This research note discusses a five-year examination of the original coding work of the Militarized Interstate Dispute (MID) project. After strictly applying MID coding rules, we recommend dropping 251 cases (or over 10% of the dataset), as either we were unable to find a militarized incident in the historical record or the dispute appeared elsewhere in the data. We found evidence linking 75 disputes to other cases, and we could not identify 19 cases in the historical record. Among the remaining disputes, we recommend major changes (changes in dispute year, fatality level, and participants) in 234 disputes and minor changes in 1,009 disputes. We use this article to examine the potential impact of our suggestions on existing studies. Though we identified several systematic problems with the original coding effort, we also find that these problems do not affect current understandings of what predicts the onset of interstate conflict. However, estimates in our replications of three recent studies of dispute escalation, dispute duration, and dispute reciprocation all witness substantial changes when using corrected data-to the point of reversing previous conclusions in some cases.

We examine a major source of heterogeneity across cases in the Correlates of War Militarized Interstate Dispute Dataset, 1816–2001, and demonstrate that this variation across cases biases most analyses of conflict. Disputes are coded using two logics—the familiar state-to-state militarized action represents one case while the second relies on sponsor governments to protest state targeting of private citizens. We show that the latter introduces additional measurement bias and does not match well the original conceptualization of what constituted a dispute. The protest-dependent cases are caused by different processes, and omitting them from analyses provides truer estimates of the effects of most conflict predictors. We find that previous controls for heterogeneity in the dispute data—such as using fatal militarized interstate disputes only—substantially underestimates the dangerous effects of contiguity and the pacifying effects of regime similarity. We also show that governments are seldom willing to risk militarized conflict for private citizens during these unique cases. We provide a list of the protest-dependent cases for future conflict analyses.

Palmer, D’Orazio, Kenwick, and McManus (PDKM) review a minority of changes we made to the CoW MID data, agree with most of those changes, but then argue that no scientific findings will be influenced by these or the thousands of changed values they ignore. We respond in two ways. First, we point out that principles of good science dictate that we discard dispute cases that either cannot be substantiated by the historical record or that do not meet dataset coding rules; we also discuss the ten-thousand-plus changed values PDKM still have not examined. Second, contrary to PDKM, we demonstrate how data differences greatly affect previous studies by returning again to the replications. There were substantial errors in both of PDKM’s replications, and these errors led them to incorrect inferences.