Civil war and the square of ethnic fractionalization
Steve Saideman wonders if the relationship between ethnic fractionalization and civil war is curvilinear, where increasing fractionalization increases the probability of civil war up to a certain point but then increasing fractionalization decreases the probability of civil war. Since I have my nose in this data right now, I’ve given this conjecture a quick probe. I find no evidence prima facie.
Below is a replication of Sambanis 2004 where I simply add the square of ethnic fractionalization. If Steve’s conjecture were true, we’d expect ef1 to be positive, ef1sq to be negative, and both of them to be significant. They are signed as expected but not significant given the controls recommended by Sambanis. I then try alternative codings of civil war and a simple equation with no controls. If Steve’s conjecture is true, it’s not obvious.
Everything below is reproducible in R-just download the replication data, easy to find with a quick search, and set your working directory to where the data is.
``` r Replicate Sambanis 2004 (Table 6 in paper, column 8, pp.845) library(foreign) sambanis<-read.dta(“SambanisJCR2004_replicationdataset.dta”) sambanis$ef1sq<-sambanis$ef1 * sambanis$ef1
model <- glm(warstnsb gdpl1 + grol1 + inst3l1 + anoc2l1 + oil2l1 +
ef1 + lpopnsl1 + mtnl1 + warnsl1, data = sambanis,
family = binomial(link = “probit”))
summary(model)
```
## Call:
## glm(formula = warstnsb ~ gdpl1 + grol1 + inst3l1 + anoc2l1 +
## oil2l1 + ef1 + lpopnsl1 + mtnl1 + warnsl1, family = binomial(link = "probit"),
## data = sambanis)
##
## Deviance Residuals:
## Min 1Q Median 3Q Max
## -0.604 -0.231 -0.173 -0.106 3.560
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -3.90226 0.46234 -8.44 < 2e-16 ***
## gdpl1 -0.09271 0.02295 -4.04 5.3e-05 ***
## grol1 -0.51380 0.49750 -1.03 0.30172
## inst3l1 0.23713 0.09634 2.46 0.01384 *
## anoc2l1 0.23792 0.08807 2.70 0.00690 **
## oil2l1 0.29680 0.11541 2.57 0.01012 *
## ef1 0.35605 0.16455 2.16 0.03049 *
## lpopnsl1 0.10503 0.02743 3.83 0.00013 ***
## mtnl1 0.00199 0.00182 1.09 0.27466
## warnsl1 -0.06609 0.10492 -0.63 0.52873
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 1133.2 on 5892 degrees of freedom
## Residual deviance: 1038.8 on 5883 degrees of freedom
## (3567 observations deleted due to missingness)
## AIC: 1059
##
## Number of Fisher Scoring iterations: 8
r Adding the square of ethnic fractionalization
model <- glm(warstnsb gdpl1 + grol1 + inst3l1 + anoc2l1 + oil2l1 + ef1 +
ef1sq + lpopnsl1 + mtnl1 + muslim + warnsl1, data = sambanis,
family = binomial(link = "probit"))
summary(model)
## Call:
## glm(formula = warstnsb ~ gdpl1 + grol1 + inst3l1 + anoc2l1 +
## oil2l1 + ef1 + ef1sq + lpopnsl1 + mtnl1 + muslim + warnsl1,
## family = binomial(link = "probit"), data = sambanis)
##
## Deviance Residuals:
## Min 1Q Median 3Q Max
## -0.595 -0.230 -0.172 -0.103 3.547
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -3.94606 0.48647 -8.11 5e-16 ***
## gdpl1 -0.09092 0.02370 -3.84 0.00012 ***
## grol1 -0.50874 0.49343 -1.03 0.30253
## inst3l1 0.23723 0.09641 2.46 0.01386 *
## anoc2l1 0.24253 0.08838 2.74 0.00607 **
## oil2l1 0.26000 0.12119 2.15 0.03192 *
## ef1 0.51893 0.72622 0.71 0.47488
## ef1sq -0.17995 0.71442 -0.25 0.80113
## lpopnsl1 0.10452 0.02775 3.77 0.00017 ***
## mtnl1 0.00184 0.00188 0.98 0.32529
## muslim 0.00104 0.00111 0.93 0.35169
## warnsl1 -0.06888 0.10501 -0.66 0.51185
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 1133.2 on 5892 degrees of freedom
## Residual deviance: 1037.8 on 5881 degrees of freedom
## (3567 observations deleted due to missingness)
## AIC: 1062
##
## Number of Fisher Scoring iterations: 8
r Using Sambanis alternative coding of civil war
model <- glm(warstns gdpl1 + grol1 + inst3l1 + anoc2l1 + oil2l1 +
ef1 + ef1sq + lpopnsl1 + mtnl1 + muslim + warnsl1,
data = sambanis, family = binomial(link = "probit"))
summary(model)
## Call:
## glm(formula = warstns ~ gdpl1 + grol1 + inst3l1 + anoc2l1 + oil2l1 +
## ef1 + ef1sq + lpopnsl1 + mtnl1 + muslim + warnsl1, family = binomial(link = "probit"),
## data = sambanis)
##
## Deviance Residuals:
## Min 1Q Median 3Q Max
## -0.532 -0.217 -0.165 -0.102 3.499
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -3.73245 0.52338 -7.13 9.9e-13 ***
## gdpl1 -0.08649 0.02469 -3.50 0.00046 ***
## grol1 -0.09783 0.48861 -0.20 0.84131
## inst3l1 0.23635 0.10911 2.17 0.03030 *
## anoc2l1 0.28191 0.09773 2.88 0.00392 **
## oil2l1 0.16875 0.14066 1.20 0.23027
## ef1 0.65104 0.77115 0.84 0.39853
## ef1sq -0.38362 0.77341 -0.50 0.61989
## lpopnsl1 0.08969 0.03039 2.95 0.00316 **
## mtnl1 0.00209 0.00205 1.02 0.30854
## muslim 0.00118 0.00124 0.95 0.34246
## warnsl1 -0.21649 0.31589 -0.69 0.49313
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 915.34 on 5161 degrees of freedom
## Residual deviance: 842.13 on 5150 degrees of freedom
## (4298 observations deleted due to missingness)
## AIC: 866.1
##
## Number of Fisher Scoring iterations: 8
r Using Fearon and Laitin 2003 coding of civil war
model <- glm(warst7b gdpl1 + grol1 + inst3l1 + anoc2l1 + oil2l1 +
ef1 + ef1sq + lpopnsl1 + mtnl1 + muslim + warnsl1,
data = sambanis, family = binomial(link = "probit"))
summary(model)
## Call:
## glm(formula = warst7b ~ gdpl1 + grol1 + inst3l1 + anoc2l1 + oil2l1 +
## ef1 + ef1sq + lpopnsl1 + mtnl1 + muslim + warnsl1, family = binomial(link = "probit"),
## data = sambanis)
##
## Deviance Residuals:
## Min 1Q Median 3Q Max
## -0.501 -0.200 -0.145 -0.086 3.590
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -4.28989 0.54194 -7.92 2.5e-15 ***
## gdpl1 -0.08807 0.02665 -3.30 0.00095 ***
## grol1 -0.27141 0.52732 -0.51 0.60677
## inst3l1 0.19536 0.10692 1.83 0.06768 .
## anoc2l1 0.25845 0.09748 2.65 0.00802 **
## oil2l1 0.13327 0.14174 0.94 0.34707
## ef1 0.44027 0.80427 0.55 0.58409
## ef1sq -0.22642 0.79784 -0.28 0.77657
## lpopnsl1 0.11930 0.03065 3.89 9.9e-05 ***
## mtnl1 0.00255 0.00204 1.25 0.21124
## muslim 0.00155 0.00122 1.26 0.20659
## warnsl1 0.01688 0.11169 0.15 0.87987
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 897.81 on 5892 degrees of freedom
## Residual deviance: 822.48 on 5881 degrees of freedom
## (3567 observations deleted due to missingness)
## AIC: 846.5
##
## Number of Fisher Scoring iterations: 9
r Only ethnic fractionalization
model <- glm(warstnsb ef1 + ef1sq, data = sambanis, family = binomial(link = "probit"))
summary(model)
## Call:
## glm(formula = warstnsb ~ ef1 + ef1sq, family = binomial(link = "probit"),
## data = sambanis)
##
## Deviance Residuals:
## Min 1Q Median 3Q Max
## -0.303 -0.227 -0.196 -0.161 2.990
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -2.302 0.131 -17.63 <2e-16 ***
## ef1 0.323 0.589 0.55 0.58
## ef1sq 0.283 0.580 0.49 0.63
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 1249.6 on 6268 degrees of freedom
## Residual deviance: 1230.5 on 6266 degrees of freedom
## (3191 observations deleted due to missingness)
## AIC: 1237
##
## Number of Fisher Scoring iterations: 7
Cite this post: RIS Citation BibTeX Entry
Murphy, Justin. 2013. "Civil war and the square of ethnic fractionalization," http://jmrphy.net/blog/2013/05/22/civil-war-and-the-square-of-ethnic-fractionalization-4/ (May 15, 2017).