Education and Militarism: Exploring the Link between Poverty and Civil Wars 1

Transcription

1 Education and Militarism: Exploring the Link between Poverty and Civil Wars 1 Moses Shayo 2 Abstract: Why are poor countries more likely to experience civil wars? This paper uses both individual level and cross country data to suggest that part of the reason could be a militaristic bias fostered by low educational attainment. We find that in a large and diverse set of countries, poorer, less educated people tend to have higher confidence in their nation's army. Popular confidence in the army is in turn positively related to civil war incidence, and can account for part of the relationship between per capita GDP and civil war incidence. The social situation of the lower strata, particularly in poorer countries with low levels of education, predisposes them to view politics as black and white, good and evil. Consequently, other things being equal, they should be more likely than other strata to prefer extremist movements which suggest easy and quick solutions to social problems and have a rigid outlook. [...] The lower class in any given country may be more authoritarian than the upper classes, but on an "absolute" scale all the classes in that country may be less authoritarian than any class in another country. In a country like Britain... even the lowest class may be less authoritarian and more "sophisticated" than the most highly educated stratum in an underdeveloped country, where immediate problems and crises impinge on every class and short-term solutions may be sought by all groups. S.M. Lipset, Political Man, 1960, p Introduction Civil wars account for most of today's armed conflicts. 3 The death toll directly resulting from civil wars between 1945 and 1999 is estimated at over 16 million five times the interstate wars toll in this period (Fearon and Laitin 2003). Civil wars in this period have occurred in 73 different countries. The enormous human suffering caused by civil wars, has prompted several attempts to empirically identify the conditions conducive to their occurrence (see Sambanis 2002 and Humphreys 2003 for recent surveys). The strongest 1 I am grateful to Roland Benabou, Anne Case, Ori Heffetz and Thomas Romer for helpful comments and suggestions. All remaining errors are my own. Financial support from the Maurice Falk Institute is gratefully acknowledged. 2 Department of Economics, The Hebrew University, Jerusalem 91905, Israel. mshayo@huji.ac.il. 3 According to Marshall and Gurr (2003), At the end of 2002 there were twelve ongoing major societal wars In addition, eleven societal wars were experiencing sporadic outbursts of violence. On the other hand, Interstate wars were uncommon after the United Nations system was established following World War II. In the 1990s, there were very few interstate wars and their magnitude and duration were mostly limited (pp.12-13). 1

2 and most robust finding in this literature is that poor societies have experienced much more civil warfare than did prosperous societies (Collier and Hoeffler 2001, Fearon and Laitin 2003, Miguel et al See Miguel et al. for a discussion of the endogeneity issue). This paper attempts to contribute to our understanding of this link. Several mechanisms have been proposed to explain why poor countries may be more prone to civil wars. Collier and Hoeffler (2001) view income per-capita as capturing the opportunity cost of joining a rebellion. Low per-capita income thus facilitates conflict by making rebellion cheaper. Fearon and Laitin (2003), on the other hand, while finding a similar relationship between GDP per-capita and civil wars, claim that low GDP percapita is related to weak states, which in turn attract insurgency: Financially, organizationally, and politically weak central governments render insurgency more feasible and attractive due to weak local policing or inept and corrupt counterinsurgency practices (pp ). The present paper examines empirically the extent to which Lipset's (1960) classic account of the conditions which promote tolerant and democratic tendencies, can help explain the link between poverty and civil wars. 4 In a nutshell, Lipset argued that due to rough upbringing, low education, social isolation and economic insecurity, the poor are more likely to have authoritarian attitudes, to seek simplistic solutions to complex problems and to attempt to resolve differences through the use of force rather than through negotiations. 5 If poor people have a militaristic bias, in that they are overconfident in the ability of their side to win a conflict militarily, and if waging and sustaining military campaigns require popular support, then resorting to force would be more likely when the disputing parties are poorer. To clarify this argument, it is useful to recall Hirshleifer's (1995) distinction between the basic sources of conflict. The two panels of Figure 1 are reproduced from his paper. The axes I B and I R represent Blue's and Red's incomes, but we may think of them as encompassing other issues under dispute (e.g. the allocation of sovereignty over a territory). The curve QQ bounds the settlement opportunity set - what the parties can jointly attain by peaceful agreement. The points P B and P R indicate, in contrast, the 4 Lipset's thesis as it relates to the emergence of democratic institutions has recently found some support in the literature on institutions and economic growth. See Glaeser et al. (2004). 2

3 parties' separate perceptions of the income distribution that would result in case of war. U B and U R are Blue's and Red's indifference curves. The shaded area is then the Potential Settlement Region (PSR). The larger this region, the more probable it is that an agreement will be reached. Figure 1 points to three sources of conflict. The first is opportunities, captured by the shape of the settlement opportunities set (and its distance from the actual expected income distribution in case of war). Panel 1a for example indicates large gains from a peaceful settlement. The second is preferences. Altruistic preferences, as in panel 1a, tend to increase the PSR, while malevolent preferences, as in panel 1b tend to shrink it. The third source of conflict is perceptions. In panel 1a the perceived incomes in the event of war are relatively small, and also agreed (P B and P R coincide). In panel 1b, the two sides have divergent and optimistic perceptions of the outcome of war, such that each believes he will do relatively better. This shrinks the Potential Settlement Region, and possibly eliminates it completely. The opportunity-cost and weak-states explanations of the link between GDP per capita and civil wars, essentially focus on the opportunities presented by war and peace. It is noteworthy that they also concentrate on the calculations of the insurgent side of the conflict and less on those of the government side. In a separate work I look at the parties' preferences, as they relate to social identities (in particular ethnic identities). As the framework for modeling social identity developed in Shayo (2005) suggests, in a dispute between ethnic groups, having an ethnic identity would tend to shrink the PSR since it would entail a concern for the relative status of one's group, and hence upward sloping indifference curves. This paper, however, focuses on perceptions. In particular, it argues that poverty is related to overly optimistic perceptions of one's party's military capabilities. Thus, the expected payoff perceived by say, Blue, is increased (or the expected losses perceived decreased): P B in Figure 1 shifts to the right and below the true expected income distribution in case of war. 6 Even if preferences and opportunities were fixed, such shifts 5 For a review of the sociological literature that examines Lipset's hypothesis see Houtman (2003). 6 To take the simplest example, normalize the utility from the issue under dispute to 1, let q be the probability of Blue winning a war and assume that war leads to a loss of and that the winner takes all. In that case if Blue holds the true probability then: 3

4 would diminish if not eliminate completely the scope for a peaceful settlement. This argument thus complements rather than competes with the other explanations offered in the literature. 7 Ideally, to empirically test this mechanism, one would use data on perceptions of the likelihood of military success by both rebel (or potentially rebellious) groups and the majority population. While data on the perceptions of rebel groups are hard to obtain, national surveys measuring the popular confidence in the armed forces in a comparable way are available from a broad range of countries, and allow for an exploratory test of the plausibility of the above argument. That is, we are able to check whether the views expressed by one side of potential conflicts the majority population are consistent with the hypothesis that poorer people are more likely to have higher confidence in their army and whether the national levels of such confidence are related to the risk of civil wars. To some extent this approach complements predominant accounts of civil wars, in that it focuses on the government side of the conflict and the population from which it draws its support rather than on the rebel side of the conflict. This paper thus augments the standard cross country analysis of the risk of civil wars in two ways. First, we include in the cross country analysis itself measures of popular confidence in the army, drawn from the World Values Survey. Second, we examine the determinants of confidence in the army at the individual level in a broad range of countries. The use of these surveys relies on the premise that the sample universe largely excludes the potentially rebellious groups in each country so that the samples are for the most part drawn from the majority population from which the government draws its support. P B E I B war,e I R war q,1 q. But if the perception of q is biased upward, P B shifts down and to the right. 7 Note however that according to the proposed explanation, poverty need not directly encourage individuals to opt for violent political actions themselves, say by lowering opportunity costs. Indeed recent research into the determinants of participation in militant and terrorist activities Hezbollah militant activities in Lebanon, attacks by Israeli Jewish settlers on Palestinians in the West Bank, suicide bombings in Israel by Palestinians has found no evidence that having a lower living standard or a lower education level is positively associated with participation in militant and terrorist activities (Krueger and Maleckova 2002, Berrebi 2003). Yet it has been repeatedly demonstrated in Europe and elsewhere, that lower social strata (poorly educated, manual workers or unemployed) are more likely to vote for extreme right parties and express hostility towards immigrants, favorable in-group attitudes and authoritarian attitudes (see e.g. Kitschelt 1996, Ignazi 2003, Lubbers et al and Shayo, 2005). This paper offers a link between these two observations and the observed correlation between GDP per capita and civil wars. 4

5 The main findings are as follows. First, at the cross country level, the extent of popular confidence in the army is shown to be strongly related to the incidence of civil wars in subsequent years, and to be partly responsible for the relationship between the latter and GDP per-capita. Such effects are not observed with respect to confidence in other measures of confidence in state institutions, which helps alleviate concerns about reverse causality. Next, we find that poorer individuals are more likely to have high confidence in the army in a large and diverse set of countries rich and poor, peaceful and belligerent. Further examination of various mechanisms underlying this relationship points at low educational attainment as the chief culprit. This is consistent with the sociological research literature on the determinants of authoritarianism, intolerance of nonconformity, and racial prejudice (Houtman 2003). We find no evidence for effects of economic insecurity as captured by unemployment, and weak evidence for an effect of social isolation as captured by living in small towns or villages. Taking these results back to the cross country level we find that the stock of national educational attainment is strongly related to the risk of civil wars, and that including it in the econometric model significantly diminishes the estimated effect of GDP per capita. The rest of the paper proceeds as follows. Section 2 replicates cross country results on the correlates of civil war, and then looks at the effect of including a measure of confidence in the armed forces in the analysis. It then uses micro data from 40 national surveys during the 1990's to examine whether the relationship between income and confidence in the army holds at the individual level within each country, and is not entirely due to aggregate country effects. Section 3 explores some of the possible mechanisms for the relationship between militarism and income. Section 4 then revisits the cross country analyses of civil war onset and civil war incidence to examine the effects of education and urbanization. Section 5 briefly concludes. 2. Income, confidence in the army and civil wars This section examines whether popular levels of confidence in the army may account for part of the association between per capita income and the risk of civil wars. We first look at the cross country patterns and then turn to examine whether confidence in the army is 5

6 consistently related to income at the individual level. 2.1 Cross-country patterns Figure 2 shows the relationship between prevalence of confidence in the armed forces and GDP per capita. The level of confidence is taken from the second and third waves of the World Values Survey (Inglehart et al. 2000, henceforth WVS), performed in the early and mid 1990's, respectively. 8 The figure suggests that some of the observed correlation between GDP per-capita and civil wars might indeed be due to the former picking up the effect of confidence in the armed forces. This section explores that possibility, using Fearon and Laitin (2003) as the point of reference. Data The data to be used are essentially those used in Fearon and Laitin, 2003 (Henceforth FL), augmented by measures of confidence in the armed forces taken form the WVS. FL code as Civil Wars (FL) conflicts between a state and organized, nonstate groups who sought to take control of a government or of a region, or to use violence to change government policies; where the conflict killed at least 1,000 over its course, with a yearly average of at least 100; and where at least 100 were killed on both sides (see appendix for details). This coding scheme results in the following countries and year(s) of survey being coded as experiencing civil war within the five years following the year at which the survey was conducted: Bangladesh (96), Britain (90), China (90), Colombia (97), India (90, 96), Pakistan (96), Philippines (96), Russia (90, 95), South Africa (90), Turkey (90, 96). 9 It is important to note that Sub-Sahara-African civil wars, which account for many of the civil wars troubling the world today, are virtually absent from the analysis to be performed here as Nigeria, Ghana and South-Africa are the only SSA countries covered. Confidence in the army is measured by the proportion of the population professing the highest level of confidence in the armed forces, taken from the WVS. Detailed information on the variables, the coding methods and the countries participating in the 8 Confidence in the army is measured on a scale of 1 to 4, where 4 represents a great deal of confidence; 3 quite a lot of confidence; 2 not very much confidence and 1 none at all. The Figure shows the proportion of the population professing the highest level of confidence. 9 The analysis is hardly affected by using only one subsequent year. Indeed the only case where civil war occured more than one year after the survey is Russia

7 analysis is given in the data appendix. Table 1 presents descriptive statistics. For comparison purposes, it also presents the same statistics from the complete FL data set at comparable years. As can readily be seen, the sample to be used here is biased towards richer and more democratic countries. It also has a lower proportion of civil wars. Again, the bias is primarily due to the under-sampling of African countries. Results We use a linear probability model (LPM), estimated by OLS. 10 The dependant variable is the incidence of civil war, namely the probability of observing either a new civil war or the continuation of an ongoing war or both (Elbadawi and Sambanis, 2002, Miguel et al., 2004). The regression error terms are allowed to be heteroskedastic and correlated across years within countries. As in most cross-country regressions, a causal link between confidence in the army (or other covariates) and civil war risk cannot be established based on these regressions alone. An obvious reason is that in countries experiencing war, support for the army may tend to rise (a rally around the flag effect), which can be reflected in the reported confidence in the army. 11 If this is the case, it would be very hard to separate such effects from the effect these attitudes in turn have on the eruption or perpetuation of the war. We address this issue first by using explanatory variables at year t=0 to explain civil war incidence in the ensuing five years t 1,2,...5. This approach obviously cannot entirely solve the problem since a rally around the flag effect may conceivably take place when people are expecting a war. 12 We hence compare the estimated effect of confidence in the army to the effects of popular confidence in other state institutions such as the central government, the police and the legal system which presumably should also 10 Maximum likelihood estimation of probit or logit models cannot identify several specifications, since some outcomes are completely determined by the various dummy variables. A second issue is that our specification probably does not include all the relevant explantory variables for the incidence of civil wars. A problem with logit and probit models in such cases is that even when the omitted variable in uncorrelated with the included, the coefficient on the included variable will be inconsistent. Finally, LPM has the benefit of readily interpretable coefficients. Where probit models could be identified (e.g. for column (1) in Table 2), the estimated marginal effects at the means were for the most part almost identical to the LPM estimates. 11 As emphasized by Bueno de Mesquita and Dickson (2004), a similar effect can also arise in the rebel population. 7

8 benefit from the same rally around the flag effect. Still, absent valid instruments for confidence in the army (i.e. instruments that satisfy the exclusion restriction), our crosscountry analysis cannot establish causality. It can however tell us whether confidence in the army may account for some of the previously observed relationship between poverty and civil war risk. It should also be kept in mind that the sample used in this section is much smaller than the samples typically used in the literature on determinants of civil war, which contain data on most of the post WWII period, and more than 150 countries. 13 Using WVS data, on the other hand, reduces these to only two periods with 30 to 40 (often different) countries in each. In section 4 we shall examine the large sample results in light of the individual level findings. The results are presented in Table 2. The baseline specification (columns 1-2) is built upon that used in FL. 14 The sample of countries with both WVS and FL data available consists of 71 country-years. To assess the comparability of the results obtained from this small sample to the larger-sample results in the literature, the first column shows the results of estimating the baseline specification using the entire FL sample ( ). Comparing the first two columns reveals a qualitatively similar pattern despite the much smaller sample in the second. The estimated effects are all in the same direction and for the most part have the expected sign (see Fearon and Laitin, 2003 for a detailed discussion of the interpretation of these effects). There are however rather large differences in the size of the estimated effects of GDP and population size. The results in column (2) are also in line with the results reported by FL, (although the dependent variable there is civil war onset). Most variables have similar qualitative effects. GDP 12 Similarly, this technique does little to establish causality of GDP per capita, which is affected by expectations (e.g. due to investment decisions anticipating war a peace. 13 Thus, Collier and Hoeffler. (2001), working with five-year intervals, use between 600 to 800 data points (country-year) in most regressions. Fearon and Laitin (2003), employing annual data, use between 5100 to 6400 data points. 14 Following Hegre et al. (2001) we also include a quadratic term for the democracy variable. On the other hand three variables are dropped from the analysis. The dummy used by FL for new states does not vary in this sample and hence is dropped. Including a dummy for prior war doesn't make much sense when the dependent variable is not the start of a new war. Finally, the dummy variable for political instability in period 0 is dropped due to the risk of endogeneity. Including this last variable however does not alter the results. The specification used here may be objected to because there is no accounting for the possible endogeneity of per capita GDP other than using lagged independent variables. But again, the main purpose 8

9 per-capita has a negative effect and population size a positive effect. Rough terrain captured by percentage of the country that is mountainous seems to have a weak positive effect on civil war incidence. 15 Being an oil exporter also seems to have a weak positive effect, consistent with Collier and Hoeffler's (2001) results. As in FL, state contiguity and levels of ethnic and religious fractionalizations are not very strongly related to the risk of civil wars, when other features in particular income are controlled for. Consistent with Hegre et al. (2001), democracy is estimated to have a hump-shape effect: semi-democracies (regimes intermediate between democracy and autocracy) have higher propensity for conflict than either extreme. Column (3) adds the variable measuring the proportion of the population expressing a great deal of confidence in the armed forces, while keeping the sample fixed. The association is positive, large and statistically significant. If we were to interpret it causally, it would mean that a percentage point increase in the population with high confidence in the army gets translated to a percentage point increase in the risk of civil war, other things equal. Note also the substantial improvement in the fit of the regression: after including the confidence in the army variable, the R 2 increases from 0.38 to The point to emphasize, however, is that the marginal effect of log GDP per-capita is cut significantly from to -0.11, and loses statistical significance. This result is consistent with the claim that part of the observed relation between income per-capita and civil war is due to the correlation between income and militarism. 16 Columns 4-5 address the possibility that the observed positive correlation is due to a rally around the flag effect. They also help us examine whether the above results capture the effect of a possible militaristic bias reflected in overconfidence in the army or of confidence in institutions in general. Column 4 replaces confidence in the army with a similar measure of confidence in the police. The estimated effect is not significantly different from zero, and it does not seem to account for any of the relationship between here is to examine whether part of the relationship between GDP and civil wars can be related to confidence in the army, not the identification of the causal effect of GDP. 15 FL interpret the rough terrain effect as capturing the ability of rebels "to hide from government forces". An alternative expalnation would have to do with the extent of trade within the country. On international trade and interstate war see Martin et al. (2005) 16 One may also note a similar outcome with respect to the coefficient on population size. Confidence in the army is generally higher in bigger countries, and omitting it from the analysis results in a higher estimated effect of population size. 9

10 civil war risk and GDP per capita. Column 5 reveals a similar result with respect to confidence in the civil service. We repeated this exercise with similar measures taken from the WVS for confidence in the churches, the press, the legal system, and major companies (results not shown). 17 None of these variables had a similarly positive effect on civil war incidence, and none undermined the effect of GDP per capita. Finally, column 6 checks the possibility that the effect of confidence in the army is related to the army's prominence as an employer. People who serve or know other people that serve in the armed forces, may be affected by that fact when stating their attitude toward the army. 18 Thus, the estimated effect of confidence in the army on civil wars may be picking up the effect of the size of the army relative to the labor force. We find no evidence of such a channel. Including the size of the military as percentage of the labor force in the regression explaining civil war incidence reveals no effect of relative army size on civil war risk in the sample with the requisite data (68 countries). 19 Further, controlling for relative size of the military does not change the estimated effects of GDP per capita nor of confidence in the army. 2.2 Poverty and militarism at the individual level Overall, the cross sectional patterns are consistent with the claim that part of the relationship between GDP per-capita and civil wars is due to popular confidence in the army in poor countries. In the remainder of this section we examine the association between income and confidence in the army using micro data. In particular we wish to ask whether it is simply the case that in rich countries people tend to have relatively 17 Data on confidence in the central government is also available in some surveys, for a sample of 51 country-years. The effect of GDP is insignificant in this sample to begin with. Including a measure of confidence in the government yields a marginally significant positive effect (p=0.09). The estimated effect of GDP is only made more negative in this regression. 18 Across countries with avilable data on the size of the military (data from WDI, based on U.S. Department of State, Bureau of Verification and Compliance, World Military Expenditures and Arms Transfers) there is indeed a positive correlation between the relative size of the army and overall levels of confidence in the army. When controlling for GDP per capita, the effect of relative army size on confidence in the army is insignificantly different from zero. 19 Repeating the regression in column 1 (using all country-years with available data) also revealed no effect of relative army size on the risk of civil war. Including measures of military personnel as percentage of the labor force in the original Fearon and Laitin (2003) specification, also revealed no effect of relative army 10

11 lower confidence in the army then in poor countries which may be due to country-level factors that affect both GDP per capita and attitudes or whether the relationship also holds at the individual level, in rich and in poor countries alike. Discussion of the possible mechanisms leading from poverty to confidence in the army is postponed to the next section. Data Data come from the WVS, second and third waves. Household income data are available for 40 national surveys conducted in a diverse set of 33 countries. Details on the procedures used to estimate household income and household-size are given in the data appendix. Table 3 presents some summary statistics. Sample sizes vary from below 600 in Finland to over 3000 in Spain 1996, with an average of 1250 observations per sample. Median confidence in the army ranges from 2 (not very much) to 4 (a great deal) and the proportion with the highest level of confidence in the army ranges from below 5% in Japan, the Baltic states and the Netherlands to over 60% in Turkey. Confidence in the army normally has within-country standard deviation of around 0.8. Data on household income is in local currency. In 1996 PPP dollars, income per household member ranges in these data from $300 to $40,000, and the country-means range from $1300 to $17, Years of schooling are calculated by subtracting 7 years from the age at which the respondent left school. The resulting national means of years of schooling in these surveys are within 2 years from the Barro and Lee (2000) educational attainment data in all countries except in Chile 90 and Venezuela where the mean calculated years of schooling are significantly higher. Finally, the average age in these samples is normally around 45 years with standard deviation of 17, and the samples are usually balanced between the sexes, except for Austria, Latvia 1990, India and the Netherlands that overrepresent women. Results We present both ordered probit and OLS results. The first technique seeks to use all the information available in the confidence-in-the-army question, which has four possible size on the likelihood of civil war onset (the point estimate is negative and insignificant). There is also no appreciable change in the estimated effect of GDP per capita. 20 This excludes East European countries in 1990, for which no reliable PPP rates are available. 11

12 answers, ordered from none at all to a great deal of confidence. One problem with this approach is that it is difficult to interpret and compare the coefficients obtained from different regressions run on different surveys. Also, the results are not directly related to those presented in the previous section, where the measure of confidence we used was the frequency of the highest level of support for the army. To address these issues, a linear probability model is also used, with an indicator variable for the highest level of confidence in the armed forces as dependant variable. The results are summarized in Table 4. Column (1) shows, for each survey, the estimated ordered probit coefficient on log household income controlling for log of household size. The point estimates are mostly negative, suggesting a negative association of household income with the level of confidence in the army. Note that there is no clear difference in this relationship between richer and poorer countries. There appears to be a strong negative relationship between income and confidence in the army in Italy and Brazil, Bulgaria and Canada. Nor is the effect limited to a particular region or culture - the relationship is strong in Austria, Taiwan, Turkey, Russia and Spain. One should note however that roughly half of the estimated coefficients are not statistically different from zero, and a few of these actually have a positive point estimate. Some of these results may perhaps be attributed to the small samples (e.g. Finland) but there are probably other factors at work as well. In India for example, although confidence in the army is generally very high, most of the variation in confidence in the army in the 1990 survey seems to come from regional differences and not from differences in income or education. 21 Column (2) presents the OLS estimated coefficients. The overall picture is similar 33 of the 40 point estimates are negative, with 22 significantly different from zero at 90% confidence or more. To gain some feeling for the size of the effect, take the average effect of This coefficient roughly means that an increase of 1% in household income (keeping household size fixed) is associated with a 0.35 percentage points decrease in the probability that the respondent would have the highest level of confidence in the army. Consider the cross-country relationship between log income per capita and 21 Residents of the north region have a significantly higher level of confidence in the army then do residents in the south, while residents of the east and west regions are in between. 12

13 the proportion of the national population with highest level of confidence presented in Figure 2. The OLS estimated effect of income per capita on the proportion with highest confidence is That is, the estimated effect of household income on confidence in the army is over a third of the effect observed at the cross-country level. This leaves plenty of room for other, national factors to simultaneously affect both average income and average support for the army. But the association at the individual level suggests that the cross country association is not all due to such factors. We now turn to examining more closely what might drive this relationship. 3. Possible mechanisms In his account of working class authoritarianism, Lipset (1960, ch. IV) points to several possible factors driving the association between low class and authoritarian attitudes. Most prominent are: 1) Harsh upbringing and authoritarian family patterns. 2) Low education, promoting a simplified view of politics and a failure to understand the rationale underlying tolerance and compromise with people one disagrees with. 3) Economic insecurity that leads to high states of tension and the search for immediate solutions. 4) Isolation from the activities and controversies of the society at large that prevents the lower strata from acquiring the complex view of the political structure which makes understandable and necessary the norms of tolerance. Subsequent sociological research on Lipset's Thesis has reached a consensus that authoritarianism, intolerance of nonconformity, and racial prejudice are closely related to poor education (see Houtman, 2003 for a recent survey). The other mechanisms proposed by Lipset remain, more controversial. However, there does seem to be evidence for the effect of harsh upbringing on criminal behavior (see Donohue and Levitt, 2001 for discussion). In this section we attempt to examine whether these mechanisms can help explain the link between income and confidence in the army (On the link between confidence in the army and authoritarian attitudes see the analyses in Fleishman 1988 and Kitschelt 1996). The analysis is performed country by country, using the data described in the previous section and in Table 3. 13

14 We have no good measure of childhood experiences or family patterns. The one available variable that may be related to this mechanism is household size. 22 The relationship however is very indirect and requires non-trivial though perhaps not unreasonable assumptions. Thus we need to assume that there is strong association between current household size and the size of the family one grew up in, and that larger household size implies that children receive less caring and attention (for instance due to parents allocating their time among more children). To measure education we use the estimate of years of schooling described above. 23 Economic insecurity is proxied by a dummy variable for current unemployment. Finally, to proxy isolation from the larger society we use the size or the town where the interview was conducted. The results presented here are from OLS regressions similar to those reported in column (2) of Table 4. This is done primarily for ease of interpretation. Using ordered probits yields qualitatively similar results. Figures examine the association between log household income and confidence in the army, and the extent to which the estimated effects of the former are modified by introducing any of the four other factors into the analysis (in those surveys where the required data exist). Figures look at the effects of these other factors on confidence in the army, controlling for income. The figures show, for each national survey, the 95% confidence interval for the coefficient of interest and the point estimates. In all regressions, the dependent variable is a dummy variable for having the highest level ( a great deal ) of confidence in the armed forces. Confidence intervals are calculated using robust standard errors. Estimations correct for the individual level sampling weights provided by the WVS. Consider first Figure 3.1 which shows the "total effect" of income without controlling for any other variable. The association with confidence in the army seems overwhelmingly negative in fact we cannot reject the hypothesis that the coefficients 22 Thus, with respect to delinquency, Donohue and Levitt (2001, p. 393) state: If women used abortion to prevent increases in family size, then abortion may indirectly lower criminality for the remaining children who will receive greater per child contributions of parental resources [Becker 1981;Barber,Axinn,and Thornton 1999].Sampson and Laub [1993,p.81]and Rasanen et al.[1999] fnd that family size signifcantly increases delinquency. 23 This is simply the age at which the respondent left school minus 7. Since we are only using the within country variation in this education measure, and will not be using it for cross country comparisons, country specific factors that affect e.g. the age people start school should not matter for estimating the effect of schooling. 14

15 are negative in all countries. There is no clear common characteristic to the countries where a zero or positive association cannot be rejected. They may be advanced or developing countries (Sweden, India), Asian, East European or Western European (Japan, The Baltic states, Germany), countries that experienced civil wars or peaceful (India, Switzerland). All of the estimated coefficients that are significantly different from zero at the 5% level 25 in total are negative. Figure 3.2 shows the results when controlling for household size. The point estimates of the effect of income on confidence in the army are still overwhelmingly negative (32 of the 39 national surveys with additional data on household size), although some lose statistical significance. The effect of household size itself is presented in Figure 4.1. There is no apparent general pattern. Most of the estimated coefficients are insignificantly different from zero, and the point estimates are if anything, negative more often then positive. We thus find no evidence that larger families are in general conducive to militarism, but of course this is hardly an optimal measure of the kind of rough upbringing and lack of nurturing that Lipset refers to. We now come to the effect of education, measured by years of schooling. According to the sociological literature, lower education is a major contributor to authoritarian predispositions among the poor. As income is usually correlated with schooling, we might therefore suspect that part of the effect observed in figure 3.1 is actually due to income picking up the effect of education. Consistent with this expectation, in all surveys except one (India 90), the estimated effect of income becomes less negative when we control for years of schooling. Indeed, as Figure 3.3 makes clear, several point estimates become positive after we control for education, and the effect of income on confidence in the army becomes insignificantly different from zero in over half the surveys. Figure 4.2 shows the association of schooling with confidence in the army. The point estimates are negative in all but one country (India), confirming the results from the sociology literature. The negative effect is statistically significant at the 5% level in just over half the surveys. The average size of the effect is , suggesting that given income, one additional year of schooling is associated with a 1 percentage point drop in the likelihood of having the highest confidence in the army. Based on the results in the previous section (column 3 of Table 2), a "back of the envelope" calculation therefore 15

16 suggests that the effect of an additional year of schooling in the adult population on civil war incidence (through the confidence in the army channel alone) should be to reduce the risk of civil war incidence by one percentage point. Next, consider the effect of unemployment. According to Lipset's thesis, economic insecurity is conducive to simplistic, authoritarian and intolerant attitudes. To the extent that current unemployment captures a history of economic insecurity, we should then expect a positive effect of unemployment on confidence in the army. We find no evidence supporting this prediction. As can readily be seen from figure 4.3, the estimated effects of unemployment (relative to full employment) show no consistent pattern and the effect is statistically insignificant in almost all of the 38 surveys with the requisite data. Finally, to the extent that the size of the town one lives in captures isolation from the rest of society, town-size should have a negative effect on militarism. The results (Figure 4.4) are only partly supportive of this notion. Point estimates are negative in over two thirds of the surveys, though they are statistically significant at the 5% level in only about a third. As figure 3.5 shows, town-size does not seem to account for the relation between income and confidence in the army. To sum up, consistent with existing research on authoritarian and non-tolerant attitudes, 24 education seems to be strongly related to militaristic attitudes. Education also seems to be a major factor underlying the association between militaristic attitudes and income. Unemployment and household size, on the other hand, show no general patterns. Finally, residing in a larger town seems somewhat conducive to less militaristic attitudes, consistent with an isolation explanation. This channel however does not seem to be responsible to any significant degree for the relation between income and confidence in the army. Using these results, we can now revisit the cross country results. If education is indeed responsible to a significant degree for the relationship between income and militarism and if militarism is part of the reason why poor countries are more likely to have civil wars, then we should find that, in reduced form, part of the effect of income per capita on civil wars is due to education levels. We now turn to this issue. 24 See e.g. Altemeyer 1988, 1996; Kitschelt 1996; Schuman, Steeh, Bobo, and Krysan,

17 4. Education, Urbanization and Civil Wars Micro data from a large set of countries provide strong support for the claim that low education is an important determinant of militaristic attitudes. Weaker support is found for the notion that social isolation, at least to the extent that it is captured by urbanization, also tends to enhance a militaristic outlook. In this section we thus revisit the relationship between GDP per capita and both the onset and the incidence of civil wars, augmenting the analysis with aggregate measures of education and urbanization. These are of course reduced form estimations: both education and urbanization are probably related to the risk of civil wars through additional channels. Aggregate education data are taken from Barro and Lee (2000), who provide (updated) estimates of the stock of education in the population, namely the average years of schooling attained by the population aged over 15. The Barro and Lee data are available for a large set of countries at five-year intervals from 1960 to For specifications that require data for intermediate years we use Linear interpolations. The results are qualitatively similar when using literacy rates (UNESCO, available from the World Development Indicators database), but the latter are available for a much smaller set of countries and years. For urbanization we use the share of the total population living in areas defined as urban in each country, available from the World Development Indicators database. Table 5 presents the results using the Fearon and Laitin (2003) baseline specification for explaining civil war onset. We maintain the same specification in terms of the variables used, adding only a quadratic term for democracy (following Hegre et al. 2001). However we estimate it using OLS rather than logit for the reasons discussed in section 2.1 above. We also allow the error term to be correlated across years for the same country. 25 The first column replicates the FL result. The next two columns restrict the sample to the 25 Using logit without clustering as in FL yields similar qualitative results in the sense that educational attainment and urbanization are estimated to have a negative effect and the estimated effect of GDP per capita is substantially reduced when controlling for either variable. However, significance levels of the effects are often different. It is highly plausible that a country with a higher than average likelihood of conflict in one year also has a higher than average likelihood of conflict in the next year (for example if we omit an important country characteristic from the analysis). Not correcting for this may produce highly biased standard errors. 17

18 country-years with available educational attainment data. The results of the baseline specification using this sample (column 2) are very similar to the results using the full sample. In particular, GDP per capita is estimated to have a negative and significant effect: a $1000 increase in GDP per capita is estimated to reduce the risk of civil war onset by 0.14 percentage points. In column 3 we add average educational attainment in the adult population. The estimated effect is negative and highly statistically significant. An increase of one year in the average schooling of the population is estimated to reduce the risk of civil war onset by one third of a percentage point. These results are also consistent with Collier and Hoeffler's (2001) finding of a significant relationship between war onset and secondary education enrollment. Further, the effect of GDP per-capita after controlling for education is no longer significantly different from zero. This corroborates the claim that part of the relationship between poverty and civil war risk has to do with educational attainment not reflected in per capita income. Columns 4 and 5 look at the effect of urbanization without controlling for education. Consistent with our predictions, more urban societies are estimated to be less likely to start a civil war. As with education, controlling for urbanization results in a sharp reduction of the estimated effect of GDP per capita. Finally, column 6 controls for both schooling and urbanization. The results suggest a significant, negative effect of education, while the estimated effects of both GDP per capita and urbanization, while negative, are small and not statistically different from zero. Table 6 looks at the incidence of civil wars, using the specification from the first column of Table 2. Average schooling is estimated to have a strong negative effect on the likelihood of civil war incidence, with and without control for urbanization (columns 6 and 3, respectively). Further, the estimated effect of GDP per capita is more than halved after controlling for education and is statistically not different from zero. Urbanization on the other hand has a zero effect in this specification, even without controlling for education. The size of the effect of schooling is in both colums 3 and 6. This effect is much larger (3.6 times) than the effect of education via militarism that our rough "back of the envelope" calculation suggested. This confirms the view that the stock of educational attainment probably affects civil war incidence in other ways. 18

19 Overall, the results from the cross country analysis are consistent with the individual level analysis: education has a strong and robust effect on the likelihood of civil wars. The effect of urbanization is not robust - it vanishes when controlling for education or when looking at war incidence rather than onset. Finally, it does not appear to be the case that the effect of militarism is only picking up some other effect of education on civil war incidence: controlling for education does not detract from the estimated effect of popular confidence in the army. Conclusion Drawing on Lipset's (1960) account of the relation between economic conditions and democratic predispositions, this paper explored a new explanation for the link between poverty and civil wars. The explanation complements existing accounts in that it focuses on the perceptions of the majority population, rather than on the opportunity set facing the rebels. Data from a large set of countries support the claim that poverty is strongly related to confidence in the army at both the individual and the aggregate level. Confidence in the army, in turn, is found to be positively related to the risk of civil wars, and can account for part of the effect of national income per capita (though we did not prove causality). Several possible mechanisms underlying the relationship between income and confidence in the army have been explored. The most important factor identified is educational attainment. Educational attainment is in turn found to be strongly related to the risk of civil wars. References Altemeyer, Bob, 1988, Enemies of freedom: Understanding Right-wing authoritarianism, San Francisco: Jossey-Bass. Altemeyer, Bob, 1996, The Authoritarian Specter, Harvard University Press. Barro, Robert J. and Jong-Wha Lee., 2000, International Data on Educational Attainment: Updates and Implications Working Paper 42. Center for International Development. Becker, Gary S Crime and Punishment: An Economic Approach, The Journal of Political Economy, 76(2), Mar. - Apr., 1968, pp Benabou, Roland, 1996, Inequality and Growth NBER Macroeconomics Annual, 1996, B. Bernanke and J. Rotemberg, eds., Berrebi, Claude, 2003, Evidence About The Link Between Education, Poverty and Terrorism Among Palestinians, mimeo, Princeton University. Bueno de Mesquita, Ethan and Dickson, Eric, 2004, The Propaganda of the Deed: Terror 19