Online Supplement to Female Participation and Civil War Relapse

Online Supplement to Female Participation and Civil War Relapse [Author Information Omitted for Review Purposes] June 6, 2014 1

Table 1: Two-way Correlations Among Right-Side Variables (Pearson s ρ) Lit. Ratio % WIP %WILF Settlement UN PKO War Duration ln(gdppc) IMR Democ Democ 2 Eth Frac Eth Frac 2 Lit. Ratio 1.000 % WIP 0.198 1.000 % WILF 0.243 0.349 1.000 Settlement 0.063 0.161 0.322 1.000 UN PKO 0.071-0.101 0.084 0.441 1.000 War Duration -0.060 0.121 0.182 0.376 0.110 1.000 ln(gdppc) 0.620 0.059-0.125-0.027 0.143 0.004 1.000 IMR -0.705-0.027 0.206 0.154-0.087 0.063-0.775 1.000 Democracy 0.379 0.075 0.171 0.249 0.085 0.131 0.442-0.231 1.000 Democracy 2 0.254 0.102-0.168-0.166 0.047-0.155 0.480-0.410 0.336 1.000 Ethnic Frac -0.267-0.025 0.057 0.168-0.048 0.177-0.318 0.427 0.002-0.153 1.000 Ethnic Frac 2-0.283-0.013 0.123 0.125-0.085 0.161-0.363 0.471 0.006-0.166 0.973 1.000 2

Robustness Check: Measuring Development via Infant Mortality Rate (IMR) Table 2: Stratified Cox Regression of Peace Duration after Civil War Variable β hazard ratio r.s.e. β hazard ratio r.s.e. Fem:Male Literacy Ratio -2.703** 0.067 1.313-2.461 0.085 1.794 % Women in Parliament -0.115** 0.892 0.046-0.115** 0.891 0.046 % Women in Labor Force 0.070*** 1.073 0.024 0.067** 1.070 0.030 Settlement 1.302*** 3.676 0.500 1.285** 3.614 0.504 War Duration -0.072 0.930 0.052-0.070 0.932 0.056 UN Peacekeepers -1.180 0.307 0.846-1.154 0.315 0.858 IMR 0.002 1.002 0.015 Democracy -0.027 0.973 0.043-0.025 0.975 0.041 Democracy 2 0.017 1.018 0.015 0.018 1.018 0.016 Ethnic Fractionalization -0.111 0.0001 5.659-0.097 0.0001 5.649 Ethnic Fractionalization 2 9.769* 17477.39 5.841 0.702* 1637.76 5.748 N 579 579 Wald χ 2 28.98*** 29.74*** Log pseudo-likelihood -33.542-33.525 *p.10, **p.05; ***p.01 (two-tailed tests). Robust standard errors are clustered on same-country observations, and are based on the nonexponentiated coefficients. 3

Alternative Measures of Social Participation Existing work uses three alternative measures of female social participation. First, Caprioli (2000, 2005) uses fertility rates, operationalized as births per woman and drawn from the World Bank s World Development Indicators (WDI). Second, Melander (2005) uses higher education attainment, operationalized as the percentage of the female population age 25 and over that has attained higher schooling divided by the percentage of the male population age 25 and over that has attained this level of schooling (Barro and Lee, 2001, 2013). Third, Gizelis (2009) uses a ratio of female-to-male life expectancy, again drawn from the WDI. As shown in Table 3, our results are robust to replacing our female:male literacy ratio with the first and second alternative measures. When we use the third (female:male life expectancy), the results retain the expected sign but fall short of statistical significance. We believe that this makes sense: as we note in the article, our sample consists only of states emerging from civil war, which by definition involve high levels of battle-deaths. Men are disproportionally represented in combat, and are also quite likely to be disproportionally represented in death tolls. Thus in our sample, male life expectancy should correlate negatively with other characteristics of war, including (for example) its severity and brutality. These factors are positively related to the likelihood of civil war relapse, and measures that covary with them including female:male life expectancy are unlikely to support our theoretical expectation. 4

Table 3: Stratified Cox Regression of Peace Duration after Civil War Variable β β β β β β (r.s.e.) (r.s.e) (r.s.e.) (r.s.e) (r.s.e.) (r.s.e) Fertility Rate 0.330** 0.028** (0.152) (0.016) Fem:Male Higher Educ Ratio -0.020* -.002 (0.016) (0.021) Fem:Male Life Expectancy Ratio -6.188-9.364 (7.678) (16.931) % Women in Parliament -0.058** -0.118** -0.084** -0.145*** -0.030-0.102*** (0.037) (0.058) (0.038) (0.059) (0.029) (0.039) Fem:Male Labor Force Ratio 0.054** 0.070* 0.064** 0.087** 0.018 0.013 (0.030) (0.048) (0.031) (0.047) (0.020) (0.045) Settlement 1.516** 1.242** 1.249*** (0.672) (0.719) (0.475) War Duration -0.092* -0.075* -0.056 (0.056) (0.053) (0.049) UN Peacekeepers -1.519** -0.908-1.169* (0.798) (1.108) (0.868) lngdp(pc) 0.161 0.050-0.368 (0.522) (0.762) (0.443) Democracy -0.022-0.084-0.018 (0.056) (0.078) (0.043) Democracy 2 0.017 0.013 0.016 (0.014) (0.018) (0.015) Ethnic Fractionalization -4.646-7.663* -8.542* (5.500) (5.664) (5.415) Ethnic Fractionalization 2 4.337 10.520* 9.814* (6.196) (6.511) (5.208) N 596 553 489 446 754 574 Wald χ 2 8.82** 22.62** 5.71* 54.10*** 2.41 52.68*** Log pseudo-likelihood -39.145-27.104-30.266-19.926-66.370-19.820 *p.10, **p.05; ***p.01 (one-tailed tests). Robust standard errors are clustered on same-country observations, and are based on the nonexponentiated coefficients. 5

Robustness Check: Measuring Economic Participation via a Ratio of Female:Male Participation in the Labor Force Table 4: Stratified Cox Regression of Peace Duration after Civil War Variable β hazard ratio r.s.e. β hazard ratio r.s.e. Fem:Male Literacy Ratio -2.259** 0.104 1.118-3.292** 0.065 1.744 % Women in Parliament -0.073** 0.929 0.033-0.125** 0.047 0.054 Fem:Male Labor Force Ratio 2.928*** 18.689 0.774 3.549*** 39.861 1.146 Settlement 1.438** 2.556 0.606 War Duration -0.077 0.048 0.052 UN Peacekeepers -1.233 0.262 0.899 lngdp(pc) 0.218 0.632 0.508 Democracy -0.035 0.044 0.046 Democracy 2 0.017 0.016 0.015 Ethnic Fractionalization -10.480* 0.0001 5.557 Ethnic Fractionalization 2 11.213* 4330.40 5.848 N 656 579 Wald χ 2 16.58*** 30.81*** Log pseudo-likelihood -46.399-32.904 *p.10, **p.05; ***p.01 (two-tailed tests). Robust standard errors are clustered on same-country observations, and are based on the nonexponentiated coefficients. 6

Schoenfeld Residuals Note: The tests presented below are conducted on the fully-specified model reported in Table 3 of the paper. Table 5: Test of Proportional Hazards Assumption Variable ρ χ 2 p > χ 2 Fem:Male Literacy Ratio -0.1026 0.34 0.5612 % Women in Parliament 0.0034 0.00 0.9866 % Women in Labor Force -0.1545 0.82 0.3650 Settlement 0.0151 0.01 0.9351 War Duration 0.0159 0.01 0.9235 UN Peacekeepers 0.0206 0.02 0.8746 ln(gdppc) -0.1357 1.04 0.3076 Democracy 0.0609 0.10 0.7553 Democracy 2 0.0391 0.12 0.7281 Ethnic Fractionalization -0.0700 0.16 0.6862 Ethnic Fractionalization 2 0.0637 0.14 0.7079 Global test 5.14 0.9244 7

The Possibility and Consequences of Selection Effects Civil war begins (onset) and lasts for some period of time (war duration). Then it ends (termination) and peace endures for some period of time (peace duration). If peace fails, it represents a new onset, and the process repeats from there. The basic selection problem arises because our sample consists only of country-years in which war has already ended, and these observations may differ in important unmeasured ways from others in which war has not ended. Thus the selection process deals with conflict termination rather than conflict onset. Is this process a problem? There are two potential selection effects. First, it is possible that civil wars with higher levels of female participation will be more likely to end than other civil wars, and so we will have a sample of observations with above-average female participation. The subset of wars that will end because of high levels of this participation will be representative of the subset of wars that might end with those same high levels. Thus, as Sartori (2003, 114) notes, this non-random aspect of the sample is what is commonly misunderstood to the problem of selection bias, but it is not a problem for empirical analysis. The second selection effect, which is the potential troublemaker, is that some wars will end even with low levels of female participation. This is because other unmeasured factors are at work here (regime type, economic development, ethnic fractionalization, etc.), which are captured in the duration model s error term. In other words, these observations enter our sample not because they have high female participation, but because they have large error terms. The problem is that whether or not female participation is correlated with the unmeasured influences in the overall population, it will be correlated in the selected sample. If these other influences do lead to longer peace duration, then we will underestimate the effect of female participation on peace duration because in the selected sample states with low female participation are different in some other meaningful ways (e.g., Heckman, 1979; Maddalla, 1983; Sartori, 2003). Accounting for this selection effect would require a two-stage model where the first stage estimated effects on war s end, and the second stage estimated effects on peace duration given that war has ended. 1 Unfortunately, the only such model that exists is the dursel model (Boehmke, Morey and Shannon, 2006), which cannot incorporate time-varying covariates in the outcome (i.e., duration) equation. Because our key independent variables change over time, dursel is not a useful option here. If or when an appropriate estimator becomes available, we will investigate the possibility of selection bias in this study. Importantly, and as explained above, the consequence of selection bias would be to underestimate the effect of female participation. If such bias does exist, we expect the estimated effects to strengthen rather than weaken or disappear. 1 When peace fails, we call it civil war relapse or recurrent civil war. Our duration models estimate the hazard or risk of relapse, which is analogous to the inverted risk of peace duration. Available estimators including the censored probit suggested by one anonymous reviewer estimate probability rather than risk in the outcome equation, and are thus inappropriate in our empirical context. 8

Alternative Dependent Variable: UCDP Civil War Readers may ask why we use the Sambanis (2004) measure of civil war rather than one drawn from the PRIO/UCDP Armed Conflict Dataset (Gleditsch et al., 2002; Themnér and Wallensteen, 2013). Certainly the PRIO/UCDP data provide an excellent measure of intrastate conflict, and we seriously and carefully considered using it before proceeding with the Sambanis data instead. Our decision was based on both substantive and practical considerations. Substantively, the question of what threshold of violence distinguishes civil war from other forms of internal armed conflict is not new. PRIO/UCDP provides a compelling argument for why 25 battle-deaths are preferable to the COW threshold of 1,000. The lower intensity threshold, for example, reduces the likelihood that we omit armed conflicts in small states that produce few deaths but are critically important for the histories of these countries and for our understanding of political violence. Nonetheless, it remains true that any numeric threshold, however high or low, is necessarily ad hoc. Any such threshold is vulnerable to measurement error (e.g., problems resulting from unreliable reporting and incomplete data). And any such threshold is based purely on battle-deaths, which leads to at least two additional weaknesses in resultant data. First, a sole focus on military casualties ignores civilian deaths due to civil war. Civilians are targeted in war, and are disproportionately affected by humanitarian crises created by combatants to hold civilian populations hostage and gain control of territory (Sambanis, 2004, 823; see also Harff, 2003; Valentino, 2004). These facts contribute to the political landscape, and we believe that they are an important consideration in the identification of civil war. Second, sole reliance on numeric thresholds overlooks the fact that termination can also be determined by discrete, easily coded events, such as peace treaties (Sambanis, 2004, 816). These weaknesses are compounded in a context like this paper, which is focused on conflict termination and recurrence rather than conflict onset. For all these reasons, we strongly prefer a measure that relies on coding rules that go beyond a strict numeric threshold, and for which extensive notes that justify the coding of each case are publicly available. Sambanis measure satisfies these preferences, relying on eleven distinct coding rules (see pages 829-831 of his paper) and providing detailed justification for each decision online (at http://pantheon.yale.edu/ ns237/civilwar.html). From a practical standpoint, the UCDP data have longer temporal coverage than the Sambanis data we utilize here. Importantly, our temporal coverage is not limited by our dependent variable. Instead, it is defined by our key independent variables: the UNESCO literacy data begin in 1970 but are largely missing until 1980 (when our temporal coverage begins), and the data on women in parliament end in 2003 (when our temporal coverage ends). As discussed in the manuscript, these are the best (most valid and reliable) data available to capture our theoretical concepts and test our empirical hypotheses. 9

Table 6: Stratified Cox Regression of Peace Duration after Civil War Variable β hazard ratio r.s.e. β hazard ratio r.s.e. Fem:Male Literacy Ratio -0.451 0.637 1.237-2.348 0.096 5.207 % Women in Parliament -0.111** 0.895 0.050-0.022* 0.978 0.015 % Women in Labor Force 0.002 1.002 0.020 0.008** 1.008 0.005 Settlement 0.245* 1.277 0.192 War Duration -0.051 0.950 0.061 UN Peacekeepers -0.435* 0.647 0.304 ln(gdppc) 1.209 3.351 1.103 Democracy -0.013 0.987 0.064 Democracy 2-0.001** 0.999 0.001 Ethnic Fractionalization -4.928 0.008 7.291 Ethnic Fractionalization 2 5.057 157.180 7.505 N 656 579 Wald χ 2 5.26* 17.82** Log pseudo-likelihood -29.707-17.928 *p.10, **p.05; ***p.01 (one-tailed tests). Robust standard errors are clustered on same-country observations, and are based on the nonexponentiated coefficients. In the fully-specified model, the following variables violate the proportional hazards assumption and are corrected (i.e., interacted with time): % women in parliament, % women in labor force, settlement, UN peacekeepers, and democracy 2. Of course, we recognize that some scholars may still prefer the UCDP data. To that end, Table 6 replicates our main models using an alternative, UCDP-based dependent variable. In particular, we use the UCDP/PRIO Armed Conflict Dataset to create a variable that equals 0 in country-years of post-civil war peace, and equals 1 (i.e., failure) in country-years when civil war recurs. Our theory suggests that women participate at higher levels when society is no longer plagued by armed conflict; for that reason, we begin recording peace spells once all armed conflicts (including intrastate and internationalized intrastate wars) are ended. In these new models, our measures of female political and economic participation maintain sign and remain statistically significant. Our measure of female social participation maintains sign, but falls short of statistical significance at conventional levels. This change may be a function of the relatively small sample size; in future iterations with larger temporal coverage, we expect this measure to regain statistical significance. 10

References Barro, Robert J. and Jong Wha Lee. 2001. Barro-Lee Education Data. http://www.barrolee.com, accessed 1 June 2014. Barro, Robert J. and Jong Wha Lee. 2013. A New Data Set of Educational Attainment in the World, 1950-2010. Journal of Development Economics 104(184-198). Boehmke, Frederick J., Daniel S. Morey and Megan Shannon. 2006. Selection Bias and Continuous-Time Duration Models: Consequences and a Proposed Solution. American Journal of Political Sciencen 50(1):192 207. Caprioli, Mary. 2000. Gendered Conflict. Journal of Peace Research 37:53 70. Caprioli, Mary. 2005. Primed for Violence: The Role of Gender Inequality in Predicting Internal Conflict. International Studies Quarterly 49:161 178. Gizelis, Theodora-Ismene. 2009. Gender Empowerment and United Nations Peacebuilding. Journal of Peace Research 46(4):505 523. Gleditsch, Nils Petter, Peter Wallensteen, Mikael Eriksson, Margareta Sollenberg and Havard Strand. 2002. Armed Conflict 1946-2001: A New Dataset. Journal of Peace Research 39(5):615 637. Harff, Barbara. 2003. No Lessons Learned from the Holocaust? Assessing Risks of Genocide and Political Mass Murder Since 1955. American Political Science Review 97(1):57 53. Heckman, J.J. 1979. 47(1):153 162. Sample Selection Bias as a Specification Error. Econometrica Maddalla, G.S. 1983. Limited Dependent and Qualitative Variables in Econometrics. Cambridge University Press. Melander, Erik. 2005. Gender Equality and Intrastate Armed Conflict. International Studies Quarterly 49:695 714. Sambanis, Nicholas. 2004. What is Civil War? Conceptual and Empirical Complexities of an Operational Definition. Journal of Conflict Resolution 48(6):814 858. Sartori, Anne E. 2003. An Estimator for Some Binary-Outcome Selection Models Without Exclusion Restrictions. Political Analysis 11(2):111 138. Themnér, Lotta and Peter Wallensteen. 2013. Armed Conflict, 1946 2012. Journal of Peace Research 50(4). Valentino, Benjamin. 2004. Final Solutions: Mass Killing and Genocide in the Twentieth Century. Ithaca, New York: Cornell University Press. 11