A Neural Network Analysis of Militarized Disputes, 1885-1992: Temporal Stability and Causal Complexity

Great progress has been made in predicting and explaining interstate conflict. Improved data, theory, and methods all deserve credit. Yet much remains to be done. First, whereas many variables (e.g., geographical proximity, relative power, alliances, political regime type, economic interdependence) have important effects, even the most successful multivariate analyses leave much of the variance in conflict behavior unaccounted for, due to inadequate data, specification, or theory, or simply random variation. Consequently, questions arise about the predictive power of such analyses. Can we identify, with enough accuracy for policy purposes, those relationships very likely or very unlikely to experience militarized disputes? Can we reduce the number of false negatives and false positives? Second, interstate conflicts are complex phenomena often displaying non-linear and non-monotonic patterns of interaction. Those complexities are hard to model. Finally are questions about whether causal or predictive relationships are stable across time and space. One such question is whether democracy reduced the risk of interstate conflict 2001) or its effect was limited to the cold war era (Gowa 1999) due to particular conditions like ideological rivalry, bipolarity, or nuclear weapons. Some early COW analyses (e.g., Singer and Small 1968) also emphasized nineteenth and twentieth century systemic differences.