Individual Submission Summary
Share...

Direct link:

Download

A Power-Law Approach to Occurrence Prediction of Political Events (Pre-Recorded)

Sat, October 2, 8:00 to 9:30am PDT (8:00 to 9:30am PDT), TBA

Abstract

Political scientists have explored more accurate methods of occurrence prediction of political events. One of the ways to improve accuracy of predictions is to modify existing statistical models to accommodate empirical features of data. In this paper, we propose to incorporate the power law, the statistics characterized by a long-tailed distribution, to statistical models that predict certain types of political events. Various phenomena in the world are known to follow the power law, urging scholars in physics, economics, linguistics, and other disciplines to utilize it in their research. Recently, the power law is used to analyze the occurrence of rare events such as stock market crashes, bank runs, and natural disasters. In political science, nevertheless, although Lewis Fly Richardson pointed out power law’s usefulness to analyze interstate conflicts more than 60 years ago (Richardson 1944, 1960), very few have since then used it in their statistical analysis. We show in this paper how the power law can be utilized in political research. In addition to power law’s better fit to some of real-world phenomena, it has properties that allow political scientists to easily combine it to existing prediction models to improve their predictability.

To illustrate how the probabilistic distribution of power law can be mathematically incorporated to existing prediction models and to enhance their predictability, we created our model making use of the properties of power law and then experimented with a dataset of interstate conflicts used in previous studies of political science (e.g., Beck, King & Zeng 2000). Here, we focused on the inter-conflict interval (ICI) as a characteristic quantity, which is defined by the period between the end of the last conflict that occurred in a specific dyad and the beginning of the next conflict in the same dyad. Totally 853 ICI samples were extracted from the dataset. Calculating the frequency of ICIs, we found that their distribution is extremely skewed to right. To identify the statistical pattern of ICIs, two types of distributions that are likely to fit such extremely rightward-skewed distributions were examined. One is an exponential distribution and the other is a power-law distribution. Our regression results indicate that ICIs fit a power-law distribution (R^2=0.8865) better than an exponential one (R^2=0.8026). The most likely interpretation of these results is that, conditioned by a conflict that occurred in any dyad, the next conflict in the same dyad will occur at the timing governed by the power law, rather than a Poisson process. The scale-free property of power law, the invariance of a power-law distribution functional form under scaling of the argument, is essential for this interpretation, making it consistent with the fact that the conflict frequency varies from dyad to dyad; different dyads correspond to different scaling. This interpretation can also be consistent with the notion that the timing of war, which is counted without specification of a dyad, obeys a Poisson process (Richardson 1944, 1960; Clause 2018).

Provided that ICIs follow a power-law distribution, we then calculated the probability that a conflict will occur in a specific dyad in the unit interval starting from the present time. We call this probability the “chance”. This probability turns out to be proportional to the inverse of the peace years i.e., the interval from the end of the last conflict that occurred in the same dyad. Because of the scale-free property of power law, the form of the chance is consistent for any dyad. In previous studies (e.g., Beck, King and Zeng 2000; Beck & Zeng 2001) that used the same data to predict conflict occurrence, prediction models were implemented by logistic regression or its augmented architecture, neural networks. The output of these models was regarded as the probability of conflict occurrence and used for prediction. Here we call this probability the “gate.” Now the probability of conflict occurrence, which will be used for our occurrence prediction, is redefined by the “chance” multiplied by the “gate”. The scale-free property of power law again justifies this prescription. If the event occurrence followed another type of statistics such as a Poisson process, this concise prescription would be no longer valid.

The prediction performance of the model implemented by logistic regression was taken as a baseline for our validation. Results of our preliminary analysis indicate that the proposed method exploiting the power-law statistics of ICI significantly and substantially outperforms the original logistic regression, with more than 20 % increase in the precision. We believe exploiting the statistical properties of power law is a powerful approach to enhancing occurrence prediction of political events.

Authors