Research Design: Panel Data

In this section, I focus on H1, pertaining to the impact of welfare on reliance on ethnicity-based social insurance.Footnote ¹⁷ I use CPHS data to estimate the effect of RBS on one outcome, whether the household has an outstanding borrowing from relatives or friends. My analysis includes 2,016 CPHS households engaged in agricultural activities. Farmers tagged as “agricultural laborers” in the CPHS data are likely landless (Ghosh and Vats Reference Ghosh and Vats2023; Shaw, Rathi, and Chakrabati Reference Shaw, Rathi and Chakrabati2023). Hence, I classify agricultural laborers as the comparison group and other farmers as the treated group. The sample includes 1,624 in the treated group and 392 in the comparison group.Footnote ¹⁸

RBS, introduced in May 2018, gave unconditional transfers to landowning farmers (treated group). The policy did not benefit landless agricultural laborers (comparison group). The analysis spans a two-year time frame, beginning 12 months prior to RBS (May 2017) and ending 12 months after its launch (April 2019), encompassing six survey waves of the CPHS. This time frame helps estimate the effects of RBS as cleanly as possible. In November 2016, the Government of India banned Rs. 500 and Rs. 1,000 notes, disrupting economic activity and rendering commonplace transactions infeasible in the short term. To ensure that the estimation isn’t confounded by such effects, I exclude the January–April 2017 wave. In April 2019, the Government of India launched an income support program called PM-KISAN, providing Rs. 2,000 to eligible farmers. I exclude data beyond April 2019 to avoid conflating the effects of RBS with PM-KISAN.

My empirical specification is a difference-in-differences design comparing treated and comparison groups before and after the introduction of the policy. The specification, given by Equation 1, includes household and survey wave fixed effects. Household fixed effects control for time-invariant heterogeneity due to differences in preferences, productivity, and other unobserved traits. Survey wave fixed effects control for time-varying differences due to seasonal variation

(1) $$ \begin{array}{rl}{Y}_{it}=\beta .{Treatment}_i\times {Post}_t+{\theta}_i+\hskip0.3em {\theta}_t+{\varepsilon}_{it}.& \end{array} $$

Here, $ {Y}_{it} $ denotes the outcome of interest measured for household i at wave t; $ {Treatment}_i $ equals one for farmers and zero for agricultural laborers; $ {Post}_t $ equals one for time starting May 2018 and zero otherwise; $ {\theta}_i $ denotes household fixed effects; $ {\theta}_t $ denotes survey wave fixed effects; and $ \beta $ is the quantity of interest. Standard errors are clustered at the district × wave level.

To assess the magnitude of the treated group’s response to RBS, I express the outcome as a fraction of the pretreatment sample average (Ghosh and Vats Reference Ghosh and Vats2023):

(2) $$ \begin{array}{cc}\begin{array}{c}\frac{Y_{it}}{Avg({Y}_{pre})}=\beta .{Treatment}_i\times {Post}_t\hskip2pt +\\ {}\hskip1em {\theta}_i+\hskip0.1em {\theta}_t+{\varepsilon}_{it}.\end{array}& \end{array} $$

The causal interpretation of $ \beta $ relies on two key assumptions. First, the treated group received RBS transfers whereas the comparison group did not. I verify this by examining income from government transfers reported by both groups. Second, outcomes for both treated and comparison groups would have evolved according to parallel trends in the absence of RBS. I investigate this by estimating a dynamic specification, given in Equation 3. Here, $ {\beta}_k $ refers to the treatment effect estimated at $ t=k $ relative to $ t=3 $ , which corresponds to the Jan–Apr 2018 survey wave conducted right before the launch of RBS:

(3) $$ \begin{array}{cc}\begin{array}{c}{Y}_{it}={\displaystyle \sum_{k=1,\hskip0.60em k\ne 3}^{k=6}}{\beta}_k.{Treatment}_i*{Post}_t\hskip0.19em +\\ {}\hskip-5em {\theta}_i+\hskip0.3em {\theta}_t+{\varepsilon}_{it}.\end{array}& \end{array} $$

To examine if any broader trends in borrowing from relatives or friends happened to coincide with RBS, I estimate Equation 3 for bordering districts in Telangana’s neighboring states. These are roughly contiguous, agrarian districts with similar climatic, geographic, and economic conditions to those in Telangana.Footnote ¹⁹

The estimator relies on three other assumptions: stability of the treated and comparison groups, homogeneity in treatment intensity, and no spillovers. The stability assumption would be violated by land sales allowing farmers to select into the treated group. However, agricultural land sales in India are infrequent and land wealth is largely inherited (Foster and Rosenzweig Reference Foster and Rosenzweig2002). The per acre price of land far exceeds RBS transfers, making it unlikely that the promise of RBS systematically incentivized land purchases.

Saez (Reference Saez2002) and Hanna and Olken (Reference Hanna and Olken2018) show that in the presence of a progressive tax schedule, cash transfers will not raise the after-tax income of all recipients by the same amount (Ghosh and Vats Reference Ghosh and Vats2023). A concern is that effective transfers vary across the income distribution, violating the homogeneity assumption. This is not an issue here because all farmers and farm-based enterprises in India are exempt from income taxes. Recognizing that not all farmers benefit equally from RBS transfers, I estimate Equation 1 with a continuous treatment variable—income from government transfers—to capture the marginal effect of one additional unit of government transfers on the outcome for eligible households.

Finally, spillovers would be a concern if RBS had indirect effects on the landless. Shaw, Rathi, and Chakrabati (Reference Shaw, Rathi and Chakrabati2023) find no evidence to suggest that the program changed the consumption or borrowing behavior of the landless. Although RBS might affect them by expanding employment opportunities in agriculture, Shaw, Rathi, and Chakrabati (Reference Shaw, Rathi and Chakrabati2023, 22) note that “policy spillovers were non-existent.” Hence, the no spillovers assumption is likely satisfied.

Results: Panel Data

The first-stage relevance of RBS is evident in Figure 2, which visualizes how government transfers received by the treated group changed relative to the comparison group.Footnote ²⁰ The estimates come from Equation 3, with April 2018 as the base month. As expected, government transfers shot up for the treated group in June 2018. A similar spike is visible in November 2018, when the second round of payments were made. Prior to RBS, the difference between the treated and comparison groups remained similar. All coefficient plots in the article display 95% confidence intervals.

Figure 2.

Income from Government Transfers

Figure 2. Long description

The Y axis is labeled Estimated coefficient and ranges from -5000 to 10000 with increments of 5000. The X axis lists months from may17 to feb19. A horizontal dashed line is at 0. A vertical dashed line is positioned between mar18 and jun18, specifically at may18.

From may17 to mar18, the data points with vertical error bars hover closely around the 0 line, with a slight dip in nov17.

Starting at the vertical dashed line in may18, the estimated coefficient begins to rise. In jun18, there is a sharp spike to approximately 7500 with a large error bar. Values in jul18, aug18, sep18, and oct18 drop back down but remain slightly above 0. In nov18, there is another significant peak at approximately 4500. The final three months, dec18, jan19, and feb19, show coefficients returning to a level slightly above 0.

Next, I examine the effect of RBS on household borrowing from relatives and friends.Footnote ²¹ Table 1 shows the treatment effect, estimated using Equation 1 and Equation 2. RBS led to a significant reduction of 11.3 percentage points in the likelihood of having a borrowing from relatives or friends. This corresponds to a 38.5% decline relative to the pre-RBS mean.Footnote ²² These effects were not significantly different for SCs, as shown in Table S2 in the Supplementary Material. RBS, hence, greatly reduced the likelihood of borrowing from caste members.

Table 1.

Effect of RBS on Borrowing from Relatives or Friends

Table 1. Long description

The table consists of three columns and four main data rows.

Column headers:

1. Empty header for variable names.

2. Borrowing from relatives or friends.

3. Borrowing from relatives or friends as a fraction of preperiod.

Data rows:

- Post times Treated: Column 2 shows minus 0.113 with three asterisks (standard error 0.0419). Column 3 shows minus 0.385 with three asterisks (standard error 0.142).

- Post: Column 2 shows minus 0.00281 (standard error 0.0651). Column 3 shows minus 0.00955 (standard error 0.221).

- N (Sample Size): 10,250 for both columns.

Footnote information: Specifications include household and survey wave fixed effects. Three asterisks indicate p is less than 0.01. Robust standard errors are clustered at the district times wave level.

Note: Specifications include household and survey wave fixed effects. The results in column 1 are based on Equation 1, while those in column 2 are based on Equation 2. The dependent variable in column 1 is a binary variable indicating whether the household had an outstanding borrowing from relatives/friends. The dependent variable in column 2 is the same outcome, expressed as a fraction of the preperiod mean. Robust standard errors clustered at the district * wave level in parentheses. *** $ p<0.01 $ , ** $ p<0.05 $ , * $ p<0.1 $ .

I use Equation 3 to investigate the parallel trends assumption and find no indication of pre-existing trends in Figure 3.Footnote ²³ The difference between the treated and comparison group in terms of borrowing from relatives or friends was quite similar up until the timing of the first round of transfers. The decline in borrowing from relatives or friends for the treated group becomes visible soon after RBS was introduced. A plot of the raw means in Figure S8 in the Supplementary Material confirms this is driven by a decline in borrowing from relatives or friends for the treated group, rather than an increase for the comparison group.Footnote ²⁴

Figure 3.

Dynamic Treatment Effects: Effect of RBS on Borrowing from Relatives or Friends

Figure 3. Long description

A line graph plotting Estimated coefficient on the Y axis against five time periods on the X axis. The Y axis ranges from negative 0.4 to 0.4 with increments of 0.1. A horizontal dashed line marks the zero baseline. A vertical dashed line between the second and third time periods indicates the start of the R B S treatment.

From left to right, the data points are as follows.

* may underscore aug 2017. The point is slightly below 0 with a vertical error bar extending from approximately negative 0.14 to 0.13.

* sep underscore dec 2017. The point is slightly above 0 at approximately 0.05 with an error bar from negative 0.13 to 0.22.

* may underscore aug 2018. Positioned immediately after the vertical treatment line, the point is at approximately negative 0.05 with an error bar from negative 0.2 to 0.1.

* sep underscore dec 2018. The point drops to approximately negative 0.13 with an error bar from negative 0.26 to 0.

* jan underscore apr 2019. The point remains at approximately negative 0.13 with an error bar from negative 0.24 to negative 0.02.

The overall trend shows coefficients near zero in the pre-treatment phase, followed by a shift toward negative values in the post-treatment phase.

Figure S9 in the Supplementary Material shows no similar decline post-RBS in districts in neighboring states that border Telangana.Footnote ²⁵ This is a set of contiguous, agrarian districts that share geographic and climatic similarities with Telangana. The null effect in these districts suggest that the above estimates cannot be attributed to nation-wide policies or some general trend that happened to coincide with RBS.

This borrowing declined owing to reduced dependence on caste members for consumption. As Table S1 in the Supplementary Material shows, RBS reduced the likelihood of borrowing from relatives or friends for consumption by 48%. I use qualitative data—presented in full in Sections A and B of the Supplementary Material—to discuss why respondents substitute out of these loans. Despite their favorable financial terms, such loans require significant social investment in co-ethnic ties. For many individuals, the social costs outweigh the financial benefits.

More generally, RBS reduced borrowing for consumption and increased borrowing for investment. This is reflected in the impact of RBS on borrowing from banks and shops. Tables S14 and S47 in the Supplementary Material show that RBS increased borrowing from banks, driven by greater investment borrowing. In contrast, it decreased borrowing from shops, driven by declining consumption borrowing. Whereas shops lend primarily for consumption, banks are an important source of credit for investment in this context. Banks’ increased lending for investment likely reflects greater confidence in farmers’ ability to repay loans with greater welfare transfers. This would be consistent with existing theory and evidence in Iversen and Rehm (Reference Iversen and Rehm2022b), showing that commercial lending increases with state-provided welfare because larger welfare transfers reduce the risk of default.

Research Design: Survey Data

In this section, I focus primarily on hypotheses 2 and 3. To test them, I use data from a household survey conducted in villages along the border between Telangana and Andhra Pradesh. Prior to the formation of Telangana, the districts of Kurnool and Mahbubnagar were in the same state. Thus, erstwhile Kurnool district (now in Andhra Pradesh) and erstwhile Mahbubnagar district (now in Telangana) have six decades of shared administrative, political, and economic institutions. They share the same language, culture, and climatic and geographic conditions. Importantly, they share the same caste hierarchy.

In 2023, five years after RBS was launched, I surveyed 3,020 households in 75 villages located along the border between the erstwhile Mahbubnagar and Kurnool districts.Footnote ²⁶ This includes 41 villages in Telangana and 34 in Andhra Pradesh.Footnote ²⁷ The location of these states and sample villages is shown in Figures 4 and 5. These villages were in the same state until 2014, but are now exposed to different welfare regimes. In terms of welfare, the major difference is that villages in Telangana benefit from RBS whereas those in Andhra Pradesh do not.Footnote ²⁸

Figure 4.

Telangana (Blue) and Andhra Pradesh (Green) in Southern India, with Sample Villages Located in the Box

Figure 5.

Sample Villages Along the Telangana–Andhra Pradesh Border

Figure 5. Long description

The map displays a concentration of green circular markers representing sample villages. These markers are densely packed along a West to East axis following the irregular black line of the state border.

* Distribution: The markers begin in the West near Raichur and Yemmiganur, following the border Eastward through a river valley area, and terminating near the Srisailam region.

* Northern Landmarks: Mahbubnagar and Devarakonda are located to the North of the sample cluster.

* Southern Landmarks: Adoni, Dhone, Nandyal, and Proddatur are situated to the South. The Nallamala Hills are visible to the Southeast of the main cluster.

* Eastern Landmarks: The Eastern Coastal Plains and cities like Vinukonda, Narasaraopet, and Kavali are located further East toward the coast.

* Infrastructure: Major roads are labeled with numbers in shields, including 18, 20, 50, 40, and 56. The Krishna River and its reservoirs are visible as blue bodies of water intersecting the border line where the markers are most concentrated.

As shown in Table S18 in the Supplementary Material based on the 2011 census and the 2013 economic survey, villages along the border are similar across most observable characteristics like literacy rate, population composition, public services, and infrastructure.Footnote ²⁹ Caste diversity and caste-based inequality in land ownership also do not differ significantly across the border, as shown in Table S19 in the Supplementary Material. During my 2023 survey, the elected head of each village assembly (known as the sarpanch) was asked for information about their village.Footnote ³⁰ The comparison between villages in Telangana and Andhra Pradesh based on these data is in Table S19 in the Supplementary Material.

Despite these similarities, a simple comparison of landowning farmers across the state border is insufficient, as farmers on either side of the border are also subject to a different administrative and policy setup since the formation of Telangana. To address this problem, I add a second difference by comparing landed and landless households within each state. Hence, my estimation strategy is to compare landed and landless households on both sides of the border, in the following difference-in-differences framework:

(4) $$ \begin{array}{l}{Y}_{ij}={\beta}_0+{\beta}_1{T}_j\times {L}_{ij}+{\beta_2T}_j+\hskip0.3em {\beta_3L}_{ij}\\ {}\hskip-2em +{X}_{ij\hskip0.3em }\gamma +{V}_{j\hskip0.3em }\delta +{\varepsilon}_{ij},\end{array} $$

where $ {Y}_{ij} $ is outcome Y for household i in village j; $ {T}_j $ is a binary variable indicating whether village j is in Telangana (eligible for RBS); $ {L}_{ij} $ indicates whether i owns land; $ {V}_{j\hskip0.3em } $ includes village-level controls for village population and number of castes; and $ {X}_{ij\hskip0.3em } $ denotes household size and welfare benefits received by the household from other programs.Footnote ³¹ Controlling for other welfare benefits accounts for differences in the welfare architecture of Telangana and Andhra Pradesh. $ {\beta}_1 $ is the quantity of interest. Standard errors $ {\varepsilon}_{ij} $ are clustered at the village level.

For this identification strategy to be valid, cross-state differences other than welfare must not differentially affect landowning cultivators. Baseline differences between border villages do not pose a problem unless they affect the landed in Telangana differently. Although pre-RBS data on intercaste interactions in the sample villages are unavailable, data on the practice of untouchability from the 2011–12 IHDS suggest that the difference between landed and landless respondents is similar in the two states (Table S46 in the Supplementary Material).Footnote ³² It is worth emphasizing that the IHDS includes respondents from across the state, whereas similarities are likely greater in border villages.

A potential concern is that the Reddy–Kamma rivalry, which shaped Telangana’s state formation, might influence intercaste dynamics in the sample. But this is unlikely because hardly any Kammas live in sample villages. There are only four Kamma households in the sample.

Reddys are the dominant landowning caste on both sides of the border. SCs constitute 23-24% of the population in sample villages on both sides, and are composed mainly of Mala and Madiga castes. A comparison of sample villages (Table S19 in the Supplementary Material) further suggests that the caste-based concentration of land resources is similar across the border, as is the prevalence of residential segregation of SCs.

Because Telangana’s creation in 2014 followed a movement for a separate state, one might wonder if subnational identity is stronger in Telangana, particularly for landed households. This is unlikely because support for the formation of a separate state was widespread across caste and class lines in Telangana (Benbabaali Reference Benbabaali2016; Janardhan and Raghavendra Reference Janardhan and Raghavendra2013; Rao Reference Rao2011). In my survey, when asked to split 10 “tokens of trust” between a family member and a random person of their state, landed respondents in Telangana assigned $ 3.3 $ tokens to a random person of their state, on average, and landless respondents assigned $ 3.5 $ —comparable to $ 3.2 $ and $ 3.4 $ , respectively, in Andhra Pradesh. Hence, subnational bonds elicit similar levels of trust regardless of state and land ownership.

A common thread here is that the attributes that differ between Telangana and Andhra Pradesh don’t change sharply upon crossing the border. Near the border, similarities outweigh differences, as evident in the data as they are to the people living in this area.Footnote ³³

Further, I observe no noteworthy cross-border migration. In the survey, 82% respondents reported that they had lived in the same village their whole life. Only nine reported moving to the village where they were surveyed in the last six years (since RBS was launched), of which four were in Andhra Pradesh.

Households were sampled using voter lists, a widely used sampling frame in rural India, comparable in representativeness to a full household listing (Joshi et al. Reference Joshi, McManus, Nagpal and Fraker2022). Voter lists for all sample villages were updated in 2022 or 2023. In each village, a random sample of 20–60 voters (depending on village size) was drawn, along with additional households sampled as replacements in the event of non-response. Landless households were over-sampled because the estimation strategy relies on a comparison of the landed and landless. Thus, almost 50% of the sample is landless, though their population percentage is likely lower.Footnote ³⁴ Table 2 summarizes the sample composition.

Table 2.

Composition of Sample Households by State and Landownership Status

Table 2. Long description

The table consists of four columns and four rows.

Column headers from left to right are: Blank, Telangana, Andhra Pradesh, and Total.

Row 1: Landed households. Telangana has 751, Andhra Pradesh has 765, for a total of 1,516.

Row 2: Landless households. Telangana has 767, Andhra Pradesh has 737, for a total of 1,504.

Row 3: Total households. Telangana has 1,518, Andhra Pradesh has 1,502, for a grand total of 3,020.

Survey questions were piloted extensively and translated in the local language by professional translators. Surveyors were thoroughly trained on this translated questionnaire, including the procedure for informed consent. To ensure confidentiality, surveyors were instructed to talk to respondents individually rather than in public spaces. Respondents were informed they could skip any question or stop the interview at any time. The data were collected using password-protected tablets, and stored in an encrypted form throughout data collection and analysis. The performance of surveyors was monitored in person by a team of supervisors based in survey areas. Further, I listened to random snippets recorded from 30% of all surveys,Footnote ³⁵ and checked the distributions of key variables daily to verify data quality.

The 3,020 households in the sample span 41 caste groups. SCs constitute 34% of the sample, 40% of whom own land. Backward Castes (BC) are 56% of the sample, 55% of whom own land. OCs account for 8%, 55% of whom own land.

I rely on four outcomes to measure intercaste ties and interactions, each of which reflects a type of social investment in non-coethnics: first, how frequently respondents reported sharing meals with someone from another caste, measured on a 1–3 scale; second, how frequently respondents believed others in the village share meals with other castes, also on a 1–3 scale; third, whether respondents reported that most or all of their friends were of the same caste; fourth, an incentivized donation choice to benefit an out-group. Respondents were informed that their name would be entered in a lottery, and they would receive Rs. 1,000 ($12) if they won. They were asked if, assuming they win the lottery, they would be willing to donate any of the prize to a charity working to educate children from marginalized castes. The outcome is the donation amount that non-SC respondents were willing to donate to an NGO working for an out-group.Footnote ³⁶

Meal sharing with non-coethnics provides an indicator of social investment in out-group members. Shared meals play a central role in building relationships and forming connections, especially in contexts where food-related customs and practices are closely tied to identity (Mintz and Du Bois Reference Mintz and Du Bois2002). The second outcome assesses perceived social norms surrounding meal sharing, which are relevant because shifts in these norms can influence individual behavior and promote intergroup interaction. Norms that allow commensality lower barriers to social and economic engagement across castes.

Caste homogeneity among friends signals relationships with out-group members, which are often developed through repeated interactions. The incentivized donation choice measures willingness to share resources with an out-group, indicating in part an openness to cooperation (Fehr and Gächter Reference Fehr and Gächter2000; Henrich et al. Reference Henrich, McElreath, Barr, Ensminger, Barrett, Bolyanatz and Cardenas2006).Footnote ³⁷ Taken together, these outcomes capture intercaste interaction, relationship, perception of norms, and resource sharing, providing evidence of investment in non-coethnic ties.

I also examine three outcomes pertaining to attitudes toward other castes: first, whether those of other castes can be trusted; second, willingness to have someone from another caste as a neighbor, measured on a 1–3 scale; and third, perceived willingness of others in the village to have a neighbor from another caste, also on a 1–3 scale. These reflect subjective evaluations rather than direct social investment in non-coethnics. Nevertheless, they indicate support for integration and hence merit analysis.

I emphasize meal sharing rather than support for residential integration among my main outcomes because productive ties are likely to be forged where the benefits of intercaste interaction are greater, which do not necessarily involve physical proximity (Varshney Reference Varshney2003).Footnote ³⁸ Productive social and economic exchange requires reaching out across group boundaries but does not depend on bringing out-group members into one’s residential space (North Reference North1990).

I also analyze respondents’ reported festival spending in the past 12 months as an indicator of social investment in co-ethnics. Festivals are typically celebrated within one’s community, which maps onto caste lines because social ties tend to be denser within castes than across them (Prillaman Reference Prillaman2023). Furthermore, as Rao (Reference Rao2001) notes, many festivals are celebrated along caste lines because they are caste-specific, serving the purpose of building cohesion within the group. Even during larger festivals in which the whole village participates, people of different castes are often assigned different roles, reinforcing caste divisions.

Results: Survey Data

Including all welfare programs of the state and national governments, landowning households in Telangana reported receiving substantially greater benefits in the 12 months prior to the survey, as shown in Table 3.

Table 3.

Welfare Amount Received from All Programs in the Last 12 Months (in USD)

Table 3. Long description

The table consists of two columns representing geographic regions and two rows representing land ownership status.

* The first column header is Telangana and the second column header is Andhra Pradesh.

* The first data row for Landed individuals shows $757 for Telangana and $338 for Andhra Pradesh.

* The second data row for Landless individuals shows $350 for Telangana and $252 for Andhra Pradesh.

I estimate the effect of welfare on borrowing from caste members, before turning to the effect of welfare on interactions and ties with people of other castes (H2). In both cases, I show that these effects vary significantly depending on whether RBS benefits reached a broad cross-section of caste groups (H3). I also examine how welfare influences investment in co-ethnics—festival spending—and attitudes toward non-coethnics.

I use Equation 4 to estimate the average effect on borrowing from caste members. Table 4 displays the effect size relative to the mean in the Andhra Pradesh sample.Footnote ³⁹ On average, it appears that welfare had a negative but statistically insignificant effect on such borrowing.

Table 4.

Average Effect of RBS on Borrowing from Caste Members, Based on Survey Data

Table 4. Long description

The table consists of two columns. The first column lists variables and the second column provides values for the Full sample.

* Telangana times Own land: minus 0.212 with a standard error of 0.162 in parentheses below it.

* Observations: 3,020.

* Andhra Pradesh mean: 0.146.

The note specifies that the dependent variable is a binary variable for borrowing from caste members in the past 12 months. Telangana and Own land are binary variables. Controls include village population, number of castes, household size, and other benefits. Standard errors are clustered at the village level.

Note: The dependent variable, borrowing from caste members, is a binary variable expressed relative to the mean of the Andhra Pradesh sample. Borrowing from caste members refers to whether the respondent reported such a borrowing in the past 12 months. Telangana is a binary variable that equals 1 if the household lives in Telangana, and 0 if the household lives in Andhra Pradesh. Own land is a binary variable indicating if the household owns any agricultural land. Specifications include Telangana and Own land as independent binary variables, in addition to controls for village population, number of castes living in the village, household size, and benefits received by the household from schemes other than RBS or Rythu Bharosa. Standard errors clustered at the village level in parentheses. *** $ p< $ 0.01, ** $ p< $ 0.05, * $ p< $ 0.1.

However, this average effect conceals heterogeneity by caste-based inequality in land ownership. RBS benefits are less likely to reach a broad cross-section of castes in villages with higher caste-based inequality in land ownership. Hence, to test H3, I estimate the impact of RBS separately in villages with lower and higher caste-based inequality in land ownership. Given the lack of detailed data on land ownership, I measure caste-based inequality in land ownership as the ratio of the share of land to share of population of the caste owning a plurality of land in the village. This helps identify villages with greater concentration of land in the hands of a dominant caste. In the median village, the ratio is 2. I categorize a village as lower inequality if this ratio is less than 2, and higher inequality otherwise.Footnote ⁴⁰ Figure S7 in the Supplementary Material shows that villages with relatively low caste-based inequality in land ownership are spread across survey areas.Footnote ⁴¹

Consistent with H3, Table 5 shows that welfare reduced borrowing from caste members by 41% in lower-inequality villages, but had no noteworthy effects in higher-inequality villages.Footnote ⁴² Welfare also reduced festival spending by 21% in lower inequality villages, while having virtually no discernible effect in higher inequality villages. Thus, it is exactly where in-group economic dependencies decline do I find evidence of reduced investment in co-ethnics.

Table 5.

The Effect of RBS on Borrowing from Caste Members and Festival Spending, by Inequality in Land Ownership

Table 5. Long description

The table is divided into two main sections based on land ownership inequality: Lower inequality and Higher inequality. Each section contains two columns: Borrowing from caste members and Festival spending.

Row 1: Telangana times Own land (interaction term).

- Lower inequality: Borrowing from caste members is minus 0.410 (significant at p < 0.1); Festival spending is minus 0.214 (significant at p < 0.1).

- Higher inequality: Borrowing from caste members is minus 0.0809; Festival spending is 0.0285.

Row 2: Standard errors (in parentheses).

- Lower inequality: 0.226 for borrowing; 0.113 for spending.

- Higher inequality: 0.231 for borrowing; 0.0972 for spending.

Row 3: N (Sample size).

- Lower inequality: 1,275 for borrowing; 1,267 for spending.

- Higher inequality: 1,704 for borrowing; 1,702 for spending.

Row 4: Andhra Pradesh mean.

- Lower inequality: 0.146 for borrowing; 9,548 for spending.

- Higher inequality: 0.146 for borrowing; 9,548 for spending.

The note indicates that dependent variables are effect sizes relative to the Andhra Pradesh mean. Standard errors are clustered at the village level.

Note: Both dependent variables are effect sizes, expressed relative to the mean of the Andhra Pradesh sample. Festival spending refers to reported spending on festivals in the past 12 months. Borrowing from caste members refers to whether the respondent reported such a borrowing in the past 12 months. Estimates grouped under “Lower inequality” are based on survey data from villages with relatively lower caste-based inequality in land ownership. Estimates grouped under “Higher inequality” are based on survey data from villages with relatively higher caste-based inequality in land ownership. Telangana is a binary variable that equals 1 if the household lives in Telangana, and 0 if the household lives in Andhra Pradesh. Own land is a binary variable indicating if the household owns any agricultural land. Specifications include Telangana and Own land as independent binary variables, in addition to controls for village population, number of castes living in the village, household size, and benefits received by the household from schemes other than RBS or Rythu Bharosa. Standard errors clustered at the village level in parentheses. *** $ p< $ 0.01, ** $ p< $ 0.05, * $ p< $ 0.1.

Next, I use Equation 4 to estimate the average effect of welfare on interactions and ties with non-coethnics. In Figure 6, I display 95% confidence intervals and point estimates, divided by the standard deviation of the Andhra Pradesh sample.Footnote ⁴³ The average effects appear to be mixed. There are significant positive effects on the frequency with which respondents report sharing meals with other castes, and their beliefs about whether others in the village share meals with other castes. The average effects are statistically indistinguishable from zero on donations by non-SCs to an NGO working for marginalized castes and whether most or all friends come from the same caste.

Figure 6.

The Effect of Welfare on Intercaste Ties

Figure 6. Long description

A coefficient plot with a horizontal x-axis labeled Effect Size (Standard Deviations) ranging from negative 0.6 to positive 0.6. A vertical dashed line marks the zero point. Four variables are listed on the y-axis from top to bottom:

* Donation (non-S C s): The point estimate is approximately positive 0.18 with a wide confidence interval spanning from negative 0.05 to positive 0.38.

* Meals with other castes: The point estimate is approximately positive 0.2 with a confidence interval from positive 0.04 to positive 0.37.

* Others have meals with other castes: The point estimate is approximately positive 0.28 with a confidence interval from positive 0.13 to positive 0.43.

* Most/all friends from own caste: The point estimate is slightly negative at approximately negative 0.04 with a confidence interval from negative 0.18 to positive 0.1.

The horizontal blue lines represent 95 percent confidence intervals, and the dots represent point estimates. The two middle variables regarding meals show significant positive effects as their intervals do not cross the zero line.

As with borrowing and festival spending, these average effects conceal heterogeneity by caste-based inequality in land ownership. The positive effects are driven mainly by villages with lower caste-based inequality in land ownership. Figure 7 shows that in these villages, welfare increased the amount that non-SCs were willing to donate to an NGO working to educate children from marginalized castes by 19%.Footnote ⁴⁴ Further, welfare reduced the likelihood of respondents reporting that most or all of their friends were of the same caste by 33%, while increasing the extent to which they reported sharing meals with other castes, and the extent to which they believed others in the village shared meals with other castes. By contrast, in villages with higher caste inequality, the impacts are smaller and mostly statistically insignificant.Footnote ⁴⁵

Figure 7.

Inequality and the Effect of Welfare on Intercaste Ties

Figure 7. Long description

A two-panel coefficient plot with a shared vertical axis and identical horizontal axes. The Y-axis lists four social indicators from top to bottom: Donation (non-S C s), Meals with other castes, Others have meals with other castes, and Most/all friends from own caste. The X-axis for both panels measures Effect Size (Standard Deviations) ranging from negative 0.8 to positive 0.8, with a dashed vertical reference line at 0.

Left Panel: Lower land inequality (blue data points).

* Donation (non-S C s): Point estimate at approximately positive 0.35 with a horizontal confidence interval from 0.05 to 0.65.

* Meals with other castes: Point estimate at approximately positive 0.35 with a confidence interval from 0.15 to 0.55.

* Others have meals with other castes: Point estimate at approximately positive 0.3 with a confidence interval from 0.1 to 0.55.

* Most/all friends from own caste: Point estimate at approximately negative 0.2 with a confidence interval from negative 0.4 to 0.

Right Panel: Higher land inequality (grey data points).

* Donation (non-S C s): Point estimate at 0 with a confidence interval from negative 0.25 to positive 0.25.

* Meals with other castes: Point estimate at approximately positive 0.1 with a confidence interval from negative 0.15 to positive 0.35.

* Others have meals with other castes: Point estimate at approximately positive 0.25 with a confidence interval from 0.05 to 0.45.

* Most/all friends from own caste: Point estimate at approximately positive 0.05 with a confidence interval from negative 0.1 to positive 0.2.

The pattern of effects is broadly similar to that in Figure 7 when using alternative thresholds to define lower-inequality villages, specifically, the mean land-to-population ratio of 3 (Figure S3 and Tables S41–S44 in the Supplementary Material) or a ratio of 1.5 (Figure S10 and Tables S53–S56 in the Supplementary Material). The effects are robust to controlling for case, that is, including a binary variable for each caste in the sample (Figure S1 and Tables S29–S32 in the Supplementary Material). The effects are qualitatively similar when controlling for public goods provision (Figure S2 and Tables S33–S36 in the Supplementary Material)Footnote ⁴⁶ or respondent characteristics (Figure S6 and Tables S37–S40 in the Supplementary Material).Footnote ⁴⁷

One concern with the donation measure is that the positive impact could reflect a general increase in generosity, which is difficult to disentangle from greater willingness to share resources with an out-group. To clarify the interpretation, I examine the impact of welfare on donations by respondents from the SCs. If welfare simply makes beneficiaries more generous, we would expect an increase in donations from SC beneficiaries as well. However, Table S57 in the Supplementary Material shows that welfare does not lead to higher donations among SC respondents. This suggests that the effect reported in Figure 7 is driven—at least in part—by stronger intergroup ties.

To clarify the hypothesized trade-off between social investments in co-ethnics and non-coethnics, I examine the correlation between festival spending and meal sharing with non-coethnics. The results in Table S52 in the Supplementary Material show that reduced festival spending is associated with increased sharing of meals with other castes. This evidence supports the proposed mechanism underlying H2: lower social investment in co-ethnics is accompanied by increased investment in non-coethnic ties.

Overall, low inequality areas witness changes across the hypothesized causal chain—welfare reduces reliance on ethnicity-based social insurance and investment in co-ethnic ties, while increasing interaction with non-coethnics. In high inequality areas, by contrast, the causal chain breaks down at the first stage. With no significant change in economic reliance on co-ethnics, as the theory would anticipate, there are no changes in intergroup ties.

The per-acre design of RBS precludes testing the counterfactual of whether welfare would have strengthened intercaste ties in high-inequality areas, had it impacted reliance on co-ethnics. Nonetheless, survey data from Andhra Pradesh show no significant differences in intercaste behavior or attitudes between lower and higher inequality areas (Tables S11 and S12 in the Supplementary Material). This suggests that such behavior and attitudes are not systematically more progressive in low-inequality villages. That is not to deny the possibility that social distances in high-inequality areas may be harder to bridge. Even if welfare were to reduce borrowing from caste members in these settings, the effects on intercaste ties might still be relatively muted.

Two alternative interpretations of the effects in low inequality areas warrant consideration. The first is an income effect, in which welfare leads to improved social relations because of greater consumption. The reduction in festival spending in low inequality areas suggests this is not the case. Moreover, survey data from Andhra Pradesh show that reported income is uncorrelated with behavior toward non-coethnics (Table S16 in the Supplementary Material). Data from the 2011–12 IHDS show that, conditional on caste category, wealthier households are more likely to practice untouchability (Table S17 in the Supplementary Material). Finally, across multiple studies, Deshpande and Spears (Reference Deshpande and Spears2016) found that people donate less to appeals featuring needy individuals from a lower caste, suggesting that charitable giving is circumscribed by caste.

Another possible interpretation is that welfare improved social relations by increasing leisure. RBS funds are substantial, equivalent to 15% of annual income for the median household. However, they are not large enough to overcome subsistence constraints in a context where most farmers are small or marginal and the median landholding is about two acres. In such settings, income support is unlikely to reduce labor supply or increase leisure (Ghatak and Muralidharan Reference Ghatak and Muralidharan2019). As reported in Tables S14 and S47 in the Supplementary Material, RBS increased borrowing for investment, suggesting that productive activity did not decline.

It is also worth discussing possible spillover effects. Did behavior changes in Telangana prompt similar changes in neighboring villages in Andhra Pradesh? Certain villages in my sample have easier access to the neighboring state because they are closer to roads connecting the two states. People in these villages likely have more opportunity for cross-border interaction. If significant cross-border spillovers existed, one would expect differential effects depending on whether a village has easier access to the other state.Footnote ⁴⁸ As shown in Table S6 in the Supplementary Material, there are no discernible heterogeneous effects along these lines. The lack of evidence for spillovers concurs with a general insight from Jodhka and Manor (Reference Jodhka and Manor2018) that intercaste dynamics are locally determined.

A related possibility is that behavior changes among RBS beneficiaries prompt similar changes among non-beneficiaries. For instance, I find that welfare increases the extent to which people believe others in the village share meals with other castes. This could reflect a shift in norms around inter-dining, extending to those who did not benefit from RBS. The research design does not allow me to estimate spillovers from beneficiaries to non-beneficiaries. If these spillovers are occurring, they lead me to underestimate the true effect on intercaste social integration.

Finally, despite changes in intercaste ties and interactions in villages with low caste-based inequality, I do not find evidence to suggest that welfare altered attitudes toward other castes. The effects shown in Figure 8 are statistically indistinguishable from zero for reported trust in other castes, reported willingness to have a neighbor from a different caste, and beliefs about whether others in the village would be willing to have a neighbor from a different caste.Footnote ⁴⁹

Figure 8.

The Effect of Welfare on Attitudes Toward Other Castes

Figure 8. Long description

The plot consists of two side-by-side panels. The x-axis for both is Effect Size in Standard Deviations, ranging from negative 0.8 to positive 0.8 with a dashed vertical line at 0. The y-axis lists three variables from top to bottom: Trust other castes, Willing to have other caste neighbor, and Others willing to have other caste neighbor.

In the left panel, titled Lower land inequality, the data points are blue dots with horizontal error bars:

* Trust other castes: Point estimate is approximately positive 0.1 with error bars spanning from negative 0.15 to positive 0.35.

* Willing to have other caste neighbor: Point estimate is approximately positive 0.1 with error bars from negative 0.1 to positive 0.3.

* Others willing to have other caste neighbor: Point estimate is approximately negative 0.15 with error bars from negative 0.35 to positive 0.05.

In the right panel, titled Higher land inequality, the data points are grey dots with horizontal error bars:

* Trust other castes: Point estimate is approximately positive 0.05 with error bars from negative 0.15 to positive 0.25.

* Willing to have other caste neighbor: Point estimate is approximately positive 0.05 with error bars from negative 0.15 to positive 0.25.

* Others willing to have other caste neighbor: Point estimate is approximately positive 0.08 with error bars from negative 0.15 to positive 0.3.

The finding that behaviors change without concomitant changes in attitudes aligns with other studies of intergroup contact (Mousa Reference Mousa2020; Paluck Reference Paluck2009; Scacco and Warren Reference Scacco and Warren2018). Scacco and Warren (Reference Scacco and Warren2018) draw on foundational theories of social psychology to posit that attitudes are slow to change and often preceded by changes in behavior. Behavior reflects a mix of strategic and social considerations beyond attitudes alone. Additionally, productive ties require interaction across group boundaries but not shared residential space, so it is unsurprising that preferences for residential segregation remain unchanged even as intercaste ties increase. Nonetheless, the lack of a discernible effect on these outcomes suggests that though welfare fosters certain forms of integration, it does not altogether unravel the boundaries that shape everyday social life.

Research Design: Qualitative Data

To understand how social insurance within castes shapes social segregation across castes, I conducted in-depth interviews with 56 respondents in 14 of the 75 surveyed villages. Eight of these villages are in Telangana and six in Andhra Pradesh, across seven subdistricts.Footnote ⁵⁰ Within each subdistrict, I visited one sample village with a relatively equitable land distribution across castes, and one with a relatively unequal land distribution.

In each village, I interviewed four respondents who routinely interacted with a broad cross-section of groups and could discuss social relations in the village. These included the elected head of the village assembly (sarpanch), the deputy sarpanch (upa sarpanch), other elected members of the village assembly, employees at the local government or panchayat office, healthcare workers (ASHA workers), and nutrition workers (Anganwadi workers), among others. I interviewed at least one respondent from the SCs and one from the backward castes in each village.

I describe this research design in more detail in Section A of the Supplementary Material, and the interview topics are outlined in Table S49 in the Supplementary Material.

Social Insurance and Investment in Co-Ethnic Ties

I have previously shown how borrowing from caste members is prevalent, substantial, and favorable in its terms. During interviews, almost all of my respondents emphasized the importance of “maintaining good relations” to borrow from one’s caste members in times of need. The emphasis on “maintaining good relations” led me to ask follow-up questions of my respondents, specifically what “good relations” meant to them and how such relations are maintained.

In Section B of the Supplementary Material, I describe the main themes that emerged in response to these questions. Most frequently, respondents alluded to reciprocity as essential for “good relations.” A second common theme was the threat of social sanctions; how borrowing and lending were accompanied by the threat of punishment for behaving in ways that one’s co-ethnics dislike. The third theme was repeated interaction. Although discussed less frequently than reciprocity or social sanctions, it is nonetheless an important social investment.

Qualitative data show that the relations that enable caste-based social insurance are cultivated through such investments. It is the possibility of future need that incentivizes people to invest more in within-caste ties. Simultaneously, the cost of sustaining these ties makes individuals more likely to substitute away from their caste safety net when they gain access to state-provided welfare.

Qualitative data also illustrate the ways in which intercaste ties can be productive, and the spaces in which intercaste interactions occur. These insights complement the survey findings that welfare reduces social investment in co-ethnics, which facilitates increased investment in non-coethnic ties.

RBS and Inequality in Land Ownership

The per-acre design of RBS implies that in villages where one caste owns a disproportionate share of land, they also corner a disproportionate share of RBS benefits. Consistent with H3, I find no discernible effects on borrowing from caste members or intercaste ties in villages with greater caste-based inequality in land ownership.

These findings are reflected in how people in high-inequality areas responded to interview questions about RBS. Several respondents expressed dissatisfaction with program design and spoke about alternatives that they thought would distribute the benefits more broadly. Some also discussed how the landless and marginal landowners perceived the program, emphasizing the government’s responsibility rather than focusing solely on larger landowners. These themes are described in Section B of the Supplementary Material.