Using Cluster Analysis as an Early Warning Indicator
Empirical Indicators of Crisis Phase in the Middle East, 1979-1995
Using Cluster Analysis to Derive Early Warning Indicators for Political Change in the Middle East, 1979-1996
These charts summarize an on-going project that is seeking to develop early warning indicators for political change in the Middle East. Our a priori cluster assignments were the following:
Discriminant Analysis of a priori Cluster Assignments
Remarks: Discriminant analysis operating on the vector of dyadic event score differentiates the phases we assigned a priori with a high level of accuracy. This indicates that there is sufficient information in the event data to differentiate between periods of political activity, despite the usual problems of source bias and coverage found in event data.
Remarks: These two charts show the first three dimensions of the discriminant space, which account for about 75% of the variance explained by the model. The first dimension ("Function 1") probably is trend -- the cluster centroids are generally in chronological order except for a couple of swaps. The second dimension is unclear, though it may differentiate the Lebanon, Palestinian and other issues (or possibly just pull out the Camp David period). The third dimension (below) is a conflict-cooperation dimension on the Palestinian issues. The periods we would expect to mark extreme points in the behavioral space -- Lebanon, the intifada, Oslo -- generally are distinct in the discriminant space.
Cluster Analysis of System Behaviors
LML and CD Indicators updated through May 1997
Both charts are consistent with a phase transition around the time of the Rabin assassination, but no distinct changes after that point (in particular, in these data the Netanyahu period is not distinct from that of Peres during 95.12 to 96.6). The CD measure has been gradually climbing over the past year, which is consistent with the possibility of a new phase in the near future, but it is currently not near the critical level.
Early Warning using Event Counts
The two charts below show the LML and early warning measures computed using a correlation metric that uses the number of events in each of the 22 two-digit WEIS categories, rather than using the scores that were aggregated using the Goldstein scale. The results are generally similar, though the correlation metric using the Goldstein score has somewhat higher variance, which may indicate a higher level of sensitivity.