Patterns of Contention: The Role of Protest Sequences in Civil Conflict Escalation

Events of political contention, such as escalation and de-escalation of protest, are dynamic phenomena driven by the interactions of government and opposition. Such repeating interactions produce recurring shapes in time series of events. Existing methods for studying the dynamics of political contention typically focus on individual data points without considering entire temporal sequences as unit of analysis. In this paper, we leverage time series clustering for regression analysis—a technique for extracting recurring patters from time series data. We argue that our method can more accurately account for complex temporal dependencies in causal inference models. Our use case is the transition from protest to civil conflict. Protest movements shift towards violent tactics if they believe that nonviolent means have failed, indicated by certain patterns in preceding series of protest events. Specific patterns in protest activity may reflect the strategic choices of protest groups and can explain whether a movement turns violent or remains peaceful, measured by the number of deaths in civil conflict. Our theoretical expectation is that recurring patterns in sequences of protest events are correlated with a change in the number of deaths in civil conflict. We introduce a novel method for validating the effect of time series shapes on an outcome of interest. Data on the input, the number of protest events, are obtained from ACLED. The level of analysis is the country-month. Using k-Means clustering and Dynamic Time Warping, our method identifies recurring patterns in sequences of protest events, represented as distinct shapes in the time series data. This process involves a two-staged clustering approach. First, we obtain cluster assignments for each country-month based on a set of cluster shapes unique to each country. This renders the algorithm sufficiently flexible to account for patterns distinct to each country. Deriving cluster shapes valid across country, we apply the same clustering algorithm on the centroids from the first step, after discarding the top ‘n’ country-months with the highest standard deviation of protest events in the next month. Time series of events are erratic, and this two-staged clustering approach serves to filter out noise while maintaining a high level of flexibility. Data on the output, the change in the number of deaths in armed conflict, are obtained from UCDP GED. Validating our theoretical expectation, we include a dummy set for the clusters of protest events in an OLS regression model to examine their influence on the change in the number of deaths in armed conflict. Conditional on static covariates, we find that certain shapes are correlated with either an increase or a decrease in the number of casualties. The corresponding shapes signal instances of escalating or de-escalating protest intensity. Our findings suggest that these patterns play a crucial role in the strategic decisions of protest groups, in particularly, whether to shift towards more violent tactics.