The Political Boundaries of Ethnic Divisions Part 4

Table C.3.1: Effects are Similar When Not Restricting to Social Conflict Administrative Unit Original District (1) Parent & Child (2) Parent Child (3) (4) Dep. Var.: Any Violence post-split post-split × ∆ ethnic polarization post-split × ∆ ethnic fractionalization post-split × ∆ religious diversity Number of District-Months Dep. Var. Mean, Pre-Split -0.007 (0.021) 0.049*** (0.016) -0.013 (0.028) 0.006 (0.012) 0.018 (0.021) 0.043*** (0.015) -0.008 (0.019) -0.008 (0.019) -0.008 (0.021) 0.054*** (0.020) -0.002 (0.025) -0.055** (0.026) 0.036 (0.027) 0.028** (0.011) 0.001 (0.027) 0.000 (0.023) 7,956 0.76 20,220 0.52 7,956 0.69 12,264 0.42 Notes: This table re-estimates our baseline specification but for all violence reported in the SNPK. Significance levels: ∗ : 10% ∗∗ : 5% ∗ ∗ ∗ : 1%. Second, we further gauge robustness to event misclassification by re-estimating our regressions for all possible combinations of the ten main categories of violence in the SNPK. Figure C.3.1 presents the distribution of the estimated coefficients on post-split × ∆P and post-split × ∆F for these 1,023 regressions with the given baseline estimate for social conflict indicated by the dashed, vertical black line. For both our baseline and each separate regression, we scale the reported coefficient by the mean of the given dependent variable, which varies across groups of categories. The magnitudes are therefore standard deviation ∆diversity effect sizes relative to the mean outcome over the sample period. Note that we are not using this data mining approach for inference purposes but rather to address concerns that our particular designation of categories as conflict was somehow spuriously generating our results. Figure C.3.1 helps to dispel such concerns and shows that our core estimated effect of ∆P on social conflict appears to be around the middle of the distribution of effect sizes across all possible combinations of violence categories. Moreover, the distribution of these coefficients seems to lie mostly above zero, which again points to the fact that changing ethnic divisions shifts most types of violence in the same direction. The takeaways are similar for ∆F . 56 Figure C.3.1: Distribution of Estimated Effects of ∆ diversity across All Possible Groupings of Violence Categories in SNPK (a) Original Districts (b) Parent and Child Districts Notes: These graphs present the distribution of estimated effects of ∆diversity across all possible groupings of the violence categories reported in the SNPK. The estimates are rescaled by the mean of the dependent variable such that the effects are standard deviations relative to the mean violence in the given grouping. The dashed line is our baseline estimate from Table 3. C.4 Main Results Broken Down by Injuries, Deaths, and Damages SNPK records injuries, deaths, and property damage. We show in Table C.4.1 that our results are robust to redefining any social conflict to include only the roughly 90% of incidents that record at least one of these outcomes. Further, we show in Table C.4.2 that our results are driven by events with injuries and are strongest for ‘any social incident with an injury or property damage’. Our results are not identifying changes in violence resulting in deaths. Table C.4.1: Effects are Similar When Restricting to Social Conflict Events with an Injury, Death, or Property Damage Administrative Unit: Original District (1) Parent & Child (2) Parent Child (3) (4) Panel A: Only events with an injury, death, or property damage post-split post-split×∆ ethnic polarization post-split×∆ ethnic fractionalization post-split×∆ religious diversity Number of District-Months Number of Districts Dep. Var. Mean, Pre-Split -0.005 (0.026) 0.029* (0.017) 0.003 (0.015) 0.018 (0.014) 0.001 (0.021) 0.032* (0.017) -0.008 (0.013) -0.008 (0.011) -0.007 (0.028) 0.031** (0.012) 0.050** (0.025) -0.003 (0.022) 0.004 (0.023) 0.044* (0.022) -0.025 (0.021) -0.010 (0.013) 7,956 52 0.53 20,220 133 0.30 7,956 52 0.43 12,264 81 0.22 Panel B: Only events with an injury or property damage post-split post-split×∆ ethnic polarization post-split×∆ ethnic fractionalization post-split×∆ religious diversity Number of District-Months Number of Districts Dep. Var. Mean, Pre-Split -0.003 (0.029) 0.034* (0.019) -0.002 (0.024) -0.004 (0.015) -0.002 (0.022) 0.037** (0.016) -0.006 (0.013) -0.007 (0.010) 0.004 (0.029) 0.029*** (0.010) 0.033 (0.024) -0.018 (0.025) -0.005 (0.023) 0.051** (0.022) -0.020 (0.021) -0.006 (0.013) 7,956 52 0.49 20,220 133 0.26 7,956 52 0.40 12,264 81 0.18 Notes: Panel A re-estimates our baseline specification only counting social conflict events with at least one recorded injury, death, or property damage. Panel B re-estimates our baseline specification only counting social conflict events with at least one recorded injury or property damage. Significance levels: ∗ : 10% ∗∗ : 5% ∗ ∗ ∗ : 1%. 58 Table C.4.2: Effects Are Driven by Social Conflict Events with Injuries and Not Deaths Administrative Unit: Original District (1) Parent & Child (2) Parent Child (3) (4) Panel A: Only events with an injury post-split post-split×∆ ethnic polarization post-split×∆ ethnic fractionalization post-split×∆ religious diversity Number of District-Months Number of Districts Dep. Var. Mean, Pre-Split 0.023 (0.028) 0.014 (0.023) 0.003 (0.023) -0.006 (0.020) 0.005 (0.021) 0.030* (0.016) -0.007 (0.013) -0.013 (0.010) 0.018 (0.027) 0.016* (0.009) 0.020 (0.024) -0.031 (0.025) -0.002 (0.022) 0.048* (0.024) -0.023 (0.021) -0.011 (0.012) 7,956 52 0.43 20,220 133 0.23 7,956 52 0.34 12,264 81 0.16 Panel B: Only events with a death post-split post-split×∆ ethnic polarization post-split×∆ ethnic fractionalization post-split×∆ religious diversity Number of District-Months Number of Districts Dep. Var. Mean, Pre-Split -0.019 (0.025) -0.003 (0.014) 0.021 (0.023) 0.061*** (0.010) -0.005 (0.013) -0.003 (0.005) -0.006 (0.009) -0.003 (0.007) -0.021 (0.021) 0.002 (0.013) 0.020 (0.025) 0.020 (0.028) 0.002 (0.014) -0.002 (0.007) -0.013 (0.013) -0.005 (0.007) 7,956 52 0.26 20,220 133 0.13 7,956 52 0.19 12,264 81 0.10 Panel C: Only events with prop. damage post-split post-split×∆ ethnic polarization post-split×∆ ethnic fractionalization post-split×∆ religious diversity Number of District-Months Number of Districts Dep. Var. Mean, Pre-Split -0.005 (0.026) 0.001 (0.009) 0.005 (0.021) 0.015 (0.011) -0.004 (0.012) 0.008 (0.006) -0.006 (0.010) 0.007 (0.007) -0.004 (0.025) 0.009 (0.009) 0.020 (0.030) 0.020 (0.015) -0.002 (0.011) 0.005 (0.006) -0.011 (0.010) 0.004 (0.008) 7,956 52 0.22 20,220 133 0.10 7,956 52 0.17 12,264 81 0.06 Notes: Panel A re-estimates our baseline specification only counting social conflict events with at least one recorded injury. Panel B re-estimates our baseline specification only counting social conflict events with at least one recorded death. Panel C re-estimates our baseline specification only counting social conflict events with at least some property damage. Significance levels: ∗ : 10% ∗∗ : 5% ∗ ∗ ∗ : 1%. 59 D Robustness Checks on the Main Results in Section 5 This Appendix discusses the main robustness checks discussed throughout the paper. D.1 Event Study Specifications We present here the event study generalization of the main equations (1) and (2) as discussed in the paper. These figures highlight both the lack of worrying pre-trends before redistricting as well as provide some insight into the post-redistricting conflict dynamics. Figure D.1.1: Event Study: Average Effects of Redistricting on Social Conflict (Table 2) (a) original district post-split → β: time relative to split .2 .1 0 -.1 -.2 > 2.5 years 24-19 12-1 7-12 19-24 31-36 43-48 55-60 67-72 30-25 18-13 1-6 13-18 25-30 37-42 49-54 61-66 >6 years months relative to split (base period = 12 months pre-split) (b) parent district post-split → .2 β: time relative to split .1 0 -.1 -.2 -.3 > 2.5 years 24-19 12-1 7-12 19-24 31-36 43-48 55-60 67-72 30-25 18-13 1-6 13-18 25-30 37-42 49-54 61-66 >6 years months relative to split (base period = 12 months pre-split) (c) child district post-split → β: time relative to split .2 .1 0 -.1 -.2 > 2.5 years 24-19 12-1 7-12 19-24 31-36 43-48 55-60 67-72 30-25 18-13 1-6 13-18 25-30 37-42 49-54 61-66 >6 years months relative to split (base period = 12 months pre-split) Notes: These figures report coefficient estimates and 95% confidence intervals from event study versions of the regression specification in  equation (1) and given by: conf lictdt = ν + 13 j=−5 βj post-splitd,t−j + θt + θd + θd × t + εdt , where j denotes 6 month bins beginning 30 months prior to splitting (i.e., j = −4 for months 30–24 before splitting) and ending 72 months after (i.e., j = 12 for months 67–72 after splitting) with an additional j = −5 for greater than 30 months before splitting (where defined) and j = 13 for all months after 72. The reference period is the 12 months just prior to splitting. The graph shows the βj coefficients. 60 Figure D.1.2: Event Study: Redistricting, Polarization, and Social Conflict (Table 3) (a) fractionalization (b) polarization .2 φ: time relative to split x ∆ ethnic polarization φ: time relative to split x ∆ ethnic fractionalization (a) Original District .1 0 -.1 -.2 .2 .1 0 -.1 -.2 > 2.5 years 24-19 12-1 7-12 19-24 31-36 43-48 55-60 67-72 30-25 18-13 1-6 13-18 25-30 37-42 49-54 61-66 >6 years > 2.5 years 24-19 12-1 7-12 19-24 31-36 43-48 55-60 67-72 30-25 18-13 1-6 13-18 25-30 37-42 49-54 61-66 >6 years months relative to split (base period = 12 months pre-split) months relative to split (base period = 12 months pre-split) (b) Parent District φ: time relative to split x ∆ ethnic fractionalization post-split → post-split → φ: time relative to split x ∆ ethnic polarization .2 .1 0 -.1 -.2 -.3 .15 .1 .05 0 -.05 -.1 > 2.5 years 24-19 12-1 7-12 19-24 31-36 43-48 55-60 67-72 30-25 18-13 1-6 13-18 25-30 37-42 49-54 61-66 >6 years > 2.5 years 24-19 12-1 7-12 19-24 31-36 43-48 55-60 67-72 30-25 18-13 1-6 13-18 25-30 37-42 49-54 61-66 >6 years months relative to split (base period = 12 months pre-split) months relative to split (base period = 12 months pre-split) post-split → post-split → .2 φ: time relative to split x ∆ ethnic polarization φ: time relative to split x ∆ ethnic fractionalization (c) Child District .1 0 -.1 -.2 .2 .1 0 -.1 > 2.5 years 24-19 12-1 7-12 19-24 31-36 43-48 55-60 67-72 30-25 18-13 1-6 13-18 25-30 37-42 49-54 61-66 >6 years > 2.5 years 24-19 12-1 7-12 19-24 31-36 43-48 55-60 67-72 30-25 18-13 1-6 13-18 25-30 37-42 49-54 61-66 >6 years months relative to split (base period = 12 months pre-split) months relative to split (base period = 12 months pre-split) Notes: These figures report coefficient estimates and 95% confidence intervals from event study versions of the regression specification in     k post-split +θ β post-split + φ × ∆k +θ equation (2) and given by: conf lictdt = ν + 13 t j d,t−j d,t−j d d +θd ×t+εdt , k∈{F,P } j j=−5 where j denotes 6 month bins beginning 30 months prior to splitting (i.e., j = −4 for months 30–24 before splitting) and ending 72 months after (i.e., j = 12 for months 67–72 after splitting) with an additional j = −5 for greater than 30 months before splitting (where defined) and j = 13 for all months after 72. The reference period is the 12 months just prior to splitting. The graph shows the φj coefficients for ethnic fractionalization on the left and polarization on the right. 61 D.2 Ruling Out Confounding Effects of Other Initial District Characteristics As discussed in Section 5.5, we take several steps to address the concern that the effects of ∆diversity on conflict are confounded by omitted district-specific characteristics that also differentially matter after redistricting. We follow the standard method of assessing omitted variable bias in heterogeneous effects DiD specifications, namely interacting treatment (post-split) with other factors besides the primary one(s) of interest (∆diversity) and assessing coefficient stability. The key question is how to select those variables. We consider two approaches: one, subjective and researcher-driven, and a second, more objective and machine-led. In both cases, we marshal a large set of variables across Census, administrative, and GIS-based data sources, mapping each measure to the district level of analysis in the given specification. All variables are time-invariant or predetermined as measured in 1999 or 2000. First, we consider groups of variables plausibly correlated with diversity and conflict based on prior literature and intuition. After reproducing our baseline estimate in column (1), Tables D.2.1–D.2.4 present results based on variables broadly capturing: (2) proximity to security forces, (3) economic development, (4) public goods, (5) demographics, (6) natural resource intensity, (7) political factors, (8) economic structure, (9) geography/topography, and (10) remoteness. Across all specifications at different administrative levels, the estimated effects of ∆diversity are statistically indistinguishable from the baseline in column 1. While reassuring, these tables are nevertheless subject to researcher degrees of freedom in which variables we include and how we combine them across different columns. Therefore, we address such concerns by taking a second, more agnostic approach to variable selection based on the double-selection post-Lasso method of Belloni et al. (2014) to identify covariates that are particularly important in explaining both diversity and social conflict. We elaborate briefly on this method here. We assume that post-split × ∆P and post-split × ∆F can be taken as exogenous, once one controls linearly for a relatively small number of variables–a simple sparsity assumption. The method uses a three-step approach to help the researcher determine which controls to include. First, we select, from the set of post-split×control variables, the covariates that predict post-split×∆P , and separately, post-split× ∆F , conditioning on the usual baseline fixed effects and post-split × ∆Religious diversity. This first step accounts for important confounding factors that are related to ∆P and ∆F . We use 65 post-split×control variables (detailed in Appendix F), drawn from key Indonesian data sources that cover 1999/2000 and are granular enough to construct controls at the eventual 2010 boundaries. Selection is accomplished using Lasso. The Lasso penalty parameter λ is a choice parameter, so we consider a range of values that yield a reasonable number of controls in the final step. In the second step, we select variables that predict the incidence of social conflict from the same set of post-split × control variables, again conditioning on the baseline specification. This step, also operationalized using Lasso, helps capture any important predictors of changes in violence intensity, which keeps residual variance small and can identify additional confounds. Finally, we estimate our baseline OLS equation including the union of selected controls from these two prior stages (hence post-lasso). Inference is uniformly valid for a large class of models under the assumed sparsity condition. Table 6 showed that our main results are unchanged when including these machine-selected covariate interactions with post-split. The fact that these machine-chosen covariates do not alter our results provides some reassurance that the relationship between post-split changes in the incidence of violence are driven by cross-district variation in ∆P and not other observable, cross-district variation. Figure D.2.1 below shows further that these results are robust to varying the penalty parameter, λ, allowing for the inclusion of more or fewer additional covariates.1 We see that the estimated effects of ∆diversity are fairly stable across λ despite large changes in the number of controls selected. In some cases, estimated effects drop and become noisier as we drop λ and grow the number of controls, which is to be expected. 1 In practice, the variable selection tends to pick variables that predict post-split × ∆P and post-split × ∆F , rather than social conflict. The full listing of included covariates in each specification, including the baseline, are available upon request. 62 Table D.2.1: Robustness to Additional Controls × Post-Split, Original District Level + controls for: post-split × ∆ ethnic polarization × ∆ ethnic fractionalization × ∆ religious polarization – (1) sec. forces (2) development (3) pub. goods (4) demog. (5) nat. res. (6) politics (7) occup. (8) geog. (9) remoteness (10) -0.012 (0.025) 0.036** (0.018) -0.003 (0.019) 0.014 (0.013) -0.017 (0.025) 0.037* (0.018) 0.005 (0.018) 0.021 (0.017) -0.009 (0.019) 0.037 (0.024) -0.017 (0.027) 0.034* (0.017) 0.000 (0.021) 0.015 (0.013) -0.011 (0.027) 0.036** (0.016) -0.003 (0.021) 0.030** (0.015) -0.014 (0.024) 0.028 (0.018) -0.021 (0.018) -0.000 (0.015) -0.014 (0.025) 0.039** (0.018) -0.007 (0.020) 0.017 (0.015) -0.014 (0.026) 0.036* (0.018) -0.010 (0.019) 0.009 (0.014) -0.013 (0.027) 0.036* (0.019) -0.013 (0.021) 0.014 (0.014) -0.012 (0.026) 0.030** (0.014) -0.022 (0.017) 0.026 (0.019) -0.018 (0.025) 0.034* (0.018) 0.004 (0.017) 0.023 (0.018) × log distance to security post × log distance to police station × nighttime light intensity -0.011 (0.018) -0.014 (0.024) × share with > primary education × distance to public market -0.007 (0.026) -0.025 (0.025) 0.018 (0.026) 0.019 (0.032) 0.055*** (0.014) -0.015 (0.028) 0.020 (0.019) × share villages with electricity × share villages with safe water × share villages with street light × share villages with transport center × health centers per capita × high schools per capita × log initial population 0.027 (0.020) 0.063* (0.034) 0.052** (0.026) × population share, 5–14 × population share, 15-49 × nat. resource revenue per capita 0.020* (0.010) 0.025 (0.022) -0.013 (0.016) × cash crop share of total ag. output × share of land area with forest × parliamentary vote polarization -0.019 (0.019) -0.014 (0.016) × fiscal transfers per capita × share in agriculture -0.012 (0.044) 0.019 (0.044) -0.007 (0.041) × share in forestry/fishing × share in other × land area 0.031** (0.015) -0.278** (0.131) -0.156** (0.076) -0.199* (0.104) -0.239** (0.115) 0.026 (0.043) -0.004 (0.026) 0.020 (0.041) 0.025 (0.033) × share villages on coast × share villages in valley × share villages on hill × share villages on flatland × shares villages in highlands × log elevation × log distance to coast × log distance to river × log distance to subdistrict capital × log distance to district capital × log distance to major roads Num. of Observations 7,956 7,956 7,956 7,956 7,956 7,956 7,956 7,956 7,956 0.020 (0.030) 0.003 (0.037) 0.019 (0.028) 7,956 Notes: This table augments our baseline specification from column 1 of Table 3 with additional interactions of post-split and potentially confounding initial district characteristics. 63 Table D.2.2: Robustness to Additional Controls × Post-Split, Parent/Child District Level + controls for: post-split × ∆ ethnic polarization × ∆ ethnic fractionalization × ∆ religious polarization – (1) sec. forces (2) development (3) pub. goods (4) demog. (5) nat. res. (6) politics (7) occup. (8) geog. (9) remoteness (10) -0.003 (0.021) 0.032 (0.019) 0.000 (0.012) -0.009 (0.011) -0.007 (0.021) 0.033* (0.019) -0.002 (0.011) -0.004 (0.010) -0.031* (0.016) 0.030* (0.017) -0.004 (0.022) 0.031 (0.019) 0.001 (0.013) -0.009 (0.011) -0.001 (0.019) 0.031** (0.015) -0.001 (0.013) -0.006 (0.013) -0.002 (0.022) 0.032 (0.020) 0.001 (0.013) -0.009 (0.012) -0.003 (0.020) 0.029 (0.021) -0.001 (0.011) -0.007 (0.011) -0.002 (0.021) 0.032 (0.019) 0.000 (0.011) -0.010 (0.011) -0.004 (0.023) 0.033 (0.020) -0.000 (0.012) -0.008 (0.011) -0.005 (0.020) 0.033 (0.021) -0.003 (0.012) -0.010 (0.012) -0.009 (0.021) 0.032* (0.019) 0.005 (0.013) -0.013 (0.011) × log distance to security post × log distance to police station × nighttime light intensity 0.003 (0.015) -0.011 (0.017) × share with > primary education × distance to public market -0.004 (0.014) -0.017 (0.015) 0.023 (0.016) 0.001 (0.018) 0.053*** (0.014) 0.007 (0.015) -0.018 (0.016) × share villages with electricity × share villages with safe water × share villages with street light × share villages with transport center × health centers per capita × high schools per capita × log initial population 0.003 (0.015) -0.001 (0.016) -0.005 (0.020) × population share, 5–14 × population share, 15-49 × nat. resource revenue per capita -0.016 (0.011) 0.011 (0.016) 0.006 (0.012) × cash crop share of total ag. output × share of land area with forest × parliamentary vote polarization 0.002 (0.013) -0.007 (0.010) × fiscal transfers per capita × share in agriculture -0.008 (0.022) 0.013 (0.023) -0.011 (0.024) × share in forestry/fishing × share in other × land area 0.008 (0.024) 0.021 (0.076) 0.019 (0.046) -0.038 (0.072) 0.009 (0.070) 0.036 (0.031) 0.005 (0.016) 0.006 (0.030) -0.013 (0.019) × share villages on coast × share villages in valley × share villages on hill × share villages on flatland × shares villages in highlands × log elevation × log distance to coast × log distance to river × log distance to subdistrict capital × log distance to district capital × log distance to major roads Num. of Observations 20,220 20,220 20,220 20,220 20,220 20,220 20,220 20,220 20,220 0.018 (0.020) 0.030 (0.024) -0.031* (0.017) 20,220 Notes: This table augments our baseline specification from column 2 of Table 3 with additional interactions of post-split and potentially confounding initial district characteristics. 64 Table D.2.3: Robustness to Additional Controls × Post-Split, Parent District Level + controls for: post-split × ∆ ethnic polarization × ∆ ethnic fractionalization × ∆ religious polarization – (1) sec. forces (2) development (3) pub. goods (4) demog. (5) nat. res. (6) politics (7) occup. (8) geog. (9) remoteness (10) 0.001 (0.026) 0.027** (0.013) 0.035 (0.026) -0.031 (0.021) -0.006 (0.024) 0.029** (0.014) 0.044* (0.024) -0.021 (0.022) -0.028 (0.021) 0.053** (0.023) -0.007 (0.028) 0.030** (0.014) 0.031 (0.027) -0.027 (0.021) -0.018 (0.026) 0.063*** (0.011) 0.056*** (0.020) -0.012 (0.018) -0.001 (0.026) 0.036** (0.016) 0.032 (0.024) -0.040* (0.020) 0.001 (0.027) 0.029 ** (0.012) 0.041 (0.026) -0.029 (0.019) -0.004 (0.027) 0.028** (0.013) 0.040 (0.025) -0.040 * (0.022) -0.002 (0.028) 0.031** (0.014) 0.033 (0.025) -0.035 (0.023) -0.006 (0.024) 0.030* (0.017) 0.045* (0.025) -0.034 (0.022) -0.003 (0.026) 0.037** (0.014) 0.038 (0.025) -0.032 (0.023) × log distance to security post × log distance to police station × nighttime light intensity -0.020 (0.021) -0.024 (0.030) × share with > primary education × distance to public market -0.060** (0.028) -0.067** (0.025) -0.010 (0.027) -0.012 (0.022) 0.076*** (0.017) 0.049*** (0.017) 0.019 (0.017) × share villages with electricity × share villages with safe water × share villages with street light × share villages with transport center × health centers per capita × high schools per capita × log initial population 0.017 (0.026) 0.058** (0.027) 0.011 (0.033) × population share, 5–14 × population share, 15-49 × nat. resource revenue per capita 0.010 (0.009) -0.005 (0.027) -0.022 (0.024) × cash crop share of total ag. output × share of land area with forest × parliamentary vote polarization -0.027 (0.019) 0.017 (0.022) × fiscal transfers per capita × share in agriculture -0.019 (0.036) 0.008 (0.041) -0.036 (0.038) × share in forestry/fishing × share in other × land area -0.018 (0.025) 0.006 (0.124) 0.041 (0.083) -0.024 (0.128) 0.003 (0.105) 0.032 (0.047) -0.029 (0.024) 0.013 (0.033) 0.041 (0.039) × share villages on coast × share villages in valley × share villages on hill × share villages on flatland × shares villages in highlands × log elevation × log distance to coast × log distance to river × log distance to subdistrict capital × log distance to district capital × log distance to major roads Num. of Observations 7,956 7,956 7,956 7,956 7,956 7,956 7,956 7,956 7,956 0.027 (0.033) -0.012 (0.039) 0.030 (0.029) 7,956 Notes: This table augments our baseline specification from column 3 of Table 3 with additional interactions of post-split and potentially confounding initial district characteristics. 65 Table D.2.4: Robustness to Additional Controls × Post-Split, Child District Level + controls for: post-split × ∆ ethnic polarization × ∆ ethnic fractionalization × ∆ religious polarization – (1) sec. forces (2) development (3) pub. goods (4) demog. (5) nat. res. (6) politics (7) occup. (8) geog. (9) remoteness (10) -0.005 (0.025) 0.043* (0.025) -0.011 (0.019) -0.005 (0.014) -0.010 (0.027) 0.048* (0.024) -0.018 (0.019) 0.000 (0.012) -0.034 (0.026) 0.027 (0.022) -0.004 (0.026) 0.042* (0.023) -0.012 (0.020) -0.006 (0.014) -0.000 (0.023) 0.027 (0.026) -0.007 (0.020) -0.006 (0.016) -0.004 (0.025) 0.045* (0.025) -0.007 (0.024) -0.004 (0.016) -0.004 (0.022) 0.035 (0.027) -0.015 (0.018) 0.002 (0.014) -0.005 (0.025) 0.046* (0.023) -0.006 (0.019) -0.011 (0.015) -0.006 (0.027) 0.044* (0.026) -0.012 (0.020) -0.004 (0.014) -0.010 (0.024) 0.055* (0.028) -0.016 (0.021) -0.008 (0.017) -0.011 (0.024) 0.049** (0.019) -0.006 (0.021) -0.020 (0.015) × log distance to security post × log distance to police station × nighttime light intensity 0.013 (0.019) -0.004 (0.026) × share with > primary education × distance to public market 0.007 (0.020) -0.008 (0.024) 0.026 (0.024) 0.011 (0.031) 0.044** (0.020) -0.002 (0.026) -0.019 (0.023) × share villages with electricity × share villages with safe water × share villages with street light × share villages with transport center × health centers per capita × high schools per capita × log initial population -0.005 (0.018) -0.019 (0.019) -0.012 (0.027) × population share, 5–14 × population share, 15-49 × nat. resource revenue per capita -0.034* (0.018) 0.029 (0.021) 0.012 (0.015) × cash crop share of total ag. output × share of land area with forest × parliamentary vote polarization 0.019 (0.017) -0.012 (0.013) × fiscal transfers per capita × share in agriculture -0.006 (0.026) 0.013 (0.022) -0.007 (0.037) × share in forestry/fishing × share in other × land area 0.006 (0.035) -0.001 (0.097) -0.025 (0.051) -0.072 (0.092) -0.020 (0.087) 0.061 (0.047) 0.007 (0.024) 0.011 (0.042) -0.036 (0.029) × share villages on coast × share villages in valley × share villages on hill × share villages on flatland × shares villages in highlands × log elevation × log distance to coast × log distance to river × log distance to subdistrict capital × log distance to district capital × log distance to major roads Num. of Observations 12,264 12,264 12,264 12,264 12,264 12,264 12,264 12,264 12,264 0.032 (0.032) 0.042 (0.030) -0.066** (0.027) 12,264 Notes: This table augments our baseline specification from column 4 of Table 3 with additional interactions of post-split and potentially confounding initial district characteristics. 66 Figure D.2.1: Varying the Penalty Parameter in Lasso Robustness Procedure 20 (b) parent/child district .08 30 10 .02 5 0 -.02 0 2500 3000 3500 Lasso penalty parameter λ (baseline=3,000) .06 25 .04 20 .02 15 0 -.02 4000 10 2000 (c) parent district .15 2500 3000 3500 Lasso penalty parameter λ (baseline=3,000) 4000 (d) child district .1 .1 20 .05 15 0 10 -.05 coefÞcient on post-split x Δ P 25 25 20 .05 15 0 10 number of variables included 2000 coefÞcient on post-split x Δ P coefÞcient on post-split x Δ P .04 number of variables included 15 number of variables included coefÞcient on post-split x Δ P .06 number of variables included (a) original district 5 -.1 -.05 2000 2500 3000 3500 Lasso penalty parameter λ (baseline=3,000) 4000 5 2000 2500 3000 3500 Lasso penalty parameter λ (baseline=3,000) 4000 Notes: This figure reports alternative estimated effects of post-split × ∆P based on varying the penalty parameter λ used to discipline variable selection in the double Lasso procedure. Table 6 in the paper reported results for λ = 3, 000 as a baseline. These figures vary that value from 2,000 to 4,000, leading to a range of variables included as seen in the red line and “x” points plotted on the right y-axis. The dashed lines are 95 percent confidence intervals on the point estimates from each individual regression. 67 D.3 Constraints on Redistricting and Changes in Ethnic Divisions Section 5.5 presented a policy exercise aimed at clarifying the sources of identifying variation in ∆diversity. This appendix provides further background on that exercise and also demonstrates further results consistent with the takeaway in the paper: The redistricting policy itself constrained the possible changes in ethnic divisions, and the associated effects on conflict can be explained by these policy constraints and the underlying ethnic geography rather than the particular way in which the boundaries were drawn. The concern is that better or worse borders were chosen by districts in manner correlated with latent conflict. We argue here that this is not consistent with the data. We construct the distribution of feasible ∆diversity based on redistricting schemes that satisfied the legal restrictions in terms of the minimum number of subdistricts (3) and basic viability proxied by contiguity. This “N P -hard” problem is challenging given the large number of possible splits.2 In order to make headway, we use a heuristic, randomized approach. Specifically, we randomly partition the district and then check to ensure the partition satisfies the contiguity requirements.3 We repeat this process until we get 1,000 valid partitions for each original district, which we achieve for all but two original districts. Within each of the valid partitions, we then compute the corresponding ∆P and ∆F , creating a distribution of feasible ∆P and ∆F for each split. When constructing ∆diversity for parent and child districts separately, we simply assign the simulated partition with the original district capital to the parent and the residual partition(s) to the child(ren).4 This procedure should provide a reasonably unbiased estimate of various moments of the distribution of ∆diversity, taking the number of splits as given. While some districts have relatively few feasible options, or many that result in very similar ∆diversity, others have a range of feasible ∆diversity. It is not obvious, in such cases, which moment of the feasible ∆diversity distribution is most appropriate. Table 6 used the mean. Results hold with the minimum or maximum. More generally, though, the key insight we derive from this exercise is that the variation across districts in feasible ∆diversity swamps variation within districts. Indeed, stacking all random draws r for each district and regressing ∆Prd on district fixed effects delivers a R2 of nearly 0.9. While some districts certainly had choices that would result in different ∆diversity, in general, regardless of their choice, their ∆diversity would differ from feasible changes in other districts. This can be seen graphically in Appendix Figure D.3.1, which plots the distribution of feasible ∆P for six districts across several major regions of Indonesia. To formally develop this intuition, we re-estimate our baseline regressions randomly assigning each of the 50 original districts to either the minimum or the maximum of their simulated feasible ∆diversity. We then repeat this a large number of times (50,000 in practice) and plot the distribution of resulting estimates for ∆P and ∆F .5 If strategic border formation is driving our results, then the baseline estimates in Table 3 should look very different for at least some of these permutations. Figure D.3.2 shows that this is not the case. In fact, the entire distribution of estimated effects of ∆P lies above zero and is roughly centered on our baseline estimate. This suggests that regardless of how local policymakers drew the borders, the constraints on redistricting and underlying geography limited the extent to which redistricting could reshape ethnic divisions.6 2 The number of possible splits of n subdistricts (of a given original district) into k new districts given by the Stirling number of the second kind (see Fryer Jr. and Holden, 2011). For example, although Aceh Tenggara only has 255 possible partitions of its 9 subdistricts into the two new districts, Kotawaringin Timur has 4.236 × 1011 possible partitions into its three new districts (see Figure 4). 3 Contiguity matrices are computed from shapefiles. We connect islands to the closest non-island. 4 If there are multiple children we use the location of the eventual capital to distinguish among them. 5 There are 250 possible ways to permute min and max ∆diversity across the districts in our regressions. Given computational constraints, we randomize this 50,000 times and appeal to the law of large numbers. 6 These graphs look similar when including the Lasso-selected control variables used in the other robustness checks. Figure D.3.1: Comparing Distribution of Feasible ∆P Across Districts 3000 Sanggau: actual ΔP=0.003 Kapuas: actual ΔP=-0.001 Maluku Tengah: actual ΔP=0.003 Poso: actual ΔP=0.003 Luwu: actual ΔP=-0.0001 2000 Density Aceh Selatan: actual ΔP=-0.0005 1000 0 -.004 -.002 0 Distribution of Feasible Δ P .002 .004 Notes: These figures plot the distribution of randomly drawn feasible ∆P for six original districts in our data. Figure D.3.2: Distribution of Estimated Effects of Randomized Min or Max ∆diversity Notes: These figures plot the distribution of estimated effect sizes on post-split × ∆diversity based on randomly assigning each district either its minimum or maximum feasible ∆diversity from the set of feasible partitions. We repeat this exercise 50,000 times and the bars reflect the density of each effect size (standard deviation change relative to mean outcome). The black solid line is our baseline effect size with actual ∆diversity, the dashed line is based on the mean ∆diversity as reported in Table 6, and the dashed lines are based on the observed min and max ∆diversity. 69 D.4 Outliers and Influential Observations: Robust Estimation and Inference This section shows details our findings in Table 4 on outliers and robust inference. The use of district fixed effects and time trends in our main, generalized DiD specification can help rule out some of the high leverage observations that might otherwise drive the results. Nevertheless, we elaborate here why a principled approach to their removal does not change our findings and, if anything, strengthens them. Point Estimates. We begin by demonstrating graphically how outliers affect the main results in Table 3. We adopt the widely used approach of Belsley et al. (2005) to identify observations with high influence as captured by a df betaki measure, which captures the difference between the regression coefficient θ for variable k when the ith observation is included versus excluded, with the difference being further scaled et al. (2005, p. 28) recommend by the estimated standard error on the regressor coefficient, θk . Belsley √ k as a rule-of-thumb to remove all observations for which |df betai | > 2/ N where N is the number of observations. Other authors recommend weaker cutoffs of 1 (Bollen and Jackman, 1990). To visualize outliers detected using this method, Figure D.4.1 plots the baseline partial regression coefficients and scatterplot of residuals for the original district, parent and child specifications in columns 1, 3, and 4 of Table 3. The red circles identify those residuals with high |df betai | for ∆P . The black lines correspond to our baseline estimate, and the gray lines are estimates based on removing the influential observations. The only regression line that seems significantly affected by the inclusion of outliers is post-split × ∆P at the original district level, which becomes more starkly positive when removing the high-influence observations. Panel B of Table 4 presented the corresponding regression results alongside our baseline estimates for reference in Panel A. Figure D.4.1: Principled Removal of Outliers from Baseline Estimates of Table 3 (a) Original District (b) Parent District (c) Child District Notes: These figures present the partial regression plots for post-split × ∆P in our baseline regressions. The black regression line and 95 percent confidence interval are the results from columns 1 (a), 3 (b), and 4 (c) of Table 3. The red observations are district-months identified by the Belsley et al. (2005) method for removing outliers described earlier. The gray regression line and 95 percent confidence interval are based on removing those observations and re-running the baseline regressions. Inference. Besides influencing point estimates and implied effect sizes, outliers and small sample sizes more generally can also affect inference. We considered in Table 4 several alternative approaches to inference in the generalized DiD panel setup. The baseline point estimates and standard errors clustered at the original district level are as suggested by the usual Bertrand et al. (2004) motivation for clustering in fixed effects DiD designs. Below those, we presented a series of standard errors or p-values. First, we considered the Conley (1999) spatial HAC estimator that allows for contemporaneous correlation in unobservables between all districts within 500 km in addition to the usual within-district correlation over time. Results are similar using other distance bandwidth. Second, we adopt the new “effective degrees of freedom” adjustment due to Young (2016), who adjusts standard errors by the effective sample size implied by the influence of each observation.7 Third, we implemented a cluster wild bootstrap procedure with Webb weights and 9,999 draws (Cameron et al., 2008). Finally, we take seriously the quasi-random timing of redistricting seen in Table 1 and implement a randomization inference procedure that randomly reassigns the set of three ∆diversity vector components across each of the districts in the given regression before estimation. We repeat this 10,000 times and recover the implied nearly exact p-values on our baseline estimates.8 Overall, the main qualitative takeaway of significant effects of ∆P remain fairly robust with the exception of the wild cluster bootstrap and the “effective degrees of freedom” adjustment. Nevertheless, as shown in Panel B of Table 4, both of these inference procedures are sensitive to outliers. Indeed, the simultaneous removal of outliers and adjustment of inference to account for remaining high influence observations delivers the most consistent evidence that ∆P exerts a significant positive effect on social conflict. 7 This novel approach to inference delivers coefficient-specific degrees-of-freedom (DoF). For example, for ∆P , the DoF across columns 1–4 are 11.3, 4.7, 6.2, and 5.6. 8 These are nearly exact as they do not recover the entire distribution of possible estimates as there 2D possible ways to reassign ∆diversity across D districts and with a relatively large number of D > 50 across all specifications, this would require far longer than necessary to identify the general shape of the distribution (and size of the tails) of estimated coefficient sizes. 71