Appendix C: Replication with other Dataset

Table 2.16: IV2SLS: Economic Growth and Civil Conict

I Ia II III

GDP Growth, t -0.676 -1.517 -2.644 *** -2.644 ***

(2.140) (1.263) (0.960) (0.960)

GDP Growth, t-1 -2.484 *** -1.545 *** -1.729 *** -1.729 ***

(0.843) (0.265) (0.356) (0.356)

GDP at 1978 3.122 3.122

(2.563) (2.563)

Quality Policy -0.001 -0.001

(0.001) (0.001)

Ethnic Fractionalization 17.195 17.195

(26.443) (26.443)

Religious Fractionalization 41.081 * 41.081 *

(22.887) (22.887)

Oil export 0.142 ** 0.142 **

(0.065) (0.065)

(Log)Population 0.006 0.006

(0.020) (0.020)

(Log)Terrain -3.202 -3.202

(3.262) (3.262)

N 1071 1071 959 959

R-square adj 0.310 0.442 0.322 0.322

* p<0.1, ** p<0.05, *** p<0.01

instance, when I drop Niger, one-point decline in (current and lagged) gdp growth increases the likelihood of civil conict by over two percentage points.

In the last two columns I have added country controls. The dierence between the two specications is that in the rst one I use country xed eects, while in the second one I use only time trend. As we can see, current and lagged economic growth is highly signicant and much more bigger than previous estimates. Hence, if I take into account country characteristics, the impact of economic variables become deep. At the same time, also oil exports and religious fractionalization show to aect civil war. Nevertheless, I have some doubts on these models, since religious fractionalization has strange values with huge standard error and because instruments are weak also in these cases. In fact, in the rst stage the structural equations are identied but the actual size of the t-test tells us that the point estimates on the endogenous variables equal zero at 25%. In the second stage, the Cragg-Donald Wald F statistic performs better, showing that the bias of the IV estimates are greater than 20% of the OLS bias.

In these two last specications, variations in the set of countries do not lead to strong changes.

Only Niger shows to aect deeply estimates, because, as before, when we remove this country, estimates change.

Each specication shows the same situation, excepts for the last two specications: models are identied, because null hypothesis that the structural equation is underidentied is rejected for the Kleibergen-Paap LM test p-value. Nevertheless, instruments are weak. In fact, both the Kleibergen-Paap and the Cragg-Donald test show high levels of p-value, far from the desired 0.05.

In the case of the last two IV2SLS, the Kleibergen-Paap LM test cannot reject the null hypothesis, meaning that the models are overidentied and have weak instruments.

−9−7−5−3−113579GDP Growth, t−1

0 10 20 30 40 50

Country

Loop by Removing Country

Fig. 2.9: Loop removing countries

and Sambanis (2004) use some variables taken from Fearon and Laitin (2003), but some other variables are conceived by the authors. For example, the dataset of Sambanis (ibidem), dier from the dataset of Fearon and Laitin (ibidem) for 5 variables: civil war onset, population, oil, ethnic and religious fractionalization.

From now on, we will call the work of Fearon and Laitin (2003) FL.

In the rst stage, where we study the causal relation between rainfall variation and GDP growth, we see substantial dierences between our results and those from FL and from Sambanis.

In Sambanis's estimates, the eect of rainfall variation seems to be much higher (compared to our output). In FL, instead, estimates are higher than our results.

All specications in both authors, show a very signicant and positive impact of rainfall variation on economic variables. About the control variables, only in Sambanis oil is highly signicant and goes in the opposite direction of our estimates. In fact we estimate a positive and non signicant relationship between oil-exporting countries and GDP growth, while in Sambanis this relationship is negative but not very high.

When we regress the probit model, in Sambanis we lose several observations because of collinearity of oil variable. First of all, in Sambanis GDP growth (both contemporaneous and lagged) is positive (though non signicant), while in FL is negative (and non signicant). This is very strange, since they use the same data for GDP and, of course, for GDP instrument.

Beside that, this output is completely dierent from our estimates; in fact, we found a negative probability of civil war related to economic variables, which has proved to be stable. In FL, this probability is doubled compared to mine but is not signicant. Moreover, country controls are insignicant in both authors and also the value is low in every variables (while in our model, we found mountainous terrain signicant and positive related to civil war).

we have a similar behaviour for the OLS; the rst OLS uses only country controls and in Sambanis contemporaneous and previous economic growth is very small and positive, while in FL the eect is negative but much more smaller than our estimation. Again, both authors do not nd signicant control variables in this rst OLS.

The second OLS uses country and time trends eect; in this case, Sambanis has the same sign and similar value of FL. Returning to the covariates, in Sambanis any control variable is not signicant, while in FL population has a positive eect on civil war (signicant at 5% condence level).

Only in the last OLS, Sambanis obtains a negative sign for GDP growth rates, which is, in turn, much more smaller than our estimate. Hence, GDP growth has a negative eect on civil war, but coecient estimates are very low and are not signicant. FL, on the contrary, estimate that a 10% variation in contemporaneous or previous GDP growth rates causes a 24% increase

in the incidence of civil war.

Finally, we run a IV 2SLS. For the rst time, both Sambanis and FL produce a negative impact of GDP growth rates on civil war. Estimates from FL are pretty similar to our esti-mates, even if in FL are lower and non signicant. In fact, the rst specication has country control variables and in FL any variable seems to impact particularly (and signicantly) civil war. In the second specication, FL have a larger impact of contemporaneous GDP growth, though insignicant. Regarding Sambanis, the rst IV 2SLS is underidentied and with weak instruments; hence, we can not rely on its estimates which have, also, very high standard errors.

Comparing these outputs with those of MSS, we can see that in the case of FL, the rst stage which estimates the eect of rainfall variation on GDP growth, has similar results. The rst OLS is the simplest, without covariates and within eects. Here, the eect is positive and signicant as in MSS, but it is lower. On the contrary, in the second OLS with control variables and time trends, the eect of rainfall variation on GDP growth is bigger and signicant. The other OLS specications provide similar results. In the case of Sambanis, the impact of rainfall variation on economic growth is even bigger and more signicant, excepts for the second specication (with country controls and time trends eect).

When we estimate the relationship between GDP and civil war, on the contrary, results dier substantially. In fact, with FL'data, both probit specication and OLS specications, estimates are much more smaller than MSS's results. Then, the size of the estimated impact of economic growth on civil war is huge; in particular, according to FL'data, a one percentage point decline in GDP increases the likelihood of civil war by over 1 point, both taking into account country characteristics or without. In any case, any variable is signicant. In the case of Sambanis, we still not have signicant variables, but the impact of economic variables seems to be smaller in every specications.

Hence, from this comparison, we can see that the denition of civil war proposed by Sambanis changes substantially the results of our study, by showing a smaller eect of the economic conditions on civil war. It is also true, that also other country characteristics do not show signicant causal relationship with civil war.

Markov Transition Probabilities: An Application to the Analysis of Civil War in Africa

85

Abstract

We develop a new indicator of civil war, based on the concept of violence escalation. Taking information from the Social Conict in Analysis Database Version 3.1 (SCAD), from the Armed Conict Location and Event Data Project 2015 (ACLED), and from the Uppsala Conict Data Program (UCDP) we design a violence indicator ranging from 1 (peace) to 3 (civil war). We use the new indicator to estimate Markov transition probabilities for 50 African countries, 479 ethnic groups and 50 non-ethnic groups in Africa from 1975 through 2014 as function of ob-servable characteristics. We implement 3 methods to construct these Markov probabilities. The

rst method is a counting method, the second predicts transition probabilities using ordered logit regression models and nally, we derive hazard rates from a non-parametric Kaplan-Meier estimator and a semi-parametric proportional hazard (Cox) model. We also test whether the Markov assumption holds.

Civil War; Africa; Markov Transition Probabilities; Survival Analaysis; Ethnic and non-ethnic groups.