Show
Do military casualties have negative consequences for the incumbent's popularity or are voters under certain circumstances more tolerant toward negative policy outcomes? Numerous observational studies have examined the effect of military casualties on public support for either military presence in a conflict (Bennett and Paletz 1994; Baum 2003; Gelpi and Feaver 2006; Berinsky 2009) or for the American president (Jentleson and Britton 1998; Meernik 2001; Mueller 1970; 2005; Ostrom Jr. and Job 1986). The negative effect of military casualties on approval rates or popularity, and the subsequent conclusion that the public acts “as a rational god of vengeance and reward” (Saris and Sniderman 2018,169), has become deeply entrenched in both civil and military elites, and stimulates casualty aversion (Gelpi, Feaver, and Reifler 2009; Lacquement 2003). In a nutshell, conventional wisdom stipulates that the general public will not accept high levels of (military) casualties. Under particular circumstances, voters tend to become more tolerant of negative policy outcomes such as military casualties (Baum and Groeling 2009; Berinsky 2009; Brader and Tucker 2012; Fordham 1998). One of these circumstances is described in the literature as the “rally ’round the flag effect.” It is characterized by the tendency of increased public support for the incumbent in times of international crises. This effect is explained by feelings of patriotism (Mueller 1970) incited by the mutual feeling of being under attack as a nation, leading to more positive opinions about the leadership (e.g., Arena and Bak 2015; Hetherington and Nelson 2003; for a review, see Levy 1996). Democratically elected political leaders (or, in multiparty systems, governing parties) all face similar pressure: public perceptions of bad policy performance can translate into poor electoral performance and endanger the political survival of the leader. So far, empirical studies have shown that military casualties can negatively influence the tenure of leaders (Bueno De Mesquita and Siverson 1995) and electoral outcomes (Gartner, Segura, and Barratt 2003) but can also change decision-making behavior, by framing an intervention as a matter of national security (Gartner and Segura 1998; Smith 1998), and the duration of peace-negotiations (Filson and Werner 2004; Pillar 1983; Werner 1998). Such a change in behavior suggests that politicians assume that voters will hold them accountable for military casualties. However, the question is whether these findings are generalizable outside the United States. To evaluate whether the results hold in general, my paper analyzes 4,055 changes in opinion polls for 104 governing parties for the post–Cold War period in ten advanced democracies (Canada, Denmark, Germany, Italy, the Netherlands, Norway, Spain, Sweden, United Kingdom, and the United States). This paper's main goal is to systematically test the implications brought forward by selectorate theory (Bueno De Mesquita, Siverson, and Morrow 2003). Assuming that politicians want to stay in power, the institutional system in which they operate shapes their expectations (e.g., getting electorally punished for negative policy outcomes) and thus behavior (Bueno De Mesquita et al. 2003; Bueno De Mesquita and Smith 2012). As a consequence, I should be able to identify similar patterns (such as casualty aversion and rally effects) across similar incentive structures operationalized by regime type. OECD countries provide “the most reasonable approximation of advanced industrialized democracies” (Dalton and Wattenberg 2002,14), which ensures that that factors brought on by the incentive structure in which the incumbent operates are held constant through the selection of countries. At the same time, these countries differ on a national level: some countries have multiparty systems (e.g., Denmark and the Netherlands), others a two-party system (e.g., the United States). The sample comprises presidential systems (e.g., the United States), semi-presidential systems (e.g., France), and parliamentary systems (e.g., Canada). Although these countries differ in specifics of government arrangement and party systems, they share broad similarities in the functioning of the democratic process (Bueno De Mesquita and Siverson 1995; Davies 2008; Lijphart 1971). Advanced Western industrialized democracies are thus particularly apt subjects for testing hypotheses on public opinion (Dalton 2000). Comparing and contrasting mechanisms of more countries allows the isolation of specific variants but also the revealing of common similarities (Moraviksic 1993; Ragin 2014). A comparative approach balances between the need for comparison with being attentive to national differences and thus will allow me to (1) identify similar patterns of behavior in different democracies, and (2) if not, assess to what extent these differences might be driven by national political structures. Since several European countries in my sample have multiparty systems and thus form coalition governments, my unit of analysis is governing parties instead of countries. Even though governing parties form coalitions, some governing parties could be punished harder than others. For instance, the German Green party clashed with its pacifist subconstituents over possible German involvement in Kosovo. Taking the average popularity of all governing parties in a coalition might cloud other mechanisms brought on, for instance, by party ideology, which might explain the effect (if there is any). All the democracies in my sample have participated in several multinational coalitions (e.g., Somalia and the First and Second Gulf Wars) and contributed troops to large-scale peacekeeping missions (e.g., former Yugoslavia) but also conducted relative small scale missions monitoring peace agreements (e.g., Georgia) for various periods of time since the Cold War. I use a monthly average of polling data from several national agencies from 1990–2014 and combine that with various datasets measuring the number of military casualties in a conflict,1 while controlling for important variables such as changes in economic conditions (misery index and consumer confidence), the early term honeymoon effect, and the number of months a country is involved in large-scale military missions. My analyses show that the rally effect lasts longer than anticipated: governing parties benefit from an increase in military casualties for at least a year but get punished, on average, from 4.5 years into the intervention. The countries in my sample were only involved for such long periods during the Afghanistan and Iraq interventions. This finding falls in line with theoretical expectations as well as earlier empirical findings (e.g., Bueno De Mesquita and Siverson 1995; Kriner and Shen 2012). Testing Casualty AversionIs the general public casualty averse? After the public outrage over the casualties of the Vietnam War, scholars had to reexamine the long-held wisdom that public opinion on foreign policy was volatile, ill-informed, and had no impact on foreign policy decision-making (Holsti 2009). The academic debate was invigorated by the negative experiences in Lebanon (the Beirut barrack bombings in 1983) and Somalia (the Black Hawk incident in 1993) and by the subsequent public outrage. Various empirical studies have examined the effect of casualties on the popularity of the incumbent (e.g., Gelpi et al. 2009; Kriner and Shen 2014; for a review, see Smith 2005). The conventional wisdom that the mass public is casualty adverse inspired scholars to describe several manifestations hereof. Examples are the “Vietnam syndrome” (reluctance to use military power abroad due to the Vietnam experience), the “Dover test” (named after Dover Air Force Base where casualties arrive, the test describes the assumed effect of US-flag-draped coffins returning), and the “CNN effect” (purported impact of images of war casualties on public opinion) (Lacquement 2003). Even though empirical studies have shown that military casualties can mobilize voters (Koch and Nicholson 2016), the question is whether this mobilization is one of opposition. Recent studies have challenged this conventional wisdom that the public is casualty adverse (Berinsky 2007; Boettcher and Cobb 2006; Gelpi et al. 2006; Kriner and Shen 2012; Larson and Savych 2005). In a survey experiment conducted by Gelpi et al. (2009), members of the general public, civilian elites, and military senior officials were asked regarding three hypothetical intervention scenarios how many casualties they would tolerate. They found members of the mass public to be more willing than the two groups of elites to accept military casualties (Gelpi et al. 2009), suggesting different levels of casualty tolerance between elites and the mass public.2 This finding can be explained by the assumption among civil and military elites that in the post–Cold War era, the public no longer tolerates military casualties (Dunlap Jr. 1999; Hyde 2000; Smith 2005). Empirical studies suggest that this conventional wisdom is widespread amongst elites from different strands: the media, politicians, and their staff (Kull and Destler 1999; Lacquement 2003). It is even widespread among Western industrial defense companies, who changed the way they advertise (Schörnig and Lembcke 2006). As for military elites, the evidence is largely anecdotal. A fair share of the literature on casualty aversion interprets the stringent military measures that were taken during the engagement in Kosovo (1999) as a sign of casualty aversion brought on by the negative experiences of the United States in both Vietnam and Somalia (e.g., Cornish 2003; Dunlap Jr. 1999). An example hereof is the US-instigated instruction for NATO forces during the air campaigns over Kosovo not to fly lower than fifteen thousand feet. This tactical decision is explained as motivated by the negative American experience in Vietnam, where the army lost around four thousand aircraft due to low-level small arms fire, and is seen as evidence of assumed casualty aversion amongst military elites (Cornish 2003). The main issues identified by comparativists in this branch of literature is that many studies are (1) either limited to anecdotal evidence from historical, high profile cases (such as the wars in Vietnam, Somalia, Afghanistan, and Iraq) and (2) mainly based on American experiences (for exceptions, see Chapman and Reiter 2004; Lai and Reiter 2005). Can these findings be generalized to advanced industrial Western democracies, and is the general public casualty adverse? My first hypothesis aims to systematically test whether this is the case by analyzing the relationship between increasing military casualties and fluctuations in governing party's popularity. I label this expectation the casualty aversion-hypothesis. H1:An increase in military casualties decreases governing parties’ popularity. Rally ’round All the Flags?Despite the conventional wisdom that military casualties negatively affect incumbent popularity, several arguments have been brought forward to explain the opposite, voter tolerance toward casualties. One branch of literature looks specifically into the mediating effect of unified elites in wartime—operationalized by the so-called rally-around-the-flag-effect. In times of crisis or war, groups display the tendency to experience increased cohesion whenever they are in conflict with an external group (Gelpi 1997, 255, with reference to Coser 1956). Increased cohesion can translate itself into patriotism and thus increased support for the incumbent, which has become known as the rally effect (for a review see Levy 1996). Besides this conflict-cohesion hypothesis (which focusses on cohesion by the public in the event of a crisis), another and related explanation for the rally effect focuses on the absence of the opposition. Opposing the administration over foreign policy decisions in times of conflict can be seen as politicizing military casualties. The absence of opposition (and thus the apparent elite consensus in the eyes of the public) can subsequently lead to more favorable opinions toward the leadership (Brody and Shapiro 1989; Lai and Reiter 2005). Here the assumption is that the mass public responds not so much to the event of a crisis but to the unified response in times of crisis by (otherwise divided political) elites.3 Another potential underlying mechanism of the rally effect, and one of which my data (unfortunately) do not allow further unpacking, is the role of the media (Aday 2010; Baum and Potter 2008). Even though empirical research is mainly conducted with US data, content analyses of newspaper editorials during the Iraq War also identified a post-9/11 rally effect (Dimitrova and Strömbäck 2005). Within the limits of available data, it cannot be concluded whether this rally effect is a reflection of the absence of opposition or whether media itself was leading (Gelpi et al. 2006). In other words, would there have been (such a strong) rally effect after 9/11 if the media would have reported completely different on the post-9/11 events? Apart from information on the intervention itself, the public gets information on how the rest of the public perceives the intervention. Perception of collective opinions can influence individual attitudes (Baum 2003; Gartner, Segura, and Barratt 2003; Rojas 2010). Again, the availability of data do not allow me to disentangle these mechanisms identified in the literature for the time period and countries of interest, so this paper will focus on whether there is such a thing as a rally effect. If there is, future research can benefit greatly from insights from empirical research on framing (Auerbach and Bloch-Elkon 2005; Boettcher and Cobb 2006), or the aforementioned elite consensus (Berinsky 2007; Druckman, Peterson, and Slothuus 2013), and help further investigation into the several (and possibly simultaneous) mechanisms that drive the rally effect. The rally effect has been identified as one of the significant determinants of the presidential (and wartime) vote in 2004 (Norpoth and Sidman 2007). It is also considered one of the many variables that explain fluctuations in party identifications (Groeling and Baum 2008; Lebo and Cassino 2007; Norrander and Wilcox 1993). Such favorable opinions toward the government in times of crisis or war are, however, short lived. Mueller found that support for military interventions in the United States peaks in the beginning of a war but steadily declines over time (Mueller 1970, 1971, 2005). Further research that focused on the duration of the rally effect confirms that governments only reap the benefits hereof in the beginning of the crisis: thirty days (Stoll 1987) up to one to two months (Lian and Oneal 1993). The relatively long-term increased public support for Bush Jr. after 9/11, over thirteen months, can be seen as an exception (Hetherington and Nelson 2003; Schubert, Stewart, and Curran 2002). Following this reasoning, if there is such a thing as a rally effect, the assumed negative effect brought on by military casualties should appear later on in the intervention, not in the beginning. I label this expectation the rally-around-the-flag hypothesis and propose that: H2:Military casualties have no effect on governing parties’ popularity in the beginning of the intervention. Exploring international determinants for government support across democratic countries is important, since it allows for systematic testing of the selectorate theory developed by Bueno De Mesquita et al. (2003). The selectorate comprises three groups that affect political leaders: (1) the nominal selectorate (e.g., in a democracy: every citizen who is eligible to vote), (2) the real selectorate (e.g., those citizens who go out to vote), and (3) the winning coalition (e.g., the voters that cast their votes for the winning candidate). The core assumption of this theory is that political leaders who want to stay in power depend on the support of the winning coalition. However, leaders are bound by different political structures—such as regime type—that shape the size of the winning coalition, thus creating different patterns of behavior. This suggests that I can expect similar patterns for democratically elected political leaders (or, in multiparty systems, governing parties), since they all face similar pressure: the public's perception of bad policy performance can translate into bad electoral performance and endanger their political survival. Even though empirical findings suggest that politicians assume that the general public will hold them accountable for military casualties (e.g., Bueno De Mesquita and Siverson 1995; Filson and Werner 2002; Gartner 2008), the question is whether this is indeed (always) the case. Design and MethodsDescription of the Dependent VariableSince I am interested in whether an increase in military casualties has an effect on popularity, and in line with empirical studies testing the rally-around-the-flag-effect, voting intention is my dependent variable (e.g., Hetherington and Nelson 2003; Lambert et al. 2010; Norpoth and Sidman 2007; Schubert et al. 2002). In contrast to individual level data such as surveys on military interventions that express levels of support for the intervention itself, voting intention is not fielded on an ad hoc basis, and polls therefore have regular intervals independent of what happens during an intervention (Feaver and Gelpi 2004). To track changes in popularity, I have built a dataset comprising monthly polling data (1990–-2014) from all the countries in my sample (Canada, Denmark, Germany, Italy, the Netherlands, Norway, Spain, Sweden, the United Kingdom, and the United States). For European countries, I used the monthly average of available polling data from several bureaus (see appendix 1). In order to make sure that changes in the polls are not related to changes in the number of polling agencies that were used to calculate a monthly average, I have created a dummy variable that indicates whether in any given month the average is made up from different polling agencies than the month before (1) or not (0). I found no significant effect (see appendix 1, table 2). Since many European countries are multiparty systems and thus several parties are simultaneously in government, my unit of analysis is governing party_month (instead of country_month). To make sure the found effects on popularity are brought on by military casualties (and not the other way around), I ran a second diagnostics on my dependent variable and tested for endogeneity. Before running a Durbin Wu-Hausman diagnostic, I have conducted a two-stage least squares (2SLS) in which I took duration as the dependent variable (Y1), military casualties as the suspected endogenous explanatory variable (Y2), and popularity as the instrumental variable. These tests reveal no reverse causality (see appendix 1, table 3). My dependent variable expresses the difference in the percentage of (potential) voters for a certain party (t) and those in the previous month (t-1). Since the rally effect lasts for one to two months on average (Eichenberg, Stoll, and Lebo 2006; Lian and Oneal 1993), it is preferable to use data on a monthly rather than a quarterly basis. Description of the Independent VariablesMy first and main independent variable of interest, military casualties, is operationalized in two different ways: (1) by the number of military casualties during missions in a given month and (2) the rate of change of military casualties from month to month. I have collected these data through open data sources, such as army archives, rolls of honor, and veterans’ websites.4 Since military casualties are a rare event, this independent variable is not normally distributed. 77.4% of all the observations are zeros, with one casualty in any given month accounting for 11.7% of the observations. In addition, there are large differences between countries in terms of the number of military casualties, with the United States accounting for every observation of more than eight-four (with a maximum of 173). Even though traditional studies (such as Mueller 1971) use logged cumulative casualties as the main independent variable, logging eliminates variation in patterns due txo casualty build up over time (Gartner 2008). I therefore use the absolute numbers of military casualties (“marginal casualties”) and control for country effects (see appendix 2.6) to ensure that the United States does not consume my model. The second operationalization, rate of change of military casualties, uses the difference in the number of military casualties (t) and those in the previous month (t-1). There are several reasons for doing so. First, rates of change take into account that a monthly casualty figure is relative (Gartner 2008; Gartner and Segura 1998; Murray 2001). For example, if there were 50 military casualties in the previous month, it would be perceived differently if the previous trend had been 150, 100, and then 50 than if the previous trend had been 0, 25, and then 50. Second, and related to the former, trends influence individuals’ estimates of not only the direction but also the possible magnitude of future losses (Gartner and Segura 1998). As a consequence, the Δcasualty growth variable can also take on negative values in a certain month (t) when there are fewer military casualties in relation to the previous month (t-1, see figure 1). Since many countries are not confronted every month with a growing number of casualties, the variable is not normally distributed. As with the previous operationalization of the military casualties variable, I have chosen not to log it, to avoid elimination of variation in patterns. In order to test the second hypothesis on the rally effect, I use this variable to operationalize a rally event. According to Mueller (1970, 1971) “specific, dramatic, and sharply focused” events during international crises can increase incumbent support. In this paper, I operationalize the event of military casualties as a rally event. Participation in active combat is highly related to the number of military casualties, and also makes it possible to account for traditional rally events, such as the initiation of war. For instance, on a US-only sample, Gelpi and Feaver found—when controlling for the initiation of war—no effect, since the “the casualties suffered during the major combat phase captures the rally event” (Gelpi and Feaver 2006, 20). In addition, many countries in my sample do not (1) often initiate wars but rather support US- or French-led interventions (e.g., Iraq or Mali) and (2) relatively rarely suffer military casualties. Therefore, the use of military casualties as a rally event is most appropriate to test both within and outside the US context. Approaching military casualties as a rally event has some theoretical implications as well. For instance, fair share of the studies on casualty tolerance have used (the log of) cumulative casualties (Gartner and Segura 1998; Mueller 1970, 1971). Some of these studies have used the theory of investment to explain varying levels of casualty tolerance (e.g., Koch 2011). Here it is argued that increasing costs in a conflict, such as military casualties, can increase the probability that voters will support the incumbent. While this paper looks into the question of whether the public responds to rally events in times of international crisis (by supporting the incumbent when faced with losses), analyses employing cumulative military casualties will help answer questions on whether mounting costs can help explain casualty tolerance. I have rerun the analysis with cumulative military casualties as a robustness check (see appendix 2.10) and found (in line with model 1 in this paper) no effects. The second independent variable, duration, is the sum of the total number of months a country was involved in missions. Since the unit of observation is the months in which countries were involved in a military intervention, the minimum number for duration is one month. This makes it possible to weigh countries that participate in several conflicts at the same time (e.g., both Afghanistan and Iraq, such as the Netherlands) differently than those that are involved in only one intervention at the time (e.g., Afghanistan alone, such as France). Countries that participated in both the Afghanistan and Iraq interventions show relatively high numbers. For instance, the number of months they were involved in any given intervention at that time can be as high as 221 (e.g., Denmark in 2011) or even 229 (e.g., United States in 2011). Figure 2 shows the total duration of the military intervention by year and conflicts. Here it becomes evident that certain interventions (e.g., Afghanistan and Iraq) took place around the same time, creating a certain degree of overlap. However, it is important to stress that the duration variable expresses the total number of months a government is involved at a given time (t). Especially since the two long lasting wars of Afghanistan and Iraq are two overlapping interventions, the total number of months can add up quickly. The data come from an original dataset I have created, which codes all the decisions (e.g., prolong mandate or increase or decrease number of troops) related to military interventions known to the public (and thus excluding covert operations) from eleven OECD countries5 in the period 1990–2014, which allowed me to clearly distinguish between (peace) enforcement and postconflict operations (such as the training missions in Afghanistan and Iraq). This is important since (peace) enforcement missions operate in (1) a higher specter of violence, translating into higher casualty rates, and (2) incorporating postconflict missions can significantly distort the image of commitment in terms of time. For instance, Operation Iraqi Freedom in 2003 can be considered a follow-up (or escalation) of Operation Southern Watch (1991–2003), which itself followed immediately after Operation Desert Storm (1991). Hence, even though the literature makes a clear distinction between the military operations in Iraq in 1991 and those in 2003, when approached from the perspective of military interventions—these can also constitute two “peaks” of military involvement during a long-term conflict with Iraq. All available literature on public opinion in wartime treats the involvement in Iraq in 1991 and 2003 as two separate events, and for good reason: separating these two provides a more accurate representation of how the public and politicians perceived the involvement in Iraq. The same goes for separating the involvement in former Yugoslavia (mainly Croatia, Bosnia, and Serbia) from the involvement in the following conflict in Kosovo. It was not until June 1998 (almost three years after the Dayton accords ending the Bosnian civil war) that American president Clinton issued a state of emergency in Kosovo, which resulted, in the following months, in the deployment of American and European troops. Even though both conflicts are a direct result of the disintegration of the Former Republic of Yugoslavia, they are treated as two separate conflicts (and interventions) in literature and thus also in this paper (see table 1). Table 1.
To ensure that if there is a rally effect, it is exclusively bound to governing parties, I have run an additional analysis for opposition parties (with a split sample) as a robustness check. The effects resulting from the analysis were indeed bound to governing parties only (see appendix 2.1). There is widespread consensus in the literature that the economic state of a country greatly matters for approval rates and thus for a government's standing (Lewis-Beck and Paldam 2000; Nadeau, Lewis-Beck, and Belanger 2012;,Nadeau and Lewis‐Beck 2001). To make sure that the effects attributed to military casualties are not an artifact of a worsening state of the economy, I have incorporated the OECD standardized unemployment rates and inflation, allowing me to create the so-called misery index on a monthly basis. This is in line with earlier empirical studies on the rally effect (Brulé and Williams 2009; DeRouen 1995; Ostrom Jr. and Job 1986). As a robustness check, I replaced the misery index with the consumer confidence index (CCI) (see appendix 2.2). The CCI measures how voters themselves experience the country's state of the economy rather than using indicators that I assume have an effect on sentiments (Eichenberg et al. 2006; O'Connor et al. 2010). Using the misery index or CCI does not influence my results. As mentioned in the theory section, besides economic performance, the so-called early term “honeymoon” effect has been found to influence polls (Cordesman 2010; Hetherington and Nelson 2003; Lai and Reiter 2005; Norpoth 1984; Powell and Whitten 1993). In order to control for this, and in line with earlier empirical studies, I have incorporated a count-down variable from the first month in office (rated with the highest score, 6) to the sixth month in office (rated with the lowest score, 1) based on the maximum duration of this effect as identified in the literature (Beck, Carr, and Walmsley 2012; Cordesman 2010; González‐Bailón, Banchs, and Kaltenbrunner 2012). Finally, I control by country and include a separate analysis that excludes the United States. To make sure the United States (the outlier in the sample) does not consume the model, I have run an extra analysis that controls for troop to population ratio, which can be found in the appendix (see appendix 2.3). Estimation MethodI conduct cross-sectional time series analyses, since this allows me to account for the dynamics between my dependent variable, the difference in popularity, and my independent variables (see table 2). Such models have issues with autocorrelation (Beck and Katz 1995). Even though lagged variables incorporate feedback over time in my model, they can introduce the issue of heteroskedastic error terms at the same time (Stimson 1985). In addition, the biases of lagged variables associated with trends within my independent variables and error term can annul the effect of the theoretical model (Plümper, Troeger, and Manow 2005). When running a fixed-effects (within) regression with a lagged dependent, the results were similar to those when applying feasible generalized least squares (FGLS) regression corrected for autocorrelated errors (see appendix 2.4). FGLS as an estimation technique is more effective, since my panel is unbalanced: not every country in my sample holds its election simultaneously, and there are some months in which polling data are missing. FGLS allows to correct for this in two ways: (1) by estimating autocorrelation within panels as well as heteroskedasticity across panels by correcting for autocorrelated errors and (2) by constructing confidence intervals around relative changes and specifying the type of standard error reported, robust to misspecification. Therefore, FGLS will be my estimation technique, since it can estimate autocorrelation (AR[1]) within panels, as well heteroskedasticity across panels. Table 2. Operationalization of main (in)dependent variables
FindingsDo military casualties influence governing parties’ popularity? Model 1 (table 3) indicates that military casualties or the rate of change of military casualties have no significant effect on popularity. Positive changes in the misery index (indicating a worsening economic condition) have a significant negative effect on the polls. An increase in the misery index leads to a −.156 drop in the polls the following month. Positive changes in the honeymoon variable translate into positive changes in the polls. Since the honeymoon variable is a count variable, this indicates that the shorter duration a government is in office, the better this is for its standing in the polls. However, and contrary to earlier empirical studies (Cordesman 2010; Hetherington and Nelson 2003; Lai and Reiter 2005; Norpoth 1984; Powell and Whitten 1993), the effect is not significant. Model 2 employs the second operationalization of military casualties by looking at the rate of change on a monthly basis. The casualty growth is positive and insignificant, suggesting there is no effect. The results for the control variables are similar. Apart from the negative significant effect of worsening economic conditions on the polls (for every one-point increase in the misery index, there is a −.155 drop in the polls the next month), all the other indicators are not significant. Table 3. Feasible generalized least squares (FGLS) regression analyses of military casualties on the change in polls for governing parties
These results indicate that military casualties do not threaten the popularity of the incumbent. They demonstrate that the public is not casualty averse, nor does it blame the government once confronted with the costs of war. This means I can reject H1. Are the effects of military casualties of a temporary nature? Do military casualties have an effect only in the beginning of the mission or only once the intervention drags on? Models 3 and 4, which table 4 presents, include the duration variable that accounts for the total number of months a country is involved in a military intervention at any given time (t). Since my hypotheses assume a different response of the public toward military casualties depending on the time, an interaction effect between the military casualties variables and duration variable is included in models 3 and 4. Both models confirm my second hypothesis: governing parties are not being punished in the polls for military casualties in the beginning of the intervention. On the contrary, every military casualty leads to an increase of .018% in the polls the following month. A small but significant effect. The effect is even larger when there is an increase in casualty growth: if the rate of military casualties increases (and more soldiers’ lives of are lost in relation to the previous month) governing parties can expect an increase of .25% in the polls. Since one standard deviation in my dependent variable is larger (1.73), the effect is significant but small. Table 4. Feasible generalized least squares (FGLS) regression analyses of military casualties on the change in polls for governing parties
Moreover, the positive effect for governing parties disappears over time and turns into a significant negative one. This indicates an asymmetrical effect: governing parties get rewarded for negative outcomes in the first period of a long-term intervention, but they get punished for the negative outcomes of their decisions as the intervention drags on. This falls in line with the expectations from the literature: the longer an intervention takes, and the more costly it becomes (in terms of military casualties), the more the general public will doubt the wisdom of the intervention (Bueno de Mesquita et al. 2003; Levy 1996). Even though both in the same direction and significant, these effect are even stronger for the operationalization of military casualty growth (see figure 2). This suggests that government popularity increases even more if the level of military casualties is higher this month than it was in the previous one. Again, this falls in line with expectations from the literature (Gartner 2008; Gartner and Segura 1998). Figure 3 shows the average marginal effects of an increase in military casualties (in absolute numbers and in terms of growth) on the polls over time. Here, I have compared several hypothetical populations (for every twenty-five months), which have the same values as the other independent variables in the model. The only difference between these hypothetical populations is time (e.g., up to twenty-five months into the intervention, up to fifty months into the intervention, etc.), so any difference can therefore be explained by a difference in time. All the values above the zero line indicate a positive effect, those below the zero a negative effect. The effects are significant if the margins remain either above or below the zero-line. Figure 2 thus illustrates how governing parties experience a positive effect (and thus go up in the polls) once they are confronted with an increase in military casualties but only in the first fifty months of the intervention. This turns into an increasingly significant negative effect from the 150th month onward. From the 150th month, for each military casualty the incumbent can expect a decrease in the polls of 0.13%, up to 0.29% from the 225th month onward. Since the duration variable expresses the total number of months a country is involved in military interventions, these high numbers are only found for those countries who participated for either (1) a long period in the war in Afghanistan (e.g., Canada, which was involved from March 2002 onward) or (2) in both the Afghanistan and Iraq Wars (e.g., the Netherlands, which by May 2009 had been involved for eighty months in Afghanistan and seventy months in Iraq). Countries that were involved in interventions for a longer (total) time period of two hundred months are limited to Denmark, the Netherlands, the United Kingdom, and the United States, with the United States as the only country which was involved for a total number of 225 months (18.75 years of involvement in Afghanistan and Iraq combined). Contrary to what the literature suggest, I see a relatively long lasting rally effect for the incumbent: the significant positive effect lasts when governments are up to a total number of fifty months caught up in intervention(s). Since there are several missions overlapping each other and are in various stages at any given time (t), it is difficult to assess how long the rally effect actually is for each separate intervention. To assess whether there are differences for those missions that started before 9/11 and after, I ran the analysis for both time periods (from January 1990 until August 2001 and from September 2001 until December 2014) as a robustness check (see appendix 2.5). The reason for doing so is twofold. First, during the 1990s, most interventions (e.g., those in former-Yugoslavia) were humanitarian missions with relatively little risk for soldiers. But after 9/11, most interventions were large-scale, large-risk ground operations to first bring about regime change and eventually quell counterinsurgencies. This led to an increase in military casualties for participating countries. Before 9/11, the countries in my sample were confronted with an average of .108 military casualties per month. After 9/11, this increased tenfold to an average of 1.11. Apart from the higher number of military casualties, public opinion toward humanitarian mission might be different from missions aimed at regime change or antiterrorism efforts. Contrary to the main consensus in the literature, there is a rally effect in the beginning of missions but only in the pre-911 period, and—in line with empirical findings on the perceived success of a mission—this seems conditional (Boettcher and Cobb 2006; Eichenberg 2005; Gelpi and Feaver 2006). For the Gulf War (the only mission in my sample not coinciding with other interventions), there is an increased positive effect from the sixth month (February 1991) onward: that month, coalition troops initiated Operation Desert Storm and moved into Iraq. This advance was swifter than anticipated, and by the end of the month, Iraqi troops began retreating from Kuwait. For the two large scale interventions after 9/11 (Afghanistan and Iraq), there is no rally effect in the beginning of the mission. However, after a total of a hundred months, governing parties get punished. It is not clear whether these differences can be explained by either the difference in the nature of these mission or the differences in casualty levels. Since a fair share of the rally-around-the-flag literature is based on US-findings (Arena and Bak 2015; Groeling and Baum 2008; Hetherington and Nelson 2003b; for an overview see Levy 1989), with the United Kingdom as an exception (Lai and Reiter 2005), the question is whether this is purely an American (or Anglo-Saxon) phenomenon. To check whether the United States does not consume the model, this country is dropped from the analysis in models 5 and 6 of which the results are shown in table 5. Again, all the effects hold, indicating that the rally-around-the-flag effect is not exclusively bound to the United States. Instead, this effect can be seen as a widespread phenomenon in mature, Western industrial democracies (see figure 4). As figure 4 shows, the significant negative effects after the 150th month of the intervention disappear. The effects (both the positive effect in the beginning and the negative effect later on in the mission) disappear altogether when using the operationalization of casualty growth. This could be explained by (1) the on average positive growth experienced by the United States (mean .178) and the on average negative growth by the other countries in my sample (mean −.074), brought on by the relatively lower military casualties non-US countries have experienced. As a robustness check (see appendix 2.6), I have rerun the analysis controlling for countries. Here, all the effects hold, suggesting that the rally effect cannot be explained by differences in country characteristics. It is worth noting that for some countries, the direction of the effect (even though insignificant) turns negative, suggesting that in these countries governing parties experience overall negative (but not significant) effects. This could be explained by the fact that the countries experiencing these negative effects, such as Canada, Spain, the Netherlands, the United Kingdom, and the United States, are relatively often “leading nations” during military missions and thus experience higher level of military casualties, but unfortunately my data do not allow me to further investigate these differences. Table 5. Feasible generalized least squares (FGLS) regression analyses of military casualties on the change in polls for governing parties
In addition to the above, in the appendix, several robustness checks of my analysis can be found with a duration dummy, a rerun of the analysis with a third operationalization of the independent variable (Δ intensity) that measures the change in the number of military casualties in relation to the total number of troops send abroad and controlled for additional operational factors such as conflict intensity and battle-related-deaths. Again, all the effects hold (see appendix 2: robustness checks). Conclusion and DiscussionDo military casualties always damage a government's popularity? No. The increasing doubt in the literature as to whether increasing death tolls during conflicts are always associated with declining political and public support—is justified on the aggregate level as well. My two hypotheses (the casualty aversion and the rally-around-the-flag hypotheses) tested whether governing parties get punished for negative consequences of their policies in the polls. By doing so, I tested a widespread assumption in the literature and among political and military elites: long term and costly interventions are bad political business for those in government (Bueno De Mesquita et al. 2003; Kriner and Shen 2012; Williams, Brulé, and Koch 2010). I found (partial) support for this assumption. In addition, I found a rally effect in the beginning of nonpeacekeeping missions that seemed to last longer than the literature anticipated. Governments can reap the benefits for at least a year into the intervention. More importantly, the rally effect is not an exclusive American phenomenon. This is perhaps the most interesting finding, since the tendency in the literature is to ascribe the rally effect almost exclusively to the United States (Arena and Bak 2015; Groeling and Baum 2008; Hetherington and Nelson 2003; Lambert et al. 2010; Norpoth and Sidman 2007; Schubert et al. 2002). Despite the relatively high opposition and diverging European governmental positions during the wars in Iraq and Afghanistan, whenever the death toll rose during the first months of the intervention, European governments too experienced increasing levels of support, most strongly connected in the literature to the rally effect (Levy, Pensky, and Torpey 2005; Schuster and Maier 2006). Moreover, evidence for punishment in the polls for long-lasting interventions are less robust than the rally effect in the beginning hereof. Debunking the widespread assumption among political and military elites that military casualties do not sit well with the electorate is important for several reasons. First, on a political level, numerous presidents felt the need to account for military policies and military casualties during their presidential campaigns, such as president George W. Bush who had to deal with antiwar sentiments during his 2004 campaign (Berinsky 2009; Karol and Miguel 2007), or were criticized for not doing so, such as presidential nominee Hillary Clinton in the 2016 campaign regarding the events in Benghazi. Second, the assumption that military casualties endanger government popularity seems to have inspired certain foreign policies and force protection rules. For instance, the early US-retreats from interventions such as Bosnia (before 1995), Rwanda (1994) and Zaire (1995) have been explained as anticipatory policies motivated by the fear of public backlash due to this aversion toward casualties (Lacquement 2003). The Weinberg/Powell doctrine further advanced the assumption that public support was difficult to obtain and easy to lose, especially when confronted with casualties. This led to overly stringent force-protection rules during the 1999 Kosovo mission, during which American troops were outfitted with every force-protection gadget imaginable in order to prevent military casualties. As one American officer noted at the time: “being dressed as a Ninja Turtle gets in the way” (Mockaitis 2004,15). During the deployment in Afghanistan, this so-called “force-protection fetishism” has affected military tactical deployment “in three ways: armorizing, employment of forward operating bases (FOBs), and the application of heavy firepower” (Gibson 2009, 4). By doing so, troops have been isolated from the populace during missions and have deflected casualties onto civilians, making counterinsurgency efforts nearly impossible (Egnell 2010; Gilmore 2011; Zambernardi 2010). In addition, and even more important, it signals to enemies that Western political leaders are casualty averse and that defeat is thus brought on by Western military casualties rather than military victory. Hence, overly focusing on force protection and the prevention of military casualties, military and political elites can paradoxically encourage suicide attacks on Western military forces to provoke withdrawal. My results demonstrate how insights from selectorate theory in combination with more recently available data allow for systematic testing using aggregate level data. Even though some phenomena are sometimes wrongfully interpreted as “typical” US (since inferences are drawn based on a US sample), these results underscore earlier findings on how regime types, in this case Western democracies, show similar mechanisms (Bueno de Mesquita et al. 2003; Kisangani and Pickering 2009). Unfortunately, my data do not allow to establish under what circumstances voters actually “defect” to another party and, if so, to which (opposition) party. This is especially difficult to establish on the aggregate level for multiparty systems where voters can often choose from more than one liberal, social-democratic, or Christian-democratic party. One way to address this would be to would be to use individual level data. Combining both aggregate and individual level data, such as was done by Koch (2011), will make it possible to model individual vote choice in response to rally events such as military casualties. Are voters more tolerant when losses incur whenever “their” party is in government, as established by US-based studies (e.g., Berinsky 2009; Fordham 1998) and a study based on American, British, and French cases (Koch and Sullivan 2010)? The limitation of aggregate-level data, as used in this study, is that it only allows us to make inferences on the entire selectorate. Future research will benefit from combining these insights with individual-level data to establish under what circumstances the winning coalition starts to crumble. A second shortcoming of this study that could be addressed in future research is the difficulties arising when trying to generalize findings from an outlier in the political, institutional, and military sense, such as the United States. Apart from the United States, there are no presidential systems in this sample, which makes the country not only an outlier in terms of the number of military casualties suffered but in institutional arrangements as well. The United States is also an outlier in terms of burden sharing in multinational military interventions, which makes it difficult to establish why other operationalizations of military casualties seem less apt. For instance, even though most US-based studies employ cumulative casualties as the main independent variable, I found no effect on incumbent popularity (see appendix 2.10). It is difficult to assess whether this is due to the limitations of my data or whether this measurement is not adequate for a larger set of democracies. It is important to note that studies employing (marginal) cumulative casualties (such as Gartner and Segura 1998) look at US interventions in Vietnam and Korea and thus marked periods of US soldiers engaged in active combat. This is different from the postconflict reconstruction phases in which European countries in my sample were involved (e.g., in former Yugoslavia). Distinguishing active fighting troops from those in support roles is difficult to disentangle on a month-to-month basis, especially during counterinsurgency operations such as Afghanistan and Iraq. My data do incorporate information on the type of troops that are deployed (e.g., marines, air force, land forces, etc.), but whether these troops are actually deployed in the theatre is not always clear. To establish this, I would need data on the actual size of the theatre (on a month-to-month basis) and the exact location of the troops (on a month-to-month basis). Unfortunately, this data is not available through open sources. Future research could benefit from operationalizing the level of burden sharing by countries, to establish the intensity of the deployment. This evidence of a cross-national rally effect concurs with the main consensus in the literature, that group membership (in this case to a country, rather than to a political party) creates a higher tolerance toward negative policy outcomes in times of crisis. Despite the high number of military casualties that are only to be found in the United States, I was able to systematically test and compare polling data and conclude that both Americans and European tend to rally around their flags. Supplemental InformationSupplemental information is available at the Foreign Policy Analysis data archive. Dieuwertje Kuijpers (1984) obtained her PhD at the Vrije Universiteit, Amsterdam and is now an independent investigative journalist writing on defense-related matters. References, , . . “When War Hits Home: The Geography of Military Losses and Support for War in Time and Space.” Journal of Conflict Resolution (): –., . . “Media Framing and Foreign Policy: The Elite Press vis-a-vis US Policy in Bosnia, 1992–95.” Journal of Peace Research (): –., . . “Shot by the Messenger: Partisan Cues and Public Opinion Regarding National Security and War.” (): –., . . “The Relationships between Mass Media, Public Opinion, and Foreign Policy: Toward a Theoretical Synthesis.” Annual Review of Political Science (): –., , . . “What Have You Done for me Lately? Charisma Attenuates the Decline in US Presidential Approval over Time.” (): –., . . “Echoes of Vietnam? Casualty Framing and Public Perceptions of Success and Failure in Iraq.” The Journal of Conflict Resolution (): –., . . “Following the Party's Lead: Party Cues, Policy Opinion, and the Power of Partisanship in Three Multiparty Systems.” (): –., . . “Policy Failure and Public Support: The Iran-Contra Affair and Public Assessment of President Reagan.” (): –., . . “Democracy and Diversion: Government Arrangements, the Economy, and Dispute Initiation.” Journal of Peace Research (): –., . . “War and the Survival of Political Leaders: A Comparative Study of Regime Types and Political Accountability.” The American Political Science Review (): –., (Eds.). . Parties Without Partisans: Political Change in Advanced Industrial Democracies .Oxford University Press on Demand .. “Inside Out or Outside In: Domestic and International Factors Affecting Iranian Foreign Policy towards the United States 1990–2004.” (): –., . . “Mission Accomplished? Framing of the Iraq War in the Elite Newspapers in Sweden and the United States.” International Communication Gazette (): –., , . . “How Elite Partisan Polarization Affects Public Opinion Formation.” American Political Science Review (): –., . . “A Bargaining Model of War and Peace: Anticipating the Onset, Duration, and Outcome of War.” American Journal of Political Science (): –., . . “Bargaining and Fighting: The Impact of Regime Type on War Onset, Duration, and Outcomes.” American Journal of Political Science (): –.. “The Multiple Effects of Casualties on Public Support for War: An Experimental Approach.” American Political Science Review (): –., , . . “War Casualties, Policy Positions, and the Fate of Legislators.” Political Research Quarterly (): –.. Hell-bent on Force Protection: Confusing Troop Welfare with Mission Accomplishment in Counterinsurgency . ., , . . “Emotions, Public Opinion, and U.S. Presidential Approval Rates: A 5-Year Analysis of Online Political Discussions.” Human Communication Research (): –., . . “Crossing the Water's Edge: Elite Rhetoric, Media Coverage, and the Rally-Round-the-Flag Phenomenon.” (): –., . . “Testing the ‘Proximate Casualties Hypothesis’: Local Troop Loss, Attention to News, and Support for Military Intervention.” Mass Communication and Society (): –., . . “Still Pretty Prudent: Post –Cold War American Public Opinion on the Use of Military Force.” The Journal of Conflict Resolution (): –., . . “The Dividends of Diversion: Mature Democracies’ Proclivity to Use Diversionary Force and the Rewards They Reap from It.” British Journal of Political Science (): –.. “Casualties and Incumbents: Do the Casualties from Interstate Conflicts Affect Incumbent Party Vote Share?” British Journal of Political Science (): –. ., . . “Death and Turnout: The Human Costs of War and Voter Participation in Democracies.” American Journal of Political Science (): –., . . “Should I Stay or Should I Go Now? Partisanship, Approval, and the Duration of Major Power Democratic Military Interventions.” (): –., . . “Rally ’Round the Union jack? Public Opinion and the Use of Force in the United Kingdom, 1948–2001.” International Studies Quarterly (): –., , , , , , . . “Rally Effects, Threat, and Attitude Change: An Integrative Approach to Understanding the Role of Emotion.” Journal of Personality and Social Psychology (): –., . . American public support for US military operations from Mogadishu to Baghdad . ., . . “The Aggregated Consequences of Motivated Reasoning and the Dynamics of Partisan Presidential Approval.” (): –., , , . . Old Europe, new Europe, core Europe: transatlantic relations after the Iraq war . .. “Integrating International and Domestic Theories of International Bargaining.” InDouble Edged Diplomacy: International Bargaining and Domestic Politics , edited by , , –. :University of California Press ., , , . . “From Tweets to Polls: Linking Text Sentiment to Public Opinion Time Series.” Proceedings of the International AAAI Conference on Weblogs and Social Media, 23–26 May 2010, Washington DC . : , –., , . . “Panel Data Analysis in Comparative Politics: Linking Method to Theory.” European Journal of Political Research (): –., . . “A Cross-national Analysis of Economic Voting: Taking Account of the Political Context.” American Journal of Political Science (): –.. “‘Corrective’ Actions in the Public Sphere: How Perceptions of Media and Media Effects Shape Political Behaviors.” International Journal of Public Opinion Research (): –., . . “The Vision of War without Casualties: On the Use of Casualty Aversion in Armament Advertisements.” The Journal of Conflict Resolution (): –., . . “The Rift: Explaining Europe's Divergent Iraq Policies in the Run‐Up of the American‐Led War on Iraq.” (): –., , . . “War Voting: Interstate Disputes, the Economy, and Electoral Outcomes.” Conflict Management and Peace Science (): –. |