For a number of decades, the Naxal movement (introduced in section 1.4.2) engaged in attacks against both civilians and the state. This hostility was the result of a long- standing commitment by the Naxals to armed struggle against the state in order to address wide-ranging grievances. Andhra Pradesh was one of the most affected Indian states during this time, and its government came under criticism for their apparent ambivalence towards the violence, failing to devise a long-term strategy that improved the security situation (Basu, 2011). At the height of hostilities in 2006, the Indian Prime Minister Manmohan Singh stated that the Naxals represented the “single largest internal security threat to India” (Basu, 2011). In recent years, the level of violence has substantially reduced; however, for much of the previous decade, violence and hostility increased periodically, often to unprecedented levels.
Data was obtained from police forces in India that detailed hostile events associ- ated with the Naxal insurgency for ten years between 2000 and 2010 in the state of Andhra Pradesh. The data consisted of official police records of Naxal-related violence or threat recorded in the 1,642 police stations within the state.
Over the course of the duration of this dataset, there is evidence to suggest that the police adopted various counterinsurgency strategies. For instance, during a period in 2004, in which various splinter groups of the Naxal movement combined to form a unified and potentially diplomatic group, counterinsurgency actions were ceased com- pletely in the hope that a diplomatic solution to the conflict could be found. During other periods, the police took up strict counterinsurgency action. Activities resulting from such police action were not detailed in the data; however, aggressive counterin- surgency activity, which involved the killing of Naxals during shootouts, were known to have been adopted as a result of fieldwork described in Belur (2010). As a result, it is assumed that events described in the dataset as an “exchange of fire” between Naxal and police, and which resulted in at least one Naxal fatality, were largely caused by strategic counterinsurgency activities. It has been claimed that using this description for Naxal fatalities is a way of legally justifying extrajudicial killings (Belur, 2010).
which at least one Naxal was killed corresponds to a counterinsurgency event under- taken by the police, it is possible to partition the dataset into events initiated by Naxals and counterinsurgency events initiated by the police. As a consequence, the dataset employed in this chapter is distinct from the data on the London riots investigated in previous chapters. During the London riots, the behaviour of the police was unknown and could not be empirically tested within the models presented. In this case, data on the activities of both adversaries in the conflict can be used to explicitly consider the ef- fect of actions of one side on the actions of the other. Moreover, the scarce availability of such datasets elsewhere implies that the model presented in this chapter provides a significant contribution to existing literature concerning the spatio-temporal modelling of sub-national conflict between two adversaries.
In total, there are 4,820 incidents in the dataset, which covers the entire state of Andhra Pradesh. For each eventi, a three-dimensional tuple is constructed, given by (ti, si, mi), where ti ∈ T denotes when event i took place, si ∈ D denotes where the
event took place, andmi ∈ {0, 1} is a mark that indicates whether incident i was ini-
tiated by Naxals (mi = 1) or was initiated by police as part of their counterinsurgency
campaign (mi = 2). The setsD and T represent the spatial and temporal domains of
the model, which are next described.
The models developed in this chapter will be continuous in time; however, the data on the Naxal conflict is discrete in time, with a daily temporal resolution. The first day included in the dataset is the 1st January 2000 and the final day is 7th August 2010 (3,872 days in total). TakingT = [0, 3872], the date of each event is translated into continuous time by initially settingti to be equal to midnight on the day on which the
event occurred. In section 5.5, concurrent events are distinguished by a randomisation procedure, which is explained in the relevant part of the text.
The domainD represents the geographic area of interest and, due to the resolution at which the data is available, is taken to be composed of the union of non-overlapping districts, as D = J [ j=1 Dj, (5.1)
where each Dj corresponds to a district in Andhra Pradesh. According to the 2011
districts, Hyderabad, just two events were recorded. Since this district comprises of the city of Hyderabad, and therefore is small in its geographic extent, and does not experience a large amount of violence, this district and the two events that occur within it are omitted from the analysis.
In 2014, the state of Andhra Pradesh was bifurcated, and the state of Telangana was formed consisting of nine districts that were previously part of Andhra Pradesh. Andhra Pradesh itself remained but now consisted of just 13 districts. The state bifur- cated to more closely align language, ethnicity and old political ties. 3,387 of the 4,820 incidents in the dataset (70%) occurred within the districts that formed the new state Telangana. For reasons of computational tractability, the domainD is initially chosen to consist of the 9 districts in Telangana, and the analysis is restricted to just these events. In particular, the models outlined below are calibrated using incidents that occur within these 9 districts. This restriction ensures that the models proposed can be calibrated over reasonable time frames. Furthermore, this restriction enables the remaining data to be used for out-of-sample model testing, in order to determine whether the model of insurgency in the Telangana state also applies in the state of Andhra Pradesh (specif- ically, only four districts are used in the out of sample data – those four districts that contained at least 100 events over the period of study). Therefore, initially,J = 9 and,
for each event i, it is sufficient to take si = si ∈ {1, 2, ..., 9}, denoting the district
within which the event took place.
The spatial distribution of both police and Naxal initiated events in Telangana and Andhra Pradesh across the entire time period of interest is shown as a thematic map in Figure 5.1. The temporal distribution of incidents occurring on each day within each district is shown in Figure 5.2. This figure also distinguishes between Naxal and police intiated events, and includes total counts of each type of event that occurred in each district. In total, there are 586 police events and 4,234 Naxal events, of which, 424 police events and 2,963 Naxal events are contained within the nine districts that make up Telangana. Examining these figures, it can be observed that the vast majority of events occur within a relatively small number of regions, with the highest number of events occurring in the Warangal district. Furthermore, the intensity of police attacks follows closely the intensity of Naxal attacks in space, suggesting that the two types of events may well be dependent upon one another (although, one should be cautious
not to confuse correlation with causation). In what follows, stochastic point process models of these events are derived by considering a variety of mechanisms that may have influenced their occurrence.