killing, but the effect is not strong or significant.
The size of the enemy’s population and both
parties having the capability to kill the adversary’s noncombatants are each positive but
insignificant, whereas being a victim of mass killing or the war occurring after 1945 are negative
yet insignificant. The explanatory power of the model is even better for mass killing than it was
for civilian targeting.
Finally, Model 8 shows that while democracies may be more prone than autocracies to
targeting civilians in costly wars of attrition, they do not appear to kill larger numbers of
noncombatants. Since this result remains consistent when using the other dependent variables
for magnitude of civilian casualties, I omit it from the remainder of the analysis.
CIVILIAN FATALITIES
ZINB. Finally, we turn to actual numbers of civilian war deaths. Models 9-10 in Table 3 display
the results of zero-inflated negative binomial (ZINB) regressions using civilian fatalities as the
dependent variable, whereas Models 11-12 show ordered logit estimates with six categories of
civilian deaths as the dependent variable (0; 1-500; 501-5,000; 5,001-50,000; 50,001-500,000;
and 500,000+). Each ZINB model contains two sets of coefficients: a logit estimate that reports
the influence of each variable on the probability of an observation taking the value of zero; and a
separate negative binomial estimate of the influence of each variable on the actual number of
casualties observed. A positive sign for the negative binomial half of the equation means that the
variable in question increases the number of civilians killed. A positive sign for the logit part of
the equation, by contrast, means that the variable makes it more likely that the number of
civilians killed will be zero. In Table 3, the negative binomial coefficient is listed in the first
column of each model (9a and 10a) with the logit estimate following in the second column (9b
and 10b).
20
An alternative coding for power, however—military personnel—performs better (not shown).
29
Convention