Country Fixed Effects and Unit Roots: A Comment on Poverty and Civil War: Revisiting the Evidence

Transcription

1 The University of Adelaide School of Economics Research Paper No March 2011 Country Fixed Effects and Unit Roots: A Comment on Poverty and Civil War: Revisiting the Evidence Markus Bruckner

2 Country Fixed Effects and Unit Roots: A Comment on Poverty and Civil War: Revisiting the Evidence by Markus Brückner* February 2011 Abstract: Djankov and Reynal Querol (2010, RESTAT) show that the level of GDP per capita has no significant effects on the risk of civil war once country fixed effects are accounted for. Therefore, they argue that the relationship between income and civil war is spurious. This paper shows that when focus is on the change, rather than on the level, of GDP per capita that the significant negative relationship between GDP per capita and an indicator variable for civil war is recovered in the country fixed effects regression. In contrast to the argument made in Djankov and Reynal Querol, the paper's findings do not support the claim that the relationship between GDP per capita and civil war is spurious due to timeinvariant omitted variables. Key words: Income, Civil War, Unbalanced Regression JEL codes: 010, O40, C23 * Department of Economics, University of Adelaide. Contact markus.bruckner@adelaide.edu.au.

3 1. Introduction There is a fierce policy debate on how, and if at all, changes in income per capita affect the likelihood of civil war (e.g. World Bank, 2003). Civil wars have killed and maimed millions of people, and they are particularly frequent in the world's poorest countries. The strong negative cross-country correlation between income per capita and the incidence of civil war has led early researchers to conclude that per capita income is a highly significant determinant of the incidence of civil war. 1 Djankov and Reynal-Querol (2010) revisit this evidence on the link between income per capita and civil war, using panel fixed effects analysis. Their main finding is that once country fixed effects are accounted for per capita income does not significantly predict the incidence of civil war. The authors conclude that increasing per capita income levels does not systematically reduce the likelihood of civil war. Taken at face value, their findings (and title) suggest that reducing poverty will not effectively reduce the incidence of war. 2 A natural question that Djankov and Reynal-Querol did not report estimates on in their paper is how changes in per capita income affect civil war. This is a natural and highly relevant question for policy makers, as policy makers are often concerned about how changes in the level of per capita income affect socio-political outcomes. This paper shows that when focus is on the change in GDP per capita rather than on the level that the significant negative relationship between civil war and per capita income is recovered in the country fixed effects regressions emphasized by DRQ. Using the DRQ data, estimates that are based on the change in GDP per capita yield that on average a one percent (permanent) increase in GDP per capita was associated with a significant decrease in the incidence of civil war by about 0.3 percentage points. These estimates are robust to using alternative civil war datasets; using the latest GDP per capita and civil war data to update and 1 See e.g. Fearon and Laitin (2003) and Collier and Hoeffler (2004). For instrumental variables studies that show that exogenous shocks to per capita income are a significant determinant of Sub-Saharan civil wars see Miguel et al. (2004) and Bruckner and Ciccone (2010). An overview of the civil war literature is provided by Blattman and Miguel (2010). 2 See also their Voxeu publication, available online at where they write: "These results indicate that policies that are directed to increase per capita income will not have any effect in reducing the probability of civil war." 1

4 extend the DRQ sample; using a conditional logit fixed effects model to address the incidental parameter problem associated with fixed effects in nonlinear probability models; using a distributed-lag model that allows to distinguish short-run from medium/long-run effects; using a dynamic panel data model that accounts for dynamics in the incidence of civil war; or using instrumental variables estimation that corrects for potential endogeneity bias. There is also a serious econometric reason for why using the change in income per capita, rather than the level, may be preferable in a country fixed effects regression: country-by-country unit root tests indicate that in over 90 percent of the countries per capita GDP follows a random walk. Civil war incidence is a bounded variable on the unit interval, and thus is stationary by construction. If GDP per capita follows a random walk, it will imply an unbalanced regression that should be avoided. Focusing in a country fixed effects regression on changes in per capita GDP is from an econometric point of view a more prudent approach as for most countries the level of GDP per capita displays extreme persistence and formal unit root tests do not reject the hypothesis of a unit root in the level of GDP per capita (while they reject for nearly all countries a unit root in the change). 3 The findings reported in this paper are also important from an economic policy point of view. In particular, the findings suggest that DRQ missed out on reporting in their paper additional results that bear an entirely different policy message than what is implied by their estimates and conclusion. While a significant effect of within-country changes in income per capita on civil war must be interpreted with caution, the conclusion drawn by DRQ based on their country fixed effects analysis (that reductions in poverty do not significantly reduce the incidence of civil war) seems to be unwarranted when taking into account that the time-series of GDP per capita displays extreme persistence and that changes in GDP per capita do yield a significant negative effect on civil war incidence in a country fixed effects regression. 3 Also more powerful panel unit root tests indicate that for a large proportion of the countries in the sample GDP per capita contains a unit root. 2

5 The remainder is organized as follows. Section 2 discusses the estimation strategy. Section 3 presents the main results. Section 4 concludes. 2. Estimation Framework DRQ estimate the following econometric model: War c,t = c t GDP c, t 1 X c,t 1 u c,t where α c are country fixed effects that capture time-invariant cross-country differences (such as for example colonial origin), and γ t are year fixed effects that capture shocks common across countries in a given year (such as the world business cycle or the end of the Cold War). War c,t is an indicator variable that is equal to unity in the event of civil war and zero else. 4 GDP c,t-1 is the log of real per capita GDP, lagged one period to reduce concerns of reverse causality, and X c,t-1 is a vector of country-specific time-varying variables (such as population size and political institutions) also lagged one period. u c,t is a residual that is clustered at the country level. The country fixed effects specification in equation (1) deserves several remarks. First, civil war is a binary variable and thus estimation of equation (1) has to take into account the nonlinear nature of the dependent variable. The presence of fixed effects complicates consistent estimation substantially due to the incidental parameter problem (e.g. Wooldridge, 2002). Consistent estimation of the slope coefficients can be obtained however by using a conditional logit fixed effects model. 5 The second issue with equation (1) is that the estimated coefficient on lagged GDP per capita, θ, captures only the short-run effect. Using lower frequency data (such as e.g. 5-year panels) 4 Summary statistics of the civil war incidence indicator variables that are used for the empirical analysis are provided in Data Appendix Table 1. The empirical analysis focuses on civil war incidence rather than civil war onset since civil war onset is a rare-event variable (e.g. King and Zeng, 2001), which requires special econometric techniques that are (computationally) difficult to implement (due to convergence problems) with country fixed effects. 5 A somewhat undesirable property of the conditional logit model is that it does not allow to compute marginal effects (since this would require knowledge of the distribution of the country fixed effects). A linear probability model may provide a first-order approximation of these marginal effects. Thus, to ensure robustness of the results estimates of the fixed effects equation are reported from both, a linear probability model and a (nonlinear) conditional logit model. 3

6 would allow to focus on more long-run effects. However, this comes at a cost of ignoring important short-run dynamics in the incidence of civil war. Moreover, using panels that are based on 5-year data substantially reduces the time-series dimension (T) of the panel, and thus calls into question the consistent estimation of the country fixed effects. Both, short-run and medium/long-run effects can be examined with annual data by using a distributed lag model: (3) k War c,t = c t i GDP c,t 1 i u c,t i=0 In the above model the θ s capture short-run effects, while their sum captures the medium/long-run effects of a permanent increase in GDP per capita on the likelihood of civil war. 6 A third and perhaps more crucial issue is the time-series property of GDP per capita. Bond et al. (2010) address this issue with great care in their panel data study of the growth effects of investment. They report both country-by-country augmented Dickey Fuller (ADF) tests as well as more powerful panel unit root tests. Their main conclusion was that the level of GDP per capita contains a unit root while the first-difference is stationary. To show that also in the sample of DRQ a similar conclusion can be reached, Table 1A reports the results of country-by-country ADF tests on the null hypothesis that the log of GDP per capita follows a random walk with drift. The main result is that in less than 4 (8) percent of the countries is this hypothesis rejected at the 99 (95) percent confidence level. On the other hand, in over 80 (84) percent of the countries is the hypothesis of non-stationarity in the change of GDP per capita rejected at the 99 (95) percent confidence level. The Im, Pesaran, Shin (2003) panel unit root test resonates the above results. The panel unit root test cannot reject at any conventional confidence level the null hypothesis of a unit root in GDP while it rejects it at the 1 percent level for the first-difference. As an identification check, Tables 1A and 1B report the results of the ADF test of nonstationarity of the civil war indicator. The civil war indicator is a bounded variable and the ADF test rejects in over half of the countries at over 95 percent confidence that the civil war indicator follows 6 The covariates are dropped in the above equation for simplicity. 4

7 a random walk. Resonating the country-by-country ADF results, the Im, Pesaran, and Shin panel unit root test rejects at the 1 percent level that the civil war indicator is a random walk. Hence, the main result of these unit root tests is that the level of GDP per capita is likely to follow a random walk, while the first-difference of GDP and the civil war indicator appear to be stationary. Since the Westerlund (2007) panel cointegration test also does not indicate a cointegration relationship between the level of GDP and civil war (see Table 1c), the preferred specification is to relate civil war to the change in GDP per capita. 7 Testing the null hypothesis that the sum of the coefficients on lagged changes in GDP (over 5-years) is equal to zero therefore amounts to testing whether a (permanent) change in the level of GDP had a non-zero effect on civil war incidence over a (5-year) period. 3. Main Results Table 2 reports the country fixed effects estimates of the relationship between civil war and GDP per capita for the DRQ annual sample. In Panel A estimates are shown where the explanatory variable is the change in real GDP per capita. The main finding in this Panel is that when the change in GDP per capita is used that there is highly significant negative effect on civil war incidence and civil conflict. This is true for both the conditional logit fixed effects model (columns (1) and (3)) and the linear probability model (columns (2) and (4)). Quantitatively, the linear probability model indicates that, approximately, a one percent increase in GDP per capita over the past five years reduced the incidence of civil war (civil conflict) by about 0.4 (0.6) percentage points on average. Panel B of Table 2 shows the results that use the level of GDP per capita as in DRQ. Here the main finding is that the average effect over five years of a (permanent) increase in GDP per capita does not significantly correlate with civil war or civil conflict. While the t-1 effect on GDP 7 Note that this point is different from the point made in Ciccone (2008), which is about transitory economic shocks (such as year-to-year variations in rainfall that are highly mean-reverting). Transitory economic shocks are (by definition) stationary, while the unit root tests indicate that GDP per capita behaves, on average, anything but like a stationary process. 5

8 per capita is significantly negative, some of the other lags are positive and significant, and offset in sum the negative and significant t-1 effect. But, because GDP per capita is highly likely to follow a random walk, this level specification should be avoided in the fixed effects regression. In fact, the offsetting coefficients in Panel B are an indication that the level specification is misspecified. 8 Table 3 shows that similar results are obtained when using DRQ's 5-year data. There is a significant effect of a change in GDP per capita on civil war incidence (Panel A), and an insignificant effect when using the level of GDP per capita. To provide further evidence that the significant negative effect of income per capita on civil war is robust to the control for country fixed effects, Tables 4 and 5 show country fixed effects estimates for the datasets of Fearon and Laitin (2003) and Collier and Hoeffler (2004). Panel A reports the estimates that use the change in GDP per capita while Panel B reports estimates that use the level. Both, Tables 4 and 5 confirm the previous results that changes in income per capita are highly significantly negatively related to the incidence of civil war. The fixed effects estimates yield that on average a one percent increase in GDP per capita over a five-year period reduced the global incidence of civil war by approximately 0.1 percentage points. Further to these robustness checks Table 6 shows that similar results are obtained when extending and updating the DRQ sample with the latest PWT GDP per capita data (Heston et al. 2009) and the latest PRIO/UPSALLA (2010) civil war data. Panel A of Table 6 shows that past changes in GDP have a highly significant negative effect on both civil conflict and civil war incidence. Quantitatively, the cumulative effect of a one percent increase in GDP implies that civil war incidence decreased on average over the period by up to 0.3 percentage points. An important final question is whether the significant negative correlation between past changes in GDP per capita and civil war reflect a causal relationship. The use of lagged GDP should 8 To see this point also intuitively, it is useful to consider the simplest possible model: (i) y t =aδx t +u t ; where x t =x t-1 +e t, and u t and e t are stationary random variables. If one estimates this model using the level of x t, such that, y t =bx t + cx t-1 +z t, then it is not be surprising that a level effects (mis-)specification does not allow to reject that there is no cumulative effect of an increase in the level of x t. This is because from (i) it follows that b=a and c=-a. 6

9 reduce concerns of substantial endogeneity bias that is due to a reverse negative effect of civil war on GDP. To be on the safe side, Table 7 reports instrumental variables estimates that treat GDP per capita as an endogenous regressor. Following Djankov and Reynal-Querol (who base their IV approach on the work of Acemoglu et al., 2008), the instrumental variable for GDP per capita is the lagged savings rate. 9 The main result is that the instrumental variables approach produces a significant effect when using the change in GDP per capita (Panel A), while the specification that uses the level of GDP per capita produces insignificant estimates. Quantitatively the estimates in Panel A imply that a percent increase in GDP per capita reduced the likelihood of civil war (civil conflict) by more than 0.1 (0.2) percentage points in the short run, and by more than 0.2 (0.3) percentage points in the long run. 10 Thus, an instrumental variables regression that corrects for potential endogeneity bias confirms that there is a significant effect of GDP per capita in the country fixed effects regression when using the first-difference of GDP per capita while the level of GDP per capita produces insignificant results. 4. Conclusion Djankov and Reynal-Querol (2010) showed that the negative cross-country correlation between income and civil war incidence disappears in a country fixed effects regression. They therefore argued that the correlation between income and civil war is spurious and that poverty is not a significant determinant of civil war. At face value, their argument (and title) suggests to policy makers that reducing poverty in the world's poorest countries will not systematically reduce the likelihood of civil war. This paper showed that the empirical analysis on which the argument of DRQ is based is incomplete: when focus is on the change of GDP per capita -- which, in contrast to the level of GDP 9 See Djankov and Reynal-Querol for a discussion of the implied exclusion restriction. The F-statistic for the savings rate in the first-stage regression is always above 10. Hence, the F-statistic exceeds the Staiger and Stock (1997) ruleof-thumb criterion for instruments to be declared weak. 10 The long-run effect is calculated as the coefficient on GDP divided by 1 minus the coefficient that is obtained on the lagged dependent variable. 7

10 per capita is more likely to be a stationary variable -- the significant negative correlation between income and civil war is recovered in the country fixed effects regression. Hence, an empirical analysis that accounts for a possible random-walk behavior of GDP per capita rejects the claim of Djankov and Reynal-Querol that the relationship between income and civil war is spuriously driven by time-invariant cross-country unobservables. 8

11 References Acemoglu, D., S. Johnson, J. Robinson, and P. Yared (2008). "Income and Democracy." American Economic Review 98: Blattman, C. and E. Miguel (2010). "Civil War." Journal of Economic Literature 48: Bond, S., A. Leblebicioglu, and F. Schiantarelli (2010). "Capital Accumulation and Growth: A New Look at the Empirical Evidence." Journal of Applied Econometrics 25: Brückner, M. and A. Ciccone (2010). "International Commodity Price Shocks, Growth, and the Outbreak of Civil War in Sub-Saharan Africa." Economic Journal 120: Burke, P. and A. Leigh (2010). "Do Output Contractions Trigger Democratic Change?" American Economic Journal: Macroeconomics 2: Ciccone, A. (2008). "Transitory Economic Shocks and Civil Conflict." CEPR Discussion Paper No Collier, P. and A. Hoeffler (2004). "Greed and Grievance in Civil War." Oxford Economic Papers 56: Djankov, S. and M. Reynal-Querol (2010). "Poverty and Civil War: Revisiting the Evidence." Review of Economics and Statistics 92: Fearon, J. and D. Laitin (2003). "Ethnicity, Insurgency and Civil War." American Political Science Review 97 (1): Heston, A., R. Summers and B. Aten (2009). "Penn World Table Version 6.3", Center for International Comparisons of Production, Income and Prices at the University of Pennsylvania, August Im, K., H. Pesaran, and Y. Shin (2003). "Testing for Unit Roots in Heterogeneous Panels." Journal of Econometrics 115: King, G. and L. Zeng (2001). "Logistic Regression in Rare Events Data." Political Analysis 9: Miguel, E., S. Satyanath, and E. Sergenti (2004). "Economic Shocks and Civil Conflict: An Instrumental Variables Approach." Journal of Political Economy 112: Staiger, D. and J. Stock (1997). "Instrumental Variables Regression with Weak Instruments." Econometrica 65: PRIO/UPSALLA (2010). Armed Conflict Database. Online database. Westerlund, J. (2007). "Testing for Error Correction in Panel Data." Oxford Bulletin of Economics and Statistics 69: Wooldridge, J. (2002). Econometric Analysis of Cross Section and Panel Data. Cambridge, Mass.: MIT Press. World Bank (2003). Breaking the Conflict Trap: Civil War and Development Policy. Oxford University Press. 9

12 Table 1A. Country-By-Country Augmented Dickey Fuller Tests 1% 5% 10% GDP D.GDP War D.War Note: Column 2 (resp. columns 3 and 4) report the share of countries for which the ADF test was able to reject the null hypothesis at the 1% level (resp. 5% and 10% level) that the series contains a unit root. The number of lagged difference terms included in the regression was set equal to 1. The data are from Djankov and Reynal-Querol (2010) and cover the period Table 1B. Im-Pesaran-Shin Test for Unit Root in Heterogeneous Panels Lag=0 Lag=1 Lag=2 GDP D.GDP War D.War Note: Column 2 (resp. columns 3 and 4) report the share of countries for which the IPS test was able to reject the null hypothesis at the 1% level (resp. 5% and 10% level) that the series contains a unit root. The number of lagged difference terms included in the regression in column 2 (resp. columns 3 and 4) was set equal to 0 (resp. 1 and 2). The data are from Djankov and Reynal-Querol (2010) and cover the period Table 1C. Westerlund Test for Cointegration in Heterogeneous Panels Lag=0 Lag=1 Lag= Note: Column 1 (resp. columns 2 and 3) report the p-value of the Westerlund (2007) panel cointegration test on the null hypothesis that the civil war indicator variable and GDP per capita are not cointegrated. The number of lagged difference terms included in the regression in column 1 (resp. columns 2 and 3) was set equal to 0 (resp. 1 and 2). 10

13 Table 2: GDP Per Capita and Civil War (Baseline Annual Data, Fixed Effects Estimates) DRQ Civil War DRQ Civil Conflict (1) (2) (3) (4) CLogit LS CLogit LS Panel A: D.GDP L.D.GDP -2.64** (-2.19) L2.D.GDP -2.97** (-2.49) L3.D.GDP (-0.35) L4.D.GDP -2.39* (-1.87) L5.D.GDP -2.29* (-1.86) -0.12*** (-2.76) -0.11*** (-2.70) (-0.63) (-1.42) -0.08* (-1.75) -3.30*** (-3.55) -1.64* (-1.72) -1.53* (-1.64) (-0.25) -2.13** (-2.27) -0.23*** (-3.32) -0.12* (-1.95) (-1.58) (-0.32) -0.14** (-2.20) Country Fixed Effects Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes Sum of Coefficients L.D.GDP+...+L5.D.GDP *** (-3.56) -0.42** (-2.35) -8.82*** (-3.56) -0.60** (-2.62) Observations Panel B: GDP L.GDP -2.34* (-1.92) L2.GDP (-0.21) L3.GDP 2.35 (1.33) L4.GDP (-1.03) L5.GDP 2.73** (2.08) -0.13*** (-2.65) 0.01 (0.44) 0.08** (1.97) (-0.94) 0.07 (1.39) -3.33*** (-3.57) 1.55 (1.21) 0.17 (0.13) 1.23 (0.98) (-0.05) -0.26*** (-3.40) 0.12** (2.23) 0.03 (0.58) 0.06 (1.11) (-0.18) Country Fixed Effects Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes Sum of Coefficients L.GDP+...+L5.GDP 0.54 (0.99) (-0.50) (-1.31) (-1.38) Observations Note: The dependent variable in columns (1) and (2) is the Djankov and Reynal-Querol (2010) civil war incidence indicator variable; in columns (3) and (4) the dependent variable is the Djankov and Reynal-Querol (2010) civil conflict incidence indicator variable. Columns (1) and (3) report estimates from a conditional logit fixed effects model; columns (2) and (4) report estimates from a fixed effects linear probability model. The variable GDP refers in the table to the log of real per capita GDP. *Significantly different from zero at 90 percent confidence, ** 95 percent confidence, *** 99 percent confidence. 11

14 Table 3. GDP Per Capita and Civil War (Robustness 5-Year Data, Fixed Effects Estimates) DRQ Civil War DRQ Civil Conflict (1) (2) (3) (4) CLogit LS CLogit LS Panel A: D.GDP L.D.GDP -2.31** (-2.31) -0.11** (-2.40) (-1.00) (-1.00) Country Fixed Effects Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes Observations Panel B: GDP L.GDP 0.09 (0.14) (-0.65) (-0.84) (-1.07) Country Fixed Effects Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes Observations Note: The dependent variable in columns (1) and (2) is the Djankov and Reynal-Querol (2010) civil war incidence indicator variable; in columns (3) and (4) the dependent variable is the Djankov and Reynal-Querol (2010) civil conflict incidence indicator variable. Columns (1) and (3) report estimates from a conditional logit fixed effects model; columns (2) and (4) report estimates from a fixed effects linear probability model. The variable GDP refers in the table to the log of real per capita GDP. *Significantly different from zero at 90 percent confidence, ** 95 percent confidence, *** 99 percent confidence. 12

15 Table 4. GDP Per Capita and Civil War (Robustness Alternative Annual Civil War Data: Fearon and Laitin (2003)) Fearon and Laitin Civil War (1) (2) CLogit LS Panel A: GDP Growth L.D.GDP -1.71** (-2.32) L2.D.GDP -2.08*** (-2.69) L3.D.GDP -1.60** (-2.12) L4.D.GDP (-0.74) L5.D.GDP 0.26 (0.34) -0.09* (-1.87) -0.11* (-1.79) (-1.26) (-0.10) 0.02 (0.26) Country Fixed Effects Yes Yes Year Fixed Effects Yes Yes Sum of Coefficients L.D.GDP+...+L5.D.GDP -3.97** (-2.34) (-0.76) Observations Panel B: GDP Level L.GDP -2.19*** (-2.94) L2.GDP (-0.25) L3.GDP 0.19 (0.19) L4.GDP 0.99 (0.99) L5.GDP (-0.18) -0.17*** (-3.02) (-0.45) 0.03 (1.07) 0.06* (1.88) (-0.62) Country Fixed Effects Yes Yes Year Fixed Effects Yes Yes Sum of Coefficients L.GDP+...+L5.GDP -1.41*** (-4.40) -0.13*** (-3.59) Observations Note: The dependent variable is the Fearon and Laitin (2003) civil war incidence indicator variable. Column (1) reports estimates from a conditional logit fixed effects model; column (2) reports estimates from a fixed effects linear probability model. The variable GDP refers in the table to the log of real per capita GDP. *Significantly different from zero at 90 percent confidence, ** 95 percent confidence, *** 99 percent confidence. 13

16 Table 5. GDP Per Capita and Civil War (Robustness Alternative 5-Year Civil War Data: Collier and Hoeffler (2003)) Collier and Hoeffler Civil War (1) (2) CLogit LS Panel A: GDP Growth L.D.GDP -1.74* (-1.84) -0.11* (-1.78) Country Fixed Effects Yes Yes Year Fixed Effects Yes Yes Observations Panel B: GDP Level L.GDP -1.65** (-2.01) (-1.18) Country Fixed Effects Yes Yes Year Fixed Effects Yes Yes Observations Note: The dependent variable is the Collier and Hoeffler (2004) civil war incidence indicator variable. Column (1) reports estimates from a conditional logit fixed effects model; column (2) reports estimates from a fixed effects linear probability model. The variable GDP refers in the table to the log of real per capita GDP. *Significantly different from zero at 90 percent confidence, ** 95 percent confidence, *** 99 percent confidence. 14

17 Table 6. GDP Per Capita and Civil War (Robustness Latest GDP Per Capita and Civil War Data) PRIO 2010 Civil War PRIO 2010 Civil Conflict Panel A: GDP Growth L.D.GDP -0.11** (-2.43) L2.D.GDP -0.11** (-2.43) L3.D.GDP (-0.47) L4.D.GDP -0.04* (-1.64) L5.D.GDP (-0.98) (1) (2) (4) (5) CLogit LS CLogit LS -2.39*** (-3.09) -2.58*** (-3.28) (-0.95) (-1.45) (-1.05) -1.92*** (-3.02) (-0.77) (-1.54) (-0.25) 0.04 (0.06) -0.16*** (-3.29) (-0.69) -0.07* (-1.95) (-0.39) Country Fixed Effects Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes Sum of Coefficients L.D.GDP+...+L5.D.GDP -0.28*** (-2.82) -6.99*** (-4.02) -3.47*** (-2.82) 0.01 (0.10) -0.28** (-2.36) Observations Panel B: GDP Level L.GDP -2.20*** (-2.81) L2.GDP (-0.26) L3.GDP 1.90 (1.61) L4.GDP (-0.36) L5.GDP 1.41 (1.62) -0.12** (-2.45) 0.01 (0.14) 0.09* (1.72) (-0.75) 0.04 (1.63) -2.08*** (-3.23) 1.49* (1.67) (-0.57) 0.80 (0.93) (-0.05) -0.18*** (-3.26) 0.14*** (2.94) (-0.94) 0.06 (1.39) (-0.05) Country Fixed Effects Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes Sum of Coefficients L.GDP+...+L5.GDP 0.42 (1.13) (-0.42) (-1.53) (-1.20) Observations Note: The dependent variable in columns (1) and (2) is the PRIO/UPSALLA (2010) civil war incidence indicator variable; in columns (3) and (4) the dependent variable is the PRIO/UPSALLA (2010) civil conflict incidence indicator variable. Columns (1) and (3) report estimates from a conditional logit fixed effects model; columns (2) and (4) report estimates from a fixed effects linear probability model. The variable GDP refers in the table to the log of real per capita GDP. *Significantly different from zero at 90 percent confidence, ** 95 percent confidence, *** 99 percent confidence. 15

18 Table 7. GDP Per Capita and Civil War (Robustness to Instrumental Variables Estimation and Conflict Dynamics) PRIO 2010 Civil War PRIO 2010 Civil Conflict Panel A: GDP Growth D.GDP * (-1.91) (1) (2) (3) (4) (5) (6) 2SLS 2SLS GMM 2SLS 2SLS GMM ** (-2.25) Lagged Dependent Variable 0.597*** (17.56) Cumulative (Long-Run) Effect of D.GDP on Dependent Variable ** (-2.20) ** (-2.23) 0.459*** (6.37) ** (-2.05) ** (-2.33) ** (-2.32) 0.603*** (20.78) ** (-2.35) ** (-2.41) 0.373*** (7.53) ** (-2.35) Country Fixed Effects Yes Yes Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes Yes Yes First-Stage F-Statistic Observations Panel B: GDP Level GDP (-0.87) (-0.73) Lagged Conflict 0.597*** (17.30) Cumulative (Long-Run) Effect of GDP on Dependent Variable (-0.74) (-0.80) 0.459*** (6.29) (-0.82) (-1.18) (-0.65) 0.603*** (20.42) (-0.66) (-0.99) 0.378*** (7.57) Country Fixed Effects Yes Yes Yes Yes Yes Yes Year Fixed Effects Yes Yes Yes Yes Yes Yes (-0.98) First-Stage F-statistic Observations Note: The dependent variable in columns (1)-(3) is the PRIO/UPSALLA (2010) civil war incidence indicator variable; in columns (4)-(6) the dependent variable is the PRIO/UPSALLA (2010) civil conflict incidence indicator variable. The method of estimation is two-stage least squares. The instrumental variable is the lagged savings rate. The variable GDP refers in the table to the log of real per capita GDP. *Significantly different from zero at 90 percent confidence, ** 95 percent confidence, *** 99 percent confidence. 16

19 Data Appendix Table 1: Sample Summary Statistics of Civil War Mean Stdv. Min Max Countries Observations DRQ (2010) Civil War DRQ (2010) Civil Conflict Fearon and Laitin (2003) Civil War Collier and Hoeffler (2003) Civil War PRIO (2010) Civil War PRIO (2010) Civil Conflict