1. Introduction
The Indian economy is characterised by a large agricultural sector that employs more than 50 per cent of the Indian workforce (World Bank, 2015). Indian agriculture is also dependent on rainfall, as access to irrigation is limited. Climate change, manifested as an increase in droughts, is detrimental to agriculture. It leads to poor harvests, farmer suicides, protests (Carleton, Reference Carleton2017), increased debt (Kandikuppa and Clark, Reference Kandikuppa and Clark2022) and low mobility for women (Afridi et al., Reference Afridi, Mahajan and Sangwan2022) in India. However, droughts are projected to be more frequent and severe in the future (Bisht et al., Reference Bisht, Sridhar, Mishra, Chatterjee and Raghuwanshi2019). Social scientists have documented various adaptation strategies to climate change, for example, informal risk sharing (Annan and Datta, Reference Annan and Datta2022; Will et al., Reference Will, Groeneveld, Lenel, Frank and Müller2023), migration (Gray and Mueller, Reference Gray and Mueller2012; Bohra-Mishra et al., Reference Bohra-Mishra, Oppenheimer and Hsiang2014; Mueller et al., Reference Mueller, Gray and Kosec2014; Maystadt et al., Reference Maystadt, Mueller and Sebastian2016; Jessoe et al., Reference Jessoe, Manning and Taylor2018) and labour reallocation (Kochar, Reference Kochar1999; Rose, Reference Rose2001; Bandyopadhyay and Koufias, Reference Bandyopadhyay and Koufias2012; Skoufias et al., Reference Skoufias, Bandyopadhyay and Olivieri2017; Emerick, Reference Emerick2018; Noack et al., Reference Noack, Riekhof and Di Falco2019; Branco and Feres, Reference Branco and Feres2021; Colmer, Reference Colmer2021; Liu et al., Reference Liu, Shamdasani and Taraz2023).
At the same time, there remain various barriers to climate change adaptation that could influence weather-induced labour reallocation in developing economies. The prior literature has focused on barriers to credit, insurance, information and transportation, and frictions arising due to possession of assets, including land that restricts allocative efficiency in labour markets (Banerjee and Duflo, Reference Banerjee, Duflo, Aghion and Durlauf2005; Karlan and Zinman, Reference Karlan and Zinman2009; Burgess and Donaldson, Reference Burgess and Donaldson2010; Gollin and Rogerson, Reference Gollin and Rogerson2010; Townsend, Reference Townsend2011; Bianchi and Bobba, Reference Bianchi and Bobba2013; Blattman et al., Reference Blattman, Fiala and Martinez2013; Viswanathan and Kumar, Reference Viswanathan and Kumar2015; Dallmann and Millock, Reference Dallmann and Millock2017; Fernando, Reference Fernando2022; Liu et al., Reference Liu, Shamdasani and Taraz2023; Nonvide, Reference Nonvide2023). In India, there are additional barriers to the sectoral labour movement related to low levels of non-agricultural skills and land ownership. There are more skilled workers in the non-agriculture sector as opposed to the agriculture sector (Herrendorf and Schoellman, Reference Herrendorf and Schoellman2018). This leads to natural barriers for agricultural workers to enter the non-agricultural sector. Moreover, the laws regarding buying and selling of farmland are very restrictive (Jin et al., Reference Jin, Deininger and Nagarajan2007), leading to limited sales and rental markets in India (Skoufias, Reference Skoufias1995; Morris and Pandey, Reference Morris and Pandey2007; Deininger et al., Reference Deininger, Jin and Nagarajan2008). There is also a strong prevalence of patrilineal land inheritance customs (Agarwal, Reference Agarwal1994). Owning and caring for one’s land is considered a mark of identity among rural Hindu households in India (Bhat and Dhruvarajan, Reference Bhat and Dhruvarajan2001; Jodhka, Reference Jodhka2006; Sharma, Reference Sharma2007). These barriers matter, as Fernando (Reference Fernando2022) finds that individuals who inherit land are significantly less likely to enter non-agricultural work in India.
Much of the empirical literature on climate change-induced labour reallocation in India has focused on district- or state-level analysis (Kochar, Reference Kochar1999; Rose, Reference Rose2001; Bandyopadhyay and Koufias, Reference Bandyopadhyay and Koufias2012; Emerick et al., Reference Emerick, de Janry, Sadoulet and Dar2016; Skoufias et al., Reference Skoufias, Bandyopadhyay and Olivieri2017). However, household attributes play a key role in occupation choices. For example, high capital investment in agriculture could deter movement out of the agriculture sector (Ahituv and Kimhi, Reference Ahituv and Kimhi2002; Matshe and Young, Reference Matshe and Young2004). Furthermore, rural households that own farmland may have incentives to stay back due to poorly defined property rights, as well as cultural norms around land ownership (Agarwal, Reference Agarwal1994; Jin et al., Reference Jin, Deininger and Nagarajan2007; Morris and Pandey, Reference Morris and Pandey2007; Fernando, Reference Fernando2022). Another key determinant of household occupation choices is skill transferability. Households primarily engaging in agriculture may not necessarily have the skill set to work in the non-agriculture sector. The literature has used repeated cross-sectional data and/or focused on a sub-national level study. Therefore, it is hard to follow how household-level attributes may influence their occupation choices over time. From a policy perspective, it is important to understand the distributional effects of climate change between different types of households.
To address these issues, we merge two rounds of the India Human Development Survey, a nationally representative household survey, with an earlier household survey of the Human Development Profile of India, allowing us to create a long household panel spanning over 20 years. We merge the household panel dataset with the weather dataset from the University of Delaware climate archives to investigate how drought, defined as rainfall deficiency, affects household occupation choices and how household attributes may mediate the impact of drought. In particular, we focus on how switching costs and skill transferability mediate the impact of drought on household occupation choices. Using a household and year fixed effects regression model, we find that there is a significant 5 per cent reduction in total household employment following a drought that is largely driven by the reduction in employment in the agricultural sector. Household employment in agriculture falls by 6 per cent, whereas employment in the non-agriculture sector falls by 1.5 per cent. Additionally, we find that households which own land tend to stick to agriculture following a drought and households where the household head possesses non-agriculture sector skills, proxied by education level, tend to leave agriculture at a higher rate.
This paper advances knowledge on the role of barriers to the sectoral labour movement in three ways. First, it provides evidence with new household-level panel data and complements earlier work on labour market frictions related to land ownership in rural India. Second, this is the first study to investigate how the weather impacts of sectoral mobility are augmented by the frictions related to land ownership. Lastly, we quantify the effect of non-agricultural skills among household members in mediating the effect of weather on labour reallocation. These findings underscore the importance and urgency of addressing human capital investment and land market reforms in India.
2. Data and methods
2.1. Data
Two sources of data are combined to examine the effect of drought, defined as rainfall deficiency, on household occupation choices. The first is a household-level dataset from the India Human Development Survey (IHDS). The second is a series of gridded temperature and precipitation datasets from Matsuura and Willmott (Reference Matsuura and Willmott2018). The IHDS is a nationally representative multi-topic household-level panel dataset resulting from two survey rounds conducted in 2004–05 and 2011–12 in 33 states and 372 districts across mainland India. The 2004–05 survey has a sample of 41,554 households with 83 per cent of the households being retained in 2011–12 (Desai and Vanneman, Reference Desai and Vanneman2015). Part of the IHDS sample is linked to data from an earlier survey, known as the Human Development Profile of India (HDPI), consisting of 33,230 rural households living in 16 states and 184 districts. We merge the two rounds of the IHDS datasets with the HDPI dataset to construct a long panel of household data for approximately 20 years from 1993 to 2012. This reduces the sample size to 8,082 rural households that were interviewed during all three rounds. These households remained in the same house for almost 20 years.Footnote 1 We focus on the households living in rural areas because agriculture is primarily concentrated in the rural parts of the country. Households that did not have any adult members (15–64 years old) were dropped from the final sample. We extract household-level data on socioeconomic characteristics, including occupation of each household member, their education levels, age, gender, income sources and levels, land holdings and access to irrigation for three waves of the data: 1994–95, 2004–05 and 2011–12.
The information on temperature and precipitation is obtained from 1900–2017 Gridded Monthly Time Series (V 5.01) data archives. Precipitation (and temperature) is measured in mm (in 
$^{\circ}$C) for each month from 1900 to 2017 for every 0.5-degree by 0.5-degree latitude/longitude grid node. We compute district-level monthly rainfall as the average precipitation levels of each 0.5 degrees by 0.5 degrees latitude/longitude grid cell that falls within the boundaries of the district for each month over the last 30 years. We convert these district-level precipitation measures into z-scores for the monsoon months of June, July, August and September during which India receives 90 per cent of the rainfall (Dimri et al., Reference Dimri, Yasunari, Kotlia, Mohanty and Sikka2016). The monsoon rainfall z-score is the average rainfall in the monsoon season of the current year, demeaned by average rainfall in the monsoon season over the last 30 years and divided by the standard deviation of the monsoon rainfall over the last 30 years. The districts with positive z-scores are re-coded as zero (i.e., no drought), and districts with negative z-scores are re-coded as the absolute value of the z-score (i.e., below average rainfall). Therefore, a higher z-score implies greater rainfall deficiency and a more severe drought (Mueller et al., Reference Mueller, Gray and Kosec2014). In the rest of the paper, ‘rainfall deficiency’ and ‘drought’ are used interchangeably, referring to the recoded monsoon rainfall z-score as defined above. To disentangle the effect of precipitation from temperature changes in a district, we control for minimum and maximum temperature in a district over the monsoon term. Instead of average monthly temperature, we include minimum and maximum monthly temperature as the variation in minimum and the maximum temperature have opposite effects on crop yields, particularly rice yields in tropical countries like China and India (Welch et al., Reference Welch, Vincent, Auffhammer, Moya, Dobermann and Dawe2010).
2.2. Measures of occupation choices
We use multiple measures of occupation choices at the household level. In order to understand how households diversify in response to drought, we use information on the occupation choices of all household members. The information on occupation categories slightly differs across the three waves of data. In wave-1, agricultural jobs included cultivation, allied agricultural activities, agricultural wage workers and cattle tending. Non-agricultural jobs included non-agricultural wage workers, artisan/independent work, petty shop/other small business, organised business/trade, salaried employment/pension, qualified profession/not classified and domestic servants. In subsequent waves, the categorisation is coarser. Agricultural jobs included farm workers, agricultural wage workers and animal tending; and non-agricultural jobs include non-agricultural wage workers, salaried workers and business people. Household (HH) members may work in more than one job. Incorporating these caveats, we calculate the following variables.
(1) The total number of HH jobs is obtained by summing the jobs of every individual member.
(2) The number of agricultural jobs is obtained by summing the jobs of each individual member in agriculture and related sectors as defined above.
(3) The number of non-agricultural jobs is obtained by summing the jobs of every individual member in the non-agriculture sectors as defined above.
(4) The fraction of agricultural jobs is defined as the number of agricultural jobs within the household divided by the total number of household jobs, expressed in percentage terms.
2.3. Summary statistics
The household dataset is merged with the gridded weather dataset to form the study sample. Figure 1 shows the geographical distribution of 8,082 households in 184 districts and 18 states in India. Table 1 summarises the key variables in the study. The number of household jobs varies between two to four jobs with at least half of the jobs in the agriculture sector. Incorporating all households’ jobs across all members, the fraction of agricultural jobs varies between 0.65 and 0.73, implying that rural households are primarily employed in the weather-dependent agriculture sector. Household agriculture and non-agriculture income, expressed in 2012 rupees, increased between wave-2 and wave-3. The average agricultural income in a rural household is much larger than the average non-agricultural income.
Figure 1. Study sample.
Table 1. Summary statistics
The number of household members has decreased from 6 members to 4 members, reflected by the decrease in the number of adults (4 to 3 members) and children (2 to 1 member) because of household splits and migration. At the same time, ageing over the last twenty years contributed to an increase in the number of seniors and a decrease in the number of children in the household. Education indicators for the household head show that wave-1 households had more educated heads. In wave-1, the head of the household had secondary education in about 11 per cent of households, which fell to 2 per cent of households by wave-3. Ageing (consequently death) and splitting of households could be the reasons for falling education levels of the household heads. Approximately 65–70 per cent of the households own farmland. However, irrigation facilities are limited. Across the three waves, about 35–40 per cent of households availed of irrigation.
The last panel of the table summarises the weather variables. The monsoon rainfall deficiency measure decreases in later waves. This implies that more households experienced drought or that the drought that households faced in wave-1 was severe compared to the next two waves or both. The minimum and maximum monthly temperatures vary between 21
$^{\circ}$C and 31
$^{\circ}$C across the three waves.
2.4. Analysis
The baseline empirical model is a reduced-form regression of annual household allocation of agricultural jobs expressed as a percentage of agricultural jobs in household i located in district d in year t (
$Y_{idt}$):
\begin{eqnarray}
Y_{idt} = \beta_{0} + \beta_{1}D_{dt-1} + \mathbf{T_{dt-1}}\beta_{2} + \mathbf{X_{idt}}\beta_{3} + \alpha_{i} + \delta_{t} + \epsilon_{idt}.
\end{eqnarray} Regression (1) includes a vector of time-varying household characteristics 
$\mathbf{X_{idt}}$ (number of household members, number of adult female members, number of adult male members (between the age of 15 to 64 years), two indicators for household head’s education level: secondary and above, college and above) for each household i located in district d in year t. This captures the altering demographics of the household composition over almost twenty years. 
$D_{dt-1}$ is the lagged drought variable which varies by district d and year t-1. Higher rainfall deficiency indicates a more severe drought. The spatial and temporal variations of rainfall deficiencies are key to the identification of the effect of drought on household labour allocation. We also include additional temperature-related variables defined in Table 1, 
$\mathbf{T_{dt-1}}$. Household (
$\alpha_{i}$) and year fixed effects (
$\delta_{t}$) capture the effect of unobservable household characteristics and time effects that may influence job allocation within a household. We cluster the standard errors at a geographical level for which drought is defined (Wooldridge, Reference Wooldridge2003; Bertrand et al., Reference Bertrand, Duflo and Mullainathan2004).
In equation (1), the effect of drought on household labour allocation is identified by spatial and temporal variation in the timing of drought experienced by individual households. The parameter of interest, 
$\beta_{1}$, is identified by the within-household variation of the timing of last year’s drought shock, conditional on weather and household covariates.
2.5. Mechanisms
The mechanisms underlying ex-post household labour allocation to short-run weather fluctuations are an important aspect of adaptation to longer-term climate change. Changes in the composition of jobs across household members could be driven by a couple of factors, including household-level attributes related to switching costs and skill transferability. We explore the relative merits of each mechanism using the following reduced-form regression specification:
\begin{eqnarray}
Y_{idt} = \beta_{0} + \beta_{1}D_{dt-1} + \beta_{2}H_{id1}D_{dt-1} + \mathbf{T_{dt-1}}\beta_{3} +\mathbf{X_{idt}}\beta_{4} + \alpha_{i} + \delta_{t}+ \epsilon_{idt}.
\end{eqnarray} Here, we introduce the variable 
$H_{id1}$ interacted with the lagged drought variable. The baseline variation in any household factor across rural households (that influences labour reallocation) identifies the heterogeneous treatment effect. 
$H_{id1}$ is the variable denoting heterogeneity in household i located in district d at baseline (1993–94). This variable captures one of the factors driving labour reallocation within the household. We describe each factor and the underlying mechanism in the following subsections. Observe that there is no un-interacted 
$H_{id1}$ term because it is time-invariant and is implicit in the household fixed effect, 
$\alpha_{i}$.
The parameter of interest is 
$\beta_{2}$, which captures the additional influence of household characteristics on the effect of drought on household labour allocation. 
$\beta_{1}+ \beta_{2}$ measures the effect of drought on the fraction of agricultural jobs for households characterised by the presence of the factor compared to those that are not. We test whether the term 
$\beta_{1}+ \beta_{2}$ is significantly different from zero using an F-statistic.
3. Results
3.1. Effect of drought on household occupation choices
Table 2 presents the results of the primary specification examining the effect of drought on household jobs. In column (1), we find a reduction in the share of agricultural jobs among households experiencing rainfall deficiency. In column (2), we find that there is a significant reduction in the total number of household employment by 5 per cent on average with one unit increase in rainfall deficiency. This average effect can be disentangled into employment in agriculture and related sectors and non-agriculture sectors. We find that employment in agriculture and related sectors decreases significantly by 6 per cent on average (in column (3)) and employment in non-agriculture decreases by 1.5 per cent (in column (4)), although the effect is insignificant. These results highlight that there is an overall reduction in household employment, driven by a reduction in the rainfall-dependent agriculture sector.Footnote 2
Table 2. Effect of drought on household jobs
Notes: the sample consists of three waves of data. The dependent variable in the first column is the fraction of household agricultural jobs, in the second column is the total number of household jobs, in the third column is the total number of agriculture jobs and in the last column is the total number of non-agriculture jobs. The independent variable, ‘Rainfall deficiency’, measures district-level monsoon rainfall deviation below historical normal using a z-score. The z-score is recoded such that zero means normal rainfall and higher positive value means greater negative deviation from average rainfall and more severe drought. The controls in each of these regressions include household-level characteristics such as number of household members, number of adult female members, number of adult male members, indicators for education level of the head and weather variables such as the minimum and maximum temperature over the monsoon term in the respective districts. The standard errors are clustered at the district level and are reported in parentheses.
These results align with previous findings in the literature. For example, weather-driven reductions in agricultural labour demand cause people to move to the manufacturing sector (Colmer, Reference Colmer2021). More evidence shows that idiosyncratic shocks affecting agriculture drive workers out of agriculture (Kochar, Reference Kochar1999; Rose, Reference Rose2001; Emerick, Reference Emerick2018). The findings in this paper are closely related to Skoufias et al. (Reference Skoufias, Bandyopadhyay and Olivieri2017). An exception is how occupational diversification is defined. Skoufias et al. (Reference Skoufias, Bandyopadhyay and Olivieri2017) defines occupational diversification for each household member as the probability that non-head member i in household j in district d has the same occupation or employment characteristics as the household head. Their sample is restricted to households whose head works in agriculture. We do not restrict the head to being employed in agriculture. We observe how the proportion of agricultural jobs within a household is changing ex-post a drought. This could potentially imply that we are capturing a more diverse population of rural households, not restricted based on the household head’s occupation.Footnote 3
Another recent work by Liu et al. (Reference Liu, Shamdasani and Taraz2023) shows that the effect of changes in decadal average temperature over the growing season increases agricultural labour and reduces non-agricultural labour in Indian districts. Several key distinctions help reconcile the contrasting results. Firstly, this paper focuses on longer-term weather patterns and structural changes in the labour market as opposed to short-term weather fluctuations and labour reallocation. Kochar (Reference Kochar1999) and Rose (Reference Rose2001) find that income shocks from bad weather can induce short-term labour reallocation away from agriculture, similar to the results in this paper. Secondly, the effect of extreme heat and rainfall deficiency on the production of crops could be different, thereby leading to different labour requirements. Lastly, the availability and effectiveness of adaptation technologies could impact productivity of workers (Emerick et al., Reference Emerick, de Janry, Sadoulet and Dar2016) and thereby impact sectoral labour reallocation differently.
3.2. Role of switching cost and skill transferability
This subsection explores the differential impacts of drought on household employment choices based on household switching costs and skill transferability. Investment in agriculture by rural households often hinders their movement out of agriculture. Households own farmland and farm equipment such as tubewells, electric pumps, diesel pumps, bullock carts, tractors, threshers and biogas plants. We use baseline land ownership status as a proxy for investment in agriculture. The first four columns of Table 3 show how high switching cost, proxied by land ownership, influences the effect of drought on household employment. We find that households that own land increase their share of agriculture jobs with a one-unit increase in monsoon rainfall deficiency. Even though there is an overall reduction in total household employment, this reduction is driven by a higher reduction in non-agriculture employment and relatively lower reduction in agriculture employment for landowning households. Households that do not own farmland lower their share of agriculture jobs, they are more mobile and increase their non-agriculture employment. In addition to high investment in agriculture, there are cultural obligations to retain land, coupled with land market transaction costs that influence occupational choices in rural India (Fernando, Reference Fernando2022). This is also common in other developing countries such as Benin (Nonvide, Reference Nonvide2023) and Ethiopia (Kosec et al., Reference Kosec, Ghebru, Holtemeyer, Mueller and Schmidt2018), where occupation-related barriers pertaining to land tenure and inheritance exist.
Table 3. Effect of drought on household jobs, heterogeneity by land ownership
Notes: the sample is for three waves of data. The dependent variable in the first column is the fraction of household agricultural jobs, in the second column is the total number of household jobs, in the third column is the total number of agriculture jobs and in the last column is the total number of non-agriculture jobs. The ‘Land ownership’ variable takes value 1 if the household owns farmland at baseline, zero otherwise. The independent variable, ‘Rainfall Deficiency’, measures district-level monsoon rainfall deviation below historical normal using a z-score. The z-score is recoded such that zero means normal rainfall and higher positive value means greater negative deviation from average rainfall and more severe drought. The controls in each of these regressions include household-level characteristics such as number of household members, number of adult female members, number of adult male members, indicators for education level of the head and weather variables such as the minimum and maximum temperature over the monsoon term in the respective districts. The standard errors are clustered at the district level and are reported in parentheses.
Skill transferability between agriculture sector jobs and non-agriculture sector jobs could be another important driver of household employment. Rural households primarily engaging in agriculture may face obstacles in finding a non-agricultural job because of the lack of the required skill set. To measure the effect of skill on the ease of labour reallocation from agricultural to non-agricultural sectors, we use education as a proxy for the skill level of household members. We use the median education level of workers in the non-agricultural sector at baseline to proxy for the skill required to work in that sector. The last four columns of Table 4 show the results of the differential impact of drought on household employment based on the non-agriculture skill level of the household head. Households where the head has non-agriculture skills tend to move out of agriculture at a higher rate compared to households where the head does not have such skills.Footnote 4
Table 4. Effect of drought on household jobs, heterogeneity by non-ag skill of HH head
Notes: the sample is for three waves of data. The dependent variable in the first column is the fraction of household agricultural jobs, in the second column is the total number of household jobs, in the third column is the total number of agriculture jobs and in the last column is the total number of non-agriculture jobs. The ‘Non-ag skill’ variable takes value 1 if the highest education level of the household head is higher than the average education level of the workers in the non-agriculture sector at baseline, zero otherwise. The independent variable, ‘Rainfall deficiency’, measures district-level monsoon rainfall deviation below historical normal using a z-score. The z-score is recoded such that zero means normal rainfall and higher positive value means greater negative deviation from average rainfall and more severe drought. The controls in each of these regressions include household-level characteristics such as number of household members, number of adult female members, number of adult male members, indicators for education level of the head and weather variables such as the minimum and maximum temperature over the monsoon term in the respective districts. The standard errors are clustered at the district level and are reported in parentheses.
4. Robustness checks
There are some potential threats to the above identification. This section explores some of the threats, performs robustness checks and discusses their implications. The first threat is that of an omitted variable bias. Some time-varying factors at the household level that are correlated with the drought shock and contribute to the reallocation of jobs could be omitted. These include changes in household-specific and location-specific characteristics that could mask the sign and magnitude of the main coefficient. One characteristic of the household that can impact household occupation choices is the composition of the household. Location-specific characteristics such as access to irrigation facilities, availability of government programs and sector-specific labour demand could also impact household occupation decisions.
Table 5 shows the results of the first set of robustness checks related to location-specific characteristics. In column (1), we find that there is an increased reduction in agricultural employment in response to rainfall deficiency for households that do not have access to irrigation. This highlights the importance of monsoon rainfall for agriculture in India. In column (2), apart from the household and year FEs, we incorporate year times region FE to account for diverse labour market conditions that can evolve differently. We still find that monsoon rainfall deficiency significantly affects household employment. Lastly, in column (3), we test whether the rolling out of a nationwide rural employment scheme, MGNREGA,Footnote 5 between wave-2 and wave-3 influenced the impact of drought on household employment. We introduce two additional interactions with the monsoon rainfall deficiency in specification (1). The two interaction terms are ‘Rainfall deficiency 
$ \times $ Wave-2’ and ‘Rainfall deficiency 
$ \times $ Wave-3’. The variables ‘Wave-2’ and ‘Wave-3’ take value 1 if the observation is for wave-2 and wave-3, respectively, zero otherwise. Therefore, the interaction term allows us to understand if there is any differential effect of drought on the percentage of agricultural jobs based on the survey year. The variable ‘Wave 3’ includes any change that could alter the effect of drought on labour reallocation. This also includes the availability of MGNREGA. We find that the effect of drought does not have a differential effect for the households in wave-3. This could plausibly imply that the exposure to MGNREGA did not affect labour allocation decisions significantly in response to drought.
Table 5. Robustness check: effect of drought on household jobs
Notes: the sample is for three waves of data of household data. The dependent variable is the fraction of household agricultural jobs. The independent variable, ‘Rainfall deficiency’, measures district-level monsoon rainfall deviation below historical normal using a z-score. The z-score is recoded such that zero means normal rainfall and higher positive value means greater negative deviation from average rainfall and more severe drought. Each regression in each column is a robustness check. The controls in each of these regressions include household-level characteristics such as number of household members, number of adult female members, number of adult male members, indicators for education level of the head and weather variables such as the minimum and maximum temperature over the monsoon term in their respective districts. In column (1), the sample includes households without access to irrigation, in column (2), the regression additionally includes region times year FE to account for diverse labour market conditions. The regions are divided into north and south and the last column shows additional effect of an employment scheme that was rolled out between wave-2 and wave-3. The variable wave-2 takes value one for survey wave-2 and the variable wave-3 takes value one for survey wave-3. Standard errors are clustered at the district level and reported in parentheses.
Another consideration is how drought affects different subgroups within a household. Table 6 shows the results for the effect of drought on employment choices of household members and working-age adults, which is further divided by gender. In column (1), the results show that a one-unit increase in the rainfall deficiency leads to a 3.1 percentage point reduction in the percentage of household members employed in agriculture. In column (2) (column (3)), the results show that an increase in rainfall deficiency by one unit leads to a reduction in the percentage of males (females) in agricultural jobs by 8 per cent (4 per cent). The last three columns, columns (4)–(6), show how household employment among the working-age members responds to rainfall deficiency. In column (4), we find that a one-unit increase in the rainfall deficiency leads to a 3.6 percentage point reduction in the percentage of working members employed in agriculture. In column (5) (column (6)), the result shows that a one-unit increase in rainfall deficiency leads to a 11.7 per cent (2.9 per cent) reduction in the percentage of female (male) working-age members in agricultural jobs. This shows that females are more likely to move out of agricultural jobs when the household experiences drought. This could mean that the non-agricultural jobs available are traditionally women-oriented. This includes manufacturing jobs in the textile industry, where 77.4 per cent of the total workforce are women (Shazli and Munir, Reference Shazli and Munir2014).
Table 6. Robustness check: effect of drought on household occupations across gender and age
Notes: the sample is for three waves of data. The dependent variable in the first column is the fraction of household members having agricultural jobs, in the second column, it is the fraction of female household members having agricultural jobs, in the third column, it is fraction of male household members having agricultural jobs, in the fourth column, it is the fraction of 15 to 64 year old household members having agricultural jobs, in the fifth column, it is fraction of female 15-64 year old household members having agricultural jobs and in the last column, it is the fraction of male 15-64 year old household members having agricultural jobs. The independent variable, ‘Rainfall deficiency’, measures district-level monsoon rainfall deviation below historical normal using a z-score. The z-score is recoded such that zero means normal rainfall and higher positive value means greater negative deviation from average rainfall and more severe drought. The controls in each of these regressions include household-level characteristics such as number of household members, number of adult female members, number of adult male members, indicators for education level of the head and weather variables such as the minimum and maximum temperature over the monsoon term in their respective districts. Standard errors are clustered at the district level and reported in parentheses.
A second threat to identification is attrition bias. Attrition in the survey occurs in the form of migration, death, or splitting of households. The IHDS survey does not track a household (or a part of a household) if it moves out of the neighbourhood (primary survey unit). However, it does track households that move out of the original household but remain in the same neighbourhood. They are assigned a split identifier. The potential reason for a household missing in the following wave is either migration or death. The composition of the neighbourhood is changed because of households moving in and out of different labour markets that could potentially be related to drought occurrence.
Differential migration across households based on their exposure to drought would have caused concern due to the selection problem. We compute the number of missing household members in each household in the sample between wave-1 and wave-2, and wave-2 and wave-3. We divide the number of missing members in each household by the total number of household members present in the last wave and express that variable in percentage terms. We use specification (1) to test whether drought significantly affects the percentage of missing members within a household. In column (1) of Table 7, we find that there is no significant effect of drought on the percentage of missing household members.
Table 7. Robustness check: attrition bias
Notes: the sample in column (1) is for three waves of data. The dependent variable in the first column is the fraction of missing members between two consecutive household panel data. The sample in column (2) is the three waves of data where household splits are consolidated into a single household. The dependent variable in column (2) is the fraction of household agricultural jobs. The independent variable, ‘Rainfall Deficiency’, measures district-level monsoon rainfall deviation below historical normal using a z-score. The z-score is recoded such that zero means normal rainfall and higher positive value means greater negative deviation from average rainfall and more severe drought. The controls in each of these regressions include household-level characteristics such as number of household members, number of adult female members, number of adult male members, indicators for education level of the head and weather variables such as the minimum and maximum temperature over the monsoon term in their respective districts. Standard errors are clustered at the district level and reported in parentheses.
Household splits are common. The younger members grow up and move to a different house. The IHDS survey was designed in such a way that the households that split but stayed in the same neighbourhood (primary survey unit) were re-interviewed in the next round of the survey. The split households were assigned an identifier that helped to link the split households to the original households. In the main analysis, we use the sample of households that remained in the same physical house in the three waves of the survey. However, previous studies have documented that household splits are non-random and could affect the results significantly (Foster and Rosenzweig, Reference Foster and Rosenzweig2002; Thomas et al., Reference Thomas, Witoelar, Frankenberg, Sikoki, Strauss, Sumantri and Suriastini2012). To understand how the diversification of labour would look if the households had not split, we construct an additional sample, namely ‘imagine no splits’ households sample. We consider all households that split in wave-2 and wave-3 from the original households in wave-1 to be part of the original household.Footnote 6 In Table 7, we find that the results are similar to the results of the main analysis. A one-unit increase in the monsoon rainfall deficiency significantly decreases the household agricultural employment share by 2 percentage points.
5. Conclusion
India’s economy, with its heavy reliance on rain-fed agriculture, remains deeply vulnerable to climate change, making adaptation to climate change a priority for rural households. However, various barriers to climate change adaptation still exist. The goal of this paper is to assess and quantify the effects of certain barriers that households face in adapting to climate change through labour reallocation. A detailed household panel dataset is merged with a high-resolution weather dataset to estimate the effect of drought on household occupation choices using a household and year fixed effects model. Two barriers to climate change, namely skill-transferability and switching costs, are investigated.
The results show that droughts, defined by rainfall deficiency, reduce household employment, which is driven by the decrease in the share of agricultural jobs. Additionally, we find that households with a head who has more than a primary education are more likely to move to the non-agriculture sector in response to drought. Human capital development is an essential component of climate change adaptation. Cultural norms associated with owning farmland, as well as lengthy bureaucratic processes related to buying and selling farmland in India, have restricted land market transactions. Landowning households respond to drought by increasing their share of agricultural jobs, reinforcing the frictions associated with owning farmland.
As of 2020, over 60 per cent of India’s population resides in rural areas, where communities are particularly vulnerable to droughts. To strengthen resilience and support structural transformation, rural households require more robust institutional frameworks, particularly in areas of land market reform and human capital development. Reforms such as digitising land titles, offering title guarantees and developing well-functioning land rental markets could help alleviate rigidities in India’s land markets and reduce barriers to occupational and spatial mobility. In parallel, greater investments in primary and secondary education, as well as on-the-job training, are essential to enhance human capital and facilitate labour reallocation. Together, these measures could play a vital role in advancing climate change adaptation across rural India.
Acknowledgements
The author is grateful for insights and suggestions from Kelly Bishop, Berthold Herrendorf, Sarah Jacobson, Nick Kuminoff, Valerie Mueller and Alvin Murphy, and also thanks the seminar audiences at AAEA, AERE, Arizona State University, OSWEET and WEAI for their helpful comments and discussions.
Competing interests
The author declares none.
Appendix
Table A.1. Effect of drought on household jobs (drought defined as rainfall below the 20th percentile of historical rainfall)
Notes: the sample is for three waves of data. The dependent variable in the first column is the fraction of household agricultural jobs, in the second column is the total number of household jobs, in the third column is the total number of agriculture jobs and in the last column is the total number of non-agriculture jobs. The key independent variable, drought, is a binary variable which takes value 1 for districts that experience monsoon rainfall below the 20th percentile of the historic monsoon rainfall, zero otherwise. Note that districts experiencing rainfall above the 80th percentile are excluded from the sample. The controls in each of these regressions include household-level characteristics such as number of household members, number of adult female members, number of adult male members, indicators for education level of the head and weather variables such as the minimum and maximum temperature over the monsoon term in their respective districts. The standard errors are clustered at the district level and are reported in parentheses.
Table A.2. Effect of drought on household jobs (using the last two rounds of the household survey)
Notes: the sample uses the two most recent waves of data. The dependent variable in the first column is the fraction of household agricultural jobs, in the second column is the total number of household jobs, in the third column is the total number of agriculture jobs and in the last column is the total number of non-agriculture jobs. The independent variable, ‘Rainfall Deficiency’, measures district-level rainfall deviation below historical normal using a z-score. The z-score is recoded such that zero means normal rainfall and higher positive value means greater negative deviation from average rainfall implying a more severe drought. The controls in each of these regressions include household-level characteristics such as number of household members, number of adult female members, number of adult male members, indicators for education level of the head and weather variables such as the minimum and maximum temperature over the monsoon term in their respective districts. The standard errors are clustered at the district level and are reported in parentheses.
Table A.3. Effect of drought on household jobs, heterogeneity by non-ag skill of working age members
Notes: the sample is for three waves of data. The dependent variable in the first column is the fraction of household agricultural jobs, in the second column is the total number of household jobs, in the third column is the total number of agriculture jobs and in the last column is the total number of non-agriculture jobs. The ‘Non-ag skill’ variable takes value 1 if the highest education among working-age members is higher than the average education level of the workers in the non-agriculture sector at baseline, zero otherwise. The independent variable, ‘Rainfall deficiency’, measures district-level rainfall deviation below historical normal using a z-score. The z-score is recoded such that zero means normal rainfall and higher positive value means greater negative deviation from average rainfall and more severe drought. The controls in each of these regressions include household-level characteristics such as number of household members, number of adult female members, number of adult male members, indicators for education level of the head and weather variables such as the minimum and maximum temperature over the monsoon term in their respective districts. The standard errors are clustered at the district level and are reported in parentheses.
Table A.4. Effect of drought on household jobs (drought defined as rainfall below 1 SD from the historical average)
Notes: the sample is for three waves of data. The dependent variable in the first column is the fraction of household agricultural jobs, in the second column is the total number of household jobs, in the third column is the total number of agriculture jobs and in the last column is the total number of non-agriculture jobs. The independent variable, ‘Rainfall deficiency’, measures district-level rainfall deviation below historical normal using a z-score. The z-score is recoded such that zero means normal rainfall and higher positive value above 1 SD means greater negative deviation from average rainfall and more severe drought. The controls in each of these regressions include household-level characteristics such as number of household members, number of adult female members, number of adult male members, indicators for education level of the head and weather variables such as the minimum and maximum temperature over the monsoon term in their respective districts. The standard errors are clustered at the district level and are reported in parentheses.
Table A.5. Effect of drought on household jobs (drought defined as rainfall below 1.25 SD from the historical average)
Notes: the sample is for three waves of data. The dependent variable in the first column is the fraction of household agricultural jobs, in the second column is the total number of household jobs, in the third column is the total number of agriculture jobs and in the last column is the total number of non-agriculture jobs. The independent variable, ‘Rainfall deficiency’, measures district-level rainfall deviation below historical normal using a z-score. The z-score is recoded such that zero means normal rainfall and higher positive value above 1.25 SD means greater negative deviation from average rainfall and more severe drought. The controls in each of these regressions include household-level characteristics such as number of household members, number of adult female members, number of adult male members, indicators for education level of the head and weather variables such as the minimum and maximum temperature over the monsoon term in their respective districts. The standard errors are clustered at the district level and are reported in parentheses.
Table A.6. Effect of drought on household jobs (drought defined as rainfall below 1.5 SD from the historical average)
Notes: the sample is for three waves of data. The dependent variable in the first column is the fraction of household agricultural jobs, in the second column is the total number of household jobs, in the third column is the total number of agriculture jobs and in the last column is the total number of non-agriculture jobs. The independent variable, ‘Rainfall deficiency’, measures district-level rainfall deviation below historical normal using a z-score. The z-score is recoded such that zero means normal rainfall and higher positive value above 1.5 SD means greater negative deviation from average rainfall and more severe drought. The controls in each of these regressions include household-level characteristics such as number of household members, number of adult female members, number of adult male members, indicators for education level of the head and weather variables such as the minimum and maximum temperature over the monsoon term in their respective districts. The standard errors are clustered at the district level and are reported in parentheses.
Open access
