Does Reliable Electricity Mean Lesser Agricultural Labor Wages? Evidence from Indian Villages

At the time of India’s independence from the British Empire, in 1947, a majority of India’s population was neither educated nor literate, with its rural population particularly worse off. In the absence of industries, rural India was overwhelmingly agrarian. In 1951, agriculture and allied sectors contributed to 51.9% of India’s GDP (Wagh and Dongre,, 2016). While that number has since fallen sharply to 17% in 2014-15, a majority of Indians continue to remain employed in agriculture. Therefore, it comes as little surprise that Government efforts in the second half of the 20th century toward electrifying villages were focused on connecting farms, rather than households.

Investments in bringing electric pumps to villages were further accelerated by the famine in Bihar in 1966-1976, and the drought in Maharashtra in 1970-1973 (Dubhashi,, 1992). In 1969, the Ministry of Power established the Rural Electrification Corporation to oversee village electrification, with the specific purpose of aiding State Electricity Boards in facilitating the adoption of electric groundwater pumps in villages to increase agricultural productivity. The droughts and famines of the 1960s and 1970s, ultimately along with the Green Revolution in the 1970s, led to an expeditious increase in the rate of village electrification in India, as can be seen in Figure 1. Despite the growth in the number of villages connected, power consumption by the agricultural sector remained low, and consumption only crossed 10,000GWh in the late 1970s (Figure 2), with an annual consumption of 12,028GWh in 1979, at the end of the Fifth 5-year Plan.²²2Reported progress at the end of the Indian financial year (31^st March) on the final year of the Plan. By the end of the Seventh Plan, in 1997, electricity consumption had grown sevenfold. The growth in agricultural power consumption and the overall installed capacity (Central Electricity Authority,, 2021) may imply that even though many villages were connected to the grid in the mid-late 1900s, the quality and quantity may have been subpar, or it may mean that it takes a longer time before households can make use of their electricity connections (Nag and Stern,, 2023).

Village electrification continued at a steady rate till the turn of the century, and by 2002, approximately 90% villages³³3According to the 2001 census, India had 593,732 inhabited villages in 2001. had been connected (Central Electricity Authority,, 2021), although a large number of households in these villages were yet to be connected as of 2004-5 (Desai et al.,, 2005). In fact, until the year 2003, power consumption by the agricultural sector exceeded domestic power consumption at the national level (Central Electricity Authority,, 2021). It is only in 2005 that the government began prioritizing household connections, and launched the Rajiv Gandhi Grameen Vidyutikaran Yojana (RGGVY) to connect the remaining unelectrified households. It is during this period that my study is set.

For my analysis, I use data from two waves of the India Human Development Survey (Desai et al.,, 2005, 2012). The first round of the survey was conducted in 2004-2005, and the second round was conducted in 2011-2012. This makes for a suitable period, as most villages had already been connected, allowing for a larger sample of villages where changes in reliability can be observed and studied. However, since a number of households would be connected in this period, my analysis would need to control for the fraction of households in a village that were connected to the grid. According to the Central Electricity Authority, (2021), transmission and distribution losses were considerably higher than average in this period (31.25% in 2004-5 which reduced to 23.65% by 2011-12). This would have an impact on the reliability of electricity and induce variation in the sample. This period is also roughly 15 years after the shock of economic liberalization in India, and the benefits to economic growth had already set in, distinct from earlier work that investigated wage rates before and after the economic liberalization of 1991 (Van de Walle et al.,, 2017).

The IHDS-I survey covered 41,554 households, and IHDS-II covered 42,152 households in total from 384 districts, 1,503 villages, and 971 urban blocks. Since I am primarily interested in village-level quantities like labor wage rates, I restrict my preliminary analysis to the village questionnaires and data. In all, I construct a panel of 1406 villages from the two survey rounds. I further use household-level data to study labor demand and supply in an attempt to explain the phenomena observed in the village-level analysis. Although there are differences in some of the questions asked in each survey, most of the variables of interest are present in both rounds.

The IHDS surveys make for an excellent data set to study the impact of electrical reliability on agricultural labor wages. The survey has detailed data on several insightful dimensions of village-level electricity. The survey tells me whether a village has electricity, what fraction of households in the village have electricity, when the village was first connected to the grid, and the average number of hours in a day that a village receives power. Similarly, the surveys also have data on the casual agricultural wage rates for sowing, and harvesting, for both men and women and for both major cropping seasons in India - Kharif (summer/monsoon) and Rabi(winter). Since the surveys also have household-level data, I can further investigate the labor demand and supply channels contributing to the trends observed in agricultural wages, at a more microscopic level, by studying changes in agricultural households.

Since I work with observed data, rather than an experiment, whether a village experienced improvements or declines in reliability need not necessarily have been determined by randomization. Therefore, it is important to check if the assignment of villages to different treatment doses i.e., different levels of improvement or declines in electricity reliability were influenced by other factors. The ideal way to check whether the treatment is “As good as random” is to see if there are parallel trends in the treatment groups in the pre-treatment periods. However, in the absence of multiple pre-treatment periods, the best I can do is make the observation that the pre-treatment levels of wages are essentially indistinguishable in the pre-treatment period (Table 1).

Another method through which one could eliminate the possibility of selection bias in assignment to treatment due to covariates is by measuring the correlations between pre-existing levels of village characteristics and the treatment dose. Table 2 presents the mean and standard deviations of village characteristics of villages that receive more reliable electricity access and those that do not along with Pearson’s correlation coefficient of the change in reliability with the 2005 levels of the variables. Overall, there are only marginal differences between most characteristics of treatment and control group villages, and the difference remains similar over time which encourages a difference-in-differences approach (Wing et al.,, 2018).

Since the correlations between pre-treatment characteristics and the change in reliability are weak, I assume that the treatment is random. This is not an unreasonable assumption as changes in reliability, are hardly conscious decisions made by policymakers based on village characteristics, and are mostly a result of transmission and distribution losses. A more formal analysis of whether the assignment to treatment is random or not is presented in Appendix A.

In 2004-5, the average hours of electricity available to villages was 11.76 hours with a standard deviation of 7.80 hours, which increased to 13.38 hours in 2011-12, with a standard deviation of 6.74 hours. The correlation between household-level reliability and village-level reliability is 0.5491 and 0.6628, in 2004-05 and 2011-12, respectively, both coefficients being exceptionally statistically significant. While there is some parity between household responses on reliability, there are large variations as well, with households often even reporting higher quality than that received at the village level, which may indicate that the data at the household level for electricity quality may not be as trustworthy.

The IHDS surveys have data on casual agriculture wage rates in villages for both major cropping seasons in India - the Kharif season where crops are sowed towards the end of the summer, around June, and harvested at the end of the monsoon around October and November, and the Rabi season in which crops are sown at the start of winter, around mid-late November, and harvested in Spring, in April. Rice and maize are the main Kharif crops in India, while wheat is the main Rabi crop. I construct the wage rates for each category by averaging the wage rates for sowing and harvesting, when available⁴⁴4In the absence of one of the two, I use the wage rate for the task that is available. Plowing wage rates were also present in the first round of the survey but I excluded plowing wage rates from the outcome variable, as plowing was relegated to the category of “Other agricultural work”, a category which may include wages for other work such as winnowing or threshing.. From the data (Table 3), I can observe that there is little seasonal variation in the wage rates. However, the wage rates are usually higher for men than for women, with men earning nearly 35% more than women in 2004-5. Even though the wage gap shrinks slightly by 2011-12, there is still a considerable disparity. Figure 3-6 shows the distribution of changes in real wage rates for agricultural labor for women (Figure 3), men (Figure 4), the gender wage gap (Figure 5), and reliability (Figure 6). Typically, in states where reliability reduces, the gender wage gap increases, while in states where reliability increases, the gender wage gap reduces (Rajasthan, Punjab, and Northeast Indian states being the only examples), with the correlation between changes in state-wise reliability and the gender wage gap being -0.33. It is noteworthy that the reliability-associated differences in the growth of wages discussed earlier in Section 1 exist in seasonal wages as well. The difference, however, is less for women’s wage rates, than men’s wage rates, suggesting that gender may play a role in the dynamics of the effect on wage rates.

I begin with a two-way fixed effects specification for village $j$:

$$y_{jt}=\alpha_{j}+\beta R_{jt}+\gamma X_{jt}+\mu_{t}+\epsilon_{jt},~{}~{}~{}t=2005,2012$$

(1)

where the outcome variable $y_{jt}$ can be either the log or level of the casual agricultural labor wage rate of the village, $R_{jt}$ is the electricity reliability, $X_{jt}$ is the vector of control variables, $\alpha_{j}$ is the village-specific fixed effect, $\mu_{t}$ is the time-specific fixed effect, and $\epsilon_{jt}$ is the idiosyncratic error term. The control variables that I use are the fraction of households in the village with electricity access, whether there are metalled roads, NGOs or development organizations, primary healthcare centers in the village, the distances to the closest bank or credit cooperative, and market, the number of government and private primary, middle, secondary and higher secondary schools, and whether the village has less than 1000 inhabitants or more than 5000 inhabitants. I am interested in the coefficient $\beta$ which quantifies the effect of reliability. Taking the first difference of equation 1 in time,

$$\Delta y_{jt}=\Delta\mu_{t}+\beta\Delta R_{jt}+\gamma\Delta X_{jt}+\Delta\epsilon_{jt},~{}~{}~{}t=2012$$

(2)

Since I have data on wage rates, reliability, and control variables, that allow me to generate a cross-section of differences as required for the above model, I can estimate the coefficients using an ordinary least squares (OLS) regression, with $\Delta\mu$ as the intercept.

At this stage, it is important to note that the reliability is constrained to be between 0 and 24 hours a day. Therefore even if the assignment is random, only villages with access to high-quality electricity in the pre-treatment period would be able to experience a large negative treatment, and similarly, only villages with poor reliability in the pre-treatment period will be able to experience large benefits. This association poses a problem to the identification, as it will be difficult to determine if the effects are attributed to the changes or rather a lagged effect arising from pre-treatment levels. Is an increase in reliability reducing wages, or are wages reducing because the village historically has low reliability? One simple way to deal with this problem is to include the pre-treatment level of reliability in the regression as well, so that the regression does not suffer from omitted variable bias, and the effects can be independently attributed to the change in reliability and the pre-treatment level of reliability. Incorporating a lagged effect of the pre-treatment level ($R_{jt-1}$), the model stands as follows:

$$\Delta y_{jt}=\Delta\mu_{t}+\beta_{1}\Delta R_{jt}+\beta_{2}R_{jt-1}+\gamma\Delta X_{jt}+\Delta\epsilon_{jt},~{}~{}~{}t=2012$$

(3)

Including the pre-treatment level in the difference equation, is the same as including a partial sum of reliability in the original two-way fixed effects specification. Thus, the correct form of equation 1 is

$$y_{jt}=\alpha_{j}+\beta_{1}R_{jt}+\beta_{2}\sum_{s=t_{0}}^{t-1}R_{js}+\gamma X_{jt}+\mu_{t}+\epsilon_{jt},~{}~{}~{}t=2005,2012$$

(4)

Clearly, the fraction of households connected, which is a control variable, has the same problem, as values range from 0 to 1, and thus I also incorporate a pre-treatment level of the fraction as a control variable. The coefficient $\beta_{1}$ in Equation 3, would give an unbiased estimate of the effect of changes in reliability.

However, there is little reason to believe that the effects of reliability would be symmetrical, i.e., the effects of a positive change in reliability need not quantitatively be the opposite of the effects of a negative change in reliability. Therefore, I consider two separate natural experiments, one to study the effect of improvements in reliability, and one to study the effect of a decline in reliability. The advantage of splitting the sample into villages that include only positive or negative samples is that I require changes in reliability – the treatment – to be randomly assigned only within each sample. For instance, there could be a selection bias in determining which villages receive positive and negative treatment, but as long as the magnitude of the treatment is random, the results will hold. In Appendix A, I investigate whether the treatment is genuinely random. We find that for villages where reliability increases, the treatment can be assumed to be as good as random, and the results of the difference-in-differences estimates will be unbiased. However, for villages where reliability reduces, there may be some endogeneity in the assignment to treatment and the results may not be as robust. Therefore, when we split the sample, as opposed to including two treatment variables in a single regression, we are able to isolate a sample of villages and a type of treatment that is as good as random.

Since there are no village-level variables for demand and supply, I use household-level data to study the demand and supply of labor, although I use village-level reliability as the treatment, as households often report considerably different levels of electricity reliability and the household data may not be as accurate.⁵⁵5Since all the electricity comes from a common distributor in the village, the variation in quality is unlikely to be as much as I find in the household data. While there may be losses that affect individual households, sometimes households report levels of electricity reliability that exceed the levels received by the village, which indicates that household reporting may be an unreliable source of information for electricity quality. Furthermore, even load shedding is typically carried out for entire villages and not individual households. Moreover, the quality of electricity available to houses may not be the same as the quality of electricity available to farms (which will be useful in studying labor demand), and in the absence of data on electricity reliability in farms, the village’s reliability is the best proxy.

There is no variable in the survey that can be directly used as a measure of labor supply. Thus, I investigate a channel through which labor supply may be affected, as this is easier to test from the data. I postulate that the supply of labor would be affected by electricity quality primarily through a reduced time burden of domestic chores. In particular, the tasks that would be especially affected by electricity disruptions are lighting and refrigeration. 4045 households in the sample (21.2%) claimed to own refrigerators in 2011-12. Long disruptions would hamper refrigeration, and the presence of good-quality electricity could reduce the time and fuel required for cooking, through refrigeration. Since the survey does not have questions related to the time spent cooking, I am constrained to restrict my analysis to fuel. About 75% households use firewood for either cooking or lighting. Over 40% households also use dung for cooking. A large number of households collect these fuels from either owned or community land, good-quality electricity could significantly reduce the time spent by members in collecting biomass. Thus, as a supply channel, I study the effect of electricity reliability on the time spent by men and women on fuel collection. I use a difference-in-difference specification, similar to equation 3. For a household $i$ in village $j$, we model the time spent in fuel collection $F_{ijt}$ by

$$\Delta F_{ijt}=\Delta\mu_{t}+\beta_{1}\Delta R_{jt}+\beta_{2}R_{jt-1}+\gamma\Delta X_{ijt}+\Delta\epsilon_{ijt},~{}~{}~{}t=2012$$

(5)

where most of the coefficients represent the effects of the same quantities as before (on time spent in fuel collection), except $\gamma^{\prime}$ which now includes coefficients for some household-level controls along with the village-level controls listed about. The household-level controls are the number of adult men and women, the presence of a source of drinking water in the house, the presence of flush toilets, and the presence of separate kitchens. Obviously, village-level electricity quality cannot have a direct effect on households not connected to the grid. Hence, I only consider those households that had electricity in both 2004-5 and 2011-12. Ideally, had there been data for how much labor individuals in a household were willing to supply I could have studied the effect of time saved in fuel collection on such a variable. But in the absence of such a variable, a proper analysis of labor supply is beyond the scope of this work.

Unlike supply, demand can be studied by considering the person-days of labor hired by agricultural households to work on their farms. Of course, I need to control for the labor wage rates as well, to avoid an omitted variable bias. Not controlling for labor wage rates could be particularly problematic, as reliability is correlated with the wage rates, and we know from theory that wage rates themselves can affect the demand for labor. In addition to the control variables listed above, we also include the change in the ownership of electric pumps ($\Delta P_{ijt}$). We use the logarithm of the person-days of hired farm labor $D_{ijt}$ as the outcome variable and formulate another difference-in-differences equation

$$\Delta D_{ijt}=\Delta\mu_{t}+\beta_{1}\Delta R_{jt}+\beta_{2}R_{jt-1}+\delta_{m}M_{jt}+\delta_{w}W_{jt}+\phi\Delta P_{ijt}+\gamma\Delta X_{ijt}+\Delta\epsilon_{ijt},~{}~{}~{}t=2012$$

(6)

$M_{jt}$ is the log men’s casual agricultural wage rate in village $j$, and $W_{jt}$ is the corresponding quantity for women’s wages. The coefficients $\delta_{m}$ and $\delta_{w}$ are the elasticities of labor demand for men and women, respectively, and tell us the fractional change in demand, for a fractional change in wages. Since pump ownership could have an impact on demand, it is also important to check whether reliability has an effect on pump ownership and is not affecting demand through pump ownership. We do so using another difference-in-difference regression:

$$\Delta P_{ijt}=\Delta\mu_{t}+\beta_{1}\Delta R_{jt}+\beta_{2}R_{jt-1}+\gamma\Delta X_{ijt}+\Delta\epsilon_{ijt},~{}~{}~{}t=2012$$

(7)

Table 4 presents the results of the difference-in-differences regressions for the impact of a change in the electricity reliability (both positive and negative) on the change in men’s and women’s casual agricultural labor wage rates. Although we present the results for both positive and negative treatments, only the positive treatments can be assumed to be as good as random, and the estimates of the effects of negative treatment may not be as robust. First, I find that reliability has a negative effect on both logs and levels of casual agricultural labor wages, for both men and women, although the effect on women’s wage rates is considerably smaller, and not statistically significant. For every additional hour’s increase in electricity reliability in a village, men’s casual agricultural wage rates reduce by nearly Rs. 2 or 1.34% in relative terms. The effect on both levels and logs is statistically significant at the 1% level. On the other hand, positive treatments have small and statistically insignificant impacts on women’s wage rates.

The effects of a decrease in quality, however, has consistently larger effects on both men’s and women’s wages. For every additional hour’s reduction in electricity availability, men’s wage rates rise by over Rs. 2 (significant at the 1% level), although the effect on the log of wage rates is small (Rs. 0.6) and not significant. Similarly, women’s wage rates increased by Rs. 1.2 (significant at the 5% level), although the effect on log wages is not statistically significant. Nevertheless, this implies that the effects of changes in reliability are asymmetrical, particularly for women’s wage rates, and wage rates are more sensitive to reductions in reliability than increases. But these estimates are not as robust as those for improvements in reliability as the assignment to treatment may be biased by initial characteristics.

Interestingly, even the pre-treatment level of reliability has a statistically significant negative effect on men’s wages– a Rs. 1.4 or a 1.04% reduction for every hour’s increase or a Rs. 1.8 or 0.91% increase for every hour’s reduction (all significant at the 1% level). From this, I can conclude that the effect of the overall level, not the change alone, is negative. However, for women’s wages, the effect is again statistically insignificant for all cases except on the levels of the wage rate in the negative treatment case, where wages increase by Rs. 1.4, significant at the 5% level. An obvious consequence of different effects on men’s and women’s wages is that better reliability leads to a smaller wage gap. From the estimated coefficients, it is clear that the wage gap increases over time for both men and women, after controlling for various factors, although villages where reliability improves would experience a smaller widening of the gender wage gap when compared to a village where reliability declines.

The effect of reliability on wage rates, however, is difficult to explain intuitively and will require studying the effects of reliability on labor demand and supply.

In order to study the possible effect of electricity reliability on the supply of labor, I estimate the impact of changes in reliability on the time spent by men and women on fuel collection. Table 5 presents the results.⁶⁶6We include only villages where reliability increases as the treatment in these villages can be assumed to be random. See Appendix A. Reliability has a negative effect on the time spent on fuel collection. In villages where reliability improved, every additional hour of power available reduces the time spent by women on fuel collection by 12.6 minutes per week, which is statistically significant at the 1% level, while men’s fuel collection time reduces by 6.1 minutes per week, significant at the 10% level. This can amount to a very large reduction of fuel collection time if the quality changes sizeably.

Clearly, women’s fuel collection times are affected more, as fuel collections like other domestic chores are typically carried out by women. Time relieved from household chores could substantially increase the supply of labor and has been hypothesized to be a channel for electricity-facilitated increase in labor participation in past studies as well (Dinkelman,, 2011). However, it may be more complicated, especially in the Indian context, and in other overwhelmingly agrarian rural economies. In general, time saved could increase women’s labor market participation, which would increase the supply of labor and may have a detrimental effect on wages, or the time could be utilized in leisure, in which case the labor market would be virtually unaffected.⁷⁷7Unfortunately, there is no variable that aptly quantifies leisure time, and I cannot test this empirically. A third possibility is that women take up more domestic chores that used to be performed by men, freeing up men’s time burden of domestic work, which could increase men’s labor market participation, which would also increase labor supply. This may explain why men’s wage rates fall more than women’s wage rates. Although that could also be if women are hired more in the labor market because farmers may want to hire cheaper labor. But this possibility only works on the assumption that men’s and women’s labor is thought of to be fungible, which is unlikely to be the case. Another alternative is that women could provide labor on household-owned businesses, or more likely in the case of rural agricultural households in India, women could supply labor to farms owned by their households, a common phenomenon in Indian farms. While this would not exactly increase labor supply in a strict sense, since it is not paid work, it could reduce the demand for hired labor. Thus, it is also important that I study if reliability, or even fuel collection time, has an impact on labor demand.

To estimate the effects of electricity quality on labor demand, I study how changes in reliability affected the number of person-days of agricultural labor hired by farms. Table 6 presents the estimated effects on the levels and logs of the number of days of hired work. Both the changes in reliability and the pre-treatment level of reliability have no statistically significant impacts on the number of days of labor hired by farms. We also do not find any statistically significant effects of changes in wage rates, either at the levels or logs, on demand. This suggests that the own elasticity of labor demand is inelastic, for both men and women.

An increase in the number of pumps, however, has a large statistically significant positive impact on the demand for labor. Although the distribution of pumps may not be random thus making the sizes difficult to interpret, the positive effect of pump ownership on log changes in the number of hired days is significant at the 1% level. Pump ownership increases the number of person-days of hired labor by 57 days or 30% in relative terms, which is close to two months for one laborer. This effect does not disappear even after controlling for baseline levels of household affluence using the 2004-5 levels of expenditure. The positive effect, signifies, that electrical equipment such as irrigation pumps do not supplant labor, and instead drive up the demand for labor. This increase in demand could be because pumps may require labor to operate, or because pumps could increase agricultural productivity, incentivizing higher investments in agriculture. It could also arise from pumps facilitating agriculture during the dry seasons, which is most likely. Thus, the possibility that electricity quality reduces labor wages through the mechanization of agriculture may be ruled out. To ensure that reliability does not affect labor demand through increased pump ownership, I test the effect of changes on reliability on pump ownership. Table 7 presents the results, and clearly there is no statistically significant increase (or decrease) in pump ownership due to increases in reliability.