The Political Boundaries of Ethnic Divisions Part 5

D.5 Validating the Conflict Measures and Ruling Out Systematic Reporting Bias Recall that the SNPK data is based on an exhaustive and carefully vetted set of local media sources across Indonesia. However, like other conflict event sdata, the SNPK still has the potential concern that it systematically underreports violence in certain areas of the country. While we control for the number of sources being used by coders in any given province–month, we can still not completely rule out the possibility that media outlets differentially report on events in (and hence reallocate resources and reporters to) more interesting locations. If “interesting” coincides with redistricting and changes in ethnic divisions, then one might worry that we are over-estimating the effects of ∆diversity on conflict. Subjective reporting is a basic fact facing all conflict research.9 We offer here one important robustness check on our own results that might also be fruitfully applied to others using similar data. In particular, we draw upon Google Trends data in an attempt to rule out confounding effects of time-varying media intensity. The idea here is that the events taking place in any given district–month in our data should attract a baseline level of interest from the (internet-using) population among whom are media actors trying to follow that interest. Once we partial out that general location-specific interest in that period, the SNPK conflict report is more likely to reflect the true likelihood of any incidents rather than just a general uptick in popular (media) attention. These Google Trends, which capture the relative frequency of searches for the given district name (original, parent, or child), are indeed highly correlated with major local events such as mayoral elections.10 More importantly, though, our core results remain qualitatively and quantitatively unchanged when controlling for these Google Trends, which we measure on a [0, 1] continuum. Table D.5.1 below shows this for our baseline results from Table 3. Results look similar for the intensive margin specifications in Table D.6.1. Table D.5.1: Table 3 Robust to Controlling for Google Trends Administrative Unit: post-split post-split × ∆ ethnic polarization post-split × ∆ ethnic fractionalization post-split × ∆ religious diversity Google trends Number of District-Months Dep. Var. Mean, Pre-Split Original District (1) Parent & Child (2) Parent Child (3) (4) -0.011 (0.025) 0.040** (0.018) -0.005 (0.019) 0.015 (0.013) 0.125* (0.067) -0.004 (0.021) 0.033* (0.019) 0.001 (0.012) -0.009 (0.011) 0.071* (0.041) 0.003 (0.026) 0.032** (0.013) 0.035 (0.026) -0.030 (0.021) 0.145* (0.073) -0.007 (0.025) 0.043* (0.025) -0.011 (0.019) -0.005 (0.014) 0.042 (0.039) 7,956 0.57 20,220 0.33 7,956 0.47 12,264 0.25 Notes: This table re-estimates the baseline specification controlling for monthly Google Trends in searches for each district’s name. This measure takes on a value ranging from 0 to 1 indicating for each district–month the relative frequency of searches for its name when compared to other benchmark searches. Significance levels: ∗ : 10% ∗∗ : 5% ∗ ∗ ∗ : 1%. 9 Similar concerns apply to nearly every study of conflict based on media reports, e.g. regions facing weather or commodity price shocks might draw media resources and reporters away from other areas of a given country. Studies at the country level suffer from similar concerns insomuch as they rely on either media reporting of deaths to define civil conflict/war or subjective assessments of conflict scholars as to the timing of conflict outbreaks and cessation (see Bazzi and Blattman, 2014). 10 A fixed effects specification suggests that parent/child district names are around 10 percent more likely to be searched for during the six month window around the direct mayoral elections. D.6 Alternative Econometric Specifications As discussed in Section 5.6, we present here several robustness checks on the econometric specifications in Table 3. D.6.1 Intensive Margin of Violence Our baseline specification focuses on the extensive margin of whether there are any social conflict events in the given district–month. This is a sensible baseline given that most district-months with any events have one or two events (see the left graphs in Figure D.6.1 below). Even at the aggregate original district level—where 63 percent of district–months have any social conflict in column 1 of Table 3—80 percent of observations with any conflict have 5 or fewer events with a very long tail up to 89 events. The skewness is even starker at the more granular parent–child district level. While each of these separate event records is meant to capture a different incident, it is of course possible that they are part of the same underlying conflict, which means that the intensive margin specification might simply introduce noise. Nevertheless, there may be substantive empirical content in the intensive margin variation of incidents. Figure D.6.1: Number of Social Conflict Incidents by Original District-Month Notes: This figure plots the distribution of the number of social conflict events by month at the original district level. The left figure is the raw data. The right figure is the inverse hyperbolic since transformation used in the regressions. Table D.6.1 presents intensive margin specifications based on the widely used hyperbolic inverse sine transformation, log #eventsdt + (#events2dt + 1)1/2 , due to Burbidge et al. (1988). This approach to dealing with zeros has much better properties than the usual method of adding a small constant inside the log and similarly can help mitigate the effect of skewness in the outcome distribution. It also allows us to maintain the basic fixed effects OLS specification. While interpreting magnitudes is less straightforward,11 the main takeaway from Table D.6.1 is that the results look very similar to the baseline extensive margin specification albeit slightly less precise. We increase precision by winsorizing the top 5th percentile of #events to further deal with the extreme skew (see the right graphs in Figure D.6.1). 11 Except for very small outcome values, the transformation can be interpreted in approximately the same way as a log dependent variable. 73 Table D.6.1: Intensive Margin Specification: Number of Conflict Events Administrative Unit: post-split post-split × ∆ ethnic polarization post-split × ∆ ethnic fractionalization post-split × ∆ religious diversity Number of District-Months Dep. Var. Mean, Pre-Split Original District (1) Parent & Child (2) Parent Child (3) (4) -0.038 (0.060) 0.021 (0.031) 0.010 (0.046) 0.031 (0.026) -0.021 (0.033) 0.041 (0.025) -0.018 (0.019) -0.006 (0.014) -0.045 (0.051) 0.030* (0.018) 0.046 (0.049) -0.005 (0.032) -0.008 (0.032) 0.060* (0.036) -0.037 (0.030) -0.005 (0.016) 7,956 0.96 20,220 0.44 7,956 0.68 12,264 0.31 Notes: The dependent variable is the hyperbolic sine transformation of the number of social conflict incidents in the given month. We winsorize at the 95th percentile of the outcome distribution. Otherwise, the specification is the same as in the baseline Table 3 with time and district FE, district-specific time trends, and standard errors clustered at the original district level. Significance levels: ∗ : 10% ∗∗ : 5% ∗ ∗ ∗ : 1%. D.6.2 Omitting Later Entrants to SNPK Data Table D.6.2 omits districts that enter the SNPK data in 2005, thereby ensuring a balanced panel. The similarity in results is reassuring insomuch as these later entrants were selected on account of policy concerns about recent violence. Table D.6.2: Alternative Time Restriction: Excluding 2005 Entrants to SNPK Administrative Unit: post-split post-split × ∆ ethnic polarization post-split × ∆ ethnic fractionalization post-split × ∆ religious diversity Number of District-Months Dep. Var. Mean, Pre-Split Original District (1) Parent & Child (2) Parent Child (3) (4) -0.015 (0.032) 0.046* (0.023) -0.006 (0.022) 0.017 (0.016) -0.018 (0.027) 0.041* (0.024) -0.008 (0.015) -0.005 (0.013) -0.013 (0.031) 0.034** (0.016) 0.025 (0.026) -0.025 (0.021) -0.022 (0.031) 0.060* (0.031) -0.024 (0.025) -0.001 (0.016) 5,196 0.52 13,020 0.31 5,196 0.43 7,824 0.24 Notes: This table drops all districts that entered the SNPK conflict data starting in 2005, thereby imposing a balanced panel. The specification is otherwise the same as in the baseline Table 3 with time and district FE, district-specific time trends, and standard errors clustered at the original district level. Significance levels: ∗ : 10% ∗∗ : 5% ∗ ∗ ∗ : 1%. 74 D.6.3 Omitting Historic Conflict Zones Table D.6.3 below excludes the two regions of Indonesia with the most intense civil conflict in the late 1990s and early 2000s at the onset of democratization. Panel A excludes districts in the Maluku islands, which saw fierce interreligious warfare from early 1999 through early 2002. Panel B excludes the entire province of Aceh, which was home to a longstanding guerilla movement to secede from Indonesia. The violent campaigned ended with a peace agreement in mid-2005. Omitting either of these two regions leaves the main takeaways intact, which is reassuring from the standpoint of generalization outside historic conflict zones per se. Table D.6.3: Excluding Civil War Regions Administrative Unit: Original District (1) Parent & Child (2) Parent Child (3) (4) Panel A: Excluding Maluku post-split post-split × ∆ ethnic polarization post-split × ∆ ethnic fractionalization post-split × ∆ religious diversity Number of District-Months Dep. Var. Mean, Pre-Split -0.025 (0.029) 0.067*** (0.021) -0.016 (0.021) 0.021 (0.013) -0.020 (0.023) 0.044*** (0.008) 0.009 (0.012) -0.012 (0.016) -0.014 (0.028) 0.043*** (0.012) 0.071** (0.035) -0.026 (0.026) -0.026 (0.029) 0.051*** (0.014) -0.002 (0.017) -0.015 (0.018) 6,900 0.64 17,100 0.38 6,900 0.53 10,200 0.29 Panel B: Excluding Aceh post-split post-split × ∆ ethnic polarization post-split × ∆ ethnic fractionalization post-split × ∆ religious diversity Number of District-Months Dep. Var. Mean, Pre-Split 0.027 (0.023) 0.026* (0.014) 0.009 (0.021) 0.032** (0.015) 0.019 (0.020) 0.032 (0.022) 0.009 (0.013) -0.001 (0.012) 0.035 (0.029) 0.024 (0.015) 0.037 (0.030) -0.017 (0.027) 0.009 (0.024) 0.048* (0.026) 0.001 (0.019) 0.001 (0.015) 6,696 0.55 17,340 0.30 6,696 0.45 10,644 0.21 Notes: Panel A excludes the districts in the Maluku islands, and B excludes districts in Aceh. The specification in both panels is otherwise the same as in the baseline Table 3 with time and district FE, district-specific time trends, and standard errors clustered at the original district level. Significance levels: ∗ : 10% ∗∗ : 5% ∗ ∗ ∗ : 1%. 75 D.6.4 Alternative Identification Strategy Table D.6.4 considers an alternative strategy that removes the district fixed effects altogether and incorporates an additional control group in those districts that never split from 2000 to 2010. In this case, the coefficients identify a mixture of differential effects of redistricting (i) among early versus late splitters and never-splitters, and (ii) splitters versus never-splitters. We include province×month fixed effects to ensure that each of these comparisons takes place between nearby districts. Note that the removal of district fixed effects means that we not longer identify within-district changes in conflict but rather cross-sectional differences (i) and (ii) within a given province akin to a nearest (geographic) neighbor matching-type design. In other words, all estimates below are with reference to the mean social conflict among never-splitters within the same province in a given month. Interestingly, as seen in Table D.6.4, this alternative identification strategy delivers estimated effects of changes in ethnic divisions among those that undergo redistricting that looks very similar to the baseline. Table D.6.4: Expanding the Counterfactual to Include Nearby Never-Splitters Administrative Unit: post-split post-split × ∆ ethnic polarization post-split × ∆ ethnic fractionalization post-split × ∆ religious diversity Original District (1) Parent & Child (2) Parent Child (3) (4) 0.067 (0.051) 0.038*** (0.011) 0.012 (0.029) 0.015 (0.016) -0.146*** (0.044) 0.030** (0.012) 0.031** (0.015) 0.011 (0.015) -0.102* (0.052) 0.057** (0.023) -0.006 (0.037) -0.039 (0.040) -0.249*** (0.050) 0.047* (0.024) 0.028 (0.023) 0.030 (0.022) 22,896 0.62 35,160 0.55 22,896 0.61 27,204 0.56 Number of District-Months Dep. Var. Mean, Pre-Split Notes: This table adds districts in the SNPK data that never split over the sample period. Relative to the baseline Table 3, the specification therefore omits district FE and district-specific time trends but includes province× month FE. Standard errors clustered at the original district level. Significance levels: ∗ : 10% ∗∗ : 5% ∗ ∗ ∗ : 1%. 76 D.6.5 Reweighting for External Validity (within Indonesia) As discussed in Section 5.6, we consider a reweighting approach to account for the two dimensions of selection in our generalized DiD sample of districts. First, Table D.6.4 above notwithstanding, our results are based solely on districts that split between 2000 and 2010. We know from prior work and our analysis that this sample of districts look different along a number of observable dimensions. One way to assess whether these differences are important for our results is to reweight the estimates by the inverse probability of redistricting, assigning greater weight to those original districts that look more like the ones that never split. We do this using a standard inverse probability weighting (IPW) approach in which we first predict the cross-sectional probability of redistricting and then apply IPW to our main estimates from Table 3. We estimate that probability based on a logit specification, using Lasso to select relevant variables from a set of 67 baseline district characteristics used for other exercises in Section 5.5. Applying the IPW in Table D.6.5 leaves our key results unchanged. Table D.6.5: Inverse Probability Weighting (IPW) for External Validity Administrative Unit: post-split post-split × ∆ ethnic polarization post-split × ∆ ethnic fractionalization post-split × ∆ religious diversity Number of District-Months Dep. Var. Mean, Pre-Split Original District (1) Parent & Child (2) Parent Child (3) (4) -0.019 (0.027) 0.036 (0.026) -0.021 (0.037) 0.001 (0.014) -0.016 (0.019) 0.037* (0.021) 0.003 (0.014) -0.014 (0.013) -0.001 (0.026) 0.030* (0.015) 0.053** (0.024) -0.033 (0.022) -0.028 (0.024) 0.057** (0.027) -0.019 (0.023) -0.013 (0.017) 7,956 0.57 20,220 0.33 7,956 0.47 12,264 0.25 Notes: This table reweights each observation in the baseline specification in Table 3 by its inverse probability of redistricting. The IPW are estimated in an initial step based on a logit specification with a battery of Lasso-selected controls. Standard errors are clustered at the original district level Significance levels: ∗ : 10% ∗∗ : 5% ∗ ∗ ∗ : 1%. 77 E E.1 Additional Evidence Supporting Political Violence Results in Section 6 Further Background on Ethnicity in Indonesian Politics Section 2.3 offered background on the changing role of ethnicity in Indonesian politics that could be broadly summarized in the following four takeaways: (i) ethnicity is an important organizing principle for political mobilization, (ii) ethnic-based clientilism and patronage networks are pervasive, (iii) decentralization and direct, majoritarian mayoral elections deepen (i) and (ii), and (iv) redistricting further amplifies all of these forces. Here, we provide additional background from the political science literature as well as fresh empirical evidence consistent with this context. Political Science Literature. Wilson (2015, p. 92) offers a helpful summary of views on ethnicity and patronage in the context of redistricting: “As local government and administrative boundaries were altered, ‘local selfishness’ was reinforced, resulting in conflicts and tensions at the local level (Firman 2013, 180). Just like national politics, local-level politics was an intense ‘arena of contestation between competing coalitions of social interests’ as networks that had relied upon central state patronage or been regime middlemen moved to establish new means to access resources (Hadiz, 2011a, 171). This contestation involved renegotiating the boundaries of collective identities, in doing so defining a social economy of who had to access to what, and under what circumstances. According to Klinken, from 1998 local elites throughout the country attempted to build ‘an exclusive discourse of ethnicity’, one that in its construction of group identity formed a ‘language with which elites compete for power by mobilising supporters’ (Klinken 2002, 68).” In the context of our case study in Section 6.4, Kobayashi (2011) notes from personal interviews that “A Dayak politician, a strong supporter of the creation of Bengkayang district, clearly explained that increase of Dayak government employees was one objective of pemekaran [redistricting]. A Dayak department head admitted that pemekaran increased job opportunities for Dayaks in government by commenting that he himself would not have been promoted to the position of department head without creation of Bengkayang. Diversity and Close Elections. Table E.1.1 demonstrates that ethnic diversity is associated with significantly closer mayoral elections. In particular, we regress the victory margin for the winning candidate on ethnic and religious diversity within the newly created parent and child districts. We consider both the first and second (when possible) quinquennial direct election after redistricting.1 The main takeaway is that greater diversity is associated with closer elections, consistent with the importance of ethnic mobilization highlighted in recent literature. Column 1 shows this when pooling across both the first and second elections taking place in the new parent and child districts. Both polarization (P ) and fractionalization (F ) matter, though the former is more precisely estimated. The effect sizes, though, are not trivial. A one standard deviation increase in P or F is associated with 10 percent lower victory margin relative to a mean of around 0.14 across all elections from 2005–2014 in these new districts. Results look similar if not slightly more pronounced for second elections. Looking separately at parent and child contests, however, reveals a difference between P and F . In parent districts, fractionalization matters much more than polarization whereas the opposite holds for child districts. These patterns line up nicely with the results in Table 7 where P does more than F to exacerbate violence around elections in child districts whereas the opposite holds for parent districts. Though perhaps consistent with differences in the stakes of political control (as noted in Section 6), these differences between P and F in Table E.1.1 are not statistically significant. The patterns for religious polarization are similarly in line with the differential effects of ∆religious diversity for parent relative to child districts in Table 7. Together, the results are suggestive about the importance of group configuration in driving electoral competition.2 1 As discussed in Appendix F, several newly created districts had not yet had their second election by the end of our study period, while others have missing data on election outcomes. 2 It is also worth noting that we can estimate the relationship between ∆diversity and ∆victory margins for 22 parent and child districts with a direct election at the original district level prior to redistricting. In particular, we find that a one standard 78 Table E.1.1: Diversity and Close Elections After Redistricting Dependent Variable: Victory Margin for Winning Mayoral Candidate in Parent/Child Parent Child All 1st 2nd All 1st 2nd All 1st (1) (2) (3) (4) (5) (6) (7) (8) Administrative Unit Which Election? ethnic polarization, new district ethnic fractionalization, new district religious polarization, new district 2nd (9) -0.013*** (0.004) -0.006 (0.010) -0.015* (0.009) -0.012* (0.007) -0.006 (0.012) -0.019* (0.010) -0.031*** (0.009) -0.006 (0.015) -0.028** (0.013) -0.012 (0.014) -0.008 (0.015) -0.042** (0.018) -0.003 (0.016) -0.029 (0.030) -0.015 (0.023) -0.014 (0.021) -0.031 (0.021) -0.071*** (0.026) -0.015*** (0.005) 0.001 (0.013) -0.002 (0.009) -0.016* (0.008) 0.004 (0.015) -0.016 (0.012) -0.029** (0.013) 0.026 (0.018) -0.001 (0.015) 113 0.14 103 0.14 68 0.13 44 0.14 34 0.14 32 0.14 69 0.14 69 0.14 36 0.12 Number of Districts Mean Victory Margin Notes: This table presents simple regressions relating ethnic diversity in the newly created parent/child districts to the victory margin in the first and second direct mayoral elections post-redistricting. Columns 1–3 pool parent/child districts, and columns 4–9 examine each separately. The ∆diversity measures are normalized, and standard errors are clustered at the original district level. Significance levels: ∗ : 10% ∗∗ : 5% ∗ ∗ ∗ : 1%. Ethnic Divisions and Preferences for Mayoral Candidates. We draw upon the Indonesia Family Life Survey (IFLS) to provide some evidence in line with these claims as they relate to border-induced changes in ethnic and religious divisions. In particular, we draw upon the 2014 round of data, which asks individuals “What factors do you consider in electing a mayor?”. We observe individuals in 40 of the parent and child districts in our main sample. In Table E.1.2 below, we control for basic demographics and relate ∆diversity to preferences over a large set of mayoral qualities. The results suggest that changes in ethnic divisions as a result of redistricting are strongly associated with preferences for mayor’s ethnicity as well as their provision of patronage. We find weaker correlations with mayoral experience, political affiliation and proposed program quality, among others. Note that this observation is at the end of the study period by which time many of these districts have had multiple mayoral elections, some of which may have been among those that witnessed violence of the sort identified in Section 6 of the paper. Table E.1.2: Changes in Ethnic Divisions and Preferences for Mayoral Candidates Dep. Var. (binary): Respondent in 2014 Believes that the Mayor’s . . . Is Important Appearance Popularity Program Political Religion Ethnicity Experience Patronage Quality Affiliation (1) (2) (3) (4) (5) (6) (7) (8) ∆ ethnic polarization ∆ ethnic fractionalization ∆ religious polarization Number of Districts Number of Districts Dep. Var. Mean 0.029** (0.013) -0.002 (0.012) 0.011 (0.009) -0.010 (0.014) -0.029** (0.013) 0.008 (0.017) -0.002 (0.010) 0.013 (0.010) -0.003 (0.009) 0.032 (0.020) 0.014 (0.016) -0.001 (0.014) 0.087** (0.037) 0.056 (0.045) 0.031 (0.026) 0.086*** (0.023) 0.044** (0.017) 0.004 (0.012) -0.012 (0.009) -0.006 (0.009) -0.013*** (0.005) 0.045*** (0.012) 0.025*** (0.006) 0.002 (0.004) 1,887 40 0.75 1,887 40 0.76 1,887 40 0.93 1,887 40 0.70 1,887 40 0.77 1,887 40 0.60 1,887 40 0.93 1,887 40 0.46 Notes: The dependent variable in each column is a binary indicator that equals one if the respondent in the 2014 IFLS agrees that the mayoral candidates’ given trait is an important factor in determining his/her vote. The regressions control for age, age squared, education level fixed effects, and gender. The ∆diversity measures are normalized, and standard errors are clustered at the district level. Significance levels: ∗ : 10% ∗∗ : 5% ∗ ∗ ∗ : 1%. deviation increase in ∆P (∆F ) is associated with a 2.2 p.p. (3.3 p.p.) reduction in ∆victory margin relative to its mean of 5.3 percent. There are only 10 prior elections and hence it is not meaningful to conduct inference, but the patterns are nevertheless supportive of the level results in Table E.1.1. 79 E.2 Close Elections and Conflict The first set of results in Table E.2.1 demonstrates that violence is more likely around new elections after redistricting when those contests are closely contested. In particular, we interact the post-split×election period indicator with the victory margin (ranging from 0.004 to 0.55). Panel A examines the baseline outcome of any social conflict, and Panel B examine the intensive margin number of conflict incidents transformed via the inverse hyperbolic sine used in baseline robustness checks in Appendix D.6.1. This latter specification allows for the possibility that the intensive margin may be differentially more important around election periods, which may be generally more intense periods of violence. Together, these results are broadly consistent with the fact that victory margins are significantly lower in more diverse, newly created districts as seen in Appendix Table E.1.1. Table E.2.1: Differential Conflict Around Close Elections After Redistricting Administrative Unit Parent Child (1) Parent Child (2) (3) Panel A: Any Social Conflict post-split post-split × 1st election period post-split × 1st election period × victory margin Number of District-Months Dep. Var. Mean 0.020 (0.029) 0.069 (0.049) -0.259 (0.155) -0.002 (0.038) -0.030 (0.063) 0.077 (0.201) 0.025 (0.036) 0.091** (0.035) -0.427** (0.206) 18,120 0.35 7,176 0.51 10,944 0.26 Panel B: # Social Conflict Events Hyperbolic Inverse Sine post-split post-split × 1st election period post-split × 1st election period × victory margin Number of District-Months Dep. Var. Mean 0.016 (0.059) 0.204** (0.097) -0.652* (0.341) -0.050 (0.105) 0.236* (0.126) -1.465*** (0.515) 0.053 (0.050) 0.106* (0.053) -0.306 (0.355) 18,120 0.52 7,176 0.84 10,944 0.34 Notes: This table examines interactions of the first mayoral election period with the victory margin in that election. The interaction of post-split and that victory margin is included but not shown. The specification is otherwise similar to the one in Table 7. Significance levels: ∗ : 10% ∗∗ : 5% ∗ ∗ ∗ : 1%. Looking across specifications, the evidence in Table E.2.1 suggests that after redistricting, violence is significantly more pronounced around close mayoral elections. These patterns are consistent with both (i) the qualitative background on election violence and incident descriptions discussed in Appendix C.2, and (ii) the conflict-amplifying effects of ∆diversity around elections seen in Table 7. While victory margins are potentially endogenous with respect to contemporaneous electoral violence, these results provide an important validation check on our interpretation. Together with the results linking ethnic diversity to closer elections , these findings paint a rich picture of how (changes in) ethnic divisions reshape conflict dynamics in settings with high returns to local political control. 80 F Data and Variable Construction We describe here the key variables and data sources used in the paper. Administrative Divisions Indonesia’s administrative divisions proceed down from the province to the district to the subdistrict to the village. These different levels of administration and our terminology for original, child and parent districts as defined below can be seen in Figure 3, which shows one of the districts in our study. Original District: This administrative unit defines all areas based on the 2000 boundaries. Child District: This represents the subdistricts that eventually become their own new district with an accompanying capital. Parent District: This represents the subdistricts that stay with the original district capital after other subdistricts split off. Post-Split: This is an indicator that turns on in the month that national parliamentary legislation first established a new district within the original district boundaries. In our main results, post-split equals one for the original district and parent district once the first child district splits off from 2000 onward. For child districts, the indicator equals one once it is ratified into law. Conflict The conflict data comes from the Indonesian National Violence Monitoring System (known by its Indonesian acronym SNPK).1 The data are reported at or below the 2011 district level, and hence we can calculate conflict within both the 2010 and 2000 borders over the years 2000–2014. Our main conflict measures are binary indicators for any conflict in a given district–month, but we also consider the number of incidents as a robustness check. Coders read articles and then assign the incident to mutually exclusive categories based on the underlying trigger. The incidents are first coded as domestic violence, violent crime, violence during law enforcement, or conflict. Eighty-two percent of incidents record some property damage, injuries, or deaths. Any Social Conflict: A dummy for whether SNPK recorded any non-crime and non-domestic violence incidents in the given month. Active Media: Using data obtained directly from SNPK managers on newspaper availability and usage by province and month, we calculate the number of papers used in any given province-month. All conflict specifications control flexibly for media availability by including dummies for the number of active papers in any given province-month. Entered 2005: SNPK coverage begins in 1998 for nine conflict-prone provinces and increases to 15 provinces plus parts of 3 provinces in greater Jakarta beginning in 2005. The data coverage is less complete and reliable for 1998 and 1999, and hence we focus on 2000–2014 for most results in the paper. We downloaded the data from http://www.snpk-indonesia.com, which is no longer active due to a recent contracting change. However, as of June 2016, the data hosted on and available through the World Bank website. A search in their Central Microdata Catalog for “Sistem Nasional Pemantauan Kekerasan” yields data from 1998 to 2014. 1 81 Diversity All measures are computed using the universal 2000 Population Census. Since this contains data at the village level, metrics can be constructed at both the 2000 and 2010 borders.  e Ethnic Fractionalization: Ethnic fractionalization in district d is given by F = N j=1 gj (1 − gj ), where Ne is the number of ethnic groups in the district, and gj is the population share of group j as reported in the 2000 Census. We observe over 1000 ethnicities and sub-ethnicities speaking over 400 languages. We also consider the related Greenberg-Gini version, which allows for non-binary distances between  e  Ne groups: G = N j=1 k=1 gj gk ηjk , where ηjk captures the linguistic distance between groups j and k as detailed below. Ne Ne 2 Ethnic Polarization: P = j=1 k=1 gj gk ηjk , where Ne , gj , and gk are as defined before, and ηjk is the distance between groups j and k. We map each ethnic group in the 2000 Census to a language in Ethnologue, which provides a full classification of the linguistic origins of each language (see the Online Appendix Section A.3 in Bazzi et al., 2016, for details). We set ηgh = 1 − sδgh , where sgh is the degree of similarity between the languages spoken by g and h as given by the ratio of common branches on the language classification tree to the maximum possible (14), and δ is a parameter that selects the level of linguistic dissimilarity to be emphasized. We set δ = 0.05 in our baseline, but consider alternate values. Ethnicities with missing languages are given province-specific average pairwise distances (η’s) between all other languages. Missing ethnic groups are necessarily grouped together, but separately from the “other” category, and also given province-specific average distances. We drop foreigners as they represent a minute fraction of the population, but we retain the ethnic Chinese.  Nr  Nr 2 Religious Polarization: Religious polarization, Relig = j=1 k=1 gj gk , where Nr is the number of religious groups, and gj (gk ) is the population share of group j (k). There are seven religions recorded in the Census, but in most districts, there is a single cleavage between a Muslim and a non-Muslim group. As a result religious polarization is effectively identical to religious fractionalization in our data (with a correlation of 0.96). ∆ Ethnic Polarization: To examine changes in diversity at the original district level, we compute the population-weighted average polarization in the new units (children and parent district) and subtract the polarization in the original district. If original district O splits into parent P and child(ren) C1 (C2 if multiple), with populations GO = GP + GC1 (+GC2 ) the change in ethnic polarization  G G GP is ∆P = G PP + GCO1 PC1 + GCO2 PC2 − PO . We construct changes in ethnic polarization at the O child/parent level analogously as: ∆P = PP − PO for the parent and ∆P = PC − PO for each child. ∆ Ethnic Fractionalization: For original district O splitting into parent P and child(ren) C1 (C2 if multiple),  with populations GO = GP + GC1 (+GC2 ) the change in ethnic fictionalization is given by G G GP ∆F = G FP + GCO1 FC1 + GCO2 FC2 − FO . We construct changes in ethnic fractionalization at the O child/parent level analogously as: ∆F = FP − FO for the parent and ∆F = FC − FO for each child. ∆ Religious Polarization: For original district O splitting into parent P and child(ren) C1 (C2 if multiple), with populations GO = GP + GC1 (+G  C2 ) the change in religious polarization is given by ∆Relig = GC1 GC2 GP GO ReligP + GO ReligC1 + GO ReligC2 − ReligO . We construct changes in ethnic fractionalization at the child/parent level analogously as: ∆Relig = ReligP − ReligO for the parent and ∆Relig = ReligC − ReligO for each child. 82 Table F.1.1: Summary Statistics for Baseline Variables Mean Std. Dev. Min. Median Max. 2000 Borders: 52 Original Districts any social conflict incidents number of social conflict incidents post-split ethnic polarization ethnic fractionalization religious polarization ∆ ethnic polariation ∆ ethnic fractionalization ∆ religious polarization 0.631 2.631 0.787 0.017 0.612 0.119 -0.000 -0.078 -0.008 0.483 5.185 0.409 0.016 0.256 0.070 0.011 0.153 0.049 0.000 0.000 0.000 0.003 0.062 0.001 -0.062 -0.677 -0.192 1.000 1.000 1.000 0.013 0.689 0.130 0.000 -0.034 -0.000 1.000 89.000 1.000 0.095 0.957 0.233 0.061 0.193 0.109 2010 Borders: 133 Parent and Child Districts any social conflict incidents number of social conflict incidents post-split ethnic polarization ethnic fractionalization religious polarization ∆ ethnic polarization ∆ ethnic fractionalization ∆ religious polarization 0.364 1.035 0.768 0.017 0.609 0.122 -0.000 -0.059 -0.008 0.481 2.941 0.422 0.016 0.258 0.067 0.005 0.083 0.020 0.000 0.000 0.000 0.003 0.062 0.001 -0.035 -0.342 -0.129 0.000 0.000 1.000 0.013 0.682 0.131 0.000 -0.032 -0.001 1.000 76.000 1.000 0.095 0.957 0.233 0.008 -0.000 0.017 Notes: At the 2000 level, there are 52 districts and 7,956 monthly observations. At the 2010 level, there are 133 Districts (52 parents and 81 children) and 20,220 monthly observations. See Appendix F for variable definitions. Voting and Elections District Head Elections: District elections occur every 5 years. Prior to 2005, district head elections were conducted by parliament and varied across districts in terms of timing. From 2005 onward, district and vice-district heads were directly elected by plurality vote contingent on that vote being at least 30 percent. If not, a second round between the top two candidates takes place. District heads directly appoint subdistrict heads. We collect data on the date of and vote shares in all direct elections from documents published by the General Election Commissions, many of which were graciously provided by Monica Martinez-Bravo, Andreas Stegmann, and Audrey Sacks. Elections in child districts typically occur 1.5–2.5 years after the split. Elections in parent districts are determined by the pre-Suharto election cycles carried over into the democratic era (see Martinez-Bravo et al., 2017). 1st Direct Election Period: Using the exact date of all direct elections, we construct an indicator that equals one in the 6 month window around the parent/child’s first direct election date. In the case of the latest splits, this can occur pre-split. 2nd Election Period: We construct an indicator that equals one in the 6 month window around the parent/child’s second direct election date. There are some children (the latest splits) for which we do not observe a second election post-split. District Head Election Victory Margins: Using the General Election Commissions records, we compute victory margins in the district head elections conducted after redistricting. This continuous measure is simply equal to the vote share for the winner minus the vote share for the loser. 83 Control Variables We list here the rich set of 65 variables from 1999 and 2000 that we interact with post − split and use as controls to ensure that the cross-district variation picked up by post − split × Diversity is not picking up other observable differences across districts. These are carefully constructed from a variety of data sources, and are generally non-missing. Several variables are missing for at most one original district, and are imputed simply using the average across districts. PODES Variables We use the 2000 administrative village census (Potensi Desa or Podes) to construct a number of control variables relating to education, public goods provision, security, and development. Each of these measures are aggregated to the district level at both the original district level, and eventual, 2010 boundaries. Health Variables: We construct a variable for the number of health care facilities (polyclinics and PHCs) per capita in 2000 at the 2000/2010 district levels. We construct the (population weighted) share of villages that say they have a midwife available. Further, we construct the (population weighted) share of villages that say they have a doctor or access to a PHC. Education Variables: We construct the number of high schools per capita in 2000 at the 2000/2010 district levels. We also construct the number of Islamic schools per capita. Public Goods: We construct the (population weighted) share of villages that have access to water from a pump or a water company; have a trash disposal system (bin/hole); have most households using gas/kerosene or electricity; and have road lighting. We also use the number of households per capita with electricity, with a telephone, and with a television. Economy: We construct the number of permanent markets per-capita and the (population weighted) average distance to the nearest market. In addition we calculate the (population weighted) share of villages with a transportation hub (airport, seaport, or bus terminal). We also construct the (population weighted) share of villages reporting good or great economic conditions and the share of villages for which agriculture is the main source of income. Finally, we construct the (population weighted) average number of natural disasters in the past 3 years. Security: We construct the (population weighted) mean distance to the nearest police post and office. We construct two variables: the logarithm of (one plus) the distance to the nearest police outfit and the logarithm of (one plus) the distance to the nearest police office (which is always larger). Geography: We construct the (population weighted) share of villages on the shore, on the coast, in a valley, on a hill, on flat land, and at high altitude. We also construct the logarithm of total land area. Importantly, we also include the logarithm of (one plus) the (population weighted) mean distance from the village to the 2000 capital and the logarithm of (one plus) the (population weighted) mean distance from the village to the sub-district capital. Census Variables Using the 2000 population micro census we construct a number of additional demographic variables. We construct each of the below at both the original district and the eventual 2010 boundaries. Population Shares: We use the Population Census in 2000 to compute the share of the population that is aged 5–14 and 15–29 at the original, child, and parent district levels. We also include the logarithm of 84 total population and mean household size. Education Shares: We compute the share of the population whose highest educational attainment is primary school, as well as the share of the population whose highest educational attainment is post-primary. Migration: We compute the share of the population who arrived from a different province in the last five years and the share arrived from a different district in the last five years. Geography: We include an indicator for the share of the population living in rural areas. Sectors of the Economy: We compute the fraction of workers in agriculture, the fraction of workers in forestry, fishing and livestock, and the fraction of workers in other sectors (industry, trade, service, and transport). Government Transfers District Revenues: District revenue figures come from the World Bank’s Indonesia Database for Policy and Economic Research (DAPOER), which in turn obtains data from the Indonesia Ministry of Finance. They are given for each district at the time of existence up to 2013. We add in the 2014 revenue data directly from the Ministry of Finance. Population data is taken from the same dataset. We construct all revenue and population variables at the original district level by aggregating up to the 2000 borders. Both the population and revenue data are missing in some cases. In our baseline, we impute these missing observations as described below, but our results are very similar if either or both variables are left as missing. Population data is missing in 2014 for all districts and in 2000 for 6 original districts. We impute population using the preceding/following year and the median growth rate of 1.5 percent. Revenue data is missing in 2000 for 4 of our original districts, and thereafter there are occasional withindistrict gaps in the data. These gaps occur between 2001–2005 and to a lesser extent between 2012–2013, never exceeding 8 missing districts. We impute missing revenues using annual median revenue growth rates. All revenue figures are adjusted for inflation using 2010 as the base year. Total district revenue comes from the general allocation grant (Dana Alokasi Umum, DAU), the special allocation grant (Dana Alokasi Khusus, DAK), shared taxes, shared natural resource rents, as well as limited own revenue, and limited revenue from other sources. We construct 5 control variables, all using the information from year 2000, that account for all of district revenues while keeping information disaggregated: grants (DAU + DAK) per capita, shared taxes per capita, shared natural resource rents per capita, own revenue per-capita, and other revenue per capita. This allows natural resources, for example, to enter separately. These are necessarily only computed at the original district level, and are included at that level in the child/parent regressions. When we examine how transfers evolve over time in Appendix A, we use the full time series of total revenues less own revenue, to capture total transfers from the central government. At the Original District level we simply use the logarithm of real total transfers. At the parent and child level, we have to make an additional assumption, since we do not observe how parent and child districts shared transfers pre-split. Specifically, we assume that parent and child districts get their initial 2000 population share of the original district transfers and use these values up to and including the year of the split. For all subsequent years, we use actual realized transfers at the lower level, imputing any missing values using the prior years value and median growth rates. 85 Light Intensity Fraction of District Area Covered by Lights: We use night lights in 2000 as a proxy for initial GDP (Henderson et al., 2012). We have data on the coverage of each village by any lights in 2000, and take the average percentage coverage across villages at the original district and eventual, 2010, borders. Village Level Light Data: For our ethnic favoritism results, we also use the village level light directly. When looking at how nighttime light intensity varies by share of residents in 2000 belonging to the largest ethnic group in the eventual child district, we use the fraction of the village area covered with any lights in each year 2000–2013. Other Variables Climate: We compute the population weighted average rainfall and temperature from 1948 to 1978 using village level information from NOAA-GPCP. GIS Data: We compute the logarithm of the population weighted average distance to the nearest road, to the coast, and to the nearest river. We also compute the logarithm of elevation (30 as), and the ruggedness of the terrain (RUGGED3). We include the population weighted average forest coverage in 2000. Finally we include detailed indicators for the slope of the terrain (slope 1– 8). See Bazzi et al. (2016) for details on the underlying sources and construction. Cash Crop Share: We use the 2003 administrative village census (Potensi Desa or Podes) to calculate the value (price × quantity) of each crop produced within the 2000 and 2010 district borders. To proxy for agricultural resources, we compute the fraction of district agricultural output that is composed of nearly 30 cash crops, the most important among which include palm oil, rubber, coffee, and cocoa. Party Vote Share Polarization: We use the 1999 parliamentary (proportional system) vote shares for all 48 political parties at the subdistrict level to construct a measure of party polarization  at the original district and eventual 2010 borders level. The measure for a given district is given by i j share2i sharej over each party i and j. The underlying data was graciously shared by Audrey Sacks. Time Varying Transfers and Distance ∆ Distance: Using PODES 2000 and PODES 2011 we calculate the population-weighted average distance (in km) to the district capital across villages within the eventual parent and child units. At the child and parent level we construct ∆ Distance as the difference in the natural logarithm of reported distance to the capital in 2011 less that in 2000. At the original district we take the average of these measures across parent and children, weighted by district population. ∆ Transfers: We use the information from DAPOER on total transfers less own revenue (which encompasses the general and specific allocation grants and all tax and natural resource sharing). As discussed above, we impute missing values using median annual growth rates and we adjust for inflation. At the original district level, we compute ∆ Transfers as the change in the logarithm of real transfers post-split. We compare the average post-split to the average pre-split (including the year of the split). We do not observe how parent and child districts shared transfers pre-split. So for the child and parent level we assume original district transfers were divided according to the child/parents population share in all pre-split years and in the year of the split. Thereafter, we use actual realized transfers at the lower level, imputing any missing values using the prior years value and median growth rates. Similar to the original district level, we then construct ∆ Transfers as the change in the average logarithm of real transfers post-split to that pre-split.