**Sensitivity Analysis**
We have already analyzed our hypotheses with both logit and rare events logit models in an effort to identify any potential bias in the data-generating process. It is possible, however, that the decision to solicit support abroad, or the decision to use non-violent or violent contention abroad are not independent of each other. After all, some organizations engage in more than one type of behavior in the same year (though this is rare in our data). Still, we might expect that each of these decisions is interrelated, which would mean that the error terms in each of our logit models are correlated, and the probabilities of each dependent variable, therefore, would be jointly distributed. To account for this possibility, we re-estimated our models as bivariate probits, pairing each of the dependent variables, and the results are listed in Table 8.
In the first column of Table 8, we assume that solicitation and non-violent protest are jointly distributed, though the results of each individual probit are nearly identical to the original logit models. Foreign state support, elections, repression, political discrimination, and inter-organizational conflict continue to drive solicitation, while protest continues to be a function of foreign state, diaspora support and participation in elections. Religious ideology significantly affects protest, but this finding is not consistent in the other model that includes protest. Inter-organizational conflict increases the probability of solicitation while decreasing the probability of protest (though weakly so). In the second column, we analyze the joint effect of our independent variables on solicitation and violence, and once again the results are comparable to our original analysis. Finally, the model analyzing both protest and violence (third column) also produces results similar to those we presented earlier. The correlation parameter (*ρ*), however is insignificant here, indicating that the errors are not correlated, and the models of protest and violence can be safely estimated independent of each other.
**Share with your friends:** |