3.1 Mechanisms

A number of mechanisms can be outlined through which climate change may affect households, depending on the principal source of income and regional factors. Within the agriculture, forestry and fishing sector, there are four key mechanisms which may influence farm income and consumption. First, higher temperatures and/or low rainfall may reduce crop yields and in turn farm productivity. Insofar as risk mitigation strategies such as availability of irrigation pumps and groundwater aquifers are available to farmers, these effects would be alleviated (Taraz, Reference Taraz2017). As crop yields decline, food prices are expected to rise, which would benefit households that are net producers, while harming those that are net consumers (Jacoby et al., Reference Jacoby, Rabassa and Skoufias2011). Hence, average effects on farm revenue and in turn household consumption are ambiguous. Second, loss of animal production due to heat stress may lead to loss of assets, particularly livestock (Walthall et al., Reference Walthall, Hatfield, Backlund, Lengnick, Marshall, Walsh, Adkins, Aillery, Ainsworth, Ammann, Anderson, Bartomeus, Baumgard, Booker, Bradley, Blumenthal, Bunce, Burkey, Dabney, Delgado, Dukes, Funk, Garrett, Glenn, Grantz, Goodrich, Hu, Izaurralde, Jones, Kim, Leaky, Lewers, Mader, McClung, Morgan, Muth, Nearing, Oosterhuis, Ort, Parmesan, Pettigrew, Polley, Rader, Rice, Rivington, Rosskopf, Salas, Sollenberger, Srygley, Stöckle, Takle, Timlin, White, Winfree, Wright-Morton and Ziska2012). In case of floods, other physical assets of the household may directly be lost or damaged. Third, farmers may respond to a changing climate and reduced crop yields by migrating to a different sector, such as manufacturing or services for employment. Lastly, farmers may migrate out of the region to another district within the state or across states in response to unfavourable climatic conditions (Dallmann and Millock, Reference Dallmann and Millock2017).

Industry would be affected by weather shocks primarily through their adverse impact on labour productivity in the form of physiological heat stress. This is particularly so for mining, manufacturing and construction activities as the strenuous working conditions of workers are compounded with climate-induced heat stress (Somanathan et al., Reference Somanathan, Somanathan, Sudarshan and Tewari2015). However, there may be an increase in demand for construction and rebuilding of infrastructure in the aftermath of extreme climate events (Hallegatte et al., Reference Hallegatte, Ranger, Mestre, Dumas, Corfee-Morlot, Herweijer and Wood2011). Hence, the net effect appears ambiguous. Higher temperatures may further result in higher energy demand, thus boosting revenues for the electricity sector (Deschênes and Greenstone, Reference Deschênes and Greenstone2011). Therefore, the overall effect on industry is ambiguous and further analysis by sub-sector would be required to understand which mechanisms dominate. Lastly, the services sector is expected to be negatively impacted by climate change through reduced worker productivity. However, the dominance of service firms in urban regions, along with the urban heat island effect, which results in higher temperatures in urban regions relative to neighbouring rural areas, may raise energy demand, leading to higher consumer expenditure, thereby also offsetting the reduced productivity (Santamouris et al., Reference Santamouris, Cartalis, Synnefa and Kolokotsa2015). Hence, the net effect of climate shocks on this sector is also unclear. Identifying the precise mechanisms driving impacts by sector involves developing sector-specific models and lies beyond the scope of this paper. Nonetheless, keeping the potential mechanisms in mind provides intuition for the observed differential climatic impacts across various sectors of the economy.

3.2 Data sources

Household-level economic dataFootnote ² are drawn from the NSSO, Ministry of Statistics and Programme Implementation (MOSPI), Government of India's Consumer Expenditure Surveys,Footnote ³ for the years 1987–88, 1999–2000, 2004–05, 2009–10 and 2011–12. The years 1983 and 1993–94 are excluded from the analysis as these rounds do not provide district codes corresponding to the households and hence cannot be mapped to the climate data. The National Sample Survey (NSS) is conducted on a quinquennial basis with a sample size of approximately 100,000 households per survey (or round). The survey is further conducted uniformly throughout the year, in the form of four sub-rounds (July–September; October–December; January–March; and April–June), following a stratified multi-stage sampling design, to ensure statistical representativeness of the population.

These surveys provide data on household-level monthly per capita consumption expenditure (MPCE) as well as expenditure incurred on a range of food and non-food items including energy (fuel and light), durable goods such as clothing, footwear, education, healthcare, recreation, transport, communication, furniture, appliances, rent, taxes and other personal items.Footnote ⁴ Real MPCE is computed by deflating the nominal MPCE values by the Consumer Price Indices for agricultural labourers for the rural sector and industrial workers for the urban sector respectively, using the base year of 1987–88, drawn from the NSS published reports (NSSO, 2010, 2012).

The 55th round (1999–2000) suffers from comparability issues with the remaining rounds due to a distinction in its design (Datta, Reference Datta2006). The survey adopted a 365-day recall period for five non-food items, namely, clothing, footwear, durable goods, educational and medical expenses, instead of the 30-day recall for these items, as in the previous rounds. Consequently, MPCE values were significantly underestimated in 1999–2000 relative to the other rounds. However, the comparability issues with the 55th round are ignored as the regression analysis is not expected to be affected by the underestimation, which is uniform throughout the distribution.

An important issue is the lack of a historical income survey in India, which is hence proxied by household-level consumption expenditure. A number of advantages exist for the use of income vis-à-vis consumption, and vice versa. Since income comprises both consumption and savings, the effects of climate shocks on assets and potential consumption smoothing behaviour by individuals can be more precisely examined. Furthermore, evidence of asset-smoothing behaviour by the poor and consumption-smoothing behaviour by the rich in climatic adversity has also been recorded, which would be easier to identify with data on both consumption and savings or assets (Olsson et al., Reference Olsson, Opondo, Tschakert, Agrawal, Eriksen, Ma, Perch, Zakieldeen, Field, Barros, Dokken and Mach2014). In addition, consumption inequality is found to be lower than income inequality through consumption-smoothing behaviour and hence the use of consumption rather than income may mask many underlying dynamics between various income groups (Attanasio and Pistaferri, Reference Attanasio and Pistaferri2016).

On the other hand, consumption may be a better measure of permanent income, as it would be less influenced by transitory shocks, in line with the permanent income hypothesis. Further, consumption may not equal income, through inter-temporal borrowing and lending among households, and hence may better reflect regular household expenditures. Moreover, wealth effects may be better captured through consumption than income, e.g., the purchase of luxury goods. In addition, consumption smoothing may occur through receipt of government transfers, which would not be captured by income. Finally, it is typically easier to accurately measure consumption for low-income households, rather than income, which may be drawn from multiple sources and economic activities, particularly in developing countries (Attanasio and Pistaferri, Reference Attanasio and Pistaferri2016). Nonetheless, both measures – income and consumption – drawn from household survey data, suffer from under-reporting, particularly at the top of the distribution. Although there is likely to be measurement error in reported consumption, this is not a concern in the regression analysis as the error is expected to be uncorrelated with the explanatory variables, temperature and precipitation. The lack of correlation ensures that the exogeneity assumption is not violated and the regression coefficients are estimated consistently (Pischke, Reference Pischke2007).

Climate data for mean temperature and precipitation over the 1983–2012 periodFootnote ⁵ are drawn from the Climatic Research Unit at the University of East Anglia (UEA) which provide data at the level of $0.5^\circ \times 0.5^\circ$ grids on a monthly frequencyFootnote ⁶ (Harris et al., Reference Harris, Jones, Osborn and Lister2014). The data were aggregated to the district level based on a simple average using R-GIS.Footnote ⁷

3.3 Methodology

The econometric equation to be estimated is:

(2)\begin{equation} lnC_{htr}=\alpha _{d}+\tau _{t}+\beta T_{dtr}+\delta P_{dtr}+\phi X_{htr}+\epsilon _{htr}, \end{equation}

where $lnC_{htr}$ is ln consumption of household $h$ in year $t$ and NSS sub-round $r$; $\alpha _{d}$ is a vector of district dummy variables; $\tau _{t}$ are survey year dummies; $X_{htr}$ is a vector of household-specific controls; and $\epsilon _{htr}$ is a normally-distributed error term. $T_{dtr}$ and $P_{dtr}$ are temperature and precipitation at the district level, expressed as four different measures: (1) absolute values of $T$ and $P$ in district $d$, year $t$ and sub-round $r$; (2) deviations of observed values from their district-sub round-specific long-term means (e.g. $T_{dtr}-\bar {T}_{dr}$); standardized z-scores averaged over the sub-round (e.g., ${(T_{dtr}-\bar {T}_{dr})}/{sd(T_{dtr})}$); and standardized z-scores corresponding to the first month of the NSS sub-round (SR), i.e., July (SR-1), October (SR-2), January (SR-3) and April (SR-4). The key identifying assumption is that within-district variation in temperature and precipitation (relative to historical means in a given year $t$ and sub-round $r$) is exogenous to household consumption expenditure, conditional on local geographic and time varying factors, controlled for by the district and year dummy variables, respectively.

Standard errors are clustered at the district level to account for correlation and heteroskedasticity among households within a district. A limitation of the analysis is that the standard errors are not corrected for spatial and temporal correlation of the climate data, typically done using the Conley (Reference Conley2008) approach, due to the large sample size of the data which proved computationally costly. While this biases the standard errors, the regression coefficients remain unaffected and hence should be asymptotically consistent.

The survey year dummy variables are important not only to capture the trend in mean consumption over the three decades of analysis but also the changing consumption distribution over time, which relies on the assumption of independent, but not identically distributed errors (Wooldridge, Reference Wooldridge2002). Hence, estimation is conducted distinctly for each time period, in which observations are i.i.d. across strata (which ensures asymptotic normality and standard hypothesis testing), with results averaged over time. In addition, the stratification dummies account for cross-sectional correlation among households within a stratum and capture unobserved heterogeneity at the district level such as socioeconomic characteristics as well as the natural geography of the region, which may be correlated. Since there are 580 districts for which the climate and NSS data merge perfectly, including district dummies should not lead to an incidental parameters problem.

Since households are sampled randomly across survey years, the estimation strategy employs pooled cross-sections, rather than household-level panel regressions. Hence, the unobserved heterogeneity is not eliminated as in the case of a fixed effects ‘within’ estimator; instead, we obtain differential intercept terms across strata. It is important to conduct household-level analysis in order to understand the sector-specific mechanisms driving differential climatic impacts. This regression model is thus estimated using OLS, with a panel fixed-effects regression model subsequently employed to corroborate the results of the pooled cross-section regressions. This is done by averaging household consumption values to the district level and matching them to the district-specific climate parameters. Lastly, the quantile regression method is used for the distributional analysis.

The quantile regression estimator is obtained by minimising the following absolute loss function, at the $\tau ^{th}$ quantile:

\[\min_{\alpha \in \mathbb{R},\beta \in \mathbb{R}^{k}}\sum_{i=1}^{n}c_{\tau }(y_{i}-\alpha -x_{i}\beta ), \]

where

\[c_{\tau }(k)=(\tau -1\cdot[k<0])k, \]

the intercept and slope coefficients, $\alpha$ and $\beta$, differ by consumption quantile $\tau$.

Dell et al. (Reference Dell, Jones and Olken2014) discuss how local climatic variables like temperature, soil quality and elevation might be correlated with economic ones such as institutional quality and the level of development. Due to many potential confounders, the cross-sectional approach typically suffers from omitted variable bias, which is circumvented by the panel fixed effects method (Mani et al., Reference Mani, Bandyopadhyay, Chonabayashi, Markandya and Mosier2018). The key identifying assumption in the current analysis is that while district-level economic and climate variables might be correlated with average temperature or rainfall, they would be uncorrelated with exogenous annual fluctuations, or deviations in these climatic parameters from their district-specific long-run mean values. Hence, studying mean deviations and standardized z-scores of climate variables rather than absolute increases in their average values over time is important for identification.

4.1 Descriptive statistics

This section presents summary statistics (table 1) for the climate and economic variables used for analysis for the 1983–2012 and 1988–2012 periods, respectively. The mean temperature in India over the 1983–2012 period was 25${{}^\circ }$C, with a standard deviation of 5.6${{}^\circ }$C, ranging from $-10{{}^\circ }$C in the northern-most, snowy state of Jammu and Kashmir (in 2011–12) to around 34${{}^\circ }$C in the more temperate and monsoonal state of Maharashtra. Mean precipitation was 111 mm over the corresponding period, with a large standard deviation of 126 mm. Many states and union territories in diverse climatic regions witnessed no precipitation on average during January to March. These include the hot, arid regions of Rajasthan and Gujarat, the temperate state of Madhya Pradesh and the equatorial regions of Maharashtra, Andhra Pradesh, Karnataka, Goa, Daman & Diu and Dadra & Nagar Haveli.Footnote ⁸ On the other hand, the state of Meghalaya, in particular the East Khasi Hills district, recorded the highest mean rainfall in the country at 781 mm over 1983–2012.Footnote ⁹

Table 1. Summary statistics

Std. temp. (1mo) and Std. pre. (1mo) refer to the first month of the NSS sub-round.

*The histogram of real MPCE and kernel density plots of real MPCE, temperature, precipitation and the regression residuals are presented in the online appendix.

Household-level real MPCE shows a steady increase from Rs. 195 in 1987–88 to Rs. 312 in 2011–12, on average. It is interesting to note that the standard deviation of consumption exceeds the mean in each year of the survey. This does not imply negative expenditures at the bottom of the distribution but simply clustering of poor households toward zero, thus highlighting the vast disparity in consumption between high- and low-income households, with the large upper tail consumption patterns raising the standard deviation. Consumption averages at different quantiles of the distribution further reflect the underlying inequality. Over 1988–2012, mean consumption of the poorest 25 per cent of individuals rose slowly, stabilising at around half of mean consumption in the country.

Similarly, the median household has always consumed less than the mean household, indicating inequality in the distribution of households, with the mean being driven by consumption levels at the top (as also discussed in Chancel and Piketty, Reference Chancel and Piketty2019). Finally, the rise in consumption levels of the top 1 per cent of households is particularly noteworthy. This subgroup witnessed a doubling of consumption over the period, compared to an increase of 62 per cent for the top 10 per cent of individuals, and only 44 per cent for the bottom 25 per cent of individuals.

The pair-wise correlation matrix (table 2) reveals a statistically significant and negative correlation between ln real MPCE and temperature, and a positive but insignificant correlation between ln real MPCE and precipitation. The negative correlation can be explained by unobserved factors which affect both consumption and local climate such as topography, elevation and institutional quality, which has typically been lower for less developed countries and more tropical climates (Dell et al., Reference Dell, Jones and Olken2014). However, this correlation does not reflect the causal effect of higher temperature and precipitation on consumption, which is now investigated through regression analysis.

Table 2. Correlation matrix for 1988–2012 (excluding 1999–2000)

$^{\ast\ast\ast}p<0.001$.

4.2 Average impacts

The pooled cross-section regression results (table 3) reveal an increase in ln real MPCE of Indian households in response to temperature shocks, by 0.1–1.2 per cent across specifications. A 1 mm increase in precipitation, on the other hand, has no detectable effect on consumption expenditure. All regressions include dummy variables for the survey year, with standard errors being clustered at the district level. The first column reports the effect of higher absolute temperature and precipitation levels on consumption expenditure, excluding district dummy variables. This regression effectively measures the correlation among the three variables, with higher temperatures and lower precipitation being associated with lower consumption expenditures. However, the inclusion of district dummy variables is important to account for unobserved heterogeneity and other omitted variables. Hence, the subsequent estimations include district dummies. The coefficient signs for both temperature and precipitation now reverse, with both higher absolute temperatures and greater variability in temperature raising consumption expenditure. On the other hand, higher precipitation and increased precipitation variability appear to lower consumption expenditure. As discussed earlier, variables that capture climate shocks, i.e., measured as deviations and standardised Z-scores better reflect unexpected fluctuations in local climate and may influence household consumption more than average increases in temperature and precipitation, which may instead be anticipated by individuals based on past trends.

Table 3. Pooled cross-section regression results (1988–2012, excluding 1993–94)

Clustered standard errors in parentheses; ** $p<0.01$, *** $p<0.001.$

Inclusion of year dummy variables.

While the point estimates obtained in the regressions are comparable in absolute value to those found in other studies, typically in the range of 1–3 per cent (reviewed by Dell et al., Reference Dell, Jones and Olken2014), most other studies perform regression analyses using income data, rather than consumption, and typically observe income declines in response to climate shocks. Income would encompass more channels than consumption alone, particularly the role of savings in either consumption- or asset-smoothing behaviour. However, the increase in consumption in response to climate shocks is interpreted as a ‘cost of adaptation’ incurred by households due to unexpected variability in temperature relative to historically observed trends in a particular three-month season within the year. This increase in consumption or the cost of adaptation may be operating through a variety of channels such as rise in energy consumption for cooling effects (as observed by Deschênes and Greenstone, Reference Deschênes and Greenstone2011) rise in food consumption due to heat stress or potential increases in health expenditures.

It is important to note that the increase in consumption is unlikely to reflect a rise in living standards, due to productivity increases. However, this mechanism could be better examined using direct income data with alternative sources, such as the India Human Development Surveys (IHDS) which are available until 2011–12.Footnote ¹⁰

The results from the pooled cross-sectional analysis are now corroborated using panel data fixed effects regressions by averaging the household data to the district level, using the household survey weights provided by the NSS. The fixed effects methodology is chosen over random effects as the district-specific unobserved heterogeneity is expected to be correlated with the explanatory variables, i.e., local climatic variability. This would not satisfy the key assumption of the random effects model, in that the unobserved heterogeneity should be uncorrelated with the regressors. Results are presented in table 4. The first key observation is that all coefficient signs are identical to those from the pooled cross-section regressions, with the temperature variables being statistically significant at the 99 per cent confidence level and precipitation variables now also becoming significant at the 90 per cent confidence level. Second, the coefficients on the temperature variables are larger in magnitude, implying higher effects of climate shocks on mean per capita consumption at the district level. Specifically, a one standard deviation increase in temperature relative to its district-specific historical average leads to a 5 per cent rise in consumption expenditure, whereas an equivalent increase in precipitation reduces consumption by 2.1 per cent. On the other hand, the effects of shocks at the beginning of the sub-round are dampened relative to effects due to the mean shock experienced over the duration of the sub-round, owing to potential adaptation by households.

Table 4. Panel fixed-effects regression results (1988–2012, excluding 1993–94)

Clustered standard errors in parentheses; **$p<0.01$, *** $p<0.001.$

Includes district and year fixed effects.

4.3 Heterogeneous impacts

The pooled cross-section results differentiated by the sector of the household's primary occupation, interacted with the region of residence, are presented in table 5. Heterogeneous effects across economic sectors and geographical areas reveal interesting disparities in climatic impacts across sub-groups of the Indian population. The dummy variables for NIC codes (sectors) and region are interacted with the climate variables to produce differentiated impacts. Agricultural and rural industrial households appear to be adversely affected by both higher temperature and its increased variability, with a one standard deviation increase in temperature causing per capita consumption declines between 1.7 and 8.3 per cent, in the two sectors respectively. On the other hand, per capita consumption levels of households employed in the services sector rise by 2.4 and 9.6 per cent in response to a one standard deviation rise in temperature, in rural and urban areas, respectively. Urban industrial households also experience consumption increases, although the effect is not statistically significant. Moreover, rural households across economic sectors witness declines in consumption expenditure in agriculture and industry, and smaller increases in consumption levels in the service sector, relative to households residing in urban areas, thus highlighting rural-urban inequalities.

Table 5. Heterogeneous effects of climate shocks by a household's primary occupation and region of residence (1988–2012)

Clustered standard errors in parentheses.

* $p<0.1$, *** $p<0.01.$

Inclusion of precipitation controls, year and district dummies.

These results are broadly in line with expectation, although one would perhaps expect consumption declines in the service sector instead of increases, on account of physiological heat stress on workers. However, as discussed in the econometric framework, a rise in energy consumption and potentially other adaptation measures such as increased use of private transport to protect oneself against the deleterious health effects of hot weather may lead to an overall increase in consumption expenditure in this sector. Another interesting and perhaps surprising result is the greater absolute decline in consumption levels of households employed in the rural industrial sector, compared to agriculture. This points to potentially differing mechanisms across sectors. While in agriculture the key channel of reduced consumption would be lower crop yields, in the industrial sector, individuals are expected to suffer greater heat stress and therefore reduced labour productivity, in addition to reduced firm output (Somanathan et al., Reference Somanathan, Somanathan, Sudarshan and Tewari2015). Consequently, their daily wages or monthly income may be lower, in turn leading to reduced consumption capacity.

The results for the services sector appear somewhat counterintuitive as the channels driving them are determined by both supply and demand factors. While labour supply might be lower due to heat stress and hence reduced productivity, the demand for certain kinds of goods might also be lower in hotter weather. For example, in the city of Roorkee within the state of Uttarakhand, rickshaw pullers have in recent years switched from manual rickshaws to battery operated ones even though the latter offer lower profits, simply due to reduced public demand for the former. They further observe a drop in their number of rides in hotter weather, particularly in the summer season, as individuals prefer to remain within their homes at peak hours of the day and on hotter days in the season.Footnote ¹¹ On the other hand, higher demand for climate control technology, such as air conditioners, electric fans and coolers, would raise revenues for service-sector households, which typically operate grocery and retail stores, repair and maintenance shops, shopping complexes and commercial centres. In turn, this may be responsible for their observed increases in consumption expenditures due to climate shocks. The differential rise in consumption levels of 9.6 in urban areas vis-à-vis 2.4 per cent in rural areas is particularly noteworthy and reflects increases in consumption inequality as a result of climate shocks.

The sectoral effects obtained above are now re-examined at various quantiles of the consumption distribution (table 6). Prima facie, impacts of climate shocks on consumption expenditure significantly diverge across the agriculture, industry and service sectors, as well as along the consumption distribution. While the lowest quartile of rural agricultural households experience consumption increases in the order of 0.4–1.4 per cent in response to a one standard deviation increase in temperature, the median household witnesses a 0.8 per cent decline in consumption, whereas those in the top quartile exhibit consumption declines of around 3.9–9.6 per cent.

Table 6. Quantile effects by sector of occupation and region of residence (1988–2012)

Clustered standard errors in parentheses; * $p<0.1$, ** $p<0.05$, *** $p<0.01.$

Inclusion of precipitation controls, year and district dummy variables.

Importantly, the increase in consumption of poor rural agricultural households may be due to higher food prices as a result of reduced crop yields, thus leading to overall higher farm income. Evidence for this is provided in Taraz (Reference Taraz2018a) and Government of India (2018). Alternatively, higher wage incomes from increased supply of casual labour by farmers, such as in MGNREGA, in response to weather shocks which reduce crop yields, may also lead to higher consumption levels. This mechanism is discussed in Taraz (Reference Taraz2018b). Although the rise in food prices temporarily benefits net agricultural producers, the decline in crop yields threatens national food security. At the upper tail, richer agricultural households are typically landowners, who might experience a drop in land values due to heat stress (evidence of which is provided in Jacoby et al., Reference Jacoby, Rabassa and Skoufias2011) and, in turn, reduced household consumption.

Urban agricultural households, on the other hand, exhibit consumption declines across the distribution, ranging from 2.2 to 3.3 per cent. While only 8 per cent of agricultural households in the sample reside in urban areas, the distribution of their economic activities across the NIC 3-digit codes, such as growing of crops, animal production and fishing, appears to be uniform across rural and urban areas, implying that the nature of urban agriculture does not differ significantly. Nevertheless, perhaps access to smaller areas for cultivation as well as limited opportunities for adaptation to higher temperatures, as opposed to those in rural areas, may be responsible for the greater consumption declines in the bottom quartile of urban households.

Analogously, rural industrial households uniformly experience adverse effects of climate shocks through reduced consumption in the order of 2.1–3.8 per cent for the bottom quartile, rising to 8 per cent for the median household and around 14.5–21.0 per cent for the top quartile of households, with substantial consumption losses for the top 1 per cent of households. In addition to heat stress which is expected to lower labour productivity, the contraction of total firm output (Somanathan et al., Reference Somanathan, Somanathan, Sudarshan and Tewari2015), due to both lower labour productivity and heat-induced efficiency losses of capital equipment, would lead to large production and revenue losses across firms. Therefore, it is plausible that workers, firm managers and owners (the latter would form the top 1 per cent of the distribution) suffer financial losses due to lower output and revenue. It is puzzling to find urban industrial households relatively unaffected by climate shocks, and more disaggregated analysis on firm-level data may be required to understand which socioeconomic strata of society comprise the broad category of urban industrial households, and the corresponding mechanisms.

In contrast to agriculture and industry, the majority of the rural and urban service sector households experience consumption increases in the range of 1.0–5.6 per cent and 6.2–10.7 per cent across the respective distributions. Further, at each consumption quantile, the expenditure increases exhibited by urban residents exceed those by rural residents, in proportionate terms. In addition, since the consumption distribution of the services sector lies to the right of that for the agricultural and industrial sectors (i.e., households employed in the former sector have higher incomes or consumption levels),Footnote ¹² it reflects a divergence in sectoral and regional consumption patterns on account of climate shocks. While many agricultural and rural industrial households experience consumption declines, households providing services witness consumption increases in response to unexpected higher temperatures, indicating an increase in consumption-based inequality due to climate shocks.

4.4 The adaptation mechanism

The divergent climatic impacts observed across the agriculture, industry and service sectors are plausibly on account of different mechanisms. Since agriculture and industry are climate-sensitive occupations, households employed in these sectors may experience declines in income and therefore consumption. On the other hand, households employed in the services sector, which is less sensitive to adverse climatic effects, may incur expenditure on adaptation to higher temperatures, with net positive effects on aggregate MPCE.Footnote ¹³ We now explore the adaptation mechanism by disaggregating the household's MPCE into an adaptation component and the remaining expenditure.

The adaptation component includes per capita expenditure on electric fans, coolers, air conditioners, refrigerators and electricity, following the 30-day recall period in the NSS survey. However, it excludes expenditure on transport, tourism, etc. which may only be applicable to a small percentage of the richest households in the sample. The effects of temperature increases, interacted with the sector of activity and region of residence, on both adaptive expenditures and non-adaptive or remaining expenditures, are presented in table 7. The results show statistically significant increases in adaptive expenditures incurred by both rural and urban service sector households, but reveal declines among rural agricultural and industrial households. The climatic effects on the non-adaptive expenditures are negative for rural agricultural and industrial households, who are typically at the lower end of the consumption distribution, but are positive for urban industrial and service sector households.

Table 7. Adaptation effects by sector of occupation in per capita terms (1987–88, 2004–05 and 2011–12)

Clustered standard errors in parentheses; * $p<0.1$, ** $p<0.05$, *** $p<0.01$.

Inclusion of precipitation controls, year and district dummies.

These results closely mirror those in table 5 and highlight that the poorest households, typically rural agricultural and industrial (largely wage labourers), do not spend on climatic adaptation measures but rather experience an overall decline in their consumption expenditures. The adaptation phenomenon, therefore, appears limited to richer households, comprising the urban industrial and service sectors.