Anticipating the Good, the Bad, and the Ugly: An Early Warning Approach for Conflict and Instability Analysis

Dr. Sean P. O’Brien

Center for Army Analysis

obrien@caa.army.mil

This document provides instructions for replicating the principal analyses in the above referenced JCR article.

The multi-nomial logit analyses displayed at www.oocities.org/seanob88 were run using SPSS version 10.0 using the variables indicated at that website.

The Fuzzy Analysis of Statistical Evidence (FASE) estimates were generated using a C++ program (version 1.2.2) developed by Dr. Yuan Yan Chen (chen@caa.army.mil) at the Center for Army Analysis, 6001 Goethals Road, Fort Belvoir, VA  22060-5230.  Dr. Chen’s program is available upon request and free of charge to anyone affiliated with the United States government.  It is available for a small fee to all others.  Please direct all inquiries and requests for the software to Dr. Chen at the above address.

1. The FASE validation results can be replicated using the data included with this documentation.
2. Use the country-year data for the period 1975-1999
3. Sort the data by year, then by country
4. Place the dependent variable in the third column of the spreadsheet (will ultimately be in first column).
5. Save the Excel file with a name
6. Replace all missing values with –999
7. Clear the entire first row of everything (will contain the 3 input fields specified in step 9)
8. Delete the first two columns (country and year)
9. The first cell in the upper left column contains the number of records to be analyzed (3671); the second cell to the right contains the number of variables to be analyzed (the dependent + all Independents=13); the third cell to the right contains the number of categories on the dependent variable (4).
10. Save that Excel file as a text file
11. Execute the FASE program
12. Locate the data text file using the “BROWSE” feature
13. Specify the proportion of records to be used for training in the % used for training field.

·        The 1975-1984 training set used 1399 of the 3671 records or .3811 for training

·        The 1975-1989 training set used 2104 of the 3671 records or .5731 for training

·        The 1975-94 training set used 2876 of the 3671 records or .7834 for training

14. For these analyses, we used a window parameter of 1, a Frank parameter of .01, and a confidence level of .8 (the default and an irrelevant parameter for these analyses).
15. Click “GO.”
16. Three files are returned:

Cf.txt (ignore)

Class.txt

Output.dat

The class.txt file contains possibility scores for each X variable observation associated with each class on the dependent variable (useful for drawing curves).  The output.dat file contains the overall possibility scores associated with each country-year’s macro-structural portfolio.

17. Normalize the possibility scores into probabilities
18. Aggregate the probabilities across the conflict event types (as described in the 6 decision rules in the JCR article) to derive the forecast level of intensity of instability.