A convention that has developed over the last couple of decades has quantitative researchers sampling only the Correlates of War Militarized Interstate Disputes (CoWMIDs) that had military fatalities. Most recognize the heterogeneity in the data set—that there are some “weird” cases—and want to analyze only those cases that truly have a chance to escalate to war.  We have found this selection method to be both theoretically and empirically fraught. Here’s why.

First, ever hear of the following conflicts?: Iran Hostage Crisis of 1979 (MIC#3020), the Alaska Boundary Dispute of 1902 (MIC#0002), the 1938 Anschluss (MIC#0011), the 1938 Sudetenland Crisis (MIC#0012), the Berlin Blockade of 1949 (MIC#0026), the 1908-1909 Bosnian Crisis (MIC#0030), or the Cuban Missile Crisis in 1962 (MIC#0061)?  None of these confrontations experienced military fatalities but each had a high likelihood of escalation to war.  Indeed, the Cuban Missile Crisis carried the real possibility of escalation to a nuclear war.  Each would be excluded from analysis with the use of a fatality threshold, and there are many, many cases like these.  In our MICnames dataset we have assigned names to more than 235 confrontations that had no fatalities.  This does not always correlate well with the likelihood of escalation to fatalities or war, but most researchers would want to examine the cases first rather than simply exclude all cases.

So what is gained by a focus on one or more battle-deaths?  There are more than 270 confrontations with between 1 and 25 fatalities in our Militarized Interstate Confrontation (MIC) data, and the vast majority did not carry the likelihood of escalation that the above confrontations did.  For example, two Turkish soldiers were killed in 1981 trying to stop a clash between Iranian Kurds and Iranian National Guard forces (MIC#3097).  Iran later apologized.  In 1965, a stray herd of sheep caused two Turkish fatalities in a clash with Syrian forces (MIC#3168).  Even earlier, in 1931, a Greek soldier crossed the border to let Bulgarian soldiers know that mules were wandering across the border.  The Bulgarians misunderstood his intentions, shot him dead, and caused a firefight; MIC#3186 is properly coded with one fatality even though the battle-death resulted from the mistaken intentions.  More recently, in 1992, two Peruvian fighters strafed an American C-130, blowing one crew member out of the plane (MIC#3550).  Apparently, the C-130 had veered off its flightpath.  Again, the data is rife with these low-level conflicts with just a few, sometimes unintentional fatalities.

Empirically, it is also important to consider how to use fatal MIDs as a selection device when fatalities are missing.  Most studies that I have seen simply ignore these cases and drop the disputes.  However, more than 10% of the CoWMID data has missing fatality cases (more than 280 MIDs).  Our MIC data has no missing fatalities, and our coding of these missing-fatality cases should at least cause some hesitation before implementing the fatal-MID selection tool. We found almost forty cases with at least 100 fatalities or more in the confrontation that were coded as missing-fatality disputes by CoWMID.  Seventeen of these confrontations had at least 500 fatalities, and eleven had more than 999 battle-deaths — these were wars! Four of these cases are wars in CoW’s own war data! 

The missingness in these cases suggests poor information available to CoW, and their presence suggests systematic bias when analyzing escalation.  Regardless, a selection on fatal-MIDs that disregards at least 10% of the cases of interest, including many cases that had a large number of fatalities, is problematic at best.

What to do?  As I describe above, and as the field implies, ample heterogeneity exists in the cases contained in either CoWMID or our MIC data.  Dealing with that heterogeneity is important, but it doesn’t have to be done with such a blunt instrument.  Instead, we have separated two datasets from the larger MIC data so researchers can begin to consider and adjust for the variation across conflict cases.  First, we have a separate set of war declarations.  Most often, these are political statements with no militarized action.  Though Jones, Bremer, and Singer (1996) code them as a militarized action, the majority of these disputes had one incident with no threat, display, or use of force.  These are cases in which the country wanted to be part of the United Nations, received political pressure from the Allies, and declared war against one or more Axis Powers in order to join.  Even for the war declarations that occurred during wars amid other militarized events, the declaration itself carried little in the way of added threat.

Second, we separately code what we call protest-dependent disputes.  We do not consider these to be militarized confrontations because the actions were always state against citizen, followed by protest.  For example, a (civilian) fishing boat is seized by another country, the fishing boat’s country protests, and JBS code it as a MID.  Another common case is one country “attacking” another country as it chases rebels across the border.  In many cases the rebel-hosting country is working with the attackers because it cannot effectively police its own territory.  We do not code state-to-state conflict in these cases—a protest is not a threat and actions against civilians are not interstate—so we code them separately, as disputes consistent with the letter of JBS’s coding rules.  See Gibler and Little (2017) for a full consideration and analysis of the protest-dependent MIDs and how analyses of these cases differ from other conflict cases.

As with the war declarations, if researchers still want to consider protest-dependent disputes, they may download our correctly-coded cases and merge them with our MIC data.  We include the MID#, and we also code why it is protest-dependent, with context variables.  Thus, researchers will know that two countries were trying to act together against rebels, even though CoWMID codes a MID with an attack by one state against the others.

I hope these points lead researchers away from focusing on fatal conflicts only.  At the very least, a study should justify why it excludes a case as potentially-dangerous as the Cuban Missile Crisis from its analyses.