World Development Vol. 39, No. 10, pp. 1820–1833, 2011 Ó 2011 Elsevier Ltd. All rights reserved. 0305-750X/$ - see front matter www.elsevier.com/locate/worlddev doi:10.1016/j.worlddev.2011.04.026

Caste Stratification and Wealth Inequality in India

AJIT ZACHARIAS Levy Economics Institute of Bard College, NY, USA

and

VAMSI VAKULABHARANAM * University of Hyderabad, India

Summary. — We analyze the relationship between wealth inequality and divisions in India using nationally representative surveys on household wealth conducted during 1991–92 and 2002–03. According to our findings, the groups in India that are generally consid- ered disadvantaged (known as Scheduled or Scheduled Tribes) have, as one would expect, substantially lower wealth than the “forward” caste groups, while the Other Backward Classes and non-Hindus occupy positions in the middle. Using the ANOGI decom- position technique, we estimate that between-caste inequality accounted for about 13% of overall wealth inequality in 2002–03. The stratification parameters indicate that the forward caste Hindus overlap little with the other caste groups, while the latter have signif- icantly higher degrees of overlap with one another and with the overall population. Using this method, we are also able to comment on the emergence and strengthening of a “,” or relatively well-off group, among the disadvantaged groups, especially the Scheduled Tribes. Ó 2011 Elsevier Ltd. All rights reserved.

Key words — caste, inequality, stratification, distribution of wealth

1. INTRODUCTION Kijima showed that both lower endowments of physical and human capital possessed by disadvantaged groups, as well as Caste is a persistent determinant of power, economic different structures of income generation, contribute equally inequality, and poverty in contemporary India. Traditionally, to the disparities among caste groups. What is remarkable mainly non-economists (anthropologists, sociologists, and his- across these studies is the persistence of systematic disparities torians) made substantial contributions to the literature on among households across different caste groups over long peri- caste relations in India (e.g., Be´teille (2007), Gupta (2000), ods of time. Thirdly, there is a growing literature that has and Srinivas (2000)). However, this trend has changed with delved into the mechanisms and belief systems that perpetuate economists contributing to this literature in the recent times, and are perpetuated by discrimination and unequal caste rela- especially over the last decade or so. As better data sources tions in diverse institutional spaces, such as labor market, have become available, many of these studies have combined schools, and the institution of marriage (e.g., Banerjee, Duflo, results from large surveys as well as from field surveys to pro- Ghatak, and Lafortune (2009) and Hoff and Pandey (2006)). duce new insights into the working of caste relations in India. Similarly, Thorat and Newman (2009) show, through an inter- The studies of caste in the economics literature can be di- esting field experiment, that the underprivileged caste groups vided into three broad categories. First, there are studies that and Muslims with the same qualifications as their privileged have used either large surveys (mainly National Sample Sur- caste counterparts are discriminated against in hiring/screen- veys [NSS] and National Family Health Surveys [NFHS]) or ing practices—a legacy of social exclusion. fieldwork-based small sample surveys to investigate caste dif- Our paper contributes to this literature by analyzing the ferentials in consumption, income, education, occupations, relationship between overall wealth inequality and caste and development indices (e.g., see Deshpande (2001), Hasan divisions 1 in India. There have been very few studies (see and Mehta (2006), Madheswaran and Attewell (2007), Mehro- Subramanian and Jayaraj (2006)) on wealth disparities (as op- tra (2006), Mohanty (2006), Munshi and Rosenzweig (2006), posed to consumption or income disparities) within and Srinivasan and Mohanty (2004), and Sundaram (2006)). The among caste groups and how these disparities contribute to near consensus in these studies is that the less privileged caste the overall inequality in India. Wealth inequality is an integral groups tend to be worse off than the others on the measured aspect of economic inequality among persons at a given point indicators across the country, although there are regional dif- in time, as well as across generations. Disparities in wealth can ferences. The second category of studies have employed the also translate into disparities in economic security. For a sub- Blinder-Oaxaca decomposition (or its variants) to separate stantial portion of the Indian population that is dependent on the structural differences (e.g., geographical, discrimination- agriculture, land is the major source of livelihood. Inequalities based) among households from the differences in endowments in the quantity and fertility of land among households are a (physical and human) that create caste disparities (e.g., see Borooah (2005) and Kijima (2006)). Borooah (2005), for in- stance, using the National Council for Applied Economic Re- * The authors acknowledge the useful comments of the participants at the search (NCAER) survey showed that about a third of the July 2007 ECINEQ conference and the URPE/ASSA meetings in January income differentials in India could be attributed to discrimina- 2008. We are also grateful to the comments and suggestions of the referees. tion in the market place. Using the NSS consumption surveys, Final revision accepted: February 3, 2011. 1820 CASTE STRATIFICATION AND WEALTH INEQUALITY IN INDIA 1821 significant determinant of economic inequality. Quality and OBC whom we call Hindu forward castes (FC); and, non-Hindus quantity of schooling accessible to the children in urban and who are not SC, ST, or OBC whom we call non-Hindus (NH). semiurban areas can vary positively with household wealth. The 2002–03 survey, therefore, allows for the classification of In addition to these considerations that point toward the the population into five caste groups. We term this classifica- intrinsic importance of wealth as an index of economic status, tion “Scheme II.” It should be noted that the OBC, SC, and it also turns out that, in general, the distribution of wealth and ST individuals might belong to any religion. consumption (the most widely used index of economic status A brief note is in order regarding the category of caste. in India) among households differs considerably. 2 Thus, the Caste in India is defined differently along the “Varna” and available evidence on caste disparities based on other indica- the “Jati” schemes. The Varna scheme has four broad tors (such as consumption) can be enriched by the results on groups—Brahmins, Kshatriyas, Vaishya, Sudras—and those wealth disparities that we analyze here to obtain a fuller pic- people outside the Varna scheme, ranked in a descending or- ture of the distribution of economic status in India. der of ritual status. Brahmins were traditionally associated The relationship between overall wealth inequality and caste with the priestly and scholarly community. Kshatriyas were is analyzed in this study using the Yitzhaki decomposition or the ruling groups. Vaishyas were identified as groups associ- ANOGI (Yitzhaki, 1994; Frick, Goebel, Schechtman, Wagner, ated with trading, moneylending, and retailing. Sudras were & Yitzhaki, 2004). 3 This allows us to separate the overall the peasants and artisans. Among those outside the Varna inequality into within-group and between-group components, scheme, the so-called “untouchables” (the present day ) rather than obtaining conditional average effects of social were mostly associated with the rural, landless laboring com- divisions via regression-based decomposition methods such munity and the tribal groups (the present day Adivasis) were as the Oaxaca-Blinder method. Furthermore, the overlapping associated with those living on the fringes of or outside the set- parameters estimated using our chosen method permits the tled agricultural society. It is generally agreed upon that this is distinction between caste-stratification and caste-inequality. a textual scheme defined by the Brahmins. This is especially important in the context of ongoing debates The Jati scheme is very different. There are thousands of in Indian political economy about the questions of affirmative Jati groups that vary spatially and temporally in terms of action and the so-called “creamy layer.” 4 their ritual rankings, socioeconomic status, and occupations. It is also important to recognize that the caste system func- tions on the ground along different Jati orderings, thus creat- 2. DATA AND DEFINITIONS ing a bewildering variety of them, as well as a system that varies greatly across space and time (see Dumont (1970), The data used in this paper are from the two rounds of the Chatterjee (1993), and Gupta (2000)). Similarly, while certain All India Debt and Investment Survey (AIDIS) conducted in occupations are traditionally associated with certain Jati 1991–92 and 2002–03. Wealth is computed as the value of total groups, this relation is subject to significant flux over time. household assets net of indebtedness. Household assets are However, it has generally been the case that those outside defined as “physical assets like land, buildings, livestock, agri- the Varna scheme and those in the lower rungs of the Jati cultural machinery and implements, non-farm business equip- system have generally suffered from relatively low socioeco- ment, all transport equipment, durable household goods, and nomic status and tended to be concentrated in the menial financial assets like dues receivable on loans advanced in cash occupations. The categories we have constructed (Schemes I or in kind, shares in companies, and cooperative societies, and II), based on the limited information available in the sur- banks, etc., national saving certificates and the like, deposits veys, are neither Varna-based nor Jati-based; yet, they reflect in companies, banks, post offices, and with individuals” some elements of both and can serve, albeit with limitations, (NSS, 2005, p. 5). Debt is defined as cash loans payable. In to assess the persistence of entrenched social hierarchies in the absence of a better deflator, the Consumer Price Index contemporary India. for agricultural workers is used to make the 1991 and 2002 The problems associated with the wealth data in the surveys rural wealth values comparable across time. Similarly, the deserve a brief recapitulation (see, e.g., Subramanian and Consumer Price Index for industrial workers is used to make Jayaraj (2006) and Jayadev et al. (2007)). First, wealth distri- urban wealth values comparable across time. butions tend to be concentrated at the very top end. Unless a The unit of analysis is the household adjusted for its size. special effort is made to oversample the very wealthy, the con- That is, the household weight is multiplied by the household centration of wealth tends to be understated. This will artifi- size to obtain a distribution among persons. We use per capita cially reduce the overall measured inequality. Second, there wealth—household wealth divided by household size—as the is a tendency among people of all wealth groups to underre- measure of wealth. The implicit equivalence scale assumed port liability or debt. This tendency is exacerbated as wealth here is that there are no “economies of scale” associated with holdings rise. This will widen the gap between those with close wealth. (For the relative advantages and disadvantages of to no wealth and those that have some wealth. Third, the re- using this method for Indian wealth data, see Jayadev, ported assets may not be correctly valued. It has been found Motiram, and Vakulabharanam (2007).) that the reported values of even recent transactions tend to The definitions of caste groups are completely dictated by be lower than the market values in India. Fourth, there is a the data and do not adequately reflect the complex and layered tendency to hide illegitimate wealth that will lead to under- reality of caste in India. Both the AIDIS rounds allow for the counting of the assets owned by the wealthy. These problems classification of the entire population into three groups, viz., add up to a state in which households in the wealthier groups the Scheduled Castes or the “Dalits” (SC), Scheduled Tribes (more prevalent among the non-SC/ST population) appear to or the “Adivasis” (ST), and everyone else whom we call Other hold lower wealth than they actually have and the less wealthy Communities (OC). We term this classification “Scheme I.” groups (especially the SC/ST groups) report higher wealth The 2002–03 survey also enumerated membership in Other than they have. This will certainly reduce the overall measured Backward Classes (OBC) 5 and religious affiliation. Cross-tab- inequality, but it will also reduce the between-caste component ulating caste and religion allows for the separation of OC into of overall inequality. These problems might be reflected in our three distinct groups: OBC; Hindus who are not SC, ST, or findings. 1822 WORLD DEVELOPMENT

3. CASTE DISPARITIES IN WEALTH place is occupied by rural SCs and STs. Historically, SCs were considered the most landless in rural areas, and this trend con- We begin with a consideration of the gaps in median and tinues into 2002. It is interesting, however, that, in the urban mean values of wealth between the caste groups in the urban areas, the distance between the urban Hindu FCs and SCs is and rural areas (Table 1). 6 In 2002, the ranking of the ten the lowest in durables, consistent with the general perception groups (Scheme II) in terms of median wealth follows a pat- that the relatively poor have also attempted to acquire con- tern that one might expect a priori: the Hindu forward castes sumer durables (e.g., TVs) during the reform period along (FC) are at the top (urban, followed by rural). Immediately with the relatively affluent groups. But the same pattern was below them are the OBC groups and urban non-Hindus who not visible in the case of productive assets such as land. have quite similar levels of median wealth. At the bottom, Comparisons between the 2 years are possible only with the we have the most disadvantaged SC and ST groups (urban, three-group schema (ST, SC, and OC), differentiated by their followed by rural). The rural non-Hindus occupy a place rural versus urban location. Between 1991 and 2002, the rela- immediately above the most disadvantaged and below every- tively disadvantaged groups (SC and ST) experienced higher one else. If we were to use the mean values to rank the groups, rates of growth in average wealth than the OC group in both the pattern shifts somewhat. The top group—urban, Hindu the urban and rural areas. However, the medians tell a differ- FC—still maintains their lead and the rural SCs and STs held ent story, especially for the ST. 8 The wealth of the average their status as the worst-off. Rural Hindu FC slips to the third person in that group rose only 7% in the urban areas (as com- place and the second place is now taken by the urban, non- pared to 42% for the urban OC) and 21% in the rural areas (vs. Hindus. Rural non-Hindus occupy the fourth place, followed 25% increase for the rural OC). In contrast, the average SC by the urban OBC, urban ST, rural OBC, and then urban person experienced a robust increase in wealth of approxi- SC. 7 The reranking of the groups is an indication of the extent mately 40% over the same period in both the urban and rural to which within-group inequalities differ, as we shall see later. areas. In spite of the increases that did occur between the Table 2 presents the average values of the three most impor- 2 years, the average SC/ST person still had a considerable tant components of wealth, viz., land, buildings, and durable wealth disadvantage in 2002 (see Table 1, panel B). goods according to Scheme II in 2002. It is clear that in all Comparison of within-group distributions reveals that caste the three components of wealth presented there, the bottom divisions and the urban-rural divide act as distinct, yet

Table 1. Wealth by caste groups 1991 2002 Percent change Mean Median Share in population Mean Median Share in population Mean Median A. Mean and median values (in thousands of 2006 Rs.) Urban ST 38.5 19.5 0.7 67.1 20.8 0.7 74% 6% Urban SC 31.1 16.1 3.1 46.0 22.6 3.8 48% 40% Urban OC 88.4 33.1 20.6 123.4 47.1 20.9 40% 42% OBC 85.5 34.8 9.0 Hindu FC 169.3 77.7 8.4 Non-Hindus 109.5 34.5 3.4 Rural ST 24.6 15.8 8.0 34.1 19.0 7.3 39% 21% Rural SC 23.3 13.0 15.3 30.9 18.0 15.9 33% 39% Rural OC 58.8 31.1 52.2 77.6 38.9 51.4 32% 25% OBC 62.0 34.2 31.3 Hindu FC 105.4 60.1 14.1 Non-Hindus 93.9 25.9 6.0 All 55.7 24.9 100.0 75.3 32.0 100.0 35% 28% 1991 2002 Mean Median Mean Median B. Ratios Urban ST 0.69 0.78 0.89 0.65 Urban SC 0.56 0.65 0.61 0.71 Urban OC 1.59 1.33 1.64 1.47 OBC 1.14 1.09 Hindu FC 2.25 2.43 Non-Hindus 1.45 1.08 Rural ST 0.44 0.63 0.45 0.59 Rural SC 0.42 0.52 0.41 0.56 Rural OC 1.06 1.25 1.03 1.22 OBC 0.82 1.07 Hindu FC 1.40 1.88 Non-Hindus 1.25 0.81 All 1.00 1.00 1.00 1.00 Source: Authors’ calculations from the All India Debt and Investment Survey (AIDIS). Key: ST = Scheduled Tribes; SC = Scheduled Castes; OC = Other Communities; OBC = Other Backward Communities; FC = Forward Castes. Note: In 2006, the PPP exchange rate (Rupee per international dollar) was 15 according to the World Development Indicators database (series name: PPP conversion factor, GDP (LCU per international $) of the World Bank (http://databank.worldbank.org/: Accessed January 28, 2011). CASTE STRATIFICATION AND WEALTH INEQUALITY IN INDIA 1823

Table 2. Principal assets by caste, 2002 Mean values 2006 Rs. Percent of overall average Land Buildings Durables Land Buildings Durables Urban ST 24,718 23,314 6,943 59.2 104.7 140.5 Urban SC 17,139 19,065 4,370 41.1 85.6 88.4 Urban OBC 36,849 30,663 8,018 88.3 137.7 162.2 Urban Hindu FC 64,442 66,635 14,231 154.4 299.2 288.0 Urban Non-Hindus 42,095 45,208 8,417 100.8 203.0 170.3 Rural ST 21,438 8,322 1,698 51.4 37.4 34.4 Rural SC 17,397 10,115 1,987 41.7 45.4 40.2 Rural OBC 39,709 15,571 3,467 95.1 69.9 70.1 Rural Hindu FC 70,836 23,026 5,298 169.7 103.4 107.2 Rural Non-Hindus 65,977 18,307 4,019 158.0 82.2 81.3 All 41,747 22,271 4,942 100.0 100.0 100.0 Source: Authors’ calculations from the All India Debt and Investment Survey (AIDIS). interrelated, influences on the overall wealth distribution bottom 60% of rural OBC enjoy higher than overall values, the (Table 3). The differences between the distributions of the indi- top 40% in their distribution have values that are increasingly vidual groups are plotted on the vertical axis in Figure 1 as lower. The opposite pattern can be observed for the rural (pij pi), which expresses the deviation between the percentile non-Hindus. This is suggestive of a much higher degree of th value of the j group (pij) from the overall percentile value (pi) inequality and polarization among the rural non-Hindus. at the ith percentile. Strikingly, only the Hindu FC stay in the The direction and amount of the urban-rural disparity with- positive territory throughout the distribution, while the SC in caste groups varies across the distribution. This can be illus- and ST groups stay in the negative territory throughout the trated by defining the following statistic for group j at distribution. The values for the former became increasingly percentile i: higher than the overall values (most markedly for the urban, u r pij pij forward caste Hindus), while for the latter they became gij ¼ r ; ð1Þ increasingly lower as we move to higher echelons of the wealth pij distribution. This suggests that the richest of the forward where the urban-rural gap in wealth is expressed as a percent- castes are richer, on the average, than the richest of any other age of the percentile values (p) in the rural area for each caste group while the richest of the SC and ST groups are, on the group (the superscripts u and r represent, respectively, the ur- average, less rich than the richest of the other groups. The ban and rural areas). OBC and non-Hindus display more complex patterns. Lower Estimates of the urban-rural gaps are shown in Figure 2 for portions of the urban OBC and non-Hindus distributions have selected percentiles, with the bold horizontal reference line percentile values that are below the values for the overall dis- representing a situation of zero urban-rural disparity. The tribution, but the higher portions have values that are higher, wealth gap is in favor of rural individuals at the bottom of especially for the non-Hindus. The rural segments of these the distribution, irrespective of their caste group. This is a communities diverge from one another markedly. While the reflection of the incidence of land ownership (however meager

Table 3. Percentiles by caste group, 2002 (in thousands of 2006 Rs.) Percentile Urban ST Urban SC Urban OBC Urban FC Urban NH Rural ST Rural SC Rural OBC Rural FC Rural NH All 5 0.8 0.7 1.1 2.5 1.0 2.7 2.1 3.4 6.1 2.4 2.4 10 1.8 2.0 3.1 6.0 2.1 4.8 4.0 6.6 12.0 4.5 5.1 15 3.3 3.7 5.8 11.0 4.7 6.3 5.5 9.5 18.1 5.9 7.6 20 5.2 5.6 8.8 17.3 8.3 7.7 7.0 12.3 22.8 7.3 10.2 25 7.0 7.9 12.3 24.5 11.2 9.3 8.6 15.2 27.7 9.5 12.9 30 9.8 10.6 16.1 33.1 14.5 11.0 10.1 18.3 33.0 11.8 16.0 35 12.0 13.4 20.1 42.2 18.0 12.6 11.9 21.6 39.2 14.8 19.3 40 14.7 16.3 24.2 52.8 22.8 14.7 13.6 25.6 45.5 17.9 23.0 45 18.2 19.4 29.0 64.6 27.9 16.6 15.8 29.7 52.5 21.4 27.3 50 20.8 22.6 34.8 77.7 34.5 19.0 18.0 34.2 60.1 25.9 32.0 55 23.6 27.2 41.4 93.1 42.0 21.6 20.3 39.1 68.9 31.0 37.6 60 28.9 31.0 49.0 111.7 52.1 24.7 23.3 45.0 78.9 38.1 44.4 65 38.0 35.1 58.4 131.7 65.3 28.5 26.6 51.3 89.9 46.8 52.3 70 49.0 41.1 70.5 155.5 84.5 32.0 31.0 60.1 104.2 59.0 62.8 75 61.4 50.4 83.9 191.8 107.7 37.3 36.2 70.8 122.4 76.7 76.3 80 77.3 63.6 106.9 235.0 139.1 43.8 43.2 83.9 145.3 104.1 94.5 85 102.9 81.8 143.2 290.9 189.9 52.8 52.2 102.8 181.2 145.1 122.2 90 144.9 109.8 192.3 379.3 273.9 71.6 66.5 136.1 230.5 220.8 170.1 95 220.6 168.8 313.1 562.9 429.0 110.4 98.3 205.9 331.0 402.3 272.3 Source: Authors’ calculations from the All India Debt and Investment Survey (AIDIS). 1824 WORLD DEVELOPMENT

325 300 275 250 Urban ST 225 200 Urban SC 175 Urban OBC 150 125 Urban FC 100 75 Urban NH 50 Rural ST 25 0 Rural SC -25 Rural OBC

Thousands of Rs -50 -75 Rural FC -100 -125 Rural NH -150 -175 -200 5 101520253035404550556065707580859095 Percentiles

Figure 1. Deviation from overall percentile values by caste at selected percentiles, 2002 (in thousands of 2006 Rs.). Source: Authors’ calculations from the All India Debt and Investment Survey (AIDIS).

120% 100% 80% 60% ST 40% SC 20% OBC 0% FC -20% 5 101520253035404550556065707580859095 NH Urban-Rural Gap -40% -60% -80% Percentile

Figure 2. Urban-rural wealth gap (as a percent of rural wealth) by caste at selected percentiles. Source: Authors’ calculations from the All India Debt and Investment Survey (AIDIS). the size of the holding might be) in the rural areas among the degree of private capital accumulation in rural Punjab gener- poor, in contrast to the greater presence of propertyless indi- ally exceeds that found in other parts of rural India, partly viduals among the urban poor, irrespective of their caste related to the historical advantage enjoyed by this area in group. Notable differences exist among the castes in the per- the so-called Green Revolution. The urban wealth gap rises centile point at which the wealth gap turns in favor of the ur- steeply within the ST group in the top portions of the distribu- ban residents (i.e., the point at which their respective curves tions, a result consistent with the well-known fact that the cross above the zero line). At one extreme are the non-Hindus, rural tribal areas fall among the most economically backward for whom the switch favoring the urban residents occurs at the areas in India. 20th percentile; at the other extreme, the switch occurs only at We now revert to Scheme I in order to examine whether any the 50th percentile for the OBC. The variation in the amount significant differences could be found among the groups in of urban-rural disparity among the castes appears to be much terms of the changes in wealth that occurred between 1991 smaller at any given percentile point below the zero-line, that and 2002 across the entire distributions ( Figure 3). Urban is, when the disparity is in favor of the rural individuals. ST is the only group in which some of the percentile cutoffs Above the zero-line, when the disparity turns in favor of the in 2002 are roughly the same as, or lower than, their 1991 lev- urban persons, the amount of disparity (at any given percentile els. In contrast, the bottom half of the urban SC group gener- point) among the castes appears to vary much more. Clearly, ally saw a much higher boost in their wealth levels than their the evidence suggests that the wealth advantage enjoyed by the counterparts in the other groups. For the upper-middle por- urban individuals within every caste becomes higher at the tion (roughly from the 50th to 80th percentile), the urban higher percentiles, indicating the concentration of wealth in OC group experienced much faster growth than their counter- urban India. As we shall see later, this would be reflected in parts in other groups. The sharpest increase in wealth between the higher levels of inequality in the urban areas. The non- 1991 and 2002 among the top 20% occurred for the urban SC/ Hindus stand out as a clear exception to this rule because ST groups. A negative correlation between the initial amount the disparity in favor of the urban individuals in this group de- of wealth and the subsequent gain could be found in the clines after the 70th percentile. This appears to be driven to a bottom half of the urban SC and OC groups, as well as the large extent by the presence of very wealthy rural Sikhs in the rural ST and SC groups. In fact, the schedule for the rural upper rungs of the wealth distribution. As is well-known, the SC group slopes downward to the right almost throughout CASTE STRATIFICATION AND WEALTH INEQUALITY IN INDIA 1825

Figure 3. Percent change in wealth at selected percentiles by caste, 1991–2002. Source: Authors’ calculations from the All India Debt and Investment Survey (AIDIS). the distribution. Finally, the rural OC group displays the most 4. WEALTH INEQUALITY AND STRATIFICATION stable pattern: their schedule remained largely flat for most of the distribution. The overall picture of changes across the (a) Yitzhaki decomposition distributions suggests a pattern of wealth accumulation that is not heavily biased in favor of those at the top within each The picture of caste disparities in India sketched out so far caste group, with the exception of the urban ST. can be made richer by relating them to an analysis of overall Gaps in consumption and wealth between individual states wealth inequality. The tools of decomposition analysis allow and occupational groups/social classes have tended to widen us to analyze the within-group and between-group inequali- after 1991. Focusing on the major states, the range in per ca- ties. Further, it would allow us to develop summary measures pita wealth among states was very wide (Table 4). We find that that would express how stratified, in terms of its wealth, a in Punjab, the wealthiest state, per capita wealth was Rs. group is from another group or from the total population. 77,051 in 2002, which is about four times higher than Bihar, The method of Gini decomposition developed originally by the least wealthy state, with a per capita wealth of Rs. Shlomo Yitzhaki offers a unified framework for addressing 19,718. The growth rates among states have been substantially these issues. 10 different. Bihar, for example, experienced a growth rate of Let G be the Gini coefficient of wealth. The Yitzhaki decom- about 0.9% per annum in per capita wealth, while position allows us to separate G into intergroup inequality (Ib) has seen the fastest growth rate at about 4.9% per annum. and a remainder (Ir) that can be interpreted as intragroup In terms of the rich, middle, and poor states, the numbers tell inequality (Yitzhaki, 1994): a stark story, with the middle-income and rich states experi- G I I 2 encing much faster asset growth rates annually than the poor ¼ b þ r ð Þ states. It is also interesting to note that growth in wealth has The amount of intergroup inequality is: been fastest in the urban areas of the middle-income states, re- 2covðl ; F ðyÞÞ gions which include dynamic urban centers such as Hyderabad I ¼ i oi ; ð3Þ and Bangalore. b l We used a seven group-schema to examine wealth gaps where y is wealth, l is mean wealth for all persons, l is mean across social classes. The categories were: urban elites, i wealth for group i, and F y is the mean rank of group i, that urban lower service workers, urban manual workers, rural oið Þ is, the average position of the members of a group in the over- elites, rural middle groups, rural workers, and rural non- all wealth distribution. 11 Thus, the amount of intergroup agricultural groups (Table 5). 9 As one would expect, the urban inequality is twice the covariance between the mean amounts and rural elites are considerably wealthier than the other of wealth and mean ranks of groups divided by the mean groups. The urban groups as a whole have seen more rapid wealth for all individuals. 12 growth in wealth accumulation. If one looks at mean values, The remainder term is calculated as: the two elite categories continue to dominate in absolute terms X but the growth rates tell a more complex story. In the urban Ir ¼ siGiOi; ð4Þ areas, there were very little intergroup differences in growth i rates, and, therefore, the disparities among them in terms of mean values remained roughly constant between 1991 and where si is the share of group i in aggregate wealth, Gi is the 2002. However, in terms of median values, the urban elites Gini coefficient of the wealth distribution within group i, increased their distance from the other urban groups. In rural and Oi is the overlapping index for group i. The Yitzhaki areas, the rural elite and the rural non-agricultural groups decomposition provides group-specific measures of overlap- have seen their wealth grow most rapidly, consistent with ping, unlike the standard decomposition of the Gini where the notion of a divergence in that sector. This is true in terms only a summary measure of overlapping by all groups can of medians as well. As such, although the interpersonal be obtained. The index of overlapping proposed by Yitzhaki wealth distribution does not show substantial acceleration in is a measure of the degree to which the range of wealth in each intergroup disparities, both class and caste groups convey a group overlaps with the range of wealth for all persons. Over- similar picture regarding the underlying dynamics of wealth lapping can thus be seen as the opposite of stratification: the distribution and the divergence between elite groups and higher the amount of overlap between a group and the popu- others. lation, the less stratified they are as a group in terms of wealth 1826 WORLD DEVELOPMENT

Table 4. Average wealth by state and sector, 1991 and 2002 (in 2006 Rs.) States 1991 2002 Implied annual growth rate Overall Rural Urban Overall Rural Urban Overall Rural Urban Andamans 15,693 15,041 17,184 34,551 32,060 39,816 7.40% 7.10% 7.90% Andhra 14,511 12,727 20,152 21,253 15,757 36,394 3.50% 2.00% 5.50% Arunachal Prades 12,196 13,030 6,096 12,938 13,118 11,662 0.50% 0.10% 6.10% Assam 11,866 10,830 24,161 15,883 14,568 28,338 2.70% 2.70% 1.50% Bihar 17,698 17,509 18,943 19,450 18,569 27,091 0.90% 0.50% 3.30% Chandigarh 30,897 17,134 32,986 67,792 35,299 70,476 7.40% 6.80% 7.10% Chattisgarh 19,800 18,613 27,206 Dadra/Nagar Hav 14,285 13,484 25,340 30,084 28,859 44,658 7.00% 7.20% 5.30% Daman/Diu 34,323 33,888 35,593 24,301 16,960 38,507 3.10% 6.10% 0.70% Delhi 69,078 75,593 68,339 56,079 26,214 62,872 1.90% 9.20% 0.80% Goa 50,227 49,108 51,728 57,316 47,855 68,722 1.20% 0.20% 2.60% Gujarat 22,974 18,890 31,345 35,711 32,310 42,154 4.10% 5.00% 2.70% Haryana 51,477 57,057 32,090 67,517 69,895 60,645 2.50% 1.90% 6.00% Himachal Pradesh 26,530 25,389 41,530 54,542 54,064 59,178 6.80% 7.10% 3.30% J&K 28,683 26,584 42,674 62,239 53,723 94,486 7.30% 6.60% 7.50% Jharkand 17,000 15,452 23,532 Karnataka 20,489 18,843 24,672 29,315 25,219 38,851 3.30% 2.70% 4.20% Kerala 37,133 35,108 43,652 61,847 56,929 76,356 4.70% 4.50% 5.20% Lakshadweep 64,376 50,605 76,155 60,499 50,812 65,826 0.60% 0.00% 1.30% Madhya Pradesh 18,098 16,834 22,816 26,547 22,816 38,570 3.50% 2.80% 4.90% Maharashtra 23,587 17,972 33,344 32,607 26,594 41,092 3.00% 3.60% 1.90% Manipur 16,932 16,335 18,593 23,234 18,721 35,479 2.90% 1.20% 6.10% Meghalaya 13,973 10,471 30,702 31,789 24,487 72,137 7.80% 8.00% 8.10% Mizoram 10,667 7,794 20,248 33,040 15,136 64,977 10.80% 6.20% 11.20% Nagaland 16,564 14,350 20,441 76,641 85,387 55,587 14.90% 17.60% 9.50% Orissa 9,564 8,701 16,489 12,307 10,556 23,878 2.30% 1.80% 3.40% Pondicherry 25,147 16,106 30,306 36,258 24,705 42,703 3.40% 4.00% 3.20% Punjab 55,510 56,905 52,410 75,645 85,599 54,230 2.90% 3.80% 0.30% Rajasthan 28,695 27,612 32,617 34,437 31,936 43,445 1.70% 1.30% 2.60% Sikkim 25,826 26,839 16,932 25,756 23,195 46,875 0.00% 1.30% 9.70% Tamil Nadu 18,844 13,975 27,739 27,748 23,241 37,033 3.60% 4.70% 2.70% Tripura 12,287 10,654 32,805 13,235 10,559 32,446 0.70% 0.10% 0.10% Uttar Pradesh 24,826 23,860 28,716 28,860 28,801 29,077 1.40% 1.70% 0.10% Uttaranchal 40,757 39,990 44,035 West Bengal 14,258 11,494 22,450 19,993 15,926 33,801 3.10% 3.00% 3.80% Addendum: Poor 21,054 20,022 25,745 25,384 23,826 32,132 1.70% 1.60% 2.00% Middle 18,670 16,569 24,879 27,699 22,968 41,220 3.70% 3.00% 4.70% Rich 29,256 27,818 32,273 42,958 42,499 43,927 3.60% 3.90% 2.80% Source: Authors’ calculations from the All India Debt and Investment Survey (AIDIS).

Table 5. Wealth by status group, 1991 and 2002 (in 2006 Rs.) Status category Mean Median 1991 2002 Ratio (2002 to 1991) 1991 2002 Ratio (2002 to 1991) Urban elite 65,397 186,328 2.85 32,149 91,010 2.83 Urban middle 31,477 91,943 2.92 17,075 45,692 2.68 Urban manual 18,612 53,232 2.86 9,501 25,120 2.64 Rural elite 64,727 190,512 2.94 51,328 147,529 2.87 Rural middle 30,529 75,136 2.46 23,775 58,359 2.45 Rural lower 9,807 25,118 2.56 6,726 16,567 2.46 Rural non agricultural l 21,583 66,277 3.07 13,893 35,924 2.59 Urban total 30,505 92,908 3.05 14,363 36,672 2.55 Rural total 20,352 53,501 2.63 12,599 26,039 2.07 Source: Authors’ calculations from the All India Debt and Investment Survey (AIDIS).

(Yitzhaki, 1994, p. 148–149). This feature of the decomposi- extent to which castes occupy or do not occupy different seg- tion is crucial for us since our objective is to ascertain the ments of the wealth distribution. CASTE STRATIFICATION AND WEALTH INEQUALITY IN INDIA 1827

The amount to which group i overlaps with the overall dis- bers of group i produced by Fi and Fji are identical (Yitzhaki, tribution is defined as: 1994, p. 150–152).

coviðy; F oiðyÞÞ Oi ¼ ; ð5Þ (b) Within-group versus between-group inequality coviðy; F iðyÞÞ where F (y) is the function that assigns to the members of We now turn to the results of the Yitzhaki decomposition oi for our data. 14 It is useful to begin with the estimates of with- group i their ranks in the overall distribution, Fi is the function that assigns to the members of group i their ranks in the in-group and between-group caste inequality (Table 6). Overall wealth inequality shows very little change between 1991 and wealth distribution within that group, and covi indicates that the covariance is according to the distribution within 2002. The share of within-group and between-group inequality 13 in overall inequality also remains roughly the same between group i . The minimum value of Oi is given by the share of group i in the population and its maximum value is equal the 2 years. The within-group inequality (the Ir term in Eqn. to 2. When the index equals the minimum possible value, it (2)) accounts for the bulk of overall inequality in both years. suggests that the group in question is a perfect stratum, that The domination of the within-group term indicates there are is, it occupies an exclusive segment of the overall wealth other wide variations in the characteristics of household mem- distribution. If a particular group has a range of wealth that bers that are also expected to contribute to wealth differentials coincides with the range of wealth of the population, then within castes—occupation, age, education, industry of the index will be equal to 1. Finally, if the index is >1, the employment, and number of earners in the household, to men- distribution of wealth within the group is much more tion a few. Additionally, we would expect product mix and fer- polarized than in the overall distribution. This can happen if tility, among other things, to also have effects on the wealth of the members of the group constitute two strata, one that farmer households. In 2002, we found that the share of within- has much higher and the other that has much lower wealth group inequality is somewhat lower (87%) under the more than l, the average wealth of all individuals in all groups elaborate Scheme II (ten groups as compared to six in Scheme (Milanovic & Yitzhaki, 2002, p. 162–163). I). In general, increasing the number of groups is likely to in- The index of overlapping defined in Eqn. (5) is constructed crease the share of between-group inequality. Since the sub- from indexes that indicate the amount by which a group over- groups included in the OC group are themselves quite laps with each of the other groups: different from each another in terms of their average wealth X and average rank, the modest increase in the share of be- Oi ¼ p þ p Oji ð6Þ tween-group inequality under Scheme II is not out of line with i j 15 j–i our expectations. where pi is the share of group i in the total population and Oji is (iii) Within-caste inequality and overlapping the index of overlapping of group j by group i. Since the over- lapping of a group by itself is equal to 1 by definition, its con- The results from decomposing the remainder term along tribution to Oi is equal to its relative size. The index of caste lines are shown in Table 7. We also plot the Gaussian overlapping of the overall distribution by a group is the kernel density estimates of wealth distribution by group in weighted sum of overlapping of each of the other groups by Figure 4. 16 In both the years, within-group inequality was that group, with the relative size of each group serving as the generally higher among the urban groups than among their weights. rural counterparts (Table 7), consistent with our findings on In turn, the group-by-group overlapping indexes are calcu- the urban-rural gaps in percentile values discussed earlier. lated as: The higher inequality in the urban areas is mainly due to the coviðy; F jiðyÞÞ fact that the principal assets in the Indian context, land and O ¼ ; ð7Þ 17 ji cov ðy; F ðyÞÞ buildings, are more unequally distributed in the urban areas. i i In 2002, the Gini of land and buildings were, respectively, 0.80 where Fji is the function that assigns members of group i their and 0.73 in the urban areas, and 0.71 and 0.58 in the rural ranks in the wealth distribution of group j. The index Oji indi- areas. In turn, the greater urban inequality in the distribution cates the extent to which the wealth of individuals in group j of these assets stems from the lower ownership rates and high- falls in the range of wealth of individuals in group i; the higher er inequality among the owners. 18 The higher inequality the fraction of group j that falls in the range of group i, the among the building owners in urban areas is quite consistent higher will be the value of Oji. For a given fraction of group with the well-known fact that urban real estate markets are j that falls in the range of group i, the closer the wealth of far more developed and feature much costlier dwellings. the individuals in that fraction are to the mean wealth of The only groups that showed a notable increase in inequal- group i, the higher will be the value of Oji. The index can take ity between 1991 and 2002 were the ST groups. This is values between 0 (no overlap) and 2. Perfect overlap occurs especially true for the urban ST and is consistent with our ear- when the index equals 1, indicating that the rankings of mem- lier finding about the big increases between the years in the

Table 6. Within-group and between-group inequality, 1991 and 2002 Gini points Percentage shares 1991 2002 1991 2002 Scheme I Scheme II Scheme I Scheme II Overall Gini 0.648 0.655 0.655 100.0 100.0 100.0 Within group 0.595 0.599 0.572 91.9 91.4 87.4 Between group 0.053 0.056 0.083 8.1 8.6 12.6 Source: Authors’ calculations from the All India Debt and Investment Survey (AIDIS). 1828 WORLD DEVELOPMENT

Table 7. Within-group inequality and overlapping, 1991 and 2002 1991 2002 Population Share Wealth Share Gini Overlap Population Share Wealth Share Gini Overlap Urban ST 0.007 0.005 0.628 1.049 0.007 0.006 0.725 1.137 Urban SC 0.031 0.017 0.627 1.056 0.038 0.023 0.632 1.051 Urban OC 0.206 0.327 0.700 0.993 0.209 0.342 0.683 0.966 Urban OBC 0.090 0.102 0.677 1.016 Urban FC 0.085 0.190 0.648 0.840 Urban NH 0.034 0.050 0.713 1.054 Rural ST 0.080 0.035 0.526 0.913 0.073 0.033 0.568 0.969 Rural SC 0.153 0.064 0.573 0.973 0.159 0.065 0.557 0.947 Rural OC 0.522 0.551 0.595 0.918 0.514 0.530 0.609 0.929 Rural OBC 0.313 0.258 0.580 0.932 Rural FC 0.141 0.197 0.563 0.791 Rural NH 0.060 0.075 0.734 1.095 All 1 1 0.648 1 1 1 0.655 1 Source: Authors’ calculations from the All India Debt and Investment Survey (AIDIS). percentile values of the upper tail of the ST wealth distribution a “nouveau rich” and growing income polarization within (Figure 3). For the urban ST group, the increase in inequality the ST groups. This is also suggested by the lengthening of appears to have been driven by two factors. First, the owner- the right tail of the ST distributions between the 2 years ship rate of buildings declined for them from 1991 to 2002 by (Figure 4, panels A and B). 23 However, being “rich” within 9% points to 68% and the Gini ratio of buildings for building the ST group is very different from belonging to the top of owners grew rapidly from 0.67 to 0.81. 19 Furthermore, the rel- the urban Indian wealth distribution. The kernel density ative importance of financial assets in the urban ST portfolio estimates (Figure 4) and the estimates of percentile cutoffs de- rose sharply from 5% to 12% during the period. Because the picted earlier (Figure 1) convey this point vividly. To be in the distribution of financial assets is extremely unequal in favor top 10% of the ST distribution in 2002 required a minimum of of the wealthy (a Gini ratio of 0.87 in both years), the expan- 144 thousand rupees (in 2006 prices); this amount, as it turns sion in its share of wealth contributed substantially to the rise out, was insufficient to belong to even the top 30% of the Hin- in inequality. 20 The more modest growth in rural ST inequal- du FC distribution. 24 ity seems to have been mainly due to the rapid growth in Apart from the index of overlapping for each group with the wealth held as land, especially in the Northeast. 21 A secondary overall population, the Yitzhaki decomposition also allows us role might also have been played by the sharp rise in the con- to estimate pair-wise indexes of overlapping among the groups centration of financial assets in the top echelons of the wealth (Eqn. (7)). The estimates of the resulting overlapping matrix distribution. 22 using Scheme II for 2002 are shown in Table 8 (panel A). Turning now to the evidence on stratification, the estimates The reference group (the caste represented by the subscript i in Table 7 indicate that the rural groups have lower values for in the overlapping index Oji) is shown in the rows of the table; their overlapping indexes than the urban groups, a result that other groups are shown in the columns (the castes represented is not surprising in light of the considerable rural–urban by the subscript j). The highest degree of overlap for the urban wealth gaps that were discussed above. When compared rela- (rural) FC was found with the rural (urban) FC indicating that tive to their shares in population, the rural OC group has a their distributions are much closer to one another than with substantially lower degree of overlapping than the rural SC/ any other group. The kernel density estimates shown in Fig- ST groups. The overlapping index for the urban OC is almost ure 4 (panels B and C) shows how little the upper tail of the 1 in 1991 and slightly lower in 2002, indicating the close sim- distribution of the FC groups overlaps with the distributions ilarity between their distribution function and the distribution of other groups (with the exception of urban OBC). 25 Over- function for the entire population. Estimates for the sub- lapping of each of the other groups by the urban ST, SC, groups included in OC in 2002 (Scheme II) show that the Hin- OBC, and the non-Hindus groups is generally high. In con- du FC is the group with the lowest amount of overlapping trast, the overlapping of each of them by the Hindu FC is among all groups, while the non-Hindus rural and urban much lower. groups take, respectively, the second and third places in terms The reason behind this apparent discrepancy can be under- of overlapping. The higher degree of overlapping by the rural stood by considering the overlapping between the urban ST non-Hindus as compared to their urban counterparts is an and urban FC. The overlapping of urban ST by urban FC is exception to the pattern observed for the other groups, and only 0.716. This reflects the fact that there are relatively few is related to our earlier discussion regarding the rich Sikh rural urban ST individuals in the urban FC wealth range. Conse- residents ( Figure 2 and related text). Both groups of non- quently, the ranks of urban FC individuals, when each of them Hindus have an overlapping index that exceeds one, indicating is placed in the wealth distribution of urban ST, did not differ wealth polarization within these groups. much from each other for a large number of them, thus reduc- The urban ST and SC groups are hardly homogenous ing the size of the covariance in the numerator of Eqn. (7).On groups. Both have values exceeding 1 for their overlapping in- the other hand, there are relatively more urban FC individuals dexes, indicating that there might be two distinct strata, one in the urban ST wealth range and, therefore, when each ST quite rich and the other extremely poor, within each of these individual placed in the FC distribution their ranks differ con- groups. This is most striking in the case of the urban ST in siderably from each other for a sizeable number of cases. This 2002. Considered in conjunction with the sharp rise in with- is reflected in the much larger (1.05) value of the overlapping in-group inequality, it appears that there is an emergence of of urban FC by urban ST. CASTE STRATIFICATION AND WEALTH INEQUALITY IN INDIA 1829

Figure 4. Kernel density estimates of wealth distribution by caste, 1991 and 2002. (A) 1991, (B) 2002, (C) Other communities (OC) by caste group, 2002. Source: Authors’ calculations from the All India Debt and Investment Survey (AIDIS).

The overlapping of rural ST and SC by each of these groups that the wealth gap in favor of the urban residents of these is higher than the overlapping of their urban counterparts by groups increased sharply at the higher percentiles (Figure 2). the same groups. This is in line with our earlier finding that For example, the overlapping of rural ST by rural SC is 1830 WORLD DEVELOPMENT

Table 8. Matrices of overlapping and ranks, 2002 Urban ST Urban SC Urban OBC Urban FC Urban NH Rural ST Rural SC Rural OBC Rural FC Rural NH A. Overlappinga Urban ST 1 1.045 1.107 1.131 1.052 1.065 1.042 1.193 1.266 1.051 Urban SC 0.938 1 1.009 0.928 0.933 1.051 1.032 1.119 1.111 0.963 Urban OBC 0.881 0.916 1 1.062 0.951 0.905 0.885 1.068 1.179 0.928 Urban FC 0.716 0.722 0.842 1 0.827 0.681 0.662 0.866 1.037 0.776 Urban NH 0.915 0.944 1.041 1.133 1 0.928 0.906 1.100 1.230 0.970 Rural ST 0.855 0.925 0.918 0.809 0.842 1 0.977 1.040 1.000 0.879 Rural SC 0.852 0.924 0.889 0.739 0.812 1.021 1 1.017 0.934 0.868 Rural OBC 0.794 0.849 0.908 0.903 0.838 0.851 0.831 1 1.070 0.826 Rural FC 0.654 0.678 0.792 0.903 0.750 0.625 0.608 0.836 1 0.697 Rural NH 0.937 0.973 1.075 1.163 1.029 0.971 0.945 1.148 1.277 1 B. Ranksb Urban ST 0.5 0.502 0.419 0.298 0.417 0.522 0.536 0.410 0.301 0.448 Urban SC 0.498 0.5 0.410 0.282 0.409 0.526 0.541 0.402 0.283 0.445 Urban OBC 0.581 0.590 0.5 0.362 0.492 0.622 0.634 0.502 0.381 0.531 Urban FC 0.701 0.718 0.638 0.5 0.619 0.748 0.758 0.653 0.544 0.656 Urban NH 0.582 0.590 0.508 0.381 0.5 0.616 0.628 0.509 0.400 0.534 Rural ST 0.478 0.473 0.378 0.251 0.383 0.5 0.517 0.361 0.237 0.423 Rural SC 0.463 0.459 0.366 0.242 0.371 0.483 0.5 0.348 0.227 0.409 Rural OBC 0.590 0.598 0.497 0.347 0.491 0.639 0.652 0.5 0.363 0.538 Rural FC 0.699 0.717 0.618 0.456 0.600 0.763 0.773 0.637 0.5 0.650 Rural NH 0.551 0.555 0.469 0.344 0.466 0.576 0.591 0.461 0.350 0.5 Source: Authors’ calculations from the All India Debt and Investment Survey (AIDIS). a Each number shows the overlapping of the group named in the column label by the group named in the row label. b Each number shows the average rank of the group named in the row label in the distribution of the group named in the column label.

1.02, while the overlapping of urban SC by rural SC is lower, other groups in terms of this indicator, too. If we look at at 0.92. Further, the overlapping of rural OBC, FC, and NH the entries under the column labeled “Urban FC,” it is evident groups by, respectively, the rural SC and ST is higher than that the average rank of all groups except rural FC is placed the overlapping of urban OBC, FC, and NH groups (e.g., below the 40th percentile of the urban FC wealth distribution; the overlapping of rural FC by rural SC was 0.934, as against the rural FC’s average rank is at the 45th percentile. Similarly, only 0.739 for urban FC). This suggests that the distributions the entries in the “Rural FC” column are also below the 40th of rural ST and SC are more similar to each other than to the percentile for all groups except, obviously, their urban coun- members of their own community in the urban areas and that terparts. 27 Viewed from another angle, this means that the they have at least some members with amounts of wealth that average ranks of all the other groups are at their lowest levels match the wealth of wealthier individuals from the rural resi- when they are placed in the distribution of forward castes. The dents of other communities. most numerous of the groups, the rural OBC, have a mean However, the rural–urban patterns of overlapping are quite rank above the 50th percentile in the distributions of all SC different for the rural OBC and FC groups. Their wealth dis- and ST groups and close to the 50th percentile for the non- tribution is more similar to the urban residents of their own Hindus and urban OBC distributions. communities than to the SC or ST in the rural areas. For The average rural ST and SC ranks are below the 40th example, the overlapping of rural SC by rural OBC is only percentile in the distributions of all other non-ST/SC groups, 0.831, while the overlapping of urban OBC by rural OBC is except for the non-Hindus, where their ranks were at the higher, at 0.908. Similarly, the overlapping of rural ST by rural 41–42nd percentile and slightly below the middle in the distri- FC is quite low at 0.625 compared to the overlapping of urban butions of their urban counterparts. Similar results are also FC by rural FC that stood at 0.903. The overlapping relation found for the urban ST and SC. Their average ranks are in between the rural OBC and rural FC, as well as that between the bottom half of the distribution of all other groups, except the rural NH and rural FC, mirrors the relationship between that of their rural counterparts, where they are slightly above urban ST and urban FC that was discussed above. the middle. The ranking is the lowest (roughly at the 30th per- The index of overlapping is sensitive to extreme values be- centile) in the FC distributions, somewhat higher (roughly at cause it depends on the ranks and amounts of wealth of indi- the 40th percentile) in the OBC distributions, and the highest viduals in each caste. 26 Hence, an examination of the ranking (roughly at the 45th percentile) in the NH distributions. of one caste in terms of another is instructive. Such an exercise can answer the following type of question: at what percentile of the forward caste wealth distribution is an average SC per- 5. CONCLUSION son located? The average rank of each caste in the distribution of other castes can be displayed in a matrix of ranks. Along The average SC/ST person in India has a substantial disad- the row labeled “Urban ST,” for example, we can read off vantage in wealth relative to people from other groups in both the average rank of an individual in that group in the wealth years of analysis. In a worrisome trend, the relative median distribution of each of the other groups. Since the ranks are wealth of the rural and urban ST was, in fact, lower in 2002 normalized to lie between 0 and 1, the average rank of a group than 1991. Among the other groups, the Hindu forward castes in its own distribution will be 0.5 (i.e., the 50th percentile). are the clear leaders in median wealth in both the rural The matrix of ranks for caste groups under Scheme II is and urban areas. For the second survey year (2002–03), the shown in Table 8 (panel B). Forward castes clearly dominate OBCs and non-Hindus occupied positions that placed them CASTE STRATIFICATION AND WEALTH INEQUALITY IN INDIA 1831 noticeably above the SC/ST groups, but significantly below and non-Hindus are quite stratified from the rural SC and the FC in terms of median wealth values. ST groups; in fact, they are less stratified from their urban A similar picture of SC/ST disadvantage and forward caste counterparts. This suggests that the intra-rural stratification advantage is evident throughout the distributions in terms of along caste lines was greater for the rural SC and ST than gaps in percentile values. Estimates of the matrix of ranks the urban-rural stratification was for the non-SC/ST groups. for caste groups also confirm the existence of sizeable wealth Strong evidence of a polarized distribution could be detected disparities between the forward castes and everyone else. Con- for the urban ST, urban NH, and rural NH (overlapping index sidered in conjunction with the findings documented in other >1). These groups also have within-group inequality that is studies regarding the considerable shortfalls of the average much higher than the overall inequality. SC/ST person in consumption, education, and development Substantial increases in within-group inequality between indices, the picture that emerges is one of comprehensive 1991 and 2002 were observed only for the urban and rural and persistent disadvantage for the SC and ST groups in con- ST. In the urban areas, the increase could be attributed to a temporary India. decline in homeownership rate and a dramatic rise in the Our decomposition analysis shows that inequality between inequality of building (mainly home) wealth. This was respon- castes (between-group inequality) accounts for as much as sible for the deterioration in the median wealth of the group 8% to 13% of overall wealth inequality. The major determi- compared to the rest of the population. In the rural areas, nant of between-group inequality is the large gap between the growth in ST inequality might be due to the rapid appre- SC/ST groups (especially rural) and the forward castes (espe- ciation of land values in the Northeast over the last decade cially urban) in average wealth. It would be interesting to com- and the sharp increase in the concentration of financial assets pare this result to the results that arise from using other at the top of the distribution. While the growth in ST inequal- variables to classify the population (e.g., age or education). ity may be suggestive of the emergence of so-called “creamy However, it is reasonable to expect that irrespective of the layer,” it should be tempered by at least two other main find- “grouping variable” used, the share of within-group inequality ings of our study. The average ST household has fallen behind is likely to be the dominant factor in overall inequality. There the overall population in terms of wealth accumulation and are, inevitably, other wide variations in the characteristics of the ST creamy layer has levels of wealth that are considerably households that, when taken together, are likely to contribute below the creamy layer of the nation and especially the creamy more than the classifying variable itself to wealth differentials layer of forward caste Hindus. The period of study of this pa- within any group. per also coincides with the increased globalization of the In- We also found that the urban and rural forward caste dian economy. Some have argued that that globalization groups are substantially stratified from the other groups in helps the disadvantaged groups to narrow the gap with the terms of their wealth distribution, especially from the SC/ST privileged groups because the rent-seeking state is pushed back groups. Among the latter, the rural SC and ST were notably by the less discriminating global markets. Our findings sug- stratified from their urban counterparts; but, there was hardly gest, however, that those that are better positioned to benefit any stratification between them, that is, the overlap of one from globally interconnected markets through access to better group by the other was almost perfect. This is indicative of education or capital or other resources tend to maintain sub- how the urban-rural divide in wealth drives a wedge within stantial advantages. these minority groups. On the other hand, the rural OBC

NOTES

1. It is widely acknowledged that caste is a highly heterogeneous 5. The determination of disadvantaged groups (OBCs) was made category. For example, tabulations based on the 1911 census for Uttar according to the 1931 census. Many changes have occurred since in the Pradesh (the largest Indian state, then known as United Provinces) socioeconomic status of these groups, but this category does not reflect showed that among the 42 castes considered in the census, “each caste these changes, in part because subsequent censuses have not collected contained landless labourers, cultivators, as well as landlords” (Chaudhury, information on caste. 2004, p. 1990). Economic differentiation within castes is the rule rather than the exception, then and now. 6. We separate the rural households from the urban, as we believe that the wealth accumulation and income generation dynamics vary signifi- 2. One method of comparing the two distributions is to observe how the cantly across this sectoral division. ranks of the households differ across the distributions. Thus, if we use quintiles to rank households, we would expect all the households in a given quintile of the consumption distribution to also belong to the same 7. When we analyze the relation between indebtedness and wealth, quintile of the wealth distribution. Conducting this exercise for the 2002 normally those groups that have higher net worth are also the ones with data showed that only 52% of those in the top wealth quintile were also in higher levels of absolute debt. However, the debt/wealth ratios are higher the top consumption quintile. For the other quintiles, the correlation was for SC and ST groups than with the other caste groups. substantially weaker and at least two-thirds of the units in a given quintile of wealth distribution were located in a different consumption quintile. 8. The difference between the outcome based on means and medians indicates the growth in inequality, as we will discuss later. 3. ANOGI stands for “Analysis of Gini.” 9. We derived the occupational/class status variable by using the NCO 4. The “creamy layer” refers to the emergence of an economically well-off classification of 1968 as well as details on landholdings and the group within castes whose average member is worse-off relative to the rest household type (see Jayadev, Motiram, and Vakulabharanam (2010) for of the population. details). 1832 WORLD DEVELOPMENT

10. In contrast, the other available methods of inequality decomposition 18. In 2002, roughly a third of the urban households did not own any by population subgroups allow only for the identification of between- land and a similar proportion did not own buildings. In contrast, only 6% group and within-group components of overall inequality. For a compre- of rural households did not own any land and there were a similar hensive overview, see Deutsch and Silber (1999). percentage of households who did not own any buildings. Among the owners, the Gini coefficient for land in the urban areas was 0.72 compared 11. For example, if the mean rank is 0.25 for SC, then the average SC to 0.69 in the rural sector, suggesting that the urban-rural difference in person’s position in the wealth distribution for all persons will be at the land inequality was mainly a function of the ownership rate differential. In 25th percentile. contrast, among building owners, the Gini coefficient for buildings was 0.63 in the urban areas, compared to only 0.55 in the rural areas, 12. In contrast, the between-group component in the standard decom- indicating that in the case of this asset, both differentials, that is, in the position would be equal to twice the covariance between the wealth of ownership rate and inequality among the owners, contributed to the each group and the rank of each group’s mean wealth divided by overall greater inequality in the urban areas. mean wealth. The Yitzhaki decomposition takes into account the ranking of each individual within each group in the overall distribution. 19. In contrast, among all urban residents, ownership rate increased by 6% points to 72% and the Gini ratio of building wealth declined sharply 13. In theory, the functions are actually cumulative distribution func- from 0.67 to 0.63. tions. However, when working with actual samples, the cumulative distribution function is estimated by the rank of the observation and hence 20. This was reflected in the sharp increase in the concentration our description of the functions as rank-assigning functions (Yitzhaki, coefficient of financial assets with respect to wealth. The ratio rose from 1994, p. 149, n.1). 0.59 to 0.71 between 1991 and 2002. It should be noted that for urban residents as a whole the share of financial assets rose from 8.6 to 10.3, a 14. Decomposition of the Gini by groups was performed using the much less dramatic change. ANOGI module for STATA (version 9). 21. The per capita average value of land rose by a whopping 165% for 15. The extent of between-caste inequality that we report here (8–13% of the rural ST in the Northeast between 1991 and 2002; the increase for the overall wealth inequality in 2002) is hard to compare with the estimates of rural ST in the rest of India was a much lower 40%. between-group inequalities reported in some other recent studies. There are a variety of obstacles to direct comparisons, including the differences 22. The share of the top decile of the rural ST wealth distribution among studies in the measure of between-group inequality (e.g., Gini increased their share of financial assets from 13% to 52% over the 1990s decomposition versus Theil decomposition), the basis of ‘grouping’ (e.g., while their share of total wealth increased “only” from 46.0% to 50.5%. caste versus ethnicity versus social class), and measure of economic status or resources (e.g., wealth versus income). With this caveat, it is interesting 23. Due to the log scale used in the figure the stretching of the tail appear to note that in a study of inequality of consumption expenditures using a to be somewhat smaller than what would be visible in a graph with the three-way classification (SC, ST and all others) in Kerala, Deshpande actual amounts. (2000) found that between-group inequality was less than 1 percent of total inequality in 1993–94. Comparing Brazil and US around 2000, 24. The p90 for the urban Hindu FC is 2.6 times the p90 for the urban Bourguignon, Ferreira, and Leite (2002), found that the household head’s ST. race accounted between 5% and 14% of overall income inequality (depending on the measure of inequality) in Brazil and 2% in the US. 25. The lack of overlap appears to be smaller due to the log scale used in the figure than it would have been in a figure with the actual amounts. 16. We have transformed the amounts of wealth into natural logarithms in this figure. Zero and negative amounts of wealth were set equal to zero 26. As noted above, the biases in the data would most likely exaggerate in the transformation to include those observations also. the overlap component between the privileged and the underprivileged caste groups. 17. In 2002, the combined share of land and buildings in total household wealth was 89.1% and 78.1%, respectively, in the rural and urban areas. 27. The sum of the average rank of group j’s rank in group i’s and the The remainder of wealth was made up of (in percent, urban areas in average rank of group i’s rank in group j’s distribution will be equal to 1. parentheses): Consumer durables, 5.3 (8.7); other real assets, 6.0 (5.8); financial assets, 2.5 (10.3); and, debt 2.9 (2.9).

REFERENCES

Banerjee, A., Duflo, E., Ghatak, M., & Lafortune, J. (2009). Marry for Chaudhury, P. (2004). The ‘creamy layer’: Political economy of reserva- what? Caste and mate selection in modern India. Cambridge, MA: MIT, tions. Economic and Political Weekly, 39(20), 1989–1991. Unpublished manuscript. Deshpande, A. (2000). Does caste still define disparity? A look at Be´teille, A. (2007). Classes and communities. Economic and Political inequality in Kerala, India. American Economic Review, 90(2), Weekly, 42(11), 945–950. 322–325. Borooah, V. K. (2005). Caste, inequality and poverty in India. Review of Deshpande, A. (2001). Caste at birth? Redefining disparity in India. Development Economics, 9(3), 399–414. Review of Development Economics, 5(1), 130–144. Bourguignon, F., Ferreira, F.H.G. & Leite, P.G. (2002). Beyond Oaxaca- Deutsch, J., & Silber, J. (1999). Inequality decomposition by population Blinder: Accounting for differences in household income distributions subgroups and the analysis of interdistributional inequality. In J. Silber across countries. Working Paper No. 2002-04, De´partement et (Ed.), Handbook of income inequality measurement (pp. 363–404). Laboratoire d’Economie The´orique et Applique´e (DELTA), Ecole Norwell, MA: Kluwer. normale supe´rieure, Paris. Dumont, L. (1970). Homo hierarchicus: The caste system and its Chatterjee, P. (1993). The Nation and its fragments: Colonial and implications (Nature of human society). London and Chicago: George postcolonial histories. Princeton, NJ: Princeton University Press. Weidenfeld and Nicholson Ltd. and University of Chicago. CASTE STRATIFICATION AND WEALTH INEQUALITY IN INDIA 1833

Frick, J. R., Goebel, J., Schechtman, E., Wagner, G. G., & Yitzhaki, S. Milanovic, B., & Yitzhaki, S. (2002). Decomposing world income (2004). Using Analysis of Gini (ANoGi) for detecting whether two sub- distribution: Does the world have a middle class?. Review of Income samples represent the same universe: The SOEP experience. Discussion and Wealth, 48(2), 155–178. Paper No. 1049, The Institute for the Study of Labor (IZA), Bonn. Mohanty, M. (2006). Social inequality, labour market dynamics and Gupta, D. (2000). Interrogating caste: Understanding hierarchy and reservation. Economic and Political Weekly, 41(46), 3777–3789. difference in Indian society. New Delhi: Penguin Books. Munshi, K., & Rosenzweig, M. (2006). Traditional institutions meet the Hasan, R., & Mehta, A. (2006). Under-representation of disadvan- modern world: Caste, gender, and schooling choice in a globalizing taged classes in India. Economic and Political Weekly, 41(35), economy. American Economic Review, 96(4), 1225–1252. 3791–3796. National Sample Survey (NSS), (2005). Household assets and liabilities in Hoff, K., & Pandey, P. (2006). Discrimination, social identity, and durable India (as on 30.06.2002). Report, November. New Delhi: Government inequalities. American Economic Review, Papers and Proceedings, May, of India, Ministry of Statistics, National Sample Survey Organization. 206–211. Srinivas, M. N. (2000). Social change in modern India. India: Orient Jayadev, A., Motiram, S., & Vakulabharanam, V. (2007). Patterns of Longman. wealth disparities in India during liberalization. Economic and Political Srinivasan, K., & Mohanty, S. K. (2004). Deprivation of basic amenities Weekly, 42(38), 3853–3863. by caste and religion: Empirical study using NFHS data. Economic and Jayadev, A., Motiram, S., & Vakulabharanam, V. (2010). Patterns of Political Weekly, 39(7), 728–735. wealth disparities in India: 1991–2002. In S. Ruparelia, S. , J. Subramanian, S., & Jayaraj, D. (2006). The distribution of household Harriss, & S. Corbridge (Eds.), Understanding India’s new political wealth in India. Paper prepared for UNU-WIDER project meeting, economy: A great transformation?. London: Routledge. Helsinki, May 4–6. Kijima, Y. (2006). Caste and tribe inequality: Evidence from India, 1983– Sundaram, K. (2006). On backwardness and fair access to higher 1999. Economic Development and Cultural Change, 54(2), 369–404. education in India: Some results from NSS 55th Round Surveys, Madheswaran, S., & Attewell, P. (2007). Caste discrimination in the 1999–2000. Economic and Political Weekly, 41(50), 5173–5182. Indian urban labour market: Evidence from the National Sample Thorat, S., & Newman, K. (2009). Blocked by caste: Economic discrim- Survey. Economic and Political Weekly, 42(41), 4146–4153. ination in modern India. New Delhi: Oxford. Mehrotra, S. (2006). Well-being and caste in Uttar Pradesh. Economic and Yitzhaki, S. (1994). Economic distance and overlapping distributions. Political Weekly, 41(40), 4261–4271. Journal of Econometrics, 61(1), 147–159.