2.1. Measuring emission intensities and deriving the household carbon footprint

This study only focuses on CO₂ emissions since it represents the largest share of GHG emissions (UNFCCC, 2010).Footnote ² To estimate an Indonesian household's carbon footprint, we follow Lenzen's (Reference Lenzen1998b) approach, which computed carbon embedded in an Australian household's final consumption. Thus we are focusing here on consumption-based (rather than territorially based) emissions.Footnote ³ We basically trace the CO₂ emitted by the final consumption element back to its intermediates and consider both the direct and indirect emissions that occur from household expenditure. Applying the expenditure approach, figure 1 shows how CO₂ intensities of goods and services in a given economy can be traced using IO analysis.Footnote ⁴

Source: Modified from Kok et al. (2006).

Figure 1. IO emission analysis: expenditure approach

In the first step, CO₂ intensities of each Indonesian IO sector (in the local currency unit, Rp) are estimated. We assume the Single Region Model, which suggests that emissions of both imported and domestic products are assumed to have the same emission intensity, implying that they are produced by the same or similarly carbon-intensive technology. One may argue that products in the developed world are produced more efficiently and may have lower emission intensities. On the other hand, imports require transport that might increase emissions. However, such issues are beyond the scope of this study.Footnote ⁵ In this study, the CO₂ emission intensities were derived using the Leontief inverse of the IO table multiplied by the carbon intensities derived from GTAP.

In the second step, the CO₂ emission intensities of each economic sector were matched to their household expenditure category. We refer to the SUSENAS questionnaire and GTAP sector classification (Huff et al., Reference Huff, McDougall and Walmsley2000) to match these sectors. Consumption expenditures from SUSENAS are then multiplied by the derived CO₂ emission intensity, and then by summing them up we get the household carbon footprint.Footnote ⁶

As the Single Region Model assumes that the domestic energy and environmental technologies used in production are the same as abroad, we just calculate direct and indirect CO₂ emissions from the final demand of sectors. First, the direct CO₂ emission intensities from final demand, CO ₂ ^fd are expressed by the following:

(1) $${CO}_{2}^{fd}=c^{\prime}E^{fd}y,$$

where c ^′, E ^fd , and y represent the inverse of the emissions coefficient vector, the matrix of energy use and the vector of final demand. The final demand vector is not disaggregated into household expenditure, exports and investment.

Secondly, the indirect emissions, CO ₂ ^ind , can be divided into three sources of emissions: (a) from domestic production for domestic final demand; (b) from imported intermediates; and (c) from imported products for domestic final demand (excluding exports). Then, the sectoral CO₂ emissions can be estimated by multiplying each sector's final demand, y, the transposed emissions coefficients, c ^′, the matrix of industrial energy use, E ^ind , and with the domestic Leontief inverse (I-A)⁻¹, as follows:

(2) $$\eqalign{{CO}_{2}^{ind}& = c^{\prime}E^{ind}\lsqb \lpar I - A \rpar^{{-1}}y_{{\ne \exp}}{+}\lpar \lpar I - A_{tot} \rpar^{{-1}} - \lpar I - A \rpar^{{-1}} \rpar y_{{\ne \exp}} \cr & \quad {+}\lpar I - A_{tot} \rpar^{{-1}}y_{{imp}_{{\ne \exp}}} \rsqb ,}$$

where $A_{tot} = A + A_{imp}$ , and $y_{tot} = y + y_{imp}$ . $y_{{\ne {\rm exp}}}$ and I represent domestic final demand and identity matrix, while A indicates the matrix of technical coefficients that reflects the intermediates’ contribution to one unit of final output.

Hence the direct and indirect CO₂ emission intensities can be calculated as follows:

(3) $${CO}_{2}={CO}_{2}^{fd} + CO_{2}^{ind}$$

(4) $$\eqalign{{CO}_{2}&=c^{\prime}\lcub {E}^{fd}y+E^{ind}\lsqb \lpar I - A\rpar^{-1}y_{\ne exp}+\lpar \!\lpar I - A_{tot}\rpar^{-1}-\lpar I - A\rpar^{-1}\!\rpar y_{\ne exp} \cr & \quad + \lpar I - A_{tot} \rpar^{-1}y_{imp\ne exp}\rsqb \rcub }$$

Finally, the above carbon intensities (in kg CO₂/Rp) of each sector are multiplied by the household consumption recorded from SUSENAS (in Rp) for the respective category and then the products from all categories are summed up for each household. The carbon footprint CO $_{2}^{hh}$ (in kg of CO₂) for each household is calculated by the following equation:

(5) $${CO}_{2i}^{hh}=\sum\limits_i^j \lpar {CO}_{2j}\ast {Exp}_{ij} \rpar ,$$

where i and j denote household and expenditure item, respectively.

2.2. Drivers of the household carbon footprint

This section will investigate how we analyze the determinants of the household carbon footprint as calculated above. The linkage between the expenditure choices and the carbon footprints will be determined from the carbon intensity of particular items consumed in Indonesia. From the list of consumption items in SUSENAS, we will analyze the determinants of particular carbon-intensive consumption preferences, including choices related to household operations such as fuel-light and transportation. The empirical analysis is postulated as follows.

(6) $$\ln {CO}_{2i}^{hh}=\alpha +\beta_{1}{lnEXP}_{i}+\beta _{2}X_{i} + \varepsilon_{i}.$$

The ordinary least squares (OLS) method will first be employed to regress the log of household carbon footprint CO $_{2}^{hh}$ per capita on log of household expenditure per capita, lnEXP, as a proxy for per capita income, and a range of other determinants, X, including region, household members, education, gender and age of household head. To capture the possible nonlinearity of expenditures on household emissions (i.e., to test for a household-level EKC), a squared term for the expenditure per capita and age will be incorporated as well.

As we derived CO₂ emissions from expenditures, one can argue that our expenditure variable could have high in-built correlation with computed CO₂ emissions by construction. Dealing with this issue, we can proxy expenditure with expenditure quintile dummies,Footnote ⁷ Q; then regression (6) could be split into two stages, as follows:

(7) $$\ln {CO}_{2i}^{hh} = \alpha +\beta_{q}\sum\limits_{q=1}^5 Q_{qi}+\varepsilon_{i}$$

and

(8) $$\varepsilon_{i}=\alpha +\beta_{1}X_{i}+\gamma_{i},$$

where $\varepsilon_{i}$ is the residual from regression (7) to study drivers that are unrelated to the expenditure–emissions link.

In other words, we regress emissions on the expenditure quintiles in (7) and then regress its residuals on other control variables (i.e., household characteristics excluding expenditure) in (8). This approach can determine the effect of characteristics of households on their emissions, over and above the expenditure–emissions link. Of particular interest is to understand the drivers of the heterogeneity of the household emissions at a given level of expenditure, and thereby to identify possible policy implications to reduce emissions without compromising the wellbeing of households.

In addition, we will also apply quantile regressions in the analysis to account for the possibility that the household emissions distribution is highly skewed and heteroscedasticity might be an issue. In this case, compared with the OLS regression, the quantile regression could be more robust to outliers, partly since it does not assume that the residuals are iid. Another reason is that we will be allowed to analyze the effect of the right-hand side variables on the location and the scale parameters in the model. Technically, while OLS minimizes the residuals sum of squares, $\sum e_{i}^{2}$ , the quantile regression minimizes the sum that gives penalties of about $\lpar {1-}q \rpar {\vert}ei{\vert}$ for overprediction and of about $q{\vert}ei{\vert}$ for underprediction (Cameron and Trivedi, Reference Cameron and Trivedi2010).

Substantively, our quantile analysis presumes that the impact of income and other determinants for lower carbon-emitting households might be different from their impact for households with a high carbon footprint. With this in mind, the quantile regression estimates the effect of a one-unit expenditure change on a particular quantile q of our dependent variable (household emissions). Technically, by linear programming, the qth quantile regression minimizes over β q:

(9) $$Q\lpar \beta_{q} \rpar =\sum\limits_{i:y\ge x'\beta}^N {q\vert y_{i}-x_{\beta}^{\prime} \vert +} \sum\limits_{i:y\le x'\beta}^N {(1-q)\vert y_{i}-x_{\beta}^{\prime} \vert }.$$

We can choose q (0 < q < 1) which uniquely estimates the value of β. Suppose choosing q= 0.9, instead of q= 0.1, indicates that more weight is to be assigned to the estimation for observations with $y_{i} \ge x_{i}^{\prime} \beta_{q}$ .

2.3. Decomposing the changes in the carbon footprint

Since we have two waves of the national household survey available, we can investigate not only the drivers of emissions across households, but also over time. One approach to doing so is given by Kaya (Reference Kaya1990), who provides an intuitive approach to the interpretation of the historical trend of CO₂ emissions. This method, which is widely known as the Kaya Identity, suggests that changes in the total emissions level can be decomposed into the changes in four inputs, i.e., population size, per capita income, energy use per unit of GDP, and CO₂ emissions per unit of energy used. While this is usually applied to decompose trends in aggregate emissions, one can also apply this approach to study household-level emissions. Using this decomposition technique, we can then directly link CO₂ emission levels to the population effect, level of economic affluence (measured by per capita expenditure), carbon emission intensity (per energy use) and energy intensity (per unit of output).Footnote ⁸ In this way, we can find the main driving forces of changes in emission levels in the periods observed.

In macro analyses, the Kaya Identity suggests that CO₂ emission levels are the product of: (i) the carbon intensity of the energy supply; (ii) the energy intensity of the economic activity; (iii) the economic per capita output, and population. However, since we do not have the data for energy intensities, in our analysis the Kaya Identity is modified as follows:

(10) $${CO}_{2_{i}} = {HHsize}_{i}\ast {{EXP}_{i}\over {HHSize}_{i}}\ast {{CO}_{2_{i}}\over {EXP}_{i}},$$

where the household CO₂ emissions level is a function of household size, HHsize, per capita expenditure, EXP/HHsize, and emission intensity, CO ₂ /EXP.

In other words, we set up an emissions equation to calculate and decompose the growth of CO₂ emissions into the population effect, a per capita expenditure effect (Rp/capita) and a carbon intensity effect (CO₂/Rp), and express the result as a percentage of the baseline CO₂ emissions level. Following Ang (Reference Ang2005), our decomposition will be employed using the Logarithmic Mean Divisia Index (LMDI), which has several advantages; apart from it being consistent in aggregation, it also gives a perfect decomposition as the results will not contain unexplained residuals. The LMDI approach is modified (10) to construct the following formula:

(11) $$\Delta {CO2_{i}}=C^{T}-C^{0}=\Delta CO2_{HHsize}+\Delta{{CO}2}_{EXP\over{HHsize}}+\Delta {CO}2_{CO2\over{EXP}}$$

Where

$$\eqalign{\Delta {CO2}_{HHsize}& =\sum\limits_i {C_{i}^{T}-C_{i}^{0}\over{lnC_{i}^{T}-{lnC}_{i}^{0}}}\ln \left({HHsize_{i}^{T}\over HHsize_{i}^{0}}\right)\cr \Delta CO2_{EXP/HHSize}&=\sum\limits_i {C_{i}^{T}-C_{i}^{0}\over{lnC_{i}^{T}-{lnC}_{i}^{0}}}\ln \left( \left({{EXP}\over{HHsize}}\right)_{i}^{T}\over{\left({{EXP}\over{HHsize}}\right)}_{i}^{0}\right)\cr \Delta {CO2}_{{CO}2/{EXP}} & = \sum\limits_i {{{C_{i}^{T}-C_{i}^{0}}\over{lnC_{i}^{T}-{lnC}_{i}^{0}}}\ln \left( {{\left({{CO}2}\over{{EXP}}\right)}_{i}^{T}}\over{{\left({{CO}2}\over{{EXP}}\right)}_{i}^{0}} \right)}}$$

where ${\Delta}{{CO}2}_{{HHsize}}$ , ${\Delta }{{CO}2}_{{EXP/HHSize}}$ , and ${\Delta}{{CO}2}_{{CO}2/{EXP}}$ represent changes in CO₂ emissions because of population, expenditure, and the carbon intensity effect, respectively.

2.4. Expenditure elasticities of emissions

Lastly, one can also study the link between expenditure and emissions using demand analysis and the expenditure elasticity of spending on particular goods. Demand analysis is generally utilized to measure the change in demand for any particular good due to the change in income or price. This demand function originates from the consumers’ utility maximization equation, which depends on the prices of goods and individuals’ income (Deaton and Muellbauer, Reference Deaton and Muellbauer1980). We modify this demand theory by replacing the demand for goods with CO₂ emissions associated with the consumption of the respective goods. By applying this, we can analyze the responsiveness of CO₂ emissions of any household consumption category to a change in household income, which is proxied by household expenditure.

As suggested by the conventional Engel curves, we should include prices as one of the independent variables. However, since there are no price data available in SUSENAS, we will estimate the impact of expenditure on the sectoral emission shares without using prices, meaning that the response of share of CO₂ emissions of a particular consumption item will only be dependent on the expenditure amount of the particular consumption item and the socio-economic level of the households. We estimate the following model:

(12) $${sCO}2_{ij}=\beta_{0}+\beta_{1_{ij}} {\ln EXP}_{i}+\beta_{2_{ij}}X_{i}+\varepsilon_{ij},$$

where ${{sCO2}}_{ij}$ represents the share of CO₂ emissions of the jth consumption category to total CO₂ emissions by the ith household, and $\ln {EXP}_{i}$ is the natural logarithm of total household i expenditure. X _i represents a vector of household characteristics and $\varepsilon_{ij}$ is an error term.Footnote ⁹

3.1. Descriptive analysis

Figure A1, in the online appendix available at https://doi.org/10.1017/S1355770X17000262 , provides an overview of the allocation of household expenditure in 2005 and 2009.Footnote ¹⁰ In general, mean household expenditure increased by 72.27 per cent (nominal) and 24.83 per cent (real). The figure also shows large differences in expenditure shares in urban and rural areas. Compared to urban households, households in rural areas have, unsurprisingly, a larger expenditure share on foods and a much smaller share on services, recreation, rents and taxes. In general, by comparing the two surveys we find that the share allocated to food expenditure declined as is to be expected in a growing economy. Moreover, the shares of telecommunication, transportation, health, education and taxes have been increasing in both the rural and urban areas. The share of spending on beverages has been increasing in urban areas as opposed to rural areas where it has been decreasing. In contrast, the share of income that is spent on housing and durable expenditures has been increasing for households in rural areas as opposed to households in urban areas where it has indeed been decreasing.

Before we begin the computation of the carbon footprint, it is very important to point out the relationship between mean consumption in SUSENAS and mean consumption in the national accounts. If we compare the two databases, we note that the expenditure computation from SUSENAS is significantly lower than consumption expenditures reported in the national accounts (this underestimation measure can also be found in other studies, e.g., Yusuf, Reference Yusuf2006, and Mishra, Reference Mishra2009). The deviation between the two measures is partly because of the computations in the national accounts which were constructed from the supply side of the economy whereas SUSENAS expenditures were taken from representative sample surveys. In addition, national accounts also include the consumption by non-households. There might also be measurement error in the survey with households understating their total consumption, an issue that has been discussed extensively in the development economics literature (e.g., Deaton and Kozel, Reference Deaton and Kozel2005).

Table A1 in the online appendix portrays the calculations of household expenditure using the national accounts and SUSENAS. Given the difference in the measurements from SUSENAS, which accounted for around 42–49 per cent of the national account measurements, we scaled up the computation of household emissions by dividing household consumption by the percentage of SUSENAS to total expenditure based on national accounts when we computed the carbon emissions (Mishra, Reference Mishra2009). However, the fact that the aggregate from SUSENAS expenditures falls short of the national accounts (including in our calculation with the scaled-up household emissions) would not imply anything about the distribution of the expenditures across households. Hence we assume that the discrepancies between expenditure items are more or less at the same amount across households.

In the next step, by incorporating the Indonesia IO table and GTAP's energy use matrix, we extract the CO₂ emission intensity level of the 175 economic sectors.Footnote ¹¹ The CO₂ emission intensity is measured in terms of kilotons per million rupiah (or gram CO₂/Rp), which captures the amount of CO₂ released from the production of goods and services in the Indonesian economy. Table A2 in the online appendix presents the 10 most and least CO₂ intensive sectors, measured by emissions per unit of expenditures. It can be seen that sectors that emit CO₂ intensively include: electricity, gas, cement, non-metallic minerals, glasses and their products, ceramics and clay products. In addition to those electric and manufacturing sectors, all transportation services are also very carbon intensive.

In contrast, the least CO₂ intensive sectors in Indonesia are associated with agricultural crops sectors, including fiber crops, grains, sweet potatoes, fruits and beans. These figures reflect the fact that these products do not use much energy in production compared to manufacturing and transportation sectors.Footnote ¹² In addition to the agricultural sectors, service sectors also have a lower CO₂ intensity, including such industries as film and distribution services, building and land rent. In general, agricultural and service-related activities emit less CO₂ compared to manufacturing sectors.

The derived CO₂ emission intensities were then matched with the consumption categories in the SUSENAS 2005 and 2009. There are around 340 consumption items in the expenditure survey and this was aggregated to represent the major household expenditures. Figure A2 in the online appendix shows the average CO₂ emissions (in kg) from major expenditure categories. It is observed that CO₂ emissions vary greatly by consumption item. The lowest CO₂ emissions were observed from the consumption of cereals, medical services, telecommunication services and recreation. On the other hand, the highest CO₂ emissions were observed from the consumption of transportation as well as fuel and light.

From 2005 to 2009, household emissions from all expenditure categories increased (by 29 per cent on average), but at a different pace. Emissions from fuel-light expenditures grew proportionately less, from 1,688 kg to 2,768 kg (19 per cent). Meanwhile, emissions from transportation, the second highest emission source, rose more proportionately, from 183 kg to 290 kg (58 per cent), pointing to the importance of transportation in emissions growth. Emissions from food-related expenditures grew around 20–36 per cent. We also note that health, transportation and tax have the fastest growth rates of emissions (albeit from a low base).

The disaggregation of the CO₂ emissions into regions and income levels is presented in figure 2. We also find a very large difference in CO₂ emissions with respect to household affluence. In more detail, the per capita emissions from the richest quintile are about seven times higher than the per capita emissions from the lowest quintile, and still about three times as high as the level from the third quintile (middle income group).

Source: Author's computation based on SUSENAS 2005–2009; IO 2005; GTAP-E 2005.

Figure 2. Carbon footprint by household expenditure quintile, location, and education of the head (2005 and 2009)

Moreover, per capita carbon emission levels rise with educational attainment of the household head. The pattern of emissions with respect to educational attainment is surely related to income levels, although the differences between are not as steep as with the affluence level. Lastly, based on location, both surveys indicated that urban household emissions are about twice the amount of rural households. Looking at change from 2005 to 2009, we observe that overall household per capita emissions grew on average from 0.70 tons (2005) to 0.90 tons (2009), an increase of about 29 per cent.Footnote ¹³

Comparing emission shares to expenditure shares (online appendix figure A3), we note first that they are very similar, suggesting a very close linkage between incomes and emissions. One can also see, however, that expenditures appear to be slightly less unequally distributed than emissions, particularly in 2005 (see also Irfany and Klasen, Reference Irfany and Klasen2016).

3.2. Drivers of household carbon footprint

The regression analysis of the determinants of household emissions is presented in table 1. Various model specifications were employed to analyze the drivers of the variation in CO₂ emissions. In Regressions I and II, we regress the log of per capita emissions with log per capita expenditure and other control variables, including dummies for different household characteristics. In the third regression, we regress the per capita carbon footprint only on income quintiles. Regressions IV and V use the residuals from Regression III as the dependent variable and household characteristics as control variables, with the cube of HH-head as an additional independent variable in Regression V.

Table 1. The determinants of household carbon footprint, 2005–2009

Notes: In Regressions I–III, the dependent variable is log of per capita carbon footprint, while in Regressions IV and V, the dependent variable is the residual from regression III. ***indicates significance at the 1% level. Province dummies are included but not reported here.

Source: Author's estimation.

From Regression I, we find that expenditure has a very high and significant relationship to emissions, with an elasticity of slightly above 1, suggesting that as per capita expenditures rise, emissions rise in equal (actually slightly higher) proportion. In Regression II, we include the square of per capita expenditures and find the square to have a negative effect. This implies an inverted U-shaped pattern of the carbon footprint with respect to expenditure. In other words, rising affluence leads to increasing CO₂ emissions, ceteris paribus, but eventually declines as per capita expenditure rises even further; this, in principle, supports a micro-level EKC, but one should also note that the turning point is far beyond the sample included here so that it is, for all practical purposes, empirically not relevant and just suggests that household emissions are increasing with expenditures at falling marginal rates. Furthermore, we also indicate that the greater the age (of the household head), if the gender (of household head) was female, if the household head was married, and if the region was an urban area, the more carbon was emitted; thus the unconditional effects we showed descriptively above still hold in a multivariate setting, although the effects of these covariates are rather small, certainly when compared to the income effect. This suggests that higher education and urban location are not just associated with higher per capita expenditures, but also with more carbon-intensive lifestyles even conditioning on expenditure; this is likely to be related to higher transportation as well as energy use in urban and more educated households. Moreover, household size has, somewhat surprisingly, a small positive impact on per capita carbon emissions, suggesting that larger households are apparently unable to economize on per capita carbon emissions and, in fact, have slightly more carbon-intensive lifestyles.

In Regression III we regress per capita emissions on affluence quintiles, which divide households into five equal parts by sorting the per capita expenditure out from lowest to highest. It is observed that households in the higher quintiles have a larger carbon footprint and the coefficients are statistically significant. Moving from the first to the second quintile increases the per capita emissions by 62 per cent, whereas moving from the first to the richest quintile increases household emissions by 231 per cent.

We then utilize the residuals from Regression III as the dependent variable of Regressions IV and V, and household characteristics as control variables. The idea is to purge the impact of incomes that would then reveal the effect of certain household characteristics on their emissions, beyond the effect of incomes. As indicated, it is not surprising that the coefficients of household characteristics (the control variables) are statistically significant and consistent with the previous specifications. The two exceptions are the effect of household size and marital status which switches from slightly positive to slightly negative. In emissions unrelated to expenditures, larger and married households now seem to have slightly lower emissions. This suggests that the previous positive effect was influenced by a correlation between household size, marital status and household expenditures; once considering the residuals of emissions unrelated to household expenditures, larger (and married) households appear to be able to (slightly) economize on per capita emissions. It is also interesting to note that, in all regressions, the dummy variable for the second year (2009) is slightly negative, suggesting that, controlling for rising expenditure, urbanization and education, per capita emissions have been falling by about 2 per cent; this could suggest slight improvements in expenditure patterns (controlling for incomes) towards less carbon-intensive products.

Lastly, in all regressions we include dummies for all of the provinces (available on request). The estimated coefficients for all control variables with and without dummies do not change significantly. However, from the province fixed-effects regression we find that the emissions of provinces in Java and Bali, Kalimantan Timur, Kalimantan Selatan, Sulawesi Selatan and Sulawesi Tenggara were higher than the amount in other provinces.

Table 2 presents quantile regression estimates using $q = 0.10; q = 0.25; 0.50; 0.75$ ; and 0.90. Apart from its statistical advantages (particularly in the case of heteroscedasticity), it helps us understand whether household affluence and other covariates might have a different effect at different quantiles of the emission distribution.

Table 2. Quantile regression estimates

Source: Author's estimation.

We find that households with low emissions seem to have slightly higher expenditure elasticities to emit: from about 1.070 (at 25 per cent quantile), the magnitudes then fall slightly to 1.044 (at median quantile), and to 1.027 (at 75 per cent quantile) and to 1.021 (at 90 per cent quantile). In other words, low-emitter household groups seem to be slightly more responsive to emit and then the effect decreases for those with higher emissions. But for all households, the emission elasticity remains above one.Footnote ¹⁴ Finally, similarly to the OLS estimation, we again observe that other household characteristics also matter as determinants of the household carbon footprint, but do not differ greatly between quantiles. Interestingly, the effect of household size is slightly negative in most quantiles.

3.3. The decomposition analysis of emission growth

Figure 3 presents the decomposition of the growth of household CO₂ emissions from 2005 to 2009. From the perspective of contributors to CO₂ emissions growth, we can clearly show that rising per capita expenditures is the largest contributor to the rise in CO₂ emissions in all quintiles. This rise in expenditures has the largest effect in the lowest quintile, which means that rising per capita expenditure of households in this quintile has increased CO₂ emissions more than the same rise in per capita expenditures of households in the upper quintiles. Moving to affluent households, the expenditure effect then decreases gradually, but the effects in all quintiles remain positive.

Note: Total expenditures used here are deflated to reflect real values. Source: Author's computation based on SUSENAS 2005–2009; IO 2005; GTAP-E 2005.

Figure 3. Decomposition of CO₂ emission growth

Moreover, moving from the lowest to the highest households, we can clearly identify that the population effect has a decreasing pattern, which has a positive effect on the first two quintiles and has a negative effect on the third to the highest quintile; here we see the effect of fertility decline which affects household size in the richer quintiles. Finally, the CO₂ intensity effect (measured as kg CO₂/Rp) has the largest negative contribution to CO₂ emissions rising in the lowest quintile. This effect has a negative sign from the first until the third quintile and has a positive sign in the highest quintile, suggesting that consumption changes in that quintile have served to significantly increase emissions. Increases in the energy expenditure share (mainly transportation) in 2009 are the most important driving factor of this positive carbon intensity effect among the richest household group (compared to a falling expenditure share in the lower income groups).Footnote ¹⁵ To sum up, richer households have lower emissions growth because of their population (household size) effect as well as a smaller increase in per capita expenditures, but this is partly offset by choosing more carbon-intensive goods due to rising affluence.

3.4. Expenditure elasticities of emissions

Due to the fact that expenditure is the most important driver of the household carbon footprint, we conduct an analysis of expenditure elasticities of CO₂ emissions that measure the responsiveness of CO₂ emissions (as a share of total household emissions) to a change of total household expenditure. There are some important issues to be taken into consideration for our analysis. First, dealing with the potential endogeneity problem, one could have a valid instrument for total expenditures, say for instance, for the asset index, and employ the instrument in a two-stage least squares procedure. However, our database unfortunately does not provide sufficient candidates as valid instruments for total expenditure, as we do not have sufficient data on assets in SUSENAS. Secondly, in addition to the national estimation, we will also analyze expenditure elasticities for both rural and urban areas, as well as computing expenditure elasticities by household quintiles.

As the demand theory suggests, a negative coefficient of expenditure elasticities reflects a decreasing share of any particular expenditure group due to rising affluence, and vice versa. Our results on expenditure elasticities on CO₂ emissions generally have the same direction as found in conventional Engle curve studies in the literature. Table A3 in the online appendix reveals some important findings. We find that inferior goods, such as vegetables and cereals, have negative signs, meaning that rising expenditure will reduce their share of CO₂ emissions of these consumption categories, due to a falling expenditure share on these goods as incomes rise. In the opposite direction, luxury goods such as health expenditures, housing, durable goods, transportation, services and rent have positive value, meaning that the rising of household affluence tends to contribute a higher share of CO₂ emissions to the total household emissions. Specifically, the rising affluence will promote carbon-intensive transportation expenditures in that a 1 per cent increase in household expenditure will increase the share of CO₂ emissions from transportation by about 0.03 per cent (both in 2005 and 2009); an even larger effect is observed for the impact of rising expenditures on housing and durable goods. Fuel and light consumption, another carbon-intensive category, has a negative elasticity, which means that a 1 per cent increase in household income will reduce the share of CO₂ emissions from these consumption items by about 0.07 per cent in 2005 (0.08 per cent in 2009).

Last but not least, most of the estimated expenditure elasticities coefficients are generally very small, but generally the directions of these expenditure elasticities to CO₂ emissions have the same signs as the conventional Engle curve. However, they have different sensitivities due to the different CO₂ intensities of the consumption categories. The small size of the expenditure elasticities indicates that the household emission change can mainly be attributed to a general volume increase in overall expenditure, and not so much to shifting the expenditure shares within the consumption basket. These findings support the previous results on the decomposition of emission growth that suggest that emission growth is mainly due to rising overall income (expenditure) level.