Near and Dear? Evaluating the Impact of Neighbor Diversity on Inter-Religious Attitudes

Near and Dear? Evaluating the Impact of Neighbor Diversity on Inter-Religious Attitudes Sharon Barnhardt, Institute for Financial Management & Research UNSW 16 September, 2011

Motivation Growing evidence ethnic diversity reduces trust and public goods; related to violence. (Alesina, Baqir, and Easterly 1999; Alesina and La Ferrara 2002; 2005; Banerjee, Iyer, and Somanathan 2005; Easterly and Levine 1997; Khwaja 2009; Miguel and Gugerty 2005) How do we design policies to manage ethnic differences? - Opposite approaches: Integration or Separation What are the consequences of mixing members of different ethnic communities into sustained, close proximity on their attitudes about each other? 2 / 34

Motivation continued Massive urbanization of poor in developing countries requires attention to prevent growth of slums Relocation programs bring low-income populations together in new buildings Potentially relevant for millions of households Similar questions in US and Europe 1950s: mixed-race housing Currently: mixed-income and native-immigrant neighborhoods 3 / 34

Issues & Contribution Issue 1: Self-selection into neighborhoods Use randomly-allocated neighbors in government housing in Hyderabad Issue 2: Explicit attitudes may suffer from self-presentation bias Add second measure: implicit associations Issue 3: Evidence we have is on college students in the US Extend into new population (Hindus & Muslims, lower income, less educated, adults, environment lacking institutions to manage conflict) 4 / 34

Preview of Results Hindus attitudes become more favorable with greater exposure to Muslims Explicit attitudes are more positive by 0.25 to 0.40 standard deviations Hindu Children s implicit associations are less negative by 0.20 to 0.57 standard deviations Muslims attitudes not consistently affected Why the difference? Suggest information available about each other differs in a Hindu-dominated society 5 / 34

Roadmap 1.Introduction 2.Related Literature 3.Setting and Identification Strategy 4.Outcome Variables 5.Estimation and Results 6.Conclusion 6 / 34

Literature - framework Mechanisms through which attitudes may become more or less favorable Preferences (taste for interaction; Becker, 1957) - Become more/less comfortable with inter-ethnic interaction Beliefs (statistics about other group; Aigner & Cain, 1977) - Information gained depends on initial priors, nature of interactions Contact Theory (Allport, 1954) 1) equal status 2) cooperation 3) common goals, and 4) institutional support - all likely to make attitudes improve 7 / 34

Literature - empirical Many laboratory experiments, result of contact tends to improve attitudes Lack of competition is key to positive result We have some experimental field evidence that integration increases empathy for white college students randomly assigned an African American roommate (Boisjoly, Duncan, Kremer et al., 2006) Also, evidence of links between attitudes and other outcomes HR managers with worse attitudes toward Muslims less likely to interview Muslims (Rooth, 2008) 8 / 34

From slum to condos India has history of Hindu-Muslim conflict 9 / 34

Religious Restrictions in the 25 most Populous Countries Pew Forum on Religion & Public Life - Global Restrictions on Religion, December 2009

From slum to condos 11/ 34

From slum to condos India has history of Hindu-Muslim conflict 11/ 34

From slum to condos India has history of Hindu-Muslim conflict Hyderabad is riot prone (Varshney) Large slum on southern side of city. Interviews indicate sorting - February 2005: Fire - April 2007: New houses completed, lottery - October 2008 - January 2009: Survey - February - April 2009: Implicit Associations Test 11/ 34

From slum to condos India has history of Hindu-Muslim conflict Hyderabad is riot prone (Varshney) Large slum on southern side of city. Interviews indicate sorting - February 2005: Fire - April 2007: New houses completed, lottery - October 2008 - January 2009: Survey - February - April 2009: Implicit Associations Test Weak property rights help maintain random assignment 11/ 34

Architecture creates proximity 112 four-story buildings 448 clusters of four households 12/ 34

Cluster types (Muslim perspective) Muslim Majority Equal Muslim Minority Orange = Hindus, Green= Muslims, White=Christians 13/ 34

Allocation 377 clusters/ 1508 units / 1441 Hindus and Muslims 1363 responded (95%) 97% Female; Average age 37 Household income Rs. 3817 ($90 in 2009) per month Hindus Muslims Own majority 584 185 Equal 199 199 Own religion is minority 57 139 Total 840 523 14/ 34

Public Lottery Interviewed Housing Corporation and Revenue Divisional Officer Government made lists of affected households on-site the day after the fire. Private company scanned irises and authenticated BPL status Stratified by ground floor Over-subscribed Randomization checks Sorting into clusters Balance by type 15/ 34

Randomization Check 1 (sorting) Run a regression like the following on administrative data: Hindu in unit 1 c =β(number Hindus in 2-4) c + γ g +ε c The 1 is a reference household in cluster of 4, γ is a ground-floor fixed effect because the lottery was stratified Age, Muslim, Hindu, Female, Widow, Christian, Unknown Religion, Backward Caste, Scheduled Caste/Tribe Following Kremer and Levy (2003), use administrative data and simulate a fair draw by randomly allocating names from list to 1792 slots Run the same regression on each simulated draw All coefficients from such regressions using actuals fall within ±2 standard deviation of mean of coefficients from regressions using simulated draws 16/ 34

Distributions of Coefficients from Simulated Lotteries Household Hindu on Number Neighbors Hindu Reference Reference Household Hindu on Number Neighbors Hindu OLS Regression Coefficients Using Simulated Lottery Data. n=448 02 24 46 68 810 Percent -.1 -.05 0.05.1Coefficient Reference Household Hindu on Number Neighbors Hindu 2 4 Percent 6 0 8 10 -.1 -.05 0.05 Coefficient OLS Regression Coefficients Using Simulated Lottery Data. n=448.1 17/ 34

Randomization Check 2 (balance) Administrative Data Mean: Hindus in Hindu Majority Cluster Hindus Difference (Equal Majority) Difference (Minority Majority) Mean: Muslims in Muslim Majority Cluster Muslims Difference (Equal Majority) Difference (Minority Majority) (1) (2) (3) (4) (5) (6) Age of beneficiary 32.144 0.094 0.580 35.297-0.525-0.859 (0.359) (0.733) (1.083) (0.863) (1.118) (1.243) Female 0.947 0.021 0.020 0.962 0.020 0.031** (0.009) (0.015) (0.025) (0.014) (0.017) (0.015) Widow 0.027 0.014 0.042 0.055-0.013-0.011 (0.007) (0.015) (0.034) (0.016) (0.022) (0.024) Backward Class/ Caste 0.398-0.005 0.036 (0.021) (0.044) (0.070) Scheduled Caste 0.421 0.016 0.018 (0.023) (0.044) (0.070) Scheduled Tribe 0.158-0.005-0.032 (0.016) (0.030) (0.046) OLS regressions. Robust standard errors in parentheses adjusted for clustering at level of 4 unit clusters. All columns contain a fixed effect for the ground floor and for the number of Christians randomly allocated to the cluster. 18/ 34

Randomization Check 2 (continued) Survey Data Mean: Hindus in Hindu Majority Cluster Hindus Difference (Equal Majority) Difference (Minority Majority) Mean: Muslims in Muslim Majority Cluster Muslims Difference (Equal Majority) Difference (Minority Majority) (1) (2) (3) (4) (5) (6) Surveyed 0.951 0.011-0.090** 0.974-0.016-0.015 (0.009) (0.017) (0.043) (0.011) (0.019) (0.020) Years lived in Hyderabad 19.261 0.678 3.957* 27.798-0.950-1.412 (0.522) (1.114) (2.011) (1.140) (1.656) (1.640) Moved to Hyderabad to earn a living 0.443-0.308-0.032 0.389-0.004-0.026 (0.166) (0.507) (0.167) (0.039) (0.052) (0.055) Years Education 1.433-0.008-0.181 1.816 0.018 0.225 (0.123) (0.251) (0.370) (0.222) (0.345) (0.372) Grew up in a village 0.719-0.022-0.123* 0.497 0.007-0.022 (0.019) (0.039) (0.068) (0.041) (0.054) (0.059) Knew any cluster neighbor before 0.089 0.027 0.017 0.222-0.083-0.113** (0.014) (0.033) (0.043) (0.046) (0.054) (0.053) OLS regressions. Robust standard errors in parentheses adjusted for clustering at level of 4 unit clusters. All columns contain a fixed effect for the ground floor and for the number of Christians randomly allocated to the cluster. ** p<.05 * p<0.1 The p-value on a chi-square test of the joint significance in predicting groups are p=.2133 for Hindus and p=0.665 for Muslims. 19/ 34

Outcome Variables 1.Explicit attitudes (index of survey questions) 2.Implicit associations (test) 3.Willingness to live together (mean of survey questions) 20/ 34

Outcome 1: Explicit Attitudes Questions How trustworthy are group X? How brave are group X? How much do Xs cheat? How peace-loving are Xs? How much do you trust Xs? Scaled from 1 (most unfavorable) to 5 (most favorable) Two approaches for aggregating explicit attitudes: Index: Standardize scores using mean & SD of other group homogenous clusters - ie, mean and SD for Muslims living in all-muslim clusters used for How brave are Muslims? - Larger negative value means less favorable attitudes. Mean = -1.9 SD = 1.2 Average effect sizes (as in Kling et al. 2004, & Clingingsmith, Khwaja, and Kremer, 2008) 21/ 34

Outcome 2: IAT IAT measures cognitive associations Put two lists of words into categories at the same time Names: Hindu or Muslim Concepts: Good or Bad Instructions change 22/ 34

Introduction Literature Setting Outcomes Results Conclusion Outcome 2: IAT 1st Double Categorization 23 / 34

Outcome 2: IAT Measurement Difference= (Mean reaction good with Hindu) - (Mean reaction good with Muslim) When the task is harder (incongruent with associations in the mind), reaction time is longer. Difference will be Negative when mental association is Hindus are good Positive when mental association is Muslims are good Use D-measure, standard in the IAT literature (Greenwald, Banaji & Nosek, 2003) D= (Mean Good H1 - Mean Good M1)/σ1 + (Mean Good H2 - Mean Good M2)/σ2 Participation fee incentivized to go as fast as possible without making mistakes Sub-sample: under 50 years old, 200 no exposure and 200 from Hindu- Muslim clusters and oldest child age 10-14, if there was one Response rate 341 adults (85%); 129 / 165 kids (78%) 2 24/ 34

Outcome 3: Living Together Two questions about preferences for inter-religious living Do you mind living next to someone from group X? - 1= Don t mind, 0 otherwise What is the best way for Hindus and Muslims to coexist? - 1= Live together and become friends, 0 otherwise Higher value indicates greater preference for interreligious living. 25/ 34

Basic Empirical Strategy Oic=β1(equalc) +β2(respondent's Group is minorityc) + β3(muslimi) + β 4 (Muslim i equal c ) + β 5 (Muslim i minority c ) +α g +γ c +ε ic Where O is an outcome, i indexes the respondent, c indexes a cluster of houses. equal is an indicator variable equaling 1 if the cluster type is equal Hindus and Muslims. Respondent's Group is the minority is an indicator equal to one if the respondent is the only one of her religion in the cluster. Fixed effects for number of Christians allocated to the cluster (γ c ) and for the ground-floor lottery strata (α g ). Standard errors adjusted for clustering at the 4-unit level. Religion is defined by lottery assignment. ITT estimates. 26/ 34

Explicit attitudes Bar chart: attitudes by cluster type (from regressions) Bigger negative value is less favorable (worse) attitude Hindu Respondent in Hindu Majority Cluster Hindu Respondent in Equal Cluster*** Hindu Respondent in Hindu Minority Cluster** -2.662-2.420-2.285 Muslim Respondent in Muslim Majority Cluster Muslim Respondent in Equal Cluster Muslim Respondent in Muslim Minority Cluster -0.757-0.768-0.741-3.0-2.3-1.5-0.8 0 Estimates from OLS regressions. FE ground floor & Number Christians. SE clustered at level of 4-housing units. 27/ 34

Explicit attitudes Robustness Treatment Effects Index Index with covariates Ave. Effect Size Bound 1 on Hindu Minority Bound 2 on Hindu Minority (1) (2) (3) (4) (5) Hindus Equal versus majority 0.242** 0.230** 0.225** 0.242** 0.242** Minority versus majority 0.377*** 0.360** 0.346** 0.340*** 0.336*** Muslims Equal versus majority -0.011 0.001-0.028 Minority versus majority 0.016 0.027-0.041 Covariates no yes no no no N 1363 1363 1363 1369 1369 Bound 1 replaces 6 un-surveyed Hindus in Hindu minority with mean index score for Hindus in Hindu majority (-2.66) Bound 2 replaces them with mean in Hindu-only clusters (-2.71) Average Effect Sizes are with SD weights.n=840 Hindus and 523 Muslims. See text section 5.1.1 for a description of average effect sizes. 28/ 34

Implicit attitudes - Adults Negative D-Measure means good & Hindu associated Hindu Respondent in Hindu Majority Cluster Hindu Respondent in Equal Cluster Hindu Respondent in Hindu Minority Cluster -0.298-0.311-0.185 Muslim Respondent in Muslim Majority Cluster Muslim Respondent in Equal Cluster Muslim Respondent in Muslim Minority Cluster~ 0.301 0.432 0.538-0.45-0.30-0.15 0 0.15 0.30 0.45 0.60 N= 256 Hindus and 83 Muslims. Estimates from OLS regressions. FE ground floor. SE clustered at level of 4-housing units. 29/ 34

Implicit attitudes - Children Negative D-Measure means good and Hindu are associated Only Hindu children affected by cluster type Hindu Respondent in Hindu Majority Cluster -0.278 Hindu Respondent in Equal Cluster** -0.02 Hindu Respondent in Hindu Minority Cluster** -0.14 Muslim Respoindent in Muslim Majority Cluster Muslim Respondent in Equal Cluster 0.371 0.36 Muslim Respondnet in Muslim Minority Cluster 0.14-0.4-0.3-0.1 0 0.1 0.3 0.4 N= 83 Hindu children & 46 Muslim Children. Estimates from OLS regressions. FE ground floor. SE clustered at level of 4-housing units. Note; Hindu in Hindu Minority Effect is robust to weighting, but when covariates are included the effect becomes smaller and is not significant. 30/ 34

Living together Only significant difference is between Hindus in Hindu Majority and Hindu Minority Clusters Hindu Respondent in Hindu Majority Cluster Hindu Respondent in Equal Cluster 0.958 0.965 Hindu Respondent in Hindu Minority Cluster*** Muslim Respondent in Muslim Majority Cluster Muslim Respondent in Equal Cluster Muslim Respondent in Muslim Minority Cluster 0.999 0.985 0.978 0.994 0.85 0.90 0.95 1.00 N=1363. Estimates from OLS regressions. FE ground floor & Number Christians. SE clustered at level of 4-housing units. 31/ 34

Why this pattern? Hindu effect does not appear to be a change in tastes for interaction Can look at behaviors - do they change along with attitudes? - Do they talk to each other? Over 90% of neighbor-pairs say they talk daily. - Do they eat in each other s houses? About 20% have eaten in neighbor s house in past month. For Hindu respondents, no effect of neighbor religion or cluster type on eating together. - Do they provide neighborly help? (index) High scores for Hindu respondents, no impact of neighbor religion or cluster type. - Who do they spend time with in the housing complex? Hindus include more Muslims on their lists of who they spend the most time with when they have more Muslim neighbors. Potential reason Muslims attitudes do not respond to more neighbors Interaction is different by social status Not getting new information since they re surrounded by a dominant Hindu culture - 70% of Hindus grew up in a village, only 50% of Muslims did 32/ 34

Summary Hindu improvement in attitudes is robust to background characteristics, index creation, inclusion of clusters with incomplete government data Consistent with results from college students in US Taste for discrimination seems unlikely channel through which Hindu attitudes change Externally valid for those who would take up highlysubsidized housing; for areas where hostilities latent 33/ 34

Thank you sharon.barnhardt@ifmr.ac.in 34/ 34