Social Identity, Behavior, and Personality: Evidence from India

Abstract

Hierarchies in social identities are integrally related to divergences in economic status. In India, caste is a significant social identity where discriminatory practices have resulted in poor outcomes for the lower castes. While there is considerable research on differences in economic outcomes along caste lines, there is limited work on behavioral preferences and personality traits that can also be adversely affected by such identity hierarchies, and that are important determinants of educational attainment and labor market performance. Combining rich data from incentivized tasks and surveys conducted among a large sample of university students, we find that the historically marginalized Scheduled Castes and Scheduled Tribes (SCSTs) and Other Backward Classes (OBCs) report lower scores than upper castes along several dimensions of economic behavior, such as competitiveness and confidence and personality traits, such as grit, locus of control, and conscientiousness. Further, socioeconomic status has a limited compensatory role in mitigating these gaps.

Keywords:

• Behavioral preferences
• personality
• caste
• experiments
• India

1. Introduction

Hierarchies in social identities are often found to be highly correlated with corresponding inter-group differences in economic and social outcomes. On average, historically marginalized and discriminated groups, perform worse on typical indicators of achievement and well-being compared to individuals from high-ranking social groups. This divergence prevails across different constructs of social identities, such as race, ethnicity, religion, gender, and caste (Akerlof & Kranton, 2010). The work on the internalization of social stereotypes suggests that this is a rather complicated problem. Especially when there is a yawning social divide along racial, ethnic, religious, or gender lines, adherence to choices that confirm identity stereotypes can influence and restrict choices of minority groups detrimentally, fostering dominated and sub-optimal outcomes (Akerlof & Kranton, 2010; Coate & Loury, 1993). Consequently, there is a vicious circle between hierarchical social structures accompanied by legacies of discrimination and stigma, and poor self-valuations due to the internalization of negative stereotypes resulting in a perpetuation of adverse outcomes (Major & O’Brien, 2005; Tajfel & Turner, 1986).

This paper contributes to the literature on discrimination and social identity by documenting gaps in behavioral preferences and personality traits among historically well-defined social groups. More specifically, we examine gaps across caste groups on a range of experimentally elicited behavioral preferences and socioemotional traits among a large sample of university students in India. These dimensions become especially important in light of recent work showing that such noncognitive skills are key determinants of labor market performance, educational attainment, and other life outcomes in both high and low-and-middle-income countries (e.g. Almlund, Duckworth, Heckman, & Kautz, 2011; Cobb-Clark, 2015; Dasgupta, Gangadharan, Maitra, Mani, & Subramanian, 2015; Deming, 2017; Nordman, Sarr, & Sharma, 2019). Unfortunately, the internalization of negative self-images has the potential to detrimentally influence exactly such characteristics. In fact, while studies – primarily from the US – have found racial gaps in these outcomes (e.g. Elder & Zhou, 2021; Urzua, 2008), evidence from developing countries is quite limited.

The Indian caste system provides a particularly relevant context for studying the issue, as it is a deeply entrenched and enduring cultural institution.¹ Caste is intricately linked to one’s economic and social outcomes in India. The lower caste groups – former untouchables (Scheduled Castes, SCs hereafter) – and indigenous tribes (Scheduled Tribes, STs hereafter) have fared worse than the upper caste groups in terms of educational and occupational attainment (Munshi & Rosenzweig, 2006), wages and consumption (Hnatkovska, Lahiri, & Paul, 2012; Kijima, 2006), and business ownership (Deshpande & Sharma, 2013, 2016). While some studies have shown that affirmative action instituted for SCs and STs in 1950 in the form of reservations in national and state legislatures, local governments, institutions of higher education, and government jobs have had positive impacts on poverty reduction, educational attainment, and public goods provision (e.g. Cassan, 2019; Chin & Prakash, 2011; Pande, 2003), impacts are not unequivocal. For example, a recent paper shows that affirmative action for SCs in local governments negatively affects public goods provision and education outcomes (Pandey & Pandey, 2022). Consequently, despite redressal policy initiatives, significant gaps remain between SCSTs and non-SCSTs, and they continue to be subjected to violence by upper castes (Sharma, 2015).

Further, the stigma associated with being low caste means that individuals are not viewed based on their own merits, but through the lens of their collective stigmatized caste identity (Shah, Mandar, Thorat, Deshpande, & Baviskar, 2006). Therefore, it is plausible that social exclusion and repeated exposure to such discrimination and differential treatment may affect one’s beliefs, perceptions, and aspirations.² However, there is no evidence on how these caste groups differ along important dimensions of behavioral preferences and socioemotional traits where the latter capture perceptions related to ‘self’ and ‘identity’ more broadly.

We use a Seemingly Unrelated Regression (SUR) framework to examine caste gaps on a range of experimentally elicited behavioral preferences (competitiveness, confidence, risk preferences, and egalitarianism), and personality traits (Big Five traits, locus of control, and grit) among college students at a large Indian university. In addition, we evaluate the role of socioeconomic status in mitigating some of the elicited differences. We present findings from a unique dataset of over 2000 college students. These data collect novel information on a variety of behavioral preferences (using incentivized experiments) and personality traits (using a survey) that are rarely found in large-scale datasets in developing countries and are non-existent, especially along the caste dimension.

Our results are striking and reveal sizable and significant gaps along caste lines in India. We find that in almost all reported measures of socioemotional and behavioral preferences, there exists a considerable gap between the discriminated SCSTs, OBCs, and the upper castes. Subjects belonging to the lower caste groups not only express lower willingness to compete and less confidence, but they also report lower scores on grit, locus of control, and Big Five measures of conscientiousness, extraversion, agreeableness, and openness to experience. However, they exhibit more egalitarian choices in the areas of social preferences. These results are robust to corrections for multiple hypotheses testing and the presence of unobservables in explaining our effects. Our finding that low caste group students evaluate themselves lower on personality traits, competitiveness, and confidence may have important implications for their academic achievement and labor market success.

Our research findings add to the important literature on social identity and preferences. For example, Benjamin, Choi, and Strickland (2010) find that making Asian-American and native-black subjects’ ethnicity salient causes them to become more patient. Hoff and Pandey (2006) show that revealing subjects’ caste affiliation in Indian villages adversely affects performance on a cognitive task and reduces willingness to compete among low castes in the presence of upper caste members. Bros (2014), using the World Values Survey, finds that even after controlling for income, education, and occupation, SC and ST respondents are more likely to believe that they belong to lower social ranks as compared to their upper caste counterparts. Deshpande and Newman (2007) find low-caste university students to have lower occupational and wage expectations than their upper-caste counterparts. In a within-village analysis, Spears (2016) finds that SCSTs and OBCs report lower life satisfaction compared to the upper castes, and the difference cannot be fully explained by caste differences in wealth and education. Our results using the lens of endogenized caste identity add an important avenue for explaining the continued differences in welfare outcomes along caste lines.

Importantly, from a policy point of view, our findings speak to the literature on parental investments in developing one’s personality and preferences (e.g. Cappelen, List, Samek, & Tungodden, 2020; Elkins & Schurer, 2020; Falk, Kosse, Pinger, Schildberg-Hörisch, & Deckers, 2021). Our student sample provides heterogeneity in terms of family background thereby allowing us to explore whether better socioeconomic status alleviates some of the disadvantages of belonging to lower caste groups. Our results suggest that higher socioeconomic status and attendance at a private high school play a limited role in reducing the caste gaps in behavioral preferences.

This paper is organized as follows. Section 2 elaborates on the study context and the data, and Section 3 lays out the estimation framework. Summary statistics, regression results, and robustness are presented in Section 4. Concluding remarks follow in Section 5.

2. Context and data

2.1. The context

We conducted our study with undergraduate students enrolled at the University of Delhi (DU). DU is one of India’s top public universities that offer three-year undergraduate programs to ∼160,000 full-time students. DU consists of 79 colleges offering degrees in multiple disciplines, such as science, commerce, arts, and humanities. College admissions for most disciplines in DU are based on entry cutoffs such that applicants whose high school exit exam scores exceed the cutoff are eligible to take admission to the college discipline.

In line with the Indian affirmative action policy of reservations (‘quotas’) in higher education institutions, DU reserves 15 and 7.5 per cent of seats for applicants belonging to the SC and ST categories, respectively, and the admission cutoffs are lower for these groups by ∼5–15 per cent. Further, 27 per cent of seats are also reserved for Other Backward Classes (OBCs) and the cutoffs can be up to 10 per cent lower than those for non-reserved applicants. However, only students with household income below a certain threshold among OBCs (‘non-creamy layer’) are eligible to take admission through the affirmative action policy.³

Recent empirical evaluations find that affirmative action has increased the representation of targeted groups in higher education and had positive downstream impacts on their educational attainment (e.g. Cassan, 2019; Desai & Kulkarni, 2008). However, several accounts reveal that higher education institutions in India are exclusionary in nature, and students from reserved groups experience discrimination at the hands of their upper caste peers and teachers based on their caste, and stigmatizing attitudes remain prevalent (Deshpande, 2019; Ovichegan, 2014; Pandey & Pandey, 2018; Subramanian, 2019). Some of the effects on socioemotional and behavioral preferences might manifest due to the (external) prejudice low caste students face on campus, in addition to internalized negative stereotypes. The continued negative behavior from upper caste groups towards low castes on campus is likely to exacerbate or contribute to the effects associated with caste identity. In a study conducted with students at a large public university in India, Deshpande (2019) finds greater support for the ‘externalization’ mechanism, indicating the prevalence of discriminatory attitudes towards beneficiaries of affirmative action, as compared to the ‘internalization’ mechanism wherein low caste students internalize their upper caste peers’ low evaluation of their abilities. Therefore, the university environment reinforces the underlying causes of marginalization, and recent reports indicate that elite higher education institutions, such as DU are not immune to these concerns (Pandey & Pandey, 2018).⁴

2.2. Data and sampling strategy

Our study was conducted in early 2014. We collected data on ∼2000 undergraduate students present across 15 daytime coeducational colleges on the day of our survey.⁵ Our survey proceeded as follows: once we identified the college, we obtained approval from the college principals and collaborated with teachers at the colleges to determine the specific session timings. We focused on second- and third-year students enrolled in economics and commerce degrees as these are among the most popular disciplines at DU. Upon arriving in the classrooms, teachers introduced the research team, and students were told that we would be conducting a decision-making study and survey, that their participation was voluntary, and that they would be monetarily compensated for their time. No student present declined to participate. In other words, the study included all students who were present in the class at the time we arrived. No efforts were made to sample based on caste. Furthermore, since the students had no a priori knowledge about the study, we do not expect caste-based differences in the desire to participate in the study to bias our findings.

Using administrative data on class sizes obtained under the Right to Information Act, we calculate the share of enrolled students who participated in our study. The average participation rate is 58 per cent in our sample. In Dasgupta, Mani, Sharma, and Singhal (2022), we show that the participation rate does not vary differentially based on college quality (as proxied by admission cutoffs). However, as we do not know caste-specific class enrolment, we are unable to calculate caste-specific participation rates in our study.

In all sessions, we first conducted a series of incentivized experiments.⁶ This was followed by a short socioeconomic survey measuring students’ demographic characteristics and socioemotional traits. Subjects’ preferences for competitiveness and confidence were elicited using a competition task adapted from Niederle and Vesterlund (2007). Subjects participated in a real-effort task that involved adding up four two-digit numbers where they were asked to predict their performances after a practice round and choose between a piece rate and tournament compensation scheme. The piece-rate scheme paid INR 10 for every correct answer; the tournament scheme paid double that amount for every correct answer if the subject out-performed a randomly selected student of DU who had solved the questions earlier.⁷ We define competitiveness as a dummy variable that takes a value of 1 if the subject chose the tournament compensation scheme and 0 if the subject chose the piece-rate compensation scheme. We define confidence as a dummy variable that takes a value of 1 if the subject predicted that her performance in the actual task would exceed those of others in the university, and 0 otherwise.

Distributional preferences were measured using the Bartling, Fehr, Maréchal, and Schunk (2009) framework that asks subjects to state their preferences over a series of four binary distributional choices that would affect their and an anonymous participant’s earnings. In all four choices, option A is an equal distribution and option B is an unequal distribution. We define egalitarianism as a dummy variable that takes a value 1 if the subject chooses the equal division (option A) in each row, and 0 otherwise.

We used the investment task of Gneezy and Potters (1997) to elicit risk attitudes. Subjects had to invest a portion of their endowment of INR 150 in a risky lottery (with an equal chance of win or loss) and set aside the remainder. If the investment was successful (based on a die roll) subjects received triple the invested amount in addition to any amount they set aside. If they lost the lottery, they only received the amount that was set aside. We define risk preference as the proportion allocated to the risky lottery with higher invested amounts indicating lower risk aversion.

Next, we administered the survey. As part of the survey, we first collected details on basic demographic and socioeconomic characteristics, such as caste group, religion, type of high school attended, and socioeconomic status (SES) measures, such as parental education and family income. We then administered standard and widely used inventories to measure socioemotional traits, such as Big Five personality traits, locus of control, and grit. The Big Five personality traits were measured using the 10-item inventory of Gosling, Rentfrow, and Swann (2003) where each item has a score between 1 (disagree strongly) and 7 (agree strongly). The Big Five traits are defined as follows: Openness to experience is the tendency to be open to new aesthetic, cultural, or intellectual experiences; Conscientiousness refers to a tendency to be organized, responsible, and hardworking; Extraversion relates to an outward orientation rather than being reserved; Agreeableness is related to the tendency to act in a cooperative and unselfish manner; and Emotional stability (opposite of Neuroticism) is predictability and consistency in emotional reactions with the absence of rapid mood changes. Additionally, we implemented the 8-item Grit scale of Duckworth and Quinn (2009) where a higher score on the Grit scale implies a greater ability to pursue long-term goals with sustained effort. Finally, Locus of control was measured using 13 items from Rotter (1966). Subjects with a high score exhibit an internal locus of control indicating that they believe that events in their life are more under their control as compared to those with an external locus of control who believe that their outcomes are determined by luck and other factors. We standardize all personality traits using the sample mean and standard deviation and use z-score constructs of these variables in the regression analyses.

The outcomes we measure capture notions of one’s preferences and beliefs and have meaningful implications for performance in education and labor market domains. For instance, competitiveness can explain gender gaps in academic track choice, job entry decisions, and wages (Buser, Niederle, & Oosterbeek, 2014; Flory, Leibbrandt, & List, 2015). Risk preferences have implications for skill accumulation and selection into entrepreneurship (Dasgupta et al., 2015). Those with an internal locus of control perceive the subjective returns to effort and investment to be higher, and this explains the positive relationship between locus of control and investments in education, job search, and health behaviors as well as entrepreneurial performance (Cobb-Clark, 2015; Sharma & Tarp, 2018).

Overall, we conducted 60 sessions, resulting in approximately 35 subjects per session. No feedback was provided between or after the experimental tasks. Each session lasted about 75 min. All subjects received a show-up fee of INR 150. The average additional payment was INR 230.⁸ All subjects participated only once in the study. To minimize wealth effects, additional payments were based on a randomly selected incentivized task.

3. Estimation strategy

As each subject makes choices in the three incentivized tasks and also scores him/herself on eight personality traits as described in Section 2.2, we estimate these equations using a Seemingly Unrelated Regression (SUR) framework that allows for these choices to be correlated. We report the correlation between the error terms obtained from estimating each of the behavioral preferences and personality traits separately using OLS (see Tables A1 and A2 in the Supplementary Appendix). We can reject the null that the outcomes are independent of the vector of elicited behavioral preferences and personality traits (p-value <0.01). This supports our choice of estimating the following SUR model that allows the errors to be correlated across equations, which also improves the precision of our estimates and reduces the risk of attaining low statistical power.

Table 1. Summary statistics by caste

Variables	Pooled (1)	Upper caste (2)	OBC (3)	SCST (4)	p-Value UC vs. OBC t-test (5)	p-Value UC vs. SCST t-test (6)	p-Value OBC vs. SCST t-test (7)
Panel A: Behavioral preferences
Competitiveness	0.31 (0.46)	0.31 (0.46)	0.32 (0.46)	0.31 (0.46)	0.91	0.98	0.95
Confidence	0.32 (0.46)	0.31 (0.46)	0.38 (0.48)	0.28 (0.45)	0.01	0.29	0.01
Risk preference	46.71 (18.85)	46.14 (18.37)	48.12 (20.15)	47.97 (19.56)	0.08	0.13	0.93
Egalitarianism	0.15 (0.36)	0.13 (0.34)	0.19 (0.39)	0.18 (0.39)	0.01	0.03	0.84
Panel B: Personality traits
Extraversion	4.62 (1.41)	4.76 (1.43)	4.28 (1.25)	4.25 (1.34)	0.00	0.00	0.77
Agreeableness	5.13 (1.18)	5.19 (1.15)	5.07 (1.22)	4.83 (1.23)	0.08	0.00	0.02
Conscientiousness	5.27 (1.27)	5.29 (1.27)	5.31 (1.29)	5.11 (1.26)	0.83	0.03	0.06
Emotional stability	4.56 (1.35)	4.52 (1.38)	4.65 (1.26)	4.62 (1.26)	0.14	0.3	0.76
Openness to experience	5.33 (1.15)	5.43 (1.13)	5.14 (1.20)	5.04 (1.15)	0.00	0.00	0.29
Locus of control	7.29 (2.02)	7.27 (2.11)	7.51 (1.76)	7.19 (1.79)	0.07	0.60	0.04
Grit	3.35 (0.57)	3.39 (0.57)	3.28 (0.55)	3.21 (0.50)	0.00	0.00	0.14
Panel C: Control variables
Female	0.49 (0.50)	0.58 (0.49)	0.28 (0.45)	0.24 (0.43)	0.00	0.00	0.27
Age	19.75 (0.96)	19.72 (0.86)	19.78 (1.17)	19.83 (1.12)	0.35	0.07	0.57
Hindu	0.91 (0.28)	0.92 (0.27)	0.87 (0.33)	0.91 (0.27)	0.006	0.88	0.07
Private school	0.70 (0.46)	0.82 (0.38)	0.52 (0.50)	0.31 (0.46)	0.00	0.00	0.00
High socioeconomic status	0.71 (0.45)	0.82 (0.38)	0.46 (0.5)	0.37 (0.48)	0.00	0.00	0.02
Raven’s test score	6.45 (2.28)	6.81 (2.13)	5.77 (2.37)	5.45 (2.45)	0.00	0.00	0.1
Sample size	2045	1433	325	287

Notes: Standard deviation in parentheses. In Columns 5, 6, and 7, we report the p-value from a two-sided t-test comparing upper castes (UCs) to OBCs, upper castes to SCSTs, and OBCs to SCSTs, respectively. In Panel B, the maximum possible score for the Big Five traits, locus of control, and grit is 7, 13, and 8, respectively. In Panel C, the maximum score for Raven’s test is 10. High socioeconomic status takes a value 1 if both parents have at least a college degree or if monthly family income exceeds INR 50,000, and 0 otherwise. Hindu takes value 1 if belonging to Hindu religion, 0 otherwise. Private school takes a value 1 if the individual was enrolled in a private high school, 0 otherwise.

Table 2. SUR estimates: behavioral preferences

	(1)	(2)	(3)	(4)
	Competitiveness	Confidence	Risk preference	Egalitarianism
SCST	−0.087** (0.036)	−0.072* (0.037)	0.367 (1.482)	0.065** (0.029)
OBC	−0.079** (0.032)	0.029 (0.033)	0.136 (1.329)	0.055** (0.026)
Female	−0.171*** (0.022)	−0.099*** (0.023)	−6.285*** (0.904)	0.006 (0.017)
Age	−0.002 (0.012)	0.004 (0.013)	0.210 (0.507)	−0.010 (0.010)
Hindu	−0.055 (0.036)	0.008 (0.037)	−1.341 (1.473)	−0.027 (0.028)
Private School	−0.053** (0.026)	−0.026 (0.027)	1.240 (1.082)	0.062*** (0.021)
High socioeconomic status	0.029 (0.026)	−0.007 (0.027)	0.248 (1.081)	−0.030 (0.021)
Raven’s test score	0.010 (0.011)	0.009 (0.011)	1.217*** (0.448)	−0.006 (0.009)
Constant	0.660** (0.264)	0.378 (0.272)	49.619*** (10.878)	0.279 (0.209)
H0: SCST = OBC (p-value)	0.04 (0.84)	6.32 (0.01)	0.02 (0.88)	0.11 (0.74)
Observations	1918	1918	1918	1918
R-squared	0.106	0.063	0.080	0.058

Notes: All regressions include session fixed effects. Robust standard errors in parentheses. See notes of Table 1 for definitions of right-hand side variables.

***p < 0.01, **p < 0.05, *p < 0.10.

We estimate the following equation: $$Y_{\mathrm{ij}}=\mathrm{ }{\mathrm{\beta }}_0+\mathrm{ }{\mathrm{\beta }}_1{\mathit{SCST}}_{\mathrm{i}}+\mathrm{ }{\mathrm{\beta }}_2{\mathit{OBC}}_{\mathrm{i}}+\mathrm{ }\sum _{\mathrm{k}=3}^{\mathrm{N}}{\mathrm{\beta }}_{\mathrm{k}}{X}_{\mathrm{ik}}+\mathrm{ }{\mathrm{\vartheta }}_{\mathrm{s}}+{\mathrm{\epsilon }}_{\mathrm{ij}}$$ where $Y_{ij}$ is the dependent variable (behavioral preferences and personality traits) observed for individual i and outcome j, SCST is a dummy that takes a value of 1 if the subject belongs to the Scheduled Caste or Scheduled Tribe group, 0 otherwise. OBC is a dummy that takes a value of 1 if the subject belongs to the Other Backward Classes, and 0 otherwise. X is a vector of family background characteristics and demographic characteristics that are included in all specifications, such as age (in years), male (takes a value of 1 if male, 0 if female), Hindu (takes a value of 1 if belonging to Hindu religion, 0 otherwise), private school (takes a value 1 if the individual was enrolled in a private high school, 0 otherwise), high socioeconomic status (takes a value 1 if both parents have at least a college degree or if monthly family income exceeds INR 50,000, and 0 otherwise), and standardized Raven’s test score as a measure of ‘fluid intelligence’. We also include session-fixed effects (${\mathrm{\vartheta }}_{\mathrm{s}}),$ which among other things, control for differences in caste and gender composition across sessions. ${\epsilon }_{ij}$ is the iid error term.

As our study is conducted with currently enrolled students, we are able to examine caste gaps between SCSTs, OBCs, and upper caste students in behavioral preferences and personality traits, conditional on their enrolment in DU. We do expect that the students enrolled in DU and in higher education in general, are positively selected on observed and unobserved characteristics relative to those who are not enrolled in university. This positive selection is expected for all caste groups – SCSTs, OBCs, and upper castes. Whether such selection effects are stronger for SCST, OBC or upper caste groups is an empirical question and beyond the scope of this study as our data are limited to those currently enrolled.

4. Results

4.1. Summary statistics

Table 1 presents the mean and standard deviation for our outcomes and control variables, for the pooled sample as well as the caste groups. Panel A summarizes the behavioral preferences. In our sample, 31 per cent of the subjects chose the competitive remuneration by deciding to enter the tournament, and we observe no caste differences in willingness to compete. About a third of our sample is confident, in that they expect themselves to perform better than other students in the university. OBCs are significantly more confident than the upper castes and SCSTs. Subjects on average invest about 47 per cent of their endowment in the risky asset with upper castes investing the smallest proportion of their endowment in the risky asset. Subjects’ preferences for equality also differ significantly by caste: lowest among upper caste (13%) and almost similar among SCSTs (18%) and OBCs (19%).

As seen in Panel B, in general, we observe some caste differences in all personality traits except emotional stability. Upper caste subjects score themselves higher on scales of extraversion, agreeableness, openness to experience, and grit relative to OBCs and SCSTs. Both upper caste and OBC students score similarly on conscientiousness, but higher than SCSTs.

Background characteristics are reported in Panel C. As expected, and in line with the literature, upper caste subjects are significantly more likely than OBCs and SCSTs to have attended a private high school and to belong to high SES families. We also find that OBCs perform significantly better than SCSTs on indicators of family SES and private school attendance. This is consistent with existing evidence that finds that the socioeconomic characteristics of OBCs lie somewhere in between those of upper castes and SCSTs (Deshpande & Ramachandran, 2019). We find a significant upper caste advantage in cognitive ability as measured by the Raven’s test. Since there are significant caste-based gaps in background characteristics, our preferred estimates on caste gaps in behavior and personality reported in Tables 2 and 3 control for these differences.

Table 3. SUR estimates: personality traits

	(1)	(2)	(3)	(4)	(5)	(6)	(7)
	Agreeableness	Extraversion	Conscientiousness	Emotional stability	Openness to experience	Locus of control	Grit
SCST	−0.202** (0.088)	−0.228*** (0.087)	−0.250*** (0.088)	−0.049 (0.089)	−0.254*** (0.086)	−0.211** (0.088)	−0.279*** (0.088)
OBC	−0.040 (0.078)	−0.264*** (0.077)	0.039 (0.078)	0.010 (0.079)	−0.192** (0.076)	−0.018 (0.078)	−0.158** (0.078)
Female	0.260*** (0.052)	0.089* (0.052)	0.128** (0.053)	−0.210*** (0.053)	0.042 (0.051)	−0.074 (0.052)	0.184*** (0.052)
Age	−0.016 (0.029)	−0.019 (0.029)	0.017 (0.029)	−0.010 (0.029)	0.016 (0.028)	0.037 (0.029)	−0.011 (0.029)
Hindu	−0.175** (0.086)	0.035 (0.085)	−0.042 (0.087)	−0.237*** (0.087)	−0.072 (0.084)	−0.024 (0.087)	−0.120 (0.086)
Private school	−0.070 (0.064)	0.128** (0.063)	−0.069 (0.065)	−0.047 (0.065)	0.041 (0.063)	−0.107* (0.064)	−0.086 (0.064)
High socioeconomic status	0.042 (0.064)	−0.021 (0.063)	−0.129** (0.065)	−0.005 (0.065)	0.051 (0.063)	−0.089 (0.064)	−0.012 (0.064)
Raven’s test score	0.023 (0.026)	0.083*** (0.026)	−0.021 (0.026)	0.031 (0.026)	0.103*** (0.026)	0.033 (0.026)	0.021 (0.026)
Constant	0.427 (0.624)	0.311 (0.618)	0.143 (0.629)	0.662 (0.631)	−0.032 (0.613)	−0.179 (0.628)	0.677 (0.627)
H0: SCST = OBC (p-value)	2.9 (0.09)	0.15 (0.7)	9.12 (0.002)	0.38 (0.54)	0.45 (0.5)	4.08 (0.04)	1.61 (0.2)
Observations	1651	1651	1651	1651	1651	1651	1651
R-squared	0.067	0.081	0.059	0.053	0.073	0.065	0.078

Notes: All regressions include session fixed effects. Robust standard errors in parentheses. See notes of Table 1 for definitions of right-hand side variables.

***p < 0.01, **p < 0.05, *p < 0.10.

4.2. Regression results

While the previous section reported unconditional differences across caste groups, in this section, we move on to regression analyses that allow us to control for various confounding factors that might be correlated with caste, thereby making these our preferred estimates. We first present caste gaps in behavioral preferences, namely, competitiveness, confidence, risk preferences, and egalitarianism in Table 2 using the SUR framework proposed in Section 3. Our main coefficients of interest in these regressions are the caste group dummies: SCST and OBC, with the upper caste serving as the reference group. SCSTs and OBCs are 8.7 percentage points and 7.9 percentage points less likely respectively to compete than the upper castes. SCSTs are also 7.2 percentage points less likely to express confidence. However, we do not find a significant difference between the expressed confidence levels of OBCs and upper castes. Risk preferences of SCSTs and OBCs do not differ significantly from those of upper castes. While SCSTs are significantly less confident than OBCs, there are no OBC-SCST differences in terms of competitiveness or risk preferences (see bottom panel of Table 2). Our findings on competition are in line with Hoff and Pandey (2006) who find that revealing subjects’ caste affiliation in Indian villages reduces willingness to compete among low castes in the presence of upper caste members.

Interestingly, SCSTs and OBCs are 5–7 percentage points more likely to prefer an equitable distribution compared to the upper castes, with no caste difference in choices between SCSTs and OBCs. This is in line with Alesina and Giuliano (2011) who find blacks to be more supportive of redistributive policies as compared to whites in the USA. There are two possible explanations for our findings. First, SCSTs and OBCs in our sample, and in general, belong to the lower socioeconomic strata, which could lead them to have stronger preferences for income redistribution. Second, since the low-caste groups have been beneficiaries of various welfare and affirmative action policies, they may be more inclined to favor an equitable distribution. In fact, in our sample, most of the SCST and OBC subjects have availed the affirmative action policy to gain admission into DU.

The coefficient estimates on the other covariates are in the expected directions. For example, females in our sample are less likely to compete, less confident, and more risk averse; and subjects with higher cognitive ability (as measured by Raven’s test score) are less risk averse.

Next, in Table 3, we examine caste gaps in personality traits. Except for emotional stability for which there are no caste differences, we find that SCSTs score themselves lower on all other traits, such as agreeableness, openness to experience, conscientiousness, extraversion, locus of control as well as grit. While OBCs also report lower grit, more introversion, and less openness to experience, they do not differ from the upper castes on other reported measures, such as locus of control, agreeableness, conscientiousness, and emotional stability. Further, we find OBC-SCST differences to be significant only for conscientiousness and locus of control and weakly significant for agreeableness.

Given the importance of traits, such as conscientiousness and locus of control in explaining labor market performance (Almlund et al., 2011), the wage and occupational disadvantage faced by SCSTs could be magnified by lower scores on such traits. For example, Borghans, ter Weel, and Weinberg (2014) argue that a stagnation of the black-white wage gap in the US may have been due to the increasing relevance of ‘people skills’ where minority groups may be at a disadvantage when interacting with the majority because of prejudice or barriers to interpersonal communication on account of racial or cultural differences.

In terms of other covariates, we find females to be more agreeable, more extrovert, and less emotionally stable, as also shown in the existing literature (Costa, Terracciano, & McCrae, 2001). Females in our sample are also grittier and more conscientious. This is also reflected in other dimensions, such as a significant gender difference in class attendance rates where we find that the proportion of female students attending classes regularly is 73 per cent as compared to 64 per cent among males (two-sided t-test, p-value <0.001).

There are two important caveats to our findings. First, while caste is a common feature of daily life in India and is routinely asked in educational and job application forms and in the marriage market, it is possible that eliciting the respondent’s caste made it more salient for the measurement of the socioemotional traits resulting in some form of implicit priming as caste was asked after completing the incentivized tasks but before the personality inventory. This could have made the observed caste-based differences in the personality traits more acute than for the experimentally elicited preferences.⁹ Second, it should be kept in mind that ${\mathrm{\beta }}_1$ identifies the average effect of belonging to the SCST group. The SCST group consists of individuals from several jatis and tribes, with varying socioeconomic statuses. While we do not have this information, it is possible that the results vary by jati/tribal affiliation within this group. This provides a promising avenue for future research.

4.3. The role of socioeconomic status

Our results thus far indicate that SCSTs and OBCs are at a disadvantage when it comes to behavioral preferences and ratings on personality traits. At the same time, there is evidence of significant convergence between SCSTs and non-SCSTs over the last few decades in terms of occupational distribution, wages, consumption, and education in India (Hnatkovska et al., 2012). An immediate question of policy interest then is the scope of parental socioeconomic status (SES) in facilitating convergence between SCSTs and non-SCSTs. For example, Falk et al. (2021) find that children from richer families and with highly educated parents are significantly more patient, less risk-seeking, and have higher IQs. Using US panel data of school-age children, Fletcher and Wolfe (2016) find family income to be an important determinant of non-cognitive skills with the disadvantages associated with low income increasing over time. Therefore, exploring whether the documented caste gaps vary by one’s SES constitutes a natural corollary to the above discussion.

To examine this, in Tables 4 and 5, we estimate the SUR models for behavior and personality, respectively, wherein the SCST and OBC variables are now interacted with a dummy for high SES. As seen in Table 1, while 82 per cent of upper castes are classified as high SES, the corresponding numbers for OBCs and SCSTs are 46 and 37 per cent, respectively. We do not find many roles for SES except we find that among behavioral preferences, high SES OBCs are less egalitarian than OBCs from lower socioeconomic backgrounds. Among the personality traits, we find that OBCs from high SES backgrounds have a more internal locus of control as compared to OBCs from less well-off families. This partial compensating effect is particularly interesting since the locus of control is a robust determinant of life outcomes (Cobb-Clark, 2015). Further, high SES SCSTs are less extroverted than low SES SCSTs.

Table 4. SUR estimates: differences in behavioral preferences by socioeconomic status

	(1)	(2)	(3)	(4)
	Competitiveness	Confidence	Risk Preference	Egalitarianism
SCST	−0.075 (0.049)	−0.072 (0.050)	0.388 (2.007)	0.091** (0.039)
OBC	−0.055 (0.048)	0.059 (0.049)	0.223 (1.966)	0.102*** (0.038)
High SES	0.042 (0.034)	0.004 (0.035)	0.285 (1.383)	−0.004 (0.027)
High SES × SCST	−0.018 (0.068)	0.013 (0.069)	−0.010 (2.779)	−0.038 (0.053)
High SES × OBC	−0.042 (0.062)	−0.058 (0.063)	−0.159 (2.541)	−0.082* (0.049)
Constant	0.655** (0.266)	0.388 (0.273)	49.627*** (10.947)	0.269 (0.210)
Observations	1918	1918	1918	1918
R-squared	0.107	0.064	0.080	0.059

Notes: Robust standard errors in parentheses. Other controls include age, gender, religion, private school, Raven’s test score, and session fixed effects.

***p < 0.01, **p < 0.05, *p < 0.10.

Table 5. SUR estimates: differences in personality traits by socioeconomic status

	(1)	(2)	(3)	(4)	(5)	(6)	(7)
	Agreeableness	Extraversion	Conscientiousness	Emotional Stability	Openness to Experience	Locus of Control	Grit
SCST	−0.227* (0.119)	−0.064 (0.117)	−0.230* (0.119)	−0.177 (0.120)	−0.209* (0.116)	−0.335*** (0.119)	−0.260** (0.119)
OBC	−0.087 (0.116)	−0.130 (0.115)	−0.015 (0.117)	−0.132 (0.117)	−0.183 (0.114)	−0.205* (0.116)	−0.115 (0.116)
High SES	0.018 (0.080)	0.078 (0.079)	−0.141* (0.081)	−0.095 (0.081)	0.069 (0.079)	−0.192** (0.081)	0.009 (0.081)
High SES × SCST	0.036 (0.167)	−0.320* (0.165)	−0.071 (0.168)	0.233 (0.168)	−0.101 (0.164)	0.204 (0.167)	−0.026 (0.168)
High SES × OBC	0.080 (0.149)	−0.199 (0.147)	0.108 (0.150)	0.226 (0.150)	−0.002 (0.146)	0.311** (0.149)	−0.074 (0.149)
Constant	0.437 (0.629)	0.178 (0.622)	0.099 (0.633)	0.751 (0.635)	−0.080 (0.617)	−0.113 (0.631)	0.672 (0.632)
Observations	1651	1651	1651	1651	1651	1651	1651
R-squared	0.067	0.084	0.060	0.055	0.073	0.067	0.079

Notes: Robust standard errors in parentheses. Other controls include age, gender, religion, private school, Raven’s test score, and session fixed effects.

***p < 0.01, **p < 0.05, *p < 0.10.

Next, in Tables 6 and 7, we examine if one’s attendance in a private high school prior to joining university helps alleviate some of the observed caste gaps in behavior and personality. Private schools are typically characterized by higher teacher-student ratios, lower teacher absenteeism, and generally better infrastructure as compared to public schools in India. In a review article, Kingdon (2020) documents that learning outcomes in private schools are higher than those in public schools. Therefore, relative to public schools, private schools may also shape the behavioral attitudes and personality traits of students differently, bridging some of the pre-existing caste gaps documented earlier. At the same time, the private school effect may also in part reflect unobserved differences in family characteristics or preferences for school type that are correlated with behavioral preferences and socioemotional traits.¹⁰

Table 6. SUR estimates: differences in behavioral preferences by private school enrollment

	(1)	(2)	(3)	(4)
	Competitiveness	Confidence	Risk Preference	Egalitarianism
SCST	−0.126*** (0.047)	−0.085* (0.048)	0.093 (1.939)	0.039 (0.037)
OBC	−0.067 (0.049)	−0.008 (0.051)	−1.935 (2.037)	0.042 (0.039)
Private school	−0.064* (0.033)	−0.042 (0.034)	0.510 (1.361)	0.047* (0.026)
Private school × SCST	0.102 (0.068)	0.015 (0.070)	−0.285 (2.809)	0.060 (0.054)
Private school × OBC	−0.027 (0.061)	0.060 (0.063)	3.464 (2.509)	0.017 (0.048)
Constant	0.666** (0.264)	0.387 (0.272)	50.060*** (10.883)	0.287 (0.209)
Observations	1918	1918	1918	1918
R-squared	0.108	0.064	0.081	0.058

Notes: Robust standard errors in parentheses. Other controls include age, gender, religion, high socioeconomic status, Raven’s test score, and session fixed effects.

***p < 0.01, **p < 0.05, *p < 0.10.

Table 7. SUR estimates: differences in personality traits by private school enrollment

	(1)	(2)	(3)	(4)	(5)	(6)	(7)
	Agreeableness	Extraversion	Conscientiousness	Emotional Stability	Openness to Experience	Locus of Control	Grit
SCST	−0.361*** (0.116)	−0.233** (0.115)	−0.326*** (0.117)	−0.253** (0.117)	−0.338*** (0.114)	−0.334*** (0.116)	−0.262** (0.116)
OBC	−0.148 (0.118)	−0.192 (0.117)	−0.126 (0.118)	−0.128 (0.119)	−0.382*** (0.115)	−0.159 (0.118)	−0.162 (0.118)
Private school	−0.159** (0.079)	0.149* (0.079)	−0.148* (0.080)	−0.161** (0.080)	−0.048 (0.078)	−0.194** (0.080)	−0.081 (0.080)
Private school × SCST	0.339** (0.167)	0.047 (0.165)	0.110 (0.168)	0.434*** (0.168)	0.116 (0.164)	0.235 (0.168)	−0.044 (0.168)
Private school × OBC	0.153 (0.147)	−0.129 (0.146)	0.273* (0.148)	0.198 (0.148)	0.315** (0.144)	0.220 (0.148)	0.011 (0.148)
Constant	0.485 (0.624)	0.289 (0.619)	0.205 (0.629)	0.737 (0.630)	0.039 (0.613)	−0.117 (0.628)	0.676 (0.628)
Observations	1651	1651	1651	1651	1651	1651	1651
R-squared	0.070	0.082	0.061	0.057	0.075	0.066	0.078

Notes: Robust standard errors in parentheses. Other controls include age, gender, religion, high socioeconomic status, Raven’s test score, and session fixed effects.

***p < 0.01, **p < 0.05, *p < 0.10.

For behavioral preferences, we do not find the caste effects to differ by the type of high school attended. However, attending a private school has compensatory effects on some personality traits. Although in the overall sample we do not find any caste differences in terms of emotional stability, SCSTs who attended private school report higher scores on emotional stability than those who did not. SCSTs that attended private school rate themselves as more agreeable than those that did not. Similarly, OBCs that attended private school expressed more conscientiousness and greater openness to experience compared to OBCs that went to a public school.

We also examine whether these caste-based differences in traits and behavior differ by gender, and do not find the caste gaps to vary by gender (see Tables A3 and A4 in the Supplementary Appendix).

4.4. Robustness checks

In this sub-section, we show that our primary results reported in Section 4.2 are robust to several checks. First, SUR models require information on all outcome variables and explanatory variables to be jointly not missing, thereby creating some extra missing observations. Therefore, we also estimate OLS/linear probability model regressions and find that the OLS estimates reported in Supplementary Appendix Tables A5 and A6 with more observations are similar to the SUR results (reported in Tables 2 and 3), ruling out any concerns related to missing data.¹¹

Second, in the event of multiple null hypotheses being tested, the probability of false rejection (i.e. Type I error) could be higher than desired. To minimize this error, it is important to consider the multiplicity of null hypotheses being tested. We use the method outlined in Anderson (2008) to correct the standard errors for multiple hypotheses. As these corrections can only account for binary treatment indicators, we construct a binary variable low caste that takes the value 1 if the subject belongs to SCST or OBC category, and 0 if the upper caste. In Tables A7 and A8 of the Supplementary Appendix, we present OLS estimates for behavioral preferences and personality traits regressed on low caste and other controls, respectively, along with unadjusted outcome-specific p-values and sharpened q-values derived using the multiple hypotheses correction.¹² Our results are robust to this correction with a minor loss in the level of significance.

Finally, we also check for the possibility that the selection of unobservables may be biasing our coefficient estimates on the low caste variable. Using the test of Oster (2019), in the lower panels of Tables A7 and A8 of the Supplementary Appendix, we report the ratio of the selection of unobservables to the selection of observables (δ) required to eliminate the caste gap, that is, to attribute the entire caste effect to selection bias, and the bias-adjusted treatment effect (β*). We report these for the outcomes where the coefficient on caste is significant.¹³ The absolute value of δ exceeds the prescribed cutoff value of 1 implies that the selection of unobservables would have to exceed the selection of observables to drive the caste gap to zero, which is unlikely. Further, assuming that unobservables matter as much as observables, the bias-adjusted caste gaps (β*) are similar to the coefficients in the controlled regressions, allaying concerns that unobservables may be driving our results.

5. Discussion and conclusion

In this paper, we examine caste-based differences in behavioral preferences and personality traits. We find that SCSTs and OBCs fare worse compared to upper castes along critical dimensions of behavior and personality that are associated with improvements in educational attainment, labor market performance, and life outcomes in general. Further, as our heterogeneity analyses show, a higher socioeconomic status or attendance in a private high school does not mitigate most of these behavioral caste gaps. This supports conclusions in other studies which note that a very large improvement in wealth status is needed to possibly overcome some of the negative self-perceptions that lower caste members harbor (Bros, 2014). While our sample is limited to students of one university, our findings are noteworthy in that we observe large caste-based differences even among students from largely urban backgrounds at an elite university. Moreover, this is consistent with overall patterns documented by the existing literature – that draws upon samples of varying representativeness from around the world – wherein minority groups tend to express lower subjective well-being due to their identity.

Currently, the earliest that members of low caste groups can avail the benefits of affirmative action is at the time of entry into higher education or employment in the public sector, by which time the preferences and traits we analyze are potentially less malleable. Given that racial gaps in cognitive and socioemotional skills emerge early on (Fryer & Levitt, 2004), from a policy perspective, there is a compelling case for targeting interventions to individuals from disadvantaged backgrounds at an earlier point in their life cycle so that they are in a better position to reap the benefits of complementary investments later on (Cunha & Heckman, 2007).

Further, while affirmative action policies have been beneficial for targeted populations in many ways, such policies may also have unintended perverse consequences by generating stereotypes about incompetence about beneficiaries and reinforcing negative self-image (Bros, 2014; Deshpande, 2019; Leslie, Mayer, & Kravitz, 2014). Our work suggests that the effects of affirmative action policies on socioemotional outcomes require further research.

At a broader level, there is also a case for using evidence from social psychology to foster greater inter-caste contact that can help reduce biases and negative stereotypes, which have repercussions on self-confidence and self-esteem among low castes (e.g. Lowe, 2021). More concerted multi-pronged policy efforts towards making public spaces inclusive and facilitating interaction on equitable terms would be a start in this direction.

Acknowledgements

We thank two anonymous reviewers, Catherine Bros, Ashwini Deshpande, Naveen Sunder, and participants at the UNU-WIDER Workshop on ‘Discrimination and Affirmative Action: What have we learnt so far?’ and ASREC Europe Meeting 2016 for comments. These institutions had no involvement in the study design, data collection, analysis, or interpretation. Neha Agarwal, Riju Bafna, Piyush Bhadani, Japneet Kaur, and Anshul Yadav provided excellent research assistance. We are grateful to the staff at the various colleges in the University of Delhi for lending their support in conducting the study.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

We acknowledge support from UNU-WIDER 2015 and IGC-India Central grant 1-R002456.

Notes

1 The caste system is an arrangement of the Hindu population into several thousand groups called ‘jatis’ (castes). These groups have emerged from the ancient ‘varna’ system according to which society was divided into initially four, later five, hereditary, endogamous, mutually exclusive, and occupation-specific groups. At the top of the varna system were the ‘Brahmins’ (priests and teachers) and the ‘Kshatriyas’ (warriors and royalty), followed by ‘Vaishyas’ (merchants and moneylenders) and finally the ‘Shudras’ (engaged in lowliest jobs). Over time, the Shudras split into two tiers, with those engaged in the most menial jobs being called the ‘Ati-Shudras’. The Ati-Shudras (Dalits) were considered untouchable and any contact with them was considered polluting. Additionally, there are the indigenous tribes (or the Adivasis) who due to geographical isolation, primitive agricultural practices and distinct social customs face large-scale exclusion from mainstream society. In 1950, the Indian Constitution extended affirmative action to Dalits and Adivasis (officially termed as Scheduled Castes and Scheduled Tribes respectively). In addition, there is a third category known as the ‘Other Backward Classes’ (OBCs). While OBCs are not historically stigmatized like the SCs and STs, they are described as a socially and educationally backward group. Reservations have been extended to OBCs since the early 1990s and remain an intensely debated issue. While these broad caste groupings were created for the administrative purposes of implementing affirmative action, the various jatis and tribes could vary in their position in the social hierarchy and extent of stigmatization. Deshpande (2011) provides an overview of the caste system in India.

2 Indeed, theoretical work by Dalton, Ghosal, and Mani (2016) and Genicot and Ray (2017) shows that aspirations affect effort and the incentives to invest, and these are determined both by personal and societal factors.

3 Applicants must provide validated caste certificates to be eligible to apply through these reserved categories.

4 https://thewire.in/caste/india-universities-caste-discrimination (accessed 28 June 2020).

5 We show in Dasgupta et al. (2022) that these 15 study colleges are representative of the remaining colleges in DU.

6 The experiment instructions and the survey instruments are available from the authors upon request.

7 We implemented a pilot version of this game where forty students from DU had participated, and their performance is used for comparison in the tournament wage scheme.

8 At the time of conducting the study, the exchange rate was USD 1 = INR 60.

9 That the observed caste gaps differ between the preferences and socioemotional skills could also be due the incentivized elicitation of behavioral preferences versus the non-incentivized elicitation of personality traits. Survey measures typically have been found to be biased upwards compared to incentivized responses (e.g. List & Gallet, 2001; Loomis, 2011).

10 Anecdotal evidence also suggests that private schools in India base admissions on entrance tests and interviews of students and parents.

11 Less than 1 per cent of choices are missing for the behavioral preferences, except egalitarianism for which 1.5 per cent of the observations are missing. Less than 5 per cent of the Big Five traits and grit are missing. About 7 per cent of the data on locus of control is missing. And 1–2 per cent of the data on all other covariates is missing.

12 The coefficient on the low caste variable is significant in all regressions where SCST and OBC variables were also independently significant.

13 For this test, we set maximum R-squared (R_max) at 1.3 times the R-squared from the regressions using controls.