Political Instability and Large-Scale Violence in the Economic North and South: Comparing Causes

Transcription

1 Political Instability and Large-Scale in the Economic North and South: Comparing Causes Benjamin E. Goldsmith University of Sydney Dimitri Semenovich Arcot Sowmya University of New South Wales Abstract This paper builds on existing scholarship in the areas of political instability, state failure, and mass violence (e.g., Harf 2003; Valentino, Huth & Balch-Lindsay 2004). It introduces two innovations which add to understanding of the processes leading to serious political instability in states, and those leading from instability to large-scale violence. First, it distinguishes between North and South in economic terms by examining the processes separately based on higher and lower per capita gross domestic product. Second, it employs 2-stage selection models to correct for potential selection bias (Heckman 1979; Sartori 2003). Thus, the quantitative analysis will answer the questions of whether the processes leading to the most serious instances of internal violence are different for richer and poorer states, and whether current understanding of this process in the quantitative literature is biased by failure to consider selection effects explicitly. Paper prepared for presentation at the International Political Science Association European Council for Political Research joint conference, February 2011, Sao Paulo, Brazil.

2 In this paper we present the results of initial analysis exploring two aspects of large-scale violence (LSV) within states: the degree to which there is a two-stage process which might obscure analysis of the causes of highly violent outcomes, and the degree to which there are different processes in richer and poorer countries. Research questions Some studies have approached the analysis of mass killing from the perspective that mass killing represents the second stage of a process which begins with other types of violence and/or political turmoil. Harf (2003) identifies internal war and regime collapse as antecedents of genocide or politicide (politically motivated mass killing). Valentino, Huth, and Balch-Lindsay (2004) identify guerilla warfare as providing the conditions under which state leaders opt for mass killing. While these important studies point to (at least) a two-stage process, neither explicitly models that process. One purpose of this paper is to assess whether understanding the second stage leading to (or avoiding) the outcome of mass killing, requires some knowledge of the first stage, which we model here as one of generic political instability. Studies of political instability have pointed to the level of development of countries as an important causal factor, with higher infant mortality in particular emerging as an important variable (Goldstone et al 2010). Given its importance in what we consider the initial stage of our models for understanding large-scale violence, we also ask whether there are different patterns or causal processes involved in richer and poorer countries that help explain this outcome, or whether the process is essentially the same. We suspect there might be different processes involved because levels of wealth affect both the capacity of the state, which is often the perpetrator of mass killing, and the motivations and concerns of individuals. In poorer countries, the state might have less capacity to carry out mass killing, but also might be more unstable and insecure due to this lack of resources. In poorer countries, civilians and members of the military might be more concerned with basic subsistence and survival, and perhaps more easily mobilized for support or actual participation in violence against out groups or others due to a sense of imminent threat to survival. In richer countries, while the repressive tools of the state might be more effective, the population might be less easily mobilized for support or actual participation in violence against out groups or others. It is suggestive that in the former Yugoslavia, with a relatively high level of wealth, mass killing seemed to be perpetrated by capable states against ethnic others, while in Rwanda, while the killing was organized by elites, much of the violence was actually committed by mobilized civilians against ethnic or political others. Regarding the two-stage process, we hypothesize that some actors involved in political instability also contemplate large-scale violence from the outset, meaning that selection bias is a danger for analysis of large-scale violence outcomes. Regarding levels of wealth, we hypothesize that there will be a difference between the causal processes for richer and poorer states: Richer states have greater capacity, so state institutions such as the military are more involved in large-scale violence (x-yugoslavia); Poorer states have lower capacity, so elites plotting or facilitating large-scale violence are more likely to rely on mobilization of civilians (Rwanda). We also expect that sufficiently wealthy states, such as developed countries, will have both more capacity to resist instability and violence, and populations with less motivation for such action. There are many limitations to our analysis, including our coding of large-scale violence as instances of civil war or genocide / politicide which incur at least 10,000 deaths (1,000 for some analyses). A more refined measure could be useful in future research. But we see this as a useful and theoretically interesting starting point, because different types of LSV might nevertheless have 2

3 many common dynamics which remain unexplored. There are also other data issues, such as very limited data for infant mortality. Research design In this section we discuss data and methods of analysis. Heckman (1979) selection models are used to model the two-stage process we hypothesize, accounting for selection effects. One set of analyses (Tables 2 and 3) focuses on onsets of political instability and large-scale violence, dropping years of ongoing instability/lsv. The idea behind this specification is that causal factors are prior to the start of the outcomes, and including years of ongoing instability or LSV risks conflating the causes of onsets of these events with causes of their duration. Another set of analyses (Tables 4 and 5) models all years of instability and LSV, but includes lagged dependent variables to account for the likelihood that instability or LSV last for multiple years, with occurrence last year increasing the chance of continuation this year. The data are pooled annual time-series for the period , with country-years the unit of observation. In the remainder of this section we describe the measures used, and data sources. Political instability For the first-stage dependent variable, selection into a situation of political instability, we use Political Instability Task Force (PITF) data for all instance s of instability, which include civil wars (revolutionary and ethnic), adverse (non-democratic) regime changes, and instances of genocide ro politicide (Esty et al. 1998). It is important to note that most instances of instability in these data include more than one of these types. We think this supports are initial steps at modeling these events together. Large-scale violence (LSV) we define as civil wars or politicides / genocides which involve at least 10,000 (1,000 for some models) deaths. We do not include adverse regime changes because we assume such events are distinct from other sorts of instability, rarely involve mass killing except in combination with civil war or politicide / genocide, and are likely to involve spontaneous bottom-up or top-down actions in which killing may result less from strategic calculations, and more from immediate circumstances, low-level decisions, and accident. As noted, we model both onsets of instability / LSV, and all years of their occurrence with lagged dependent variables. We now briefly describe the variables. Political Instability Political Instability Onset t+1 is coded 1 for the first year of each episode of political instability as coded by PITF. The subscript time notation t+1 indicates that the variable is forward lagged by 1 year to aid causal inference. Years of ongoing instability for the same incident are dropped, and years with no instability are coded 0. Political Instability is coded 1 for each country-year of instability as coded by PITF. Large-scale violence The variable Large-Scale Outcome with 1,000 or more deaths t+1 is coded 1 for each country-year in which a distinct instance of political instability begins, provided that instance 3

4 eventually leads to at least 1,000 deaths as recorded in the PITF dataset. The variable is forward lagged by 1 year. Ongoing country-years for the same incident are dropped, and country-years with no such LSV incident are coded 0. Large-Scale Outcome with 10,000 or more deaths t+1 is similar to the variable noted above, but uses the threshold of 10,000 deaths. The variable Large-Scale with 10,000 or more deaths t+1,2 is similar to the variable above, but measures whether LSV with 10,000 deaths occurs in year t+1 or t+2. Thus, it does not record onsets only, but all instances in which violence reaches that level either in the next year or after 2 years. Large-Scale with 10,000 or more deathst+1,2,3 is similar to the variable above, but includes LSV occurrences at the 10,000-death level in year t+3 as well. In effect, these can be used to forecast LSV over the next 2 or 3 years, respectively, from the time that the independent variables were measured, and 1 and 2 years, respectively, from the time Political Instability t+1 is measured. Causal variables The causal or independent variables in the models will not be discussed here in detail. They are for the most part standard measures used in much political science research. We descrive each briefly and note the data sources. Regime Type is the polity scale ranging from -10 for fully autocratic regimes to +10 for fully democratic ones. Full Democracy is coded 1 for a polity score of +7 or higher. Full Autocracy is coded 1 for a polity score of -7 or lower. The data source is the Polity IV dataset (Marshall and Jaggers 2007). The variable Elections is coded 1 for each country-year experiencing both legislative and executive elections (e.g., a general election in a parliamentary system or a presidential election year in a U.S.-style presidential system), 0 otherwise. Data are from the Institutions and Elections Project (Regan and Clark n.d.) User s Manual for the IAEP Dataset GDP per capita data are from the World Bank s World Development Indicators (WDI), measured in inflation-adjusted real dollars with 2000 as the benchmark year. Regime durability is a variable which codes the number of years since there was a regime shift of at least 3 polity points on the 21-point scale referenced above. The squared and cubed values are also included to account for temporal dependence in the models (Carter and Signorino 2010). Variables at outcome level only include New Regime, Military Personnel, and Total Population. New Regime is 1/Durability. Military Personnel is taken from the correlates of war (COW) dataset (Ghosn and Faten 2007). Total population is from World Bank WDI. Results Although our finding regarding the differences between richer and poorer countries are limited, we point out that above $17,000 per capita, there are no instances of large-scale violence for us to study. 4

5 Clearly, wealthier states are less susceptible to instability and mass killing. This is evident in Table 1, and Table A1 in the appendix, which lists all states with instances of instability. Table 1. Mean Values for Instability and Large-scale Measures, GDPpc range Number Infant of states Mortality GDPpc Instability Instability Onset Years 1,000 Onset 1,000 Years 10,000 Onset 10,000 Years Revolutonary War Ethnic War Adverse Regime Gen- / Change Politicide <$ $ $ $1,000-1, $2,000-3, $4,000-17, We first present results for three models of the onset of political instability and large-scale violence incurring at least 1,000 deaths. Variables which do not approach at least the.10 significance level at some point in the analyses in Tables 2-5 have usually been dropped. Table 2 contains some intriguing results for the first stage of the models, explaining the onset of political instability. For countries with less than $4000 GDPpc, higher levels of democracy increase the likelihood of the onset of instability, but the effect is non-linear because we also include a measure of full democracy, which is significant and negative. Thus we find that partial democracy is more conducive to instability, but it is important to note that a measure for democracy with factions is not significant, which challenges recent findings (Goldstone et al 2010). We also find some weak evidence that elections might reduce the chance of political instability. We feel this is worth further investigation with more precisely specified models. Elections might provide both an outlet for elites who would seek power, and, perhaps at the LSV outcome stage, a rallying event for masses and elites who would challenge entrenched regimes (e.g., Iran), perhaps refusing to accept election results perceived as rigged, or with regimes refusing to accept elections results which they lose (e.g., Algeria, Myanmar) For richer countries, both full authoritarian and full democratic regimes are less likely to experience the onset of instability, while the polity scale (increasing democracy) is not significant. This is consistent with our expectation, although falls short of direct evidence, that richer states will have greater capacity to control unrest, and that citizens in those states will be less motivated to challenge the regime. We also find a significant negative effect for GDP per capita, but only among poorer states (or when all states are analyzed together). This further supports our contentions. At the stage of LSV outcomes, we note that only regime type variables generally proved significant, which in itself is of considerable interest. Democracy as measured by the polity scale is robustly associated with a lower likelihood of the onset of LSV. New regimes in poorer states show some weaker evidence (p =.096) of being more susceptible to LSV. 5

6 Table 2. Selection Models: Onsets of Political Instability and Large-Scale, Poorer States Richer States All States coef s.e. sig. coef s.e. sig. coef s.e. sig. Large-Scale Outcome with 1,000 or more deaths t+1 Regime Type New Regime constant Political Instability Onset t+1 Political Instability Regime Type Full Democracy Full Autoc racy Elections t Elections t GDPpc Regime durability Regime durability Regime durability constant rho Number of obs Uncensored obs Wald chi2(2, 1, 2 df) Note: Heckman probit models. Bold font indicates.05 significance or better. Poorer states have GDP per capita below US$4,000 (inflation-adjusted to year 2000), richer states are all those at or above this level. Country-years of onging instability dropped. Our subsequent analyses using onsets of LSV at the 10,000-death threshold (Table 3), and all years of instability and LSV with 1-2 year forecasts (Table 4) and 1-3 year forecasts (Table 5) provide very similar findings. We find this robustness of results encouraging at this initial stage. We therefore only highlight results which are different from those in Table 2 in the remainder of this section. For onsets of LSV at 10,000 or more deaths, we do find that the size of the military, controlling for total population is significant for poor countries, as shown in Table 3. Why this would not also be the case for richer states is interesting, and we feel deserves further investigation. Our expectations might lead to the supposition that richer states have more capable militaries, and so larger number of soldiers are not necessary to commit large-scale violence. 6

7 Table 3. Selection Models: Onsets of Political Instability and Large-Scale, Poorer States Richer States All States coef s.e. sig. coef s.e. sig. coef s.e. sig. Large-Scale Outcome with 10,000 or more deaths t+1 Regime Type New Regime Military Personnel Total Population constant Political Instability Onset t+1 Political Instability Regime Type Full Democracy Full Autoc racy Elections t Elections t GDPpc Regime durability Regime durability Regime durability constant rho Number of obs Uncensored obs Wald chi2(4, 1, 4 df) Note: Heckman probit models. Bold font indicates.05 significance or better. Poorer states have GDP per capita below US$4,000 (inflation-adjusted to year 2000), richer states are all those at or above this level. Country-years of onging instability dropped. Although Tables 4 and 5 present models with a considerably different coding of the outcomes, and using powerful lagged dependent variables which often overshadow other effects, the results are highly consistent with those in Table 1 and with those in Table 2 except the effect of military personnel. We find this robustness of the models encouraging. 7

8 Table 4. Selection Models: Country-Years of Political Instability and Large-Scale at t+1 or t+2, Poorer States Richer States All States coef s.e. sig. coef s.e. sig. coef s.e. sig. Large-Scale with 10,000 or more deaths t+1,2 Large-Scale10, Regime Type New Regime Military Personnel Total Population constant Political Instability t+1 Political Instability Regime Type Full Democracy Full Autoc racy Elections t Elections t GDPpc Regime durability Regime durability Regime durability constant rho Number of obs Uncensored obs Wald chi2(5, 2, 5 df) Note: Heckman probit models. Bold font indicates.05 significance or better. Poorer states have GDP per capita below US$4,000 (inflation-adjusted to year 2000), richer states are all those at or above this level. Examining Tables 2-5 together, we note that the rho statistics for poorer states and for all states modelled together are highly significant in Tables 2-3, and significant at the.10 level in Table 5. While we cannot claim that selection bias is always clearly present, we can argue that the weight of the evidence is strongly in favor of selection effects in the process of instability and LSV. 8

9 Table 5. Selection Models: Country-Years of Political Instability and Large-Scale at t+1, t+2, or t+3, Poorer States Richer States All States coef s.e. sig. coef s.e. sig. coef s.e. sig. Large-Scale with 10,000 or more deaths t+1,2,3 Large-Scale10, Regime Type New Regime Military Personnel Total Population constant Political Instability t+1 Political Instability Regime Type Full Democracy Full Autoc racy Elections t Elections t GDPpc Regime durability Regime durability Regime durability constant rho Number of obs Uncensored obs Wald chi2(4, 2, 4 df) Note: Heckman probit models. Bold font indicates.05 significance or better. Poorer states have GDP per capita below US$4,000 (inflation-adjusted to year 2000), richer states are all those at or above this level. - statistics not produced in Stata 11. Conclusions Both our negative and positive findings are suggestive for future research. We see these models as initial steps in a large project, and anticipate further theoretical elaboration as well as more finely specified models, in particular regarding wealth, state capacity, and our two stages of interest. But, the robust results do lead to some suggestive findings, which we expect may hold in future analyses. Our negative findings of interest include the lack of robustly significant effects for Democracy with factions, limited effects for Elections, no effect of GDPpc at the stage of LSV (once its role in general political instability is considered). We also found that newer regimes are not generally more 9

10 susceptible to LSV, once regime durability (longevity) is considered at stage of general political instability. Our positive findings include, the general observation that selection bias is definitely a danger in models of LSV. Analysts would be wise to consider the antecedents creating the conditions for LSV, rather than ignoring them. Regarding differences between rich and poor, we are more equivocal: while richer countries are less susceptible to instability or LSV, it is not yet clear that there are different processes involved. But, our results regarding specific variables are suggestive. Democratic countries above $4,000 GDPpc (2000 base year) seem strongly resistant to instability, while more democratic regimes below that level seem more susceptible to instability than lessdemocratic regimes Poorer autocratic regimes are not resistant to instability, while richer ones are. Larger militaries in poorer countries may increase the chance of LSV, but no evidence for this in richer countries. On the other hand, regardless of wealth, fully democratic states are less susceptible to instability for richer and poorer states. Democracy is robustly associated with a lower likelihood of LSV for richer and poorer states Instead of an assertive conclusion in favor of one hypothesis or another, we consider it more appropriate to address possible next steps in the analysis. These include: closer attention to wealth levels and their substantive implications for state capacity; more thinking about 2-stage process and strategic actors; modeling of interaction effects with wealth and/or regime type; more complete data (which are being collected); out-of sample tests for selection models versus other approaches; and qualitative case research. 10

11 Appendix Table A1. Frequency of Instances of Political Instability by Country, sorted by Level of Wealth, ,000 Onset 1,000 Years 10,000 Onset 11 10,000 Years Adverse Infant Instability Instability Revolutonary Ethnic Regime Gen- / ccode abbrev country Mortality GDPpc Onset Years War War Change Politicide AFG Afghanistan LAO Laos SOM Somalia BUI Burundi ETH Ethiopia GNB Guinea-Bissau EQG Equatorial Guinea BFO Burkina Faso NIR Niger TAJ Tajikistan MZM Mozambique CHA Chad MLI Mali LBR Liberia SIE Sierra Leone UGA Uganda DRC Congo Kinshasa NEP Nepal RWA Rwanda BNG Bangladesh CAM Cambodia GHA Ghana SUD Sudan MYA Myanmar (Burma) JOR Jordan BEN Benin IND India NIG Nigeria GAM Gambia LES Lesotho KEN Kenya CHN China GUI Guinea COM Comoros PAK Pakistan BOS Bosnia YEM Yemen INS Indonesia HAI Haiti SEN Senegal ARM Armenia MLD Moldova ZAM Zambia THI Thailand ZIM Zimbabwe SRI Sri Lanka ANG Angola AZE Azerbaijan SWA Swaziland PNG Papua New Guinea SYR Syria GRG Georgia GUY Guyana IRQ Iraq PHI Philippines CON Congo Brazzaville BLR Belarus ECU Ecuador ALB Albania DOM Dominican Rep NIC Nicaragua SOL Solomon Islands MOR Morocco DJI Djibouti MAL Malaysia KOS EGY Egypt GUA Guatemala IRN Iran BRA Brazil ROK Korea South ALG Algeria RUS Russia SAL El Salvador FIJ Fiji YUG Serbia and Montenegro PER Peru ROM Romania CYP Cyprus CHL Chile COL Colombia CUB Cuba PAN Panama SIN Singapore SAF South Africa TUR Turkey LEB Lebanon CRO Croatia OMA Oman URU Uruguay GRC Greece ARG Argentina UKG United Kingdom ISR Israel

12 References Carter, David and Curtis S. Signorino Back to the Future: Modeling Time Dependence in Binary Data, Political Analysis 18: Esty, D.C., J. Goldstone, T.R. Gurr, B. Harff, P.T. Surko, A.N. Unger, & R.S. Chen The State Failure Task Force Report: Phase II findings. McLean, VA: Science Applications International Corporation. Ghosn, Faten and Scott Bennett Codebook for the Dyadic Militarized Interstate Dispute Data, Version 3.10, Correlates of War2 Project, The Pennsylvania State University, Goldstone, J. A., R.H. Bates, D.L. Epstein, T.R. Gurr, M.B. Lustik, M.G. Marshall, J. Ulfelder, and M. Woodward A Global Forecasting Model of Political Instability, American Journal of Political Science 54. 1: Harff, B No Lessons Learned from the Holocaust? Assessing risks of genocide and political mass murder since 1955, American Political Science Review 97, 1: Heckman, J Sample selection bias as a specification error, Econometrica 47, Marshall, Monty G. and Keith Jaggers Polity IV Project: Political Regime Characteristics and Transitions, , Dataset Users Manual, v.2007, Colorado State University, Regan, Patrick and David Clark. n.d. User s Manual for the IAEP Dataset SUNY Binghamton, Sartori, A.E An Estimator for Some Binary-Outcome Selection Models Without Exclusion Restrictions, Political Analysis 11: Valentino, B. P. Huth, & D. Balch-Lindsay Draining the sea : Mass killing and guerrilla warfare, International Organization 58 (2):