Supplementary Information ’If a fight starts, watch the crowd’ The effect of violence on popular support for social movements. Jordi Muñoz∗1 and Eva Anduiza2 1 Department of Political Science, University of Barcelona 2 Department of Political Science and Public Law, Universitat Autònoma de Barcelona November 1, 2018 ∗ Corresponding author: jordi.munoz@ub.edu 1 Appendices A Balance section*.1 2 A.1 Entropy balancing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 B Full tables with interaction 2 C Robustness Checks 3 C.1 Reachability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 C.2 Time trends . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 C.3 Gràcia effect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 C.4 Alternative operationalization of treatment . . . . . . . . . . . . . . . 8 C.5 Alternative outcomes and Placebo Tests . . . . . . . . . . . . . . . . 9 C.6 Non-response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 D Newspapers 11 A Balance A.1 Entropy balancing As a complement to the multivariate regression approach, we also applied the data pre-processing method known as entropy balancing, as described by Hainmueller (2012) and Hainmueller & Xu (2013). Entropy balancing produces a set of weights that balance the treatment and control distributions on a vector of covariates. Cru- cially, it balances them in the first, second and third moments of the distribution. It reduces model dependence, as compared to the multivariate regression framework in which several assumptions of the functional forms of the relationships are made. Table A.1 shows the treatment and control moments of the distribution across a range of variables. The main, statistically significant unbalance found is related to the share of weak supporters in the treatment group, that is higher than in the control group. There are also marginally significant imbalances in gender composi- tion and latitude. Table A.2 compares treatment and control after the balancing, showing full balance. Table A.1. Pre-weighting Balance Pre-riots Post-riots mean variance skewness mean variance skewness Core Supporters 0.30 0.21 0.89 0.28 0.20 0.96 Weak Supporters 0.34 0.22 0.68 0.36 0.23 0.57 Opposers 0.29 0.21 0.93 0.30 0.21 0.85 Age 51.83 300.50 -0.03 50.89 352.90 0.02 Refusals 5.73 68.65 3.12 6.32 94.90 5.24 Gender 0.55 0.25 -0.21 0.53 0.25 -0.10 Gracia 0.10 0.09 2.70 0.07 0.07 3.31 lat 41.41 0.00 0.19 41.41 0.00 0.19 lon 2.17 0.00 0.07 2.17 0.00 0.04 B Full tables with interaction Here we present the full table of the interaction on which we base our estimation of the heterogeneous effects. It shows how, relative to the core supporters, the effect is significantly larger among the weak supporters and the non-aligned. Moreover, use two alternative moderators that are allegedly more stable than vote choice: party identification and left-right self-placement. While there are reasons to think that these are stable attitudes unlikely to be affected by such an event, we privilege in 2 Table A.2. Post-weighting balance Pre-riots Post-riots mean variance skewness mean variance skewness Core Supporters 0.30 0.21 0.89 0.30 0.21 0.89 Weak Supporters 0.34 0.22 0.68 0.34 0.22 0.68 Opposers 0.29 0.21 0.93 0.29 0.21 0.93 Age 51.83 300.50 -0.03 51.82 300.50 -0.02 Refusals 5.73 68.65 3.12 5.73 68.65 4.14 Gender 0.55 0.25 -0.21 0.55 0.25 -0.21 Gracia 0.10 0.09 2.70 0.10 0.09 2.70 lat 41.41 0.00 0.19 41.40 0.00 0.34 lon 2.17 0.00 0.07 2.17 0.00 0.04 the main text vote recall in order to avoid potential post-treatment bias. With party identification we use a similar classification scheme as with vote recall, but including CUP supporters to the core group (in the previous case we cannot use them because the CUP does not run in Spanish elections). Finally, with left-right, we distinguish between those in the leftmost positions of the scale (0-2), those in the center-left (3-4), center(5), center-right (6-7) and far-right (9-10). Figure B.1 presents the average marginal effects using partisanship as moderator, and figure B.2 shows the results with self-reported ideology as a measure of predispo- sition towards the movement. Results are generally consistent across specifications, with the effect being always small -and in these cases, non significant) for the core supporters, and stronger for the weak supporters. C Robustness Checks C.1 Reachability This is a serious threat, especially in the nominal samples in which a specific individ- ual is selected and contacted repeated times until it can be effectively interviewed. In this case, however, we sampled addresses and not individuals, and the fieldwork protocol did not include multiple attempts at interviewing specific individuals. Re- fusals were substituted by close neighbors using random routes departing from the selected address. In spite of this it might still be the case that some confounding effect of reachability biases the estimates. While this is, by definition, difficult to assess, we can proxy for it using the number of refusals collected before every completed survey. 3 Table B.1. Heterogeneous Treatment Effects (1) (2) (3) Past vote Partisanship Ideology Post -0.04 -0.02 -0.08* (0.03) (0.04) (0.05) Weak Supporters -0.27*** -0.24*** (0.04) (0.05) Opposers -0.54*** -0.50*** (0.05) (0.05) Non-aligned -0.34*** -0.34*** (0.13) (0.05) Post×Weak -0.12* -0.16** (0.07) (0.08) Post×Opposers -0.09 -0.15** (0.07) (0.07) Post×Non-aligned -0.30* -0.06 (0.16) (0.07) Center-left -0.07 (0.05) Center -0.38*** (0.06) Center-right -0.46*** (0.07) Right -0.47*** (0.10) Post×Center-left -0.06 (0.07) Post×Center 0.09 (0.08) Post×Center-right -0.05 (0.10) Post×Right -0.14 (0.13) Constant 0.94*** 0.92*** 0.84*** (0.02) (0.02) (0.03) R-squared 0.247 0.185 0.152 N 848 1173 1102 * p<.1. ** p<.05. *** p<.01 Controls include age, gender, ideology Reference category: Core As it can be seen in figure C.1, the average number of refusals is relatively constant around 5 with three exceptions in specific days: one before the riot outbreak, one after, and the final day of the fieldwork in which it took 32 attempts to complete the final two questionnaires. The other two peaks correspond to Saturdays, while the final one might be related to the quota sampling procedure. We will take this into 4 Figure B.1. Treatment effects, by partisanship Effects of Riot exposure on support for the 15M, by partisanship ,1 0 -,1 -,2 -,3 Core Weak Opposers Non-aligned Figure B.2. Treatment Effects, by Ideology Effects of Riot exposure on support for the 15M, by ideology ,2 0 -,2 -,4 -,6 Far Left Center-Left Center Center-Right Far Right account in the estimation. In any case, the difference between the average number of refusals per survey pre-riots (5.6) and post-riots (6.16) is not statistically significant (t = −1.25; p = 0.211) so as far as we can test, there isn’t a serious problem of imbalance on reachability. 5 Figure C.1. Interviews and refusals by day 100 80 60 40 20 Refusals per interview Completed interviews 0 29apr2016 13may2016 27may2016 10jun2016 C.2 Time trends A possible concern with our research design is the existence of unrelated time trends that covary with our treatment of interest and that might be driving our results. In table C.1 we present some robustness checks that allow us to rule out this possibility. Column 1 re-estimates the main model with a linear time trend (centered around the first day of riots) and its interaction with the treatment. Both the linear trend and the interaction have non-significant and negligible effects on our outcome, while the treatment maintains its effect. In columns 2 and 3 of this table we test the effect of the trend in the pre and post-riot samples, to show that there is no apparent time trend within each of these groups. Therefore, the net effect of time on our outcome of interest can be attributed to the treatment. In columns 4 to 6, we further deal with the issue of time trends using alternative treatments. The model in column 4 uses an arbitrary placebo date located at the empirical median of the control group (as suggested by Imbens & Lemieux (2008) for the case of regression discontinuity designs) and estimates its effect for the control group. This partition of the sample does not show any significant difference between the pre and post-placebo groups. In columns 5 and 6 we use a less arbitrary date: the 5th anniversary of the 15M movement, that occurred during the survey, one week before the riot outbreak. This 6 event got some (scant) attention in the press, as shown by our front page analysis in section D. As this could be a potential confounder, we test the effect of this alternative date. In column 5 we show how the anniversary does not have any effect, once we condition for the riot outbreak. In column 6 we investigate the effect of the anniversary only within the control group, and find no effect whatsoever. Table C.1. Time trends and placebo dates Linear trend Placebo dates (1) (2) (3) (4) (5) (6) Interaction Pre Post Median control 15M Anniversary Post-riot -0.14* -0.11** (0.08) (0.05) Days -0.00 -0.00 0.01 (0.00) (0.00) (0.01) Post×Days 0.01 (0.01) Placebo date -0.04 (0.05) 15M anniversary -0.03 -0.03 (0.05) (0.05) Constant 0.63*** 0.63*** 0.49*** 0.70*** 0.69*** 0.69*** (0.05) (0.05) (0.07) (0.03) (0.03) (0.03) R-squared 0.020 0.005 0.003 0.002 0.017 0.001 N 848 464 384 464 848 464 * p<.1, ** p<.05, *** p<.01 All models with entropy balancing weights as defined previously, and precinct-clustered SE C.3 Gràcia effect One obvious way of providing additional credibility to our estimated effect is to check whether it varies with intensity of exposure. If our treatment is capturing exposure to the riots, and therefore we are indeed estimating the effect of the use of violence on support for the 15-M movement, we should expect the treatment to have a stronger effect on those that were more exposed to it, namely the residents of the district of Gràcia. We have 94 Gràcia residents in our sample, 52 interviewed before the riots and 42 after the beginning of the violent outbreak in the neighborhood. In table C.2 we present the results of a model in which treatment is interacted with residence in Gràcia. We use the weights produced by the entropy balancing to keep balance in observables. The interaction term is statistically significant, and indeed the marginal effects for Gràcia predict a drop of support of 39 percentage points, far stronger than the 9 points for the rest of the city. 7 Table C.2. Gràcia Conditional Average Treatment Effect (1) Post-riot -0.10*** (0.04) Gràcia 0.11 (0.07) Post×Gràcia -0.30*** (0.10) Constant 0.67*** (0.03) N 848 * p<.1, ** p<.05, *** p<.01 Estimates after entropy balancing Table C.3. Alternative operationalizations of treatment (1) (2) Post-start -0.10*** (0.03) During 0.06 (0.06) After -0.13*** (0.03) Constant 0.68*** 0.68*** (0.02) (0.02) N 906 906 * p<.1. ** p<.05. *** p<.01 C.4 Alternative operationalization of treatment In the main text we have used a dichotomous operationalization of the treatment variable, that divides the sample between those that were interviewed before and after the riots. However, the conflict extended over four nights, so the treatment is not as sharp and the effects might not be immediate. To account for that, in table C.3 we use two alternative operationalizations: one based on a dummy variable that takes value one if the respondent was interviewed after the start of the riots and zero otherwise, and another one based on three categories: one for those interviewed before the start of the riots, one for those interviewed during the riots (N=98) and one for those interviewed after the riots (N=631). Results show that the effect is concentrated after the end of the rioting week, which might point to the fact that it took some time to materialize. 8 C.5 Alternative outcomes and Placebo Tests To lend additional credibility to our causal identification strategy, in this section we present a set of placebo tests. Results are shown in table C.4. Following the same model specification as in the main set of results, we show how the treatment (ex- posure to the riots) does not affect support for other social movements, not linked to the riots: the AAVV neighborhood associations (quite relevant in Barcelona), the anti-eviction movement PAH or the large-scale pro-independence movement Na- tional Assembly of Catalonia (ANC). We also show how the riots did not have any significant effect on external political efficacy, as measured by the question "To what extent would you say that the Spanish political system allows people like you to have an political impact?". Although we are limited by the questions included in the survey, this is suggestive evidence that the riots and the police response did not affect the citizens’ mood towards the political system in general. We interpret this, with caution, as an indication that our treatment is primarily capturing the effect of the riots. Table C.4. Alternative outcomes (1) (2) (3) (4) AAVV ANC PAH Ext. Efficacy Post-Riots -0.04 -0.04 0.01 -0.04 (0.04) (0.06) (0.05) (0.06) Constant 0.55*** 0.87*** 0.99*** 0.85*** (0.03) (0.04) (0.03) (0.04) N 732 721 801 924 * p<.1, ** p<.05, *** p<.01 Entropy balancing weights and precinct-clustered SE C.6 Non-response A potential threat to our identification strategy might come from non-response. Some respondents may refrain from answering the outcome variable precisely be- cause of the treatment. This creates a situation of attrition that might bias our estimates, because the missigness is related to the potential outcomes (Gerber & Green, 2012). Overall, we have a 17.1% of respondents who declared not to know the 15M move- ment, and therefore, they did not answer the question on sympathy. In the pre-riots group we have a 15.8% of Don’t knows and in the post-riots group the figure grows up to 18.5%. A two-sample test of proportions reveals that the difference is only 9 marginally significant in a one-tailed test (P r(Z > z) = 0.07). However, the effect disappears when we introduce controls, especially the neighborhood fixed-effects. This suggests that the differences in levels of don’t knows are related to compo- sitional sampling effects rather than a by-product of the riots. The very limited nature of the riot outbreak makes it highly implausible that the street agitation created an environment in which some respondents refrained from expressing their opinion towards the movement. Table C.5. Missing values: Don’t know the movement (1) (2) (3) (4) Post-riots -0.08 -1.03*** -0.81** -0.88 (0.15) (0.31) (0.40) (0.64) Political knowledge 0.94*** 1.20*** 0.94*** (0.19) (0.37) (0.19) Lef-right self placement -0.18** -0.18** -0.16 (0.08) (0.08) (0.11) Post×Knowledge -0.36 (0.42) Post×Ideology -0.04 (0.13) Age -0.02** -0.02** -0.02** (0.01) (0.01) (0.01) Gender 0.32 0.32 0.32 (0.30) (0.30) (0.30) Weak Supporters 0.11 0.12 0.11 (0.43) (0.43) (0.43) Opposers 0.09 0.11 0.10 (0.49) (0.49) (0.49) Non-aligned -0.65 -0.62 -0.63 (0.60) (0.60) (0.60) Constant 1.83*** 3.56*** 3.44*** 3.46*** (0.10) (1.01) (1.02) (1.07) Neighborhood FE YES YES Pseudo-R2 0.000 0.160 0.162 0.160 N 1402 887 887 887 * p<.1, ** p<.05, *** p<.01 If we analyze the patterns of missigness, we observe that the share of respondents that declare not to know the movement is higher among the less politically knowl- edgeable and also higher among the citizens located at the right. However, as shown in table C.5, these predictors of missigness do not interact with our treatment: right- ist, or less politically aware respondents are not any more likely than the rest of the sample, to declare that they do not know the movement after the start of the riots. Given this coavariates of missingness, we can expect that, if anything, the higher 10 rate of don’t knows in the post-riots group provokes a downward bias in our estimates given that those that declare not to know the movement are more similar to those that oppose it. We can use multiple imputation to check the robustness of our results. Using the relevant predictors of support for the 15M movement (ideology, partisanship, age, gender and neighborhood) we have imputed the missing values in our dependent variable. Results are presented in table C.6 and show how the effect is robust to the imputation of missing values. Table C.6. Multiple imputation models (1) (2) (3) mi_1 mi_2 mi_3 Post-Riots -0.10*** -0.10*** -0.11*** (0.03) (0.03) (0.03) Controls YES YES District FE YES Constant 0.65*** 0.70*** 0.58*** (0.02) (0.09) (0.12) R-squared . . . N 1244 887 887 * p<.1, ** p<.05, *** p<.01 D Newspapers We reviewed and coded the frontpages of the three main Barcelona-based newspapers (La Vanguardia, El Periódico and Ara) during the whole period of the fieldwork, in order to make sure that the event was indeed salient in the news, and that during the period there were no simultaneous events potentially affecting our outcome. In figure D.1 we present a summary of the contents day-by-day. The riots (clear blue) made an appearance during a few days in the frontpages, but other than that there was no other reference to the movement except for a small piece on its anniversary on the 15M in one newspaper (el Periódico). The reference was a small announcement of a special edition of the lifestyle section of the newspaper, that was devoted to the legacy of the 15M movement (in a generally positive tone). No other news during the period covered topics that would plausibily affect citizens’ attitudes towards the movement. The full sample of frontpages, and the detailed coding is available upon request. 11 Figure D.1. Frontpage topics References Gerber, Alan S & Donald P Green (2012) Field experiments: Design, analysis, and interpretation. WW Norton. Hainmueller, Jens (2012) Entropy balancing for causal effects: A multivariate reweighting method to produce balanced samples in observational studies. Polit- ical Analysis 20(1): 25–46. Hainmueller, Jens & Yiqing Xu (2013) Ebalance: A stata package for entropy bal- ancing. Journal of Statistical Software 54(7). Imbens, Guido W & Thomas Lemieux (2008) Regression discontinuity designs: A guide to practice. Journal of Econometrics 142(2): 615–635. 12