the temperature decreased by 14 degree Celsius from June to the aforementioned winter months.

The main objectives of the empirical analysis are to examine (1) whether households choosing different energy source combinations increase CO₂ emissions at different speeds as the temperature drops and (2) what types of household characteristics influence winter energy consumption. Before reporting the results of the CO₂ emissions analysis, we summarize the estimation results for the energy source combination. See the results in Table 3.

In urban grid areas, Combination 2 of electricity and city gas is chosen as the base combination. On the other hand, in rural non-grid areas, Combination 6 of electricity, LP gas, and kerosene is chosen as the base combination. The impacts of the covariates on the likelihood to be chosen are compared between the base and alternative combinations. The positive (negative) coefficient means that the likelihood of an alternative combination increases (decreases) as the size of the covariate increases.

HRDs became positive in all combinations in urban grid areas and became statistically significant in four combinations at less than the 5% level. This result implies that the base combination of electricity and city gas is less likely to be chosen in cold areas. The comparison of the size of the estimated coefficients suggests that Combination 1 of full electrification is less popular in cold areas. In ural non-grid areas, HRDs became negative in Combinations 1 and 3 but positive in Combination 4 at less than the 5% level. This result implies that households in cold regions are more likely to use kerosene in winter.

In urban grid areas, household income became negative and statistically significant at less than a 5% level in Combinations 3–6. The results show that wealthy households tend to choose urban gas rather than LP gas and tend not to choose kerosene. In rural non-grid areas, household income became positive and statistically significant at less than the 5% level in Combinations 1 and 4. This result implies that full electrification is the choice of wealthy households, but LP gas is the choice for less wealthy households.

The floor area and age of the house are also associated with energy source choices. The positive coefficient for floor area in urban grid areas suggests that households living in small houses are more likely to choose Combination 2 of electricity and city gas. On the other hand, floor area became negative in Combination 3 but positive in Combination 4, indicating that households living in large homes do not choose LP gas. Age of house became negative in Combination 1 of full electrification; this result means that full electrified houses are those that have recently been built. Age of house became negative in all three combinations in rural non-grid areas, while it became positive in Combinations 5 and 6. This result suggests that households that recently constructed a house are using a more limited number of energy sources than households that constructed a house at an earlier date.

Table 4 is our main estimation result and shows increases in CO₂ emissions with reductions in temperature. We report the estimation results obtained from the model proposed by Dubin and McFadden (1984) here¹⁰.

The temperature variable became positive and statistically significant at less than the 5% level in all cases, which implies that households increase their CO₂ emissions with decreases in temperature. Although the reported coefficients of temperature show that CO₂ emissions increase (kg) for every one-degree Celsius decrease in temperature, comparison of the coefficients suggests that households using different energy source combinations increase CO₂ emissions at different speeds. The results show that fully electrified households increase CO₂ emissions at the fastest speeds. In contrast, households using gas increase CO₂ emissions at slower speeds. Although households living in fully electrified houses increases CO₂ emissions per month by 29.55 kg for every one-degree Celsius decrease, those combining electricity and city gas increases it only by 12.59 kg. These results suggest that the choice of energy source combination substantially influences CO₂ emissions.

The abovementioned tendencies are observed in both urban grid and rural non-grid areas. However, the size of the temperature coefficient for rural non-grid areas is smaller than for urban grid areas except for Combination 6. This means that households in urban areas increase CO₂ emissions at faster speeds than those in rural areas.

We obtained expected results for the interaction variables. The positive signs for the Income^*temperature variable show that wealthy households increase CO₂ emissions at a faster speed. However, considering the unit of income (JPY 10,000) and the size of the estimated coefficients, we can state that household income does not make much difference in winter CO₂ emissions. In other words, both wealthy and less wealthy households increase CO₂ emissions at similar speeds corresponding to decreases in temperature. This leads to a similar conclusion about the housing size effect because the unit of the floor area of the house is m².

The interaction variable of Persons^*temperature became positive, and therefore, CO₂ emissions increase faster among larger households. The family size effect is much larger than the income effect. However, the results show that household CO₂ emissions during winter increase by <10% even if the number of family members increases by one. This result suggests that household members can share a substantial part of winter energy services; namely, scale economy is important for winter energy use.

Finally, we find a positive sign for the interaction variable of Ownership^*temperature. This result suggests that households living in their own houses increase CO₂ emissions faster with temperature decreases. It seems that this result is opposite to the prediction by previous studies such as those of Levinson and Scott (2004) and Ramos et al. (2016); energy consumption becomes greater among households living in rented houses since landlords do not sufficiently invest in energy efficiency. Perhaps the difference in the housing ownership effect between our study and previous studies arises from differences in the model structure. In previous studies, energy consumption between housing owners and renters was compared directly. In this study, we took two-step procedures; we initially evaluated the effect of housing ownership on the choice of the energy source combination and subsequently evaluated the effect of housing ownership on energy consumption. To fully grasp the effect of housing ownership, we need to consider the effect of housing ownership on the choice of the combination of energy sources.

The study reveals that households living in fully electrified houses emit more CO₂ in winter. The result arises from two conditions. The first is that the emissions factor for electricity is high¹¹. Japan still relies heavily on thermal power generation. The second condition is that much electricity is required to obtain adequate heat during the winter months. This is one reason that kerosene is popular among households in cold regions. Although full electrification of houses has been promoted in the past few decades, it seems necessary to reconsider such a policy from the perspective of global warming countermeasures.

According to the estimate by the Fuji Keizai Group (2016), the number of fully electrified houses will reach 9.39 million in 2025 in Japan, accounting for 17.9% of all homes. However, only 5.24 million houses will be equipped with power generation facilities such as solar panels. It is expected that more households will abandon LP gas use and will move to full electrification in order to lower energy cost. Such energy transition will increse CO₂ emissions unless the power source composition is changed and solar panels are widely adopted.

The study reveals that households using gas emit less CO₂ in winter. Households using gas seem to go through the winter with less energy than households not using it. Perhaps such households warm rooms and supply hot water more efficiently. Given this fact, policies that encourage gas use would be desirable.

The carbon tax rate currently introduced in Japan is kept low at 289 yen per ton. Under such a low carbon tax, households compare exclusively the price of energy itself. The current low carbon tax encourages households to use kerosene and hesitates to use gas. For global warming countermeasures, it will be necessary to reflect the CO₂ emission factor in the energy sales price.

Although we have chosen Japan as a case study country, the general message from this study is valid for other countries as well. The efficiency of household energy consumption must be assessed together with household energy source selection. Otherwise, we could make an incorrect policy recommendation.