Changing Wage Structure in India in the Post-Reform Era: 1993-2011

This paper documents the changing structure of wages in India over the post-reform era, the roughly two-decade period since 1993. To investigate the factors underlying these changes, a supply-demand framework is applied at the level of the Indian state. While real wages have risen across India over the past two decades, the increase has been greater in rural areas and, especially, for unskilled workers. The analysis finds that, in rural areas, the changing wage structure has been driven largely by relative supply factors, such as increased overall education levels and falling female labor force participation. Relative wage changes between rural and urban areas have been driven largely by shifts in employment, notably into unskilled-intensive sectors like construction.


I. Introduction
We investigate changes in the structure of wages in India over the post-reform era, the roughly two-decade period since 1993, paying particular attention to recent trends in the wages of rural workers, especially the unskilled. Poverty reduction, much of it concentrated in rural areas, has accelerated over the last few years, largely due to increased earnings from non-agricultural wage employment (Balcazar et al. 2015). An exploration of the fine-grained details of India's labor market transformation will thus help us to better understand this poverty decline.
Our approach hews closely to the Supply-Demand-Institutions (SDI) framework pioneered by Katz and Murphy (1992) and Bound and Johnson (1992). The idea is to divide the workforce into imperfectly substitutable demographic groups; e.g., by gender, education, and age. The twist, in our case, is to also cut the data by rural/urban, recognizing that, to a large extent, rural and urban India constitute distinct labor markets, or at least are far from being perfectly integrated. Thus, our apparatus allows us to investigate, for example, changes in wages of the rural unskilled relative to their urban counterparts.
A second point of departure from conventional SDI analysis is its application at the state level, treating each Indian state (or group of states) as having separate urban and rural labor markets. A state-level approach provides the requisite degrees of freedom for econometric analysis (see Juhn and Kim, 1999, for a related study of US states). In particular, SDI decomposes wage changes for a group into supply shifts (changing group employment shares), demand shifts (changing industrial composition biased for or against a group), and wage-premia shifts (essentially, movements into or out of structurally low-paying jobs). We then take the analysis a step further by investigating the key state-level drivers of recent relative wage trends; i.e., we ask what types of supply or demand shifts were particularly influential in explaining the changing wage structure in India over the last decade.
There is a modest literature exploring India's wage structure using data from NSS's Employment-Unemployment surveys. Hnatkovska and Lahiri (2013) consider rural-urban wage convergence in India from 1983-2009 using a model of long-run structural transformation, but they do not decompose supply and demand factors behind the more recent wage trends. While Chamarbagwala (2006), like us, uses a supply-demand decomposition, it is focused on the impact of trade liberalization over the earlier 1983-99 period (see also Azam, 2010).
The organization of the paper is as follows. We begin in Section II by defining our groups and industries and then documenting how real and relative wages in India have changed over the past two decades. In Section III, we review the SDI framework and apply it to the national level. Next, we turn to the state-level SDI analysis in Sections IV and V, followed by conclusions in Section VI.

II. Preliminaries A. Definitions: groups and industries
We analyze three rounds of the NSS, the 50th (1993-94), 61st (2004-05), and 68th , thus covering an 18 year span. Workers between 12 and 65 years of age are divided into 8 demographic groups, consisting of the 2x2x2 interaction of male/female, educated (completed secondary level or above)/uneducated (less than completed secondary), young (12-29)/old . In addition, we construct aggregates for these 8 demographic groups by sector (urban/rural), yielding 16 groups in total.
Wage earners are defined as those engaged in "gainful activities", as recorded in their "usual principal status" in the NSS, but not self-employed. Usual principal status also serves as our basis for categorizing individuals into industry groups below. We focus on principal status because this accounts for the preponderance of the reference period of 365 days preceding the date of survey. 1 Figure 1 shows the extent to which each demographic group is represented among sectorspecific wage-earners. Of particular note is the increasing share of educated, which is much more pronounced in rural than in urban areas. Despite this trend, uneducated males remain the dominant group in rural wage labor markets.

Figure 1: Relative shares of demographic groups among wage earners
1 Subsidiary status is much more temporary in nature and, as NSSO suggests, only about 1.3% in the rural and 0.1% in the urban areas had participated in two subsidiary economic activities during the period of one year before the date of survey in round 55 (NSSO, 2008 We next define five broad "industry" or occupational categories: (1) agriculture (inclusive of forestry and fishery); (2) construction; (3) manufacturing (inclusive of mining and utilities); (4) Professional (including public administration); (5) services (inclusive of wholesale/retail trade and domestic service). 2 SDI analyses using developed country data, and even Chamarbagwala's (2006) study of urban Indian wages, typically use a much more finegrained industrial classification. However, sample size considerations constrain us to only five. Rural India is predominately agricultural; manufacturing, in particular, has until very recently accounted for much less than ten percent of rural employment. Given the typical NSS sample, there would simply not be enough wage-earners in each category to support a very detailed classification. This concern is only reinforced in our state-level analysis, where state-wise wage-earner samples are much smaller.

1993-94
2004-05 2011-12  Patterns of industrial employment have changed rather dramatically in rural India over the past two decades. As seen in Figure 2, from around three-quarters in the early-1990s, the share of rural labor employed in agriculture had, by 2011, declined to around one-half. The two main rural growth industries are services and construction, with the latter's employment share more than quadrupling over the last two decades. By contrast, the urban picture is one of relative stasis, with more modest expansions of services and construction over the same period. Representation of the 8 demographic groups in each industry is shown for both rural and urban areas in Figure 3. Educated workers, obviously, predominate in the professions, whereas wage-jobs in rural construction are largely held by unskilled males, even more so than in agriculture and quite substantially more so than in services. Both rural and urban areas have seen a gradual up-skilling of the workforce over the past two decades, across all industrial sectors.

B. Changes in real wages
Information on weekly wage earnings and days worked per week is available for regular and casual workers. In the case of those who perform multiple jobs in the week, we calculate average daily wages by dividing weekly wage income from all sources by total number of days worked. To compute real wages, we use the state-level Consumer Price Index for Agriculture (CPI-AL) and Industrial Workers (CPI-IW). Originally, the CPI-AL was available with base-year 1986-87 and CPI-IW with base-year 1982. We converted these indices to have a uniform base-year 2004-05. CPI-AL is used to deflate wages in rural areas and CPI-IW in urban India. Because these deflators are not available for some of the small states, we used available information for larger states either adjacent to them or from which they had been split (see Appendix Table B.2 for details on the CPI calculation for the smaller states).  Mean annualized changes in log real wages by group are shown in Figure 4 across each of the two sub-periods. Evidently, wages have been rising in real terms over the past two decades for all groups and especially in rural areas. There has also been a marked acceleration in wage-growth in recent years, most pronounced in urban India as well as among the unskilled (those with less than secondary education). Looking across states ( Figure 5), we see big realwage gains for unskilled workers in the south and east of the country, with W. Bengal being a notable exception. The remainder of our analysis will largely ignore this overall rising-tide to focus on why some "boats" have risen faster than others.

C. Changes in relative wages within and across sectors
For ease of presentation under the first major column heading of Table 1, we aggregate mean relative wage changes across pairs of demographic groups using the respective (base-year) shares of wage-earners as weights. Thus, for example, the change in the wage for rural educated males relative to rural uneducated males is computed as a weighted average of the corresponding mean wage changes for old and young rural males in each of these educational categories.  The first two rows of Table 1 (denominated in log changes) indicate that the wages of uneducated rural workers rose relative to those of educated rural workers (hence the negative sign) for both males and females. Much of these relative gains occurred in the most recent decade . Note that the urban unskilled also experienced relative wage gains in the second period, but not quite as much as their rural counterparts. Overall, rural females (especially the unskilled) gained ground on rural males in the last decade.

Table 2: SDI decomposition of relative wage changes between rural and urban India
Looking across the urban-rural divide in Table 2, the striking pattern is wage convergence, albeit skewed toward the unskilled. Overall, wages for uneducated males rose by around 47% relative to their urban counterparts; the corresponding figure for uneducated females is 37%. However, much of these gains occurred in the earlier decade of the post-reform era, especially for males. Similar, but substantially smaller, relative gains were experienced by educated rural workers. Aggregating across groups, rural wages rose a modest 9% relative to urban wages over the last decade, following a 27% increase in the first decade.

III. Supply, Demand, Institutions A. Conceptual framework
Suppose we have a CES production function for aggregate output that depends on just two types of labor (ignore capital), types a and b. Katz and Autor (1999), e.g., show that where are wages for type i in time t, is an index of relative demand shifts favoring group a, and is an index of relative supply shifts favoring group a. The parameter represents the aggregate elasticity of substitution in production between labor of type a and b. A key implication of the model is that only net demand shifts (i.e., net of supply shifts) matter for relative wages. Differencing equation (1) over time, using the notation ∆x x x , delivers ∆log ∆ ∆ . (2) Thus, on the left-hand side of equation (2) we have a difference-in-difference in mean logwage for two groups over time. 3 These diff-in-diffs are precisely what is reported in Tables  1 and 2 for, respectively, within and between sector contrasts.
We may write the relative supply for group i in sector s at time t as log (3) where is the group's employment in the sector and is total employment in the sector. Note that employment includes self-employment in agriculture or in a household enterprise and hence the employed are a much larger set than wage-earners, especially in the rural sector. Shifts in supply, ∆ , are assumed to be predetermined; that is, not caused by changes in relative wages. In the state-level analysis we will have the opportunity to test this assumption.
Theoretically consistent measurement of demand shifts is a complicated issue (see Katz and Autor, 1999;Bound and Johnson, 1992). We follow Juhn and Kim (1999), who use the between (industrial) sector demand shift measure of Katz and Murphy (1992) where indexes industry. So, the first term in the sum is the share of demographic group in industry 's employment and the second term is the growth rate in the share of industry employment in overall sectoral employment. Intuitively, ∆ is larger when demographic group (initially) predominates in relatively fast-growing industries. As with supply shifts, ∆ is taken as exogenous with respect to changes in relative wage structure; again, this is testable.
The "institutions" component of SDI boils down to allowing for industry wage premia. A wage premium measures the extent to which a given type of worker (demographic group) is paid more (or less) when working in a particular industry. Labor market institutions matter insofar as wages are not determined solely by the interaction of skill endowments and skill prices-i.e., by the competitive market for skills. A salient example in the case of India is agricultural labor. On average, jobs in agriculture pay around a-third less than those outside of agriculture, holding location and type of worker constant. Why this premium arises is beyond the scope of the present investigation, but it may have something to do with the fact that a higher proportion of agricultural than nonagricultural workers in India are hired on a casual daily basis (see Appendix Table B.3 for details).
Following Bound & Johnson (1992), then, let where is the competitive market wage given group skills, is the industry wage premium for group at time , and ⁄ is the proportion of group workers in industry . Based on estimated from wage regressions, the institutions index (∆ ) for group is the change in the entire wage-premium term or Returning to the case of the negative wage premium in India's agricultural sector, we can see that a group with a higher ∆ is one which is moving out of agriculture relatively quickly.
Comparison of group-shares across NSS rounds, as shown in Figure 6, indicates that uneducated rural males (young and old) are shifting out of agriculture most rapidly.  Table 1. So, why did the wages of educated workers decline relative to the wages of uneducated workers in rural India? First off, there was a substantial increase in relative supply of educated workers, especially for females, spread rather evenly across the two sub-periods. Meanwhile, relative demand for educated workers fell, especially for males. And, finally, there were modest declines in the institutions index for educated relative to uneducated workers. In other words, uneducated workers moved out of (low-paid) agricultural labor faster than educated workers. Similar, but less pronounced, patterns are seen for educated vs. uneducated workers in urban India (rows 6 and 7).
In Table 2, we compute urban vs. rural SDI changes. Focusing on unskilled labor, we see that shifts in relative supply were not a decisive factor behind the wage gains of uneducated workers vis-a-vis the educated. There were, however, big drops in relative demand for

2011-12
Agri.+ Forestry+Fishing Construction Mining, Manufacturing and Utilities Professional Services unskilled male labor in urban areas, with smaller declines in the case of females. The institutions index also moved against the urban unskilled. The story of wage gains by the rural unskilled relative to their urban counterparts is, therefore, one of changing patterns of industrial employment rather than one of changing relative supplies (as was the case within the rural sector).

IV. State-level SDI Analysis
We now compute changes in mean log wages, ∆ , ∆ , and ∆ separately for each major state or group of adjacent states. 5 Our "data set", therefore, consists of 448 = 2 x 2 x 8 x 14 observations for 2 decadal intervals, 2 sectors (rural/urban), 8 demographic groups, and 14 states. Note that in treating a state as, for all intents and purposes, a distinct labor market, we are assuming that changes in, say, labor supply within a given state are not driven by inter-state migration. This assumption seems reasonable as a first approximation given India's historically low mobility (see Hnatkovska and Lahiri, 2013).  Bivariate scatterplots ( fig. 7) reveal that increases in supply are strongly associated with wage declines in each period. Increases in demand, by contrast, are associated with wage increases (fig. 8). This is all as it should be, but to properly assess the SDI framework we need to control for both supply and demand shifts simultaneously.
To do so, we run a series of regressions of state mean log-wage changes on the SDI shift variables. The first such regression, shown in Table 3, uses the full dataset, thus including log-wage changes between 1993-2004 and 2004-2011. Among the independent variables is a dummy for the second decadal change. Results in the first column of Table 3 show that increases in supply lead to lower wages, conditional on the demand shift. Likewise, increases in demand increase wages, conditional on the supply shift. Moreover, we cannot reject the null hypothesis that the coefficient on supply is equal to minus the coefficient on demand; i.e., that only net demand shifts matter for wages (cf., equation (1)). Next, we address the simultaneity between wage changes on the one hand and demand and/or supply shifts on the other. Do ∆ , ∆ , and, for that matter, ∆ cause wages to change, or is it the other way around? Arguably, the supply of skills and the structure of industrial employment are slow to adjust and may reasonably be thought of as predetermined. However, to test this proposition, we instrument ∆ , ∆ and ∆ by their lagged values ∆ , ∆ and ∆ . The idea here is that lagged changes reflect long-run trends, uncontaminated by contemporaneous wage shocks. Of course, using lags as instruments requires us to drop the first decadal change, which corresponds to half our sample. Hence, in column 2 we replicate our original OLS specification on the sample of second-decadal changes, with very similar results. IV estimates are shown in column 3. There is little evidence of endogeneity bias; to be sure, the coefficient on demand shifts more than doubles from its OLS magnitude, but this could be due to chance. And, the null hypothesis of the SDI framework fares extremely well in this specification. Thus, in column (4), we report the same IV specification but with the SDI restriction imposed, which is to say that only net demand shifts ∆ ∆ are now included along with ∆ . In all specifications, the coefficient on the institutions index ∆ is not significantly different from zero. Finally, we run the same set of regressions with state fixed effects and obtain very similar results (see Appendix Table B.4).

V. SDI Drivers across States
The diagnostics of the previous section suggest that the SDI framework does a reasonably good job explaining wage growth of the past decade across both demographic groups and states. But what are the key structural trends underlying these changes? Five candidates for consideration are: (1) Urbanization; (2) NREGA; (3) the rural construction "boom"; (4) falling rural female LFP; (5) Rising agricultural prices.
We begin by predicting log-wage changes from 2004-2011 for each group x state observation using the results in Table 3, column 4; i.e., ∆ ∆ ∆ ∆ .
Next, we construct predicted differences-in-differences across groups i and j within a sector as follows or across sectors within group i using where subscripts u and r denote, respectively, urban and rural. Finally, we examine the bivariate associations between the predicted D-in-Ds and each of the five structural wage drivers mentioned above.

A. Within rural India
We look first at rural areas and, in particular, at wages of educated rural workers (old/young and male/female taken together) relative to uneducated. Each panel of figure 9 shows a scatterplot of ∆ , ∆ against a relevant driver. Having now aggregated wage changes across all 8 demographic groups, we end up with 14 data points, which is to say one ∆ , ∆ for each state-group.
Consider the change in the employment share of construction in rural areas of each of the 14 state-groups. The top left panel of Figure 9 shows that higher construction shares are strongly positively associated with the predicted growth in wages for the uneducated relative to educated. Indeed, differences in construction industry growth explain about two-thirds of the variation in the relative wage growth predicted by the SDI framework. The same exercise using the rural services share, an industry which also employs significant numbers of unskilled workers and which also expanded in relative terms over the last decade, shows a similar pattern but a weaker association with wages. In sum, the rural construction boom appears to have been an important, if not the main, driver of unskilled relative wage-growth within rural India.

Figure 9: Drivers of changes in educated vs. uneducated wages within rural India
It is interesting to contrast the labor market impacts of the above compositional shifts to those of NREG (National Rural Employment Guarantee). Phase-in of NREG began at around the mid-point of our 2004-2011 window. Analyses of NSS data preceding the 68 th (2011-12) round provide mixed evidence as to the rural wage impacts of NREG expansion (see Azam, 2012;Zimmerman, 2013;Imbert and Papp, 2015). However, NSS68, for the first time, provides individual level data on NREG registration (job-card holding) and take-up (i.e., NREG employment in the last 12 months). This allows us to construct, for each state, the proportion of each demographic group that are job-card holders or who have worked in NREG.  Looking across state-groups in figure 10, there are huge differences in NREG registration rates, with Rajasthan and MP topping the list, although rates of participation in this massive public works program are actually highest in the far east of India ("Seven Sister" states). Also relevant for our analysis is the large registration and participation gap between the educated and uneducated, with much higher NREG involvement among the latter ( figure 11). Thus, we have in the two bottom panels of figure 9, plots of the predicted log-wage D-in-D against the state-wise differences in NREG participation shares (job-card on the left; worker on the Share of rural labor force working in NREG Share rural labor force holding NREG job card right) between educated and uneducated groups. Given Figure 11, all of the NREG share differences are negative (educated have lower registration and take-up than uneducated). What we do not see is much of a relationship between NREG participation and wage growth (the slopes are positive, but the R 2 s are essentially zero). Put differently, states in which NREG has (presumably) expanded relative employment opportunities for unskilled labor more do not appear to have experienced differential growth in net demand for unskilled labor. This is, of course, not to say that NREG has been ineffectual as a safety-net for the poor, only that it is evidently too small of a labor market intervention to have detectable general equilibrium effects. 6 Next, using the same approach, we consider what has been driving changes in relative wages of men versus women in rural India over the last decade. In this case, we compute ∆ , ∆ by aggregating wage changes for all male (m) and female (f) demographic groups within the rural sector of each state. Here we introduce another potentially relevant factor, the change in female labor force participation (LFP). Figure 12 shows massive declines in female LFP in rural areas of most states, whereas figure 13 shows much more muted ones in the corresponding urban areas. 6 We have done a similar analysis using "raw", as opposed to predicted (by SDI), wage changes with the same result. The top left panel of figure 14 provides striking confirmation that this recent movement of women out of the rural labor force explains much of the predicted increase in their wages relative to those of men; the R 2 of the associated bivariate regression is 0.84. By contrast, changes in the rural construction share (top right panel) or in women's participation in NREG relative to men's (bottom panels) explain next to nothing.

B. Urban vs. rural India
In the remainder of our analysis, we contrast urban and rural wage changes for unskilled labor. In particular, we use equation (9)   For males, the construction sector stands out as the key relative wage driver, with higher construction growth strongly associated with higher wage growth (R 2 =0.34), whereas for females the corresponding association is actually negative, albeit weak (R 2 =0.05). Relative growth in the service sector, by contrast, bears little relationship to relative wage changes for either males or females. As for female LFP, we again see a strong correlation with wage growth. In states where women have withdrawn from the labor force faster in the countryside than in cities, rural wages of females have risen faster than urban wages (R 2 =0.31), a pattern essentially absent with respect to male wages (R 2 =0.04).
Next, we ask whether the growth of cities has in and of itself led to changes in SDI at the state level. By far the fastest urbanization over the last decade occurred in Kerala, which is clearly an outlier in the bottom right panels of figures 15 and 16. Nevertheless, even with Kerala excluded, the story is clear. Faster urbanization is associated with greater urban wage growth relative to rural areas for both genders, but especially for females. Moreover, this latter effect is not driven merely by correlation between falling female LFP and urbanization; it survives virtually intact after controlling for the relative change in female LFP. Thus, it appears that in rapidly urbanizing states the demand for female labor, as reflected in their wages, has been growing faster in cities than in the countryside. As a final exercise, we turn to the agricultural commodity price boom of recent years as an explanation for the relative rise in rural wages. Jacoby (2014) uses variation across Indian districts in the shares of different crops in production to show that districts experiencing relatively higher agricultural prices over the 2004-09 period also saw higher wages for unskilled labor. Adapting this approach to the state-level analysis of this section and extending the price data to 2011-12, we construct the following measure of differential agricultural price change where is the initial (i.e., 2004-05) share of labor in agriculture for a state in sector ( , ), is the share of crop c in the total value of state agricultural production in base-year 2003-04, and ∆log is the change in log-price of crop c between the 2004-05 and 2011-12 crop marketing years for the 18 top field crops of India. 7 Intuitively, the labor market response to changes in agricultural prices is modulated by the output share of agriculture in the overall economy of the sector; if production is Cobb-Douglas, this output share is equivalent to the labor share.
The relationship between differential urban-rural agricultural price changes, as reflected in ∆ ∆ , and relative wage changes, as reflected by ∆ ∆ , is complicated by the fact that the agricultural labor share differential affects both quantities independently. Referring to equations (4) and (7), one can see that and ∆ ∆ are mechanically related. In particular, since unskilled workers shifted out of agriculture into construction and other services over the last decade, the demand index for unskilled workers is dominated by a weighted average of the proportion of each of these industry's share of unskilled labor, where the weights are, essentially, the growth rates of employment in the respective industries. In a state where agriculture had a larger initial employment share, the growth rate of agriculture employment tends to be smaller and, hence, there appears to be a greater increase in demand for unskilled labor. The upshot is that, in considering the bivariate relationship between ∆ ∆ and ∆ ∆ , we must partial out this mechanical correlation with . Figure 17 thus plots the residuals of ∆ ∆ against those of ∆ ∆ in regressions on across the 14 state-groups. Consistent with Jacoby (2014), the figure shows that rural wages of the unskilled (males and females combined) have risen faster relative to urban wages in states where the terms of trade for agriculture have improved by more. Evidently, this is due to the fact that in states benefitting differentially from the agricultural commodity boom, the secular decline in agriculture has been attenuated and, as a result, the demand for unskilled labor has not fallen as much due to structural transformation.

VI. Conclusions
Real wages have risen across India in the past two decades, but the increase has been greater in rural areas and, especially, for unskilled workers. Broadly speaking, the changing wage structure within rural areas has been driven largely by relative supply factors, such as increased overall education levels and falling female LFP, whereas the changing wage structure between rural and urban areas has been driven largely by shifts in employment, notably into unskilled-intensive sectors like construction. Notwithstanding the rural construction boom, the recent expansion of the national public-works program (NREG) throughout rural India does not appear to be associated with shifts in the structure of wages (i.e., to the advantage of the unskilled) over the last decade. Finally, while structural transformation-the gradual movement of labor out of agriculture-has been the dominant trend of the last two decades in rural India, our evidence suggests that the recent upturn in agriculture's terms of trade may have muted the commensurate decline in demand for unskilled rural labor, contributing to growth in wages for the rural unskilled relative to their urban counterparts.   Note: Industry premia sum to zero by construction. Industry share of casual labor is the % share (weighted) of each industry in total casual labor force.