The Determinants of Low-Intensity Intergroup Violence: The Case of Northern Ireland Online Appendix Laia Balcells (Duke University), Lesley-Ann Daniels (Institut Barcelona d Estudis Internacionals & Universitat Pompeu Fabra), Abel Escribà-Folch (Universitat Pompeu Fabra) Table A1. Descriptive statistics of core variables in the models (2005-2012) Variable Observations Mean Std. deviation Min. Max. Sectarian Violence 4656 1.88 4.67 0 78 Log square kilometers 4656 2.21 1.49 0.17 5.33 Log population 4656 7.95 0.37 6.57 9.19 Urban ward 4656 0.52 0.49 0 1 Border with Rep. of Ireland 4656 0.05 0.21 0 1 Males 16-39 (%) 4656 16.38 2.66 9 37.07 Unemployment (%) 4656 3.72 2.53 0.3 17.1 Housing benefits (%) 4656 7.87 5.99 0.28 35.28 Conflict deaths (log) 4656 1.14 1.09 0 5.03 Parity 4656 0.60 0.30 0.03 1 Polarization 4656 0.61 0.28 0.04 0.99 Catholics (%) 4656 43.35 31.93 0.90 99.04 Protestants (%) 4656 52.98 30.32 0.66 97.05 Catholic ward (>50%) 4656 0.39 0.49 0 1 Nº of Prot. (>50%) neighboring wards 4656 3.03 2.18 0 11 Catholic ward (>60%) 4656 0.33 0.47 0 1 Nº of Prot. (>60%) neighboring wards 4656 2.57 2.21 0 11 Avg. % of Prot. in neighboring wards 4656 53.49 24.13 1.54 94.72 Avg. % of Cath. in neighboring wards 4656 42.85 25.77 2.53 98.04 Max % of Prot. in neighboring wards 4656 73.05 22.72 2.11 97.05 Avg. nº of incidents in neighboring 4656 1.86 3.02 0 44 wards Number of neighboring wards 4656 5.23 1.74 1 11 Log peace lines 4656 0.06 0.28 0 2.08 Log other crimes (not sectarian) 4656 4.83 0.82 1.79 8.45 % of seats Republican parties (LGD) 4656 39.05 24.56 0 80 % of seats Loyalist parties (LGD) 4656 51.67 19.90 13.33 83.33
Robustness checks The main findings from our baseline models (Table I Model 5 and Table II Model 3) are robust to a number of changes to the specification which we briefly described in the paper but that we report and further discuss in this Appendix. First, to ensure that our results are not dependent on the specific model specifications reported in the paper, we estimate our baseline models without control variables and year and LGD fixed-effects. The results are shown in Table A2 and reveal that the main findings are not dependent on the inclusion of any specific control or fixedeffects. Models 1-3 report the results for parity (Table I column 5) without controls and fixed-effects, without controls but including year and LGD fixed-effects, and with controls but without year and LGD fixed-effects. Models 4-6 do the same for the model testing inter-ward dynamics (Table II column 3). Second, in Table A3 we report a series of models with some additional control variables and additional specifications. First, the positive impact of intra-ward parity is robust to the inclusion of two extra control variables. The first is a dummy variable indicating wards that are urban (compared to rural or mixed ). The second is an alternative measure of economic scarcity, namely, the number of people claiming welfare assistance for housing (as a percentage of a ward s population). 1 Our main results remain unchanged (see Model 4 of Table A3). The urban dummy is highly correlated with the size of wards in square kilometers (ρ = -0.87) and it is not significant. When the latter is excluded from the model, the urban dummy is positive and significant. The impact of the alternative poverty measure is positive and significant. However, the housing benefits variable is highly correlated with unemployment (ρ = 0.77), and that is why we exclude it from the main models. The baseline model testing the existence of segregation and cross-ward violence (Table II column 3) is also robust to the inclusion of these two extra controls (urban dummy and housing benefits claimants) plus a dummy variable indicating if a ward borders with the Republic of Ireland to capture the possibility of cross-border group activity (Model 5 Table A3). Bordering the Republic of Ireland has been pinpointed as an explanatory factor of violence during the Troubles (Mueller, Rohner & Schoenholzer, 2013) as British 1 Both variables are obtained from the NISRA.
forces were attacked in border control points. Interestingly, this variable has no significant effect for the post-conflict period under study which confirm that the spatial distribution of violence has somewhat changed. The inclusion of these three controls does not alter our main result. Further, Model 8 (Table A3) reports the results of this same baseline model including two control variables capturing sectarian parties electoral results at the LGD level. Concretely, we include the percentage of seats won by nationalist parties (i.e., Sinn Féin (SF), and Social Democratic and Labour Party (SDLP)) and the percentage won by unionist parties (i.e., Democratic Unionist Party (DUP), Ulster Unionist Party (UUP), Progressive Unionist Party (PUP), Ulster Democratic Party (UDP), United Kingdom Unionist Party (UKUP), and United Unionist Assembly Party (UUAP)). The inclusion of these controls does not affect our main result and suggests that political representation does not reduce violence at the ward level. 2 Some additional specifications and results mentioned in the paper are reported in Table A3. In Model 1 we use the polarization index instead of parity. Both are highly correlated in two-group contexts and, so, the results are extremely similar. Model 2 includes an interaction between parity and the percentage of young males to test the impact of youth unemployed bulges and shows that the impact of parity does not significantly increase as the percentage of young males increases in a ward. Model 3 includes an interaction between parity and unemployment. The results show that the impact of parity increases with unemployment, which increases the stakes of competition and reduces the opportunity costs of violence. Model 6 shows the stronger result obtained if we use a 60% threshold to identify and code mostly Catholic and Protestant wards instead of 50% (as we did in Model 5 in Table II). In Model 7 we use the percentage of Protestants in a ward and interact it with the average percentage of Catholics in the neighboring wards to show that the effect of neighbors ethnic composition is also present in Protestant areas surrounded by mostly Catholic wards. In other words, we show that the effects are symmetric to those presented in the paper and explained by the existence of interface areas. Finally, Model 9 shows the inter-ward effect by including also the interaction between the percentage of 2 The impact of these electoral variables should however be interpreted with caution due to potential reversed causality.
Protestants in neighboring wards and the squared percentage of Catholics in a ward, to capture non-linearity. Additionally, Table A4 reports models that show the impact of the interaction between parity and two rough proxies for segregation at the ward level: the number of peace-lines and the number of deaths occurred during the Troubles. Segregation exists at the within-ward level but it is difficult to capture. A ward can be diverse but still highly segregated internally. The results indicate that parity and segregation contribute to more violence at the ward level. The coefficients of the interactions are positive and significant. Third, we show that our findings are not driven by a few influential cases. In particular, we have re-run our baseline models excluding Derry and Belfast, which are the two most populated cities in Northern Ireland and where sectarian violence has been traditionally higher. Table A5 reports the results and shows that the main findings are robust to the exclusion of these cases from the sample. Fourth, the likelihood ratio tests of the models in the paper show that the overdispersion parameter is non-zero, which clearly indicate that the negative binomial model is more appropriate than the Poisson model. Given that standard negative binomial regression models already account for overdispersion and that we do not consider zero counts to be caused by a separate process, we do not use zero-inflated models in our main estimations. However, to check the robustness of our results, Table A6 reports our main models estimated using zero-inflated negative binomial regression. Again, our main findings remain largely unaltered. Fifth, we report models estimated using negative binomial models for panel data with ward random- and fixed-effects. Further, the baseline models have been also re-run using ward dummies instead of LGD fixed-effects. 3 These models are however highly problematic in dealing with unit effects because our main independent variables barely vary over time. 4 Nonetheless, we report these models in the Appendix to prove that the main findings are not model dependent. Indeed, the reported results in Table A7 show that, despite these caveats, the impact of parity and the interaction between the percentage of 3 Besides, in these models (3 and 6), errors are clustered at ward level. 4 Recall that census data is only collected every 10 years (2001 and 2011 within the period for which sectarian violence data is available).
Catholics in a ward and the average percentage of Protestants in neighboring wards are both positive and significant. As expected, the estimated effects are smaller though. Finally, Table A8 shows the results for parity at a more aggregated level of analysis, namely the Local Government District (LGD). Model 1 does not include year fixed-effects while Model 2 does. The results reveal that the impact of parity is still strong and positive at the LGD level, especially if year fixed-effects that control for temporal and common shocks are included. The estimated coefficient is significant (at the 0.05 level) when the year fixed-effects are included and very similar in size to that estimated in Model 5 (Table I). This result indicates that group parity is also associated with more violent incidents at the LGD level.
Table A2. Robustness tests: Excluding controls as well as year and LGD fixed-effects DV: Number of incidents of sectarian violence (1) (2) (3) (4) (5) (6) Sq. Km. (log) -0.33** -0.23** (0.030) (0.036) Population (log) 0.31* 0.083 (0.14) (0.14) Males 16-39 (%) 0.032 0.060** (0.020) (0.022) Unemployment (%) 0.085** 0.066** (0.016) (0.016) Conflict deaths (log) 0.35** 0.11* (0.047) (0.050) Parity 0.91** 1.54** 1.85** (0.28) (0.15) (0.16) Incidents in neighbors (avg.) 0.15** (0.018) Neighbors (nº) 0.033 (0.028) Catholics (%) -0.034** -0.035** -0.033** (0.0057) (0.0038) (0.0036) Avg. % of Protestants in neighboring wards -0.037** -0.034** -0.027** (0.0052) (0.0047) (0.0041) Catholics(%)*Avg. Prot. 0.00078** 0.00079** 0.00067** (0.00010) (0.000071) (0.000066) Constant 0.046-0.19-3.91** 2.60** 2.59** -0.11 (0.21) (0.23) (1.09) (0.40) (0.37) (1.03) Log(dispersion) 1.03** 0.35** 0.42** 0.90** 0.30** 0.23** (0.075) (0.064) (0.060) (0.085) (0.070) (0.069) LGD fixed-effects N Y N N Y N Year fixed-effects N Y N N Y N N 4656 4656 4656 4656 4656 4656 Log-likelihood -7997.9-7218.7-7343.2-7834.3-7185.1-7120.5 Clustered standard errors in parentheses. + p < 0.10, * p < 0.05, ** p < 0.01.
Table A3. Robustness tests: Alternative specifications DV: Number of incidents of sectarian violence (1) (2) (3) (4) (5) (6) (7) (8) (9) Sq. Km. (log) -0.21** -0.20** -0.19** -0.12* -0.20** -0.21** -0.25** -0.25** -0.23** (0.028) (0.028) (0.028) (0.050) (0.066) (0.038) (0.035) (0.035) (0.034) Population (log) 0.71** 0.68** 0.70** 0.82** 0.71** 0.50** 0.65** 0.62** 0.75** (0.16) (0.16) (0.16) (0.16) (0.16) (0.19) (0.17) (0.17) (0.15) Males 16-39 (%) -0.023-0.040-0.026-0.016 0.015 0.036* 0.015 0.021-0.0057 (0.015) (0.039) (0.016) (0.016) (0.018) (0.018) (0.019) (0.018) (0.017) Unemployment (%) 0.24** 0.24** 0.18** 0.068* 0.074* 0.14** 0.17** 0.15** 0.21** (0.021) (0.022) (0.027) (0.029) (0.031) (0.027) (0.025) (0.023) (0.021) Conflict deaths (log) 0.20** 0.19** 0.20** 0.15** 0.12* 0.16** 0.15** 0.14** 0.16** (0.047) (0.047) (0.048) (0.045) (0.050) (0.059) (0.050) (0.050) (0.047) Polarization 2.30** (0.14) Parity 1.76* 1.59** 2.11** (0.82) (0.19) (0.12) Parity*Males 16-39 0.020 (0.049) Parity*Unemployment 0.12** (0.038) Urban ward 0.011 0.024 (0.14) (0.16) Housing benefits (%) 0.077** 0.044** (0.011) (0.011) Border with Rep. of Ireland -0.083 (0.21) Peace-lines (log) 0.23 (0.18) Incidents in neighbors (avg.) 0.053** 0.10** 0.057** 0.052** 0.068** (0.014) (0.016) (0.014) (0.014) (0.012) Neighbors (nº) 0.045+ 0.073* 0.045* 0.046* 0.034+ (0.024) (0.033) (0.023) (0.023) (0.020) Catholics (%) -0.040** -0.042** 0.075** (0.0036) (0.0037) (0.016) Avg. % of Prot. in neighboring -0.035** -0.035** -0.0048 wards (0.0047) (0.0047) (0.0052) Catholics*Avg. Prot. 0.00079** 0.00079** -0.00038 (0.000065) (0.000065) (0.00025) Catholics(%) ^2-0.00093** (0.00014) Catholics^2*Avg. Prot. 0.0000069** (0.0000024) Catholic ward (>60%) -1.05** (0.13) Nº Prot. neighbors (>60%) -0.10** (0.034) Catholic ward*prot. neighbors 0.37** (0.079) Protestants (%) -0.036** Avg. % of Catholics in neighboring wards (0.0041) -0.042** (0.0049) Protestants*Avg. Cath. 0.00081** (0.000065) Republican parties seats (%) 0.019 (0.014) Loyalist parties seats (%) 0.027* (0.012) Constant -5.98** -5.39** -5.46** -7.20** -3.45** -3.81** -2.71* -4.76** -5.84** (1.18) (1.27) (1.18) (1.22) (1.20) (1.41) (1.29) (1.62) (1.15) Log(dispersion) -0.085-0.076-0.091-0.14* -0.075 0.081-0.066-0.070-0.21** (0.061) (0.062) (0.062) (0.062) (0.071) (0.067) (0.068) (0.068) (0.070) LGD fixed-effects Y Y Y Y Y Y Y Y Y Year fixed-effects Y Y Y Y Y Y Y Y Y N 4656 4656 4656 4656 4656 4656 4656 4656 4656 Log-likelihood -6848.4-6855.2-6845.9-6800.5-6826.4-6978.8-6837.1-6835.6-6741.4 Clustered standard errors in parentheses. + p < 0.10, * p < 0.05, ** p < 0.01.
Table A4. Parity and internal segregation proxies DV: Number of incidents of sectarian violence (1) (2) Sq. Km. (log) -0.069+ -0.072* (0.036) (0.036) Population (log) 0.18 0.20 (0.15) (0.16) Males 16-39 (%) -0.014-0.015 (0.016) (0.017) Unemployment (%) 0.11** 0.12** (0.020) (0.021) Conflict deaths (log) 0.036-0.061 (0.047) (0.067) Parity 1.63** 1.48** (0.12) (0.16) Parity*Conflict deaths 0.15+ (0.082) Peace-lines (log) 0.25+ 0.47** (0.15) (0.11) Parity*Peace-lines 0.45+ (0.27) Other crimes (log) 0.50** 0.49** (0.065) (0.066) Incidents in neighbors (avg.) 0.082** 0.081** (0.011) (0.011) Neighbors (nº) 0.0057 0.0046 (0.020) (0.020) Constant -4.35** -4.36** (1.06) (1.07) Log(dispersion) -0.27** -0.27** (0.068) (0.068) LGD fixed-effects Y Y Year fixed-effects Y Y N 4656 4656 Log-likelihood -6698.4-6698.4 Clustered standard errors in parentheses. + p < 0.10, * p < 0.05, ** p < 0.01.
Table A5. Robustness tests: Excluding Belfast and Derry DV: Number of incidents of sectarian violence (1) (2) Sq. Km. (log) -0.22** -0.25** (0.030) (0.036) Population (log) 0.56** 0.65** (0.17) (0.18) Males 16-39 (%) 0.016 0.036 (0.022) (0.024) Unemployment (%) 0.18** 0.18** (0.025) (0.026) Conflict deaths (log) 0.099+ 0.14** (0.052) (0.053) Parity 1.88** (0.13) Incidents in neighbors (avg.) 0.100** (0.020) Neighbors (nº) 0.052* (0.024) Catholics (%) -0.047** (0.0042) Average % of Protestants in neighboring wards -0.041** (0.0048) Catholics*Avg. Prot. 0.00087** (0.000079) Constant -4.99** -2.94* (1.23) (1.31) Log(dispersion) -0.042-0.072 (0.068) (0.075) LGD fixed-effects Y Y Year fixed-effects Y Y Sample Excluding Belfast and Derry N 4008 4008 Log-likelihood -5243.2-5221.2 Clustered standard errors in parentheses. + p < 0.10, * p < 0.05, ** p < 0.01. Excluding Belfast and Derry
Table A6. Robustness tests: Zero-inflated negative binomial regression models DV: Number of incidents of sectarian violence (1) (2) (3) Sq. Km. (log) -0.33** -0.22** -0.21** (0.030) (0.049) (0.042) Population (log) 0.29* 0.060 0.42+ (0.14) (0.16) (0.22) Males 16-39 (%) 0.034+ 0.039 0.014 (0.021) (0.026) (0.017) Unemployment (%) 0.086** 0.048** 0.18** (0.016) (0.017) (0.026) Conflict deaths (log) 0.36** 0.12* 0.10+ (0.047) (0.061) (0.053) Parity 1.75** (0.17) Incidents in neighbors (avg.) 0.13** 0.049** (0.017) (0.014) Neighbors (nº) 0.026 0.046* (0.034) (0.023) Catholics (%) -0.025** -0.034** (0.0051) (0.0050) Average % of Protestants in neighboring wards -0.020** -0.032** (0.0045) (0.0049) Catholics*Avg. Prot. 0.00053** 0.00069** (0.000092) (0.000080) Constant -3.69** 0.23-1.17 (1.10) (1.17) (1.59) Inflate equation Conflict deaths (log) -0.24 0.0028-1.00* (0.16) (0.23) (0.45) Parity -27.7** (8.04) Sq. Km. (log) 0.073 0.82* (0.20) (0.37) Population (log) -0.30-5.32 (0.42) (4.57) Males 16-39 (%) -0.19 (0.17) Unemployment (%) -0.23* 0.088 (0.10) (0.18) Incidents in neighbors (avg.) -1.13** -0.42** (0.32) (0.15) Neighbors (nº) 0.0019 (0.090) Catholics (%) 0.049+ 0.12* (0.026) (0.053) Average % of Protestants in neighboring wards 0.040* 0.068 (0.019) (0.050) Catholics*Avg. Prot. -0.00077* -0.0015** (0.00033) (0.00043) Constant 1.67+ 2.20 17.2 (0.94) (3.49) (37.0) Log(dispersion) 0.39** 0.032-0.21* (0.060) (0.092) (0.080) LGD fixed-effects N N Y Year fixed-effects N N Y N 4656 4656 4656 Log-likelihood -7333.6-7078.6-6779.4 Clustered standard errors in parentheses. + p < 0.10, * p < 0.05, ** p < 0.01.
Table A7. Robustness tests: Ward Random- and Fixed-effects DV: Number of incidents of sectarian violence (1) (2) (3) (4) (5) (6) Sq. Km. (log) -0.22** -0.15** -0.73** -0.21** -0.14** -1.02** (0.025) (0.048) (0.081) (0.026) (0.052) (0.24) Population (log) 0.34** 0.15 0.43 0.23* 0.051 0.16 (0.099) (0.17) (0.48) (0.096) (0.17) (0.47) Males 16-39 (%) 0.0044 0.0076-0.016 0.015 0.0099 0.00080 (0.012) (0.016) (0.027) (0.011) (0.016) (0.024) Unemployment (%) 0.13** 0.094** 0.085** 0.13** 0.086** 0.099** (0.015) (0.019) (0.029) (0.015) (0.019) (0.025) Conflict deaths (log) 0.15** 0.047 0.74+ 0.028-0.099 1.87+ (0.036) (0.059) (0.39) (0.036) (0.065) (1.00) Parity 1.43** 0.67** 1.21** (0.11) (0.19) (0.39) Incidents in neighbors (avg.) 0.059** 0.057** 0.083** (0.0054) (0.0062) (0.014) Neighbors (nº) 0.048* 0.016-0.16 (0.020) (0.041) (0.16) Catholics (%) -0.036** -0.019** -0.038* (0.0029) (0.0053) (0.015) Avg. % of Prot. in neighboring wards -0.033** -0.023** -0.021 (0.0032) (0.0056) (0.015) Catholics*Avg. Prot. 0.00068** 0.00037** 0.00041+ (0.000047) (0.000086) (0.00022) Constant -3.10** -0.99-2.88 0.26 1.50 2.75 (0.78) (1.30) (3.88) (0.75) (1.33) (4.21) Log(dispersion) -1.11** -1.20** (0.086) (0.094) Unit effects RE FE FE (ward dummies) RE FE FE (ward dummies) Year fixed-effects Y Y Y Y Y Y N 4656 4344 4656 4656 4344 4656 Log-likelihood -6810.0-4861.5-6038.8-6710.8-4814.0-5991.9 Standard errors in parentheses. + p < 0.10, * p < 0.05, ** p < 0.01. In Models 3 and 6 errors are clustered at ward level.
Table A8. Sectarian incidents and parity at LGD level DV: Number of incidents of sectarian violence at LGD level (1) (2) Sq. Km (log) -0.081-0.058 (0.22) (0.18) Population (log) 1.36** 1.40** (0.48) (0.43) Males 16-39 (%) 0.17 0.085 (0.12) (0.11) Unemployment (%) 0.066 0.33** (0.049) (0.12) Conflict deaths (log) -0.20-0.30 (0.23) (0.20) Parity (LGD level) 1.31 2.04* (1.03) (0.80) Constant -14.4** -14.2** (5.45) (4.98) Log(dispersion) -0.67** -0.89** (0.17) (0.16) Year fixed-effects N Y N 208 208 Log-likelihood -898.3-874.9 + p < 0.10, * p < 0.05, ** p < 0.01. Standard errors clustered on LGD in parentheses. Year fixed-effects included but not reported in Model 2.
Authors elaboration. Reproduced with permission of Land and Property Services Crown Copyright 2015. Conflict deaths data source: Visualising the Conflict project, CAIN, http://www.cain.ulst.ac.uk/victims/gis/index.html Figure A1. Number of deaths during the Troubles at the ward level (1969-1998)
Log of avg. number of sectarian crimes (2005-12) 0 1 2 3 4 0 1 2 3 4 5 Log of number of deaths during conflict Including Belfast and Derry Excluding Belfast and Derry Fitted values Fitted values Figure A2. Past deaths and current sectarian crimes: The figure shows the relationship between the (logged) number of deaths during the conflict and the (logged) average number of sectarian crimes for all wards both including and excluding Belfast and Derry s wards from the sample.
Catholics (%) 0 20 40 60 80 100 0 20 40 60 80 100 Protestants (%) Parity 0.2.4.6.8 1 0.2.4.6.8 1 Polarization Figure A3. Scatter plot: Relationship between the percentage of Catholics and the percentage of Protestants within wards (top panel), and between the parity and polarization indexes (bottom panel)
Parity 0.2.4.6.8 1 Parity 0.2.4.6.8 1 0 20 40 60 80 100 Catholics (%) 0 20 40 60 80 100 Protestants (%) Figure A4. Scatter plots: Parity and percentages of Catholics and Protestants at ward level