18
extent. Hence, there is no support for a mirror image behavior where the violence by
one party is a response to the other party’s behavior.
Taking a look at the logit coefficients in the lower part of the table, we find a
different result than in model 1. Both rebel deaths and rebel deaths squared now
have the power to successfully predict zero-events, in line with the results from the
negative binomial regression above. Rebel deaths, which is the substantially
interesting variable here, thus has a negative effect, suggesting that rebel losses
decreases the likelihood of civilians not being killed. This is in line with the initial
finding under descriptive statistics – that there seems to be a correlation between
rebel losses and the occurrence of violence against civilians. Further, the number of
rebel groups also decreases the probability of non-violence, or in simpler terms, rebel
groups that fight governments that are at the same time involved in armed conflict
with other rebel groups tend to be more prone to target civilians.
In the last model a subset including only those rebel groups that at least once
during the period of observation killed civilians. The reason is that there may be
different types of rebel groups. The data shows that 32 percent (19 groups) of the
rebel groups included in the data set never kill civilians during the observed time
period. The explanation to this does not seem to be that these groups never lose on
the battlefield, because they do. A more plausible explanation is that there are
different types of rebel groups, those that consider killing civilians as option, and
those that do not. The latter type could be groups that have taken a policy decision
to only fight the government through regular military means and not to target
civilians. This raises the question how it affects the argument posed in this paper.
Reasonably, the theoretical argument should be stronger when focusing only on the
groups that can consider killing civilians. Therefore, the regression is run without
those groups that clearly never target civilians, in order to see how it affects the
results, which are presented in table 1, model 3. Just like the previous models, the
first hypothesis is supported. But interestingly, conflict duration now has a
significantly negative effect, which is in line with hypothesis 2. Thus, we can
conclude that when focusing on the groups that at least occasionally do target
civilians, the theoretical argument applies well. Turning to the logit coefficients,
there are two changes from model 2: rebel deaths and rebel deaths squared are no
longer significant, but instead territory is. Territory, indicating a conflict where the
rebels make territorial claims, increases the likelihood of no killings of civilians.
Convention