From Steve Miller | Dyad-year models of conflict onset and conflict escalation are the most common way for researchers to explore the causes of conflict between states. Using some data set of dyad years in the Correlates of War state system (e.g. politically relevant, contiguous, universal), a researcher can explore the determinants of the onset of a new conflict episode in a given dyad-year. Selecting on conflict onsets, researchers can look for what correlates with particularly “severe” conflicts involving the use of force or even armed conflict reaching 1,000+ fatalities (i.e. war). The proliferation of the Heckman selection model among researchers in the late 1990s and early 2000s allowed for both processes to be modeled jointly, as Senese and Vasquez do (2005, 2008) in their treatises on the “steps of war.”

There is nothing wrong with these approaches on paper, but they are limited in what they can communicate. These limitations are born from the data available to researchers (i.e. the Correlates of War Militarized Interstate Dispute [CoWMID] data set) because these data are primarily defined by confrontation and confrontation-participant summaries. These limitations suggest major assumptions of the data-generating and modeling process that are often ignored by researchers, mostly because no better alternative was available to them (until now).

First, the nature of CoWMID’s limited information on conflict episodes meant researchers focused on unique confrontation episodes that started in a given year (i.e. onsets). Either one was observed (a 1) or not observed (a 0) in a given dyadic panel with an implicit assumption that the observation of a 1 meant only one new thing happened in a given dyad. This might work well in how we think about the onset of the second world war in 1939; that issue was the only major point of contention among Germany, France, and the United Kingdom to become militarized that year. However, it betrays that dyads can have multiple new onsets in a given calendar year either over the same issue or over multiple issues.

Consider the case of France and Italy in 1860, which had three separate MID onsets that year (MID#0112, MID#0113, MID#0306), as illustrative of the problem. In a model that jointly analyzes conflict onset and conflict escalation, which one of these does a researcher choose? Clearly at least one thing happened, so a “1” in a binary onset variable communicates that at least one new conflict started. However, there is only one France-Italy dyad-year for 1860, which cannot duplicate to account for these. Perhaps a researcher can look at the underlying summaries for these disputes to determine that France is mostly a third party in the first two MIDs to wars related to Italian unification, but that sidesteps the question of what confrontation is being selected to be summarized. It would also mean having to repeat this process for the over 500 instances in which it appears in naive extensions of the CoWMID data. Whereas onsets are the most obvious confrontation-level feature of a data set best defined by summary information, privileging them can obscure as much as it hopes to reveal.

Once a researcher has selected a confrontation to be summarized by some outcome (most commonly its severity or highest action obtained), the best they can do is by reference to some indicator observed only after the confrontation has concluded. Thus, a dyad-year model of the Iran-Iraq War has to observe the onset at 1979 and wait until 1988 to conclude it was a war with at least 1,000 battle-related deaths befitting a highest action of 20. This might be a simple case where we can plausibly impute that at least 1,000 troops were dying each year in combat, but there are more complex cases like the so-called Kargil War between India and Pakistan. This case in the CoWMID data started in 1993 and became a war (albeit only for Pakistan at the participant-level), but that summary judgement would belie a lot of details. First, by all accounts, this is a “war” only because of a long-running accumulation of small engagements resulting in a series of skirmishes in 1999 that push the sum total to around 1,000. There were only a few dozen fatalities from 1993 to 1996 and the severity of “war” observed in 1999 comes with as few as 352 fatalities between both sides. However, the presence of only confrontation-level summaries from CoWMID mean a researcher’s best guess is that the escalatory process of the Iran-Iraq War is functionally equal to what we observe in the Kargil War, an assumption that is built into the statistical model the researcher uses.

We believe that our data means researchers are no longer compelled to engage in these modeling practices as they have for decades. You no longer have to privilege unique confrontation onsets in coding 1s and 0s in a dyad-year model because, with our data, you have more than just confrontation-level summaries over some (undefined) issue that have to be coerced into a dyad-year data set. Assuming the selection of some confrontation to summarize, you no longer have to define it only by the highest action observed at some point over the confrontation, or a fatality estimate observed only at the confrontation’s end. Our event data, in particular, tells you who did what to whom and when. They come with reasonable estimates of fatalities that can be aggregated by year and not just confrontation episode.

Thus, we believe our data allows researchers of inter-state conflict to be flexible with dyad-year models of conflict onset and severity. You can define “onset” in a given year by the observation of an event, and not just a confrontation that summarizes a string of events. You can define “severity” by highest action observed in a year, and not just the highest action observed in a confrontation irrespective of what year it was actually observed.

We think it important to note there are always assumptions that a researcher makes in building data for dyad-year models of inter-state conflict. However, we also think our suite of data allows researchers more insight into what these assumptions are, and have long been, in the study of inter-state conflict.