The author apologizes for an error in the coding of the variable “Percentage of Capabilities Held by State A.” This variable appears in all of the main regressions and, therefore, correcting it changes all of these regression results. Most of these changes do not affect the substantive conclusions, but this corrigendum presents the correct results for all tables and graphs in the order of the paper and appendix.
Table 2 should be replaced with the following:
Table 2. Main results
Note: Models 1–2 are probits predicting MID Initiation, with standard errors clustered by dyad. Models 3–4 are probits predicting Reciprocation, with standard errors clustered by State A. The madness reputation variables are lagged by one year. *p < 0.10, **p < 0.05, ***p < 0.01.
One change to the text is necessary due to the changes to the models in Table 2. In the discussion of the results, the second part of the sentence, “In contrast, leaders with slight madness reputations are only 1.4 times more likely to be targeted than leaders with no madness reputation, and this difference is only significant at the 94 per cent confidence level,” should be removed because this difference now surpasses the 95 per cent confidence threshold.
The predicted probabilities in Figure 1 also change, although the changes are difficult to discern with the human eye. The corrected Figure 1 appears below:
Figure 1. Predicted probabilities from Models 2 and 4
Note: the figure shows average predicted probabilities, produced by calculating the predicted probability for every observation and averaging. The lines represent 95 per cent confidence bounds.
The changes to Figure 2 are much more substantive. Figure 2 should be replaced with the following:
Figure 2. Marginal effects from interactions with relative capabilities
Note: these are average marginal effects. The dotted lines represent 95 per cent confidence bounds. The histograms in the background show the distribution of relative military capabilities.
As a result of this correction, the text in the “Military capabilities interaction” section should be replaced with the following:
“Hypothesis 4 can be tested by interacting the madness reputation indicators with relative military capabilities – specifically, with the percentage of military capabilities in the dyad held by State A. Figure 2 shows marginal effects plots from the interactions. First, analyzing the deterrence regression plots in the top row of Figure 2, the marginal effects of both Strong Madness Reputation and Slight Madness Reputation on the probability of deterrence failure increase as Leader B's relative power weakens (i.e., the percentage of capabilities held by State A increases). This does not support Hypothesis 4.
Turning to the bottom row of Figure 2, we analyze the crisis bargaining regression. The marginal effects from this regression provide support for Hypothesis 4. As State A's military capabilities increase relative to State B, Strong Madness Reputation has an increasingly large detrimental effect on the probability of reciprocation, while the effect of Slight Madness Reputation changes from helpful to insignificant. Therefore, the support for Hypothesis 4 is mixed, which might reflect a greater commitment problem within dyads that are already experiencing conflict.
Figure 2 also shows that perceived madness can actually be beneficial, as the bottom right graph shows that Slight Madness Reputation has a negative and significant marginal effect on the probability of MID reciprocation when State A holds less than about 40 percent of capabilities in the dyad.”
Due to the changes to Figure 2, some changes to the text in the introduction and conclusion are also necessary. In the introduction, the text beginning with, “I find that perceived madness is clearly harmful to general deterrence,” should be amended as follows (with changes and additions in bold):
“I find that perceived madness is clearly harmful to general deterrence and typically has a harmful or insignificant effect in crisis bargaining. However, it may be helpful in crisis bargaining under certain conditions, particularly when the reputation for madness is slight and is coupled with relatively weak military power. My analysis suggests that the harmful effect of perceived madness, at least in the context of crisis bargaining, results from a commitment problem.”
The first two paragraphs of the conclusion should be amended as follows (with changes and additions in bold):
“Overall, my findings suggest little support for the Madman Theory. For general deterrence, the effect of perceived madness is purely harmful. In crisis bargaining, the effect of a strong madness reputation seems to be generally harmful or at least unhelpful, but it does appear that a slight madness reputation can be beneficial when a leader's country does not have too much relative military power. In sum, therefore, the effect of perceived madness is more often harmful than helpful. The main apparent beneficiaries of a madness reputation are militarily weak leaders who are perceived as only slightly mad – not necessarily those we would be most likely to think of as ‘madmen’.
Why does a reputation for madness often undermine coercive success? My findings suggest that the inability of perceived madmen to make credible commitments to peace may be key. I find that greater relative military power, which increases the commitment problem, causes the impact of a madness reputation to become more detrimental in crisis bargaining (although not for general deterrence). This suggests that when a reputation for madness prevents a leader from credibly committing not to attack in the future, adversaries are more likely to resist the leader firmly. Thus, my findings are in line with research that emphasizes mistrust and the commitment problem as causes of war.”
Finally, most of the tables in the appendix have changed slightly, although none of these changes affect the conclusions drawn in the text of the article. The corrected tables appear below in order:
Table A2: Summary Statistics from General Deterrence (MID Initiation) Dataset
Table A3: Summary Statistics for Crisis Bargaining (MID Reciprocation) Dataset
Table A5: Observations with the Highest Values of Cook's D in the Initiation Regression
Table A6: Observations with the Highest Values of Cook's D in the Reciprocation Regression
Table A7: Alternate Indicator Cutoffs and Dropping Outliers (Initiation Model)
Table A8: Address Regional and Time Bias (Initiation Model)
Table A9: Address Pro-Western Bias and Strategic Use of Madness Adjectives (Initiation Model)
Table A10: Address Potentially Confounding Leader and Country Characteristics (Initiation Model)
Table A11: Adjustments to the Madness Measure (Initiation Model)
Table A12: Different Sample and Dependent Variables (Initiation Model)
Table A13: Interaction (Initiation Model)
Table A14: Alternate Indicator Cutoffs and Dropping Outliers (Recip Model)
Table A15: Address Regional and Time Bias (Recip Model)
Table A16: Address Pro-Western Bias and Strategic Use of Madness Adjectives (Recip Model)
Table A17: Address Potentially Confounding Leader and Country Characteristics (Recip Model)
Table A18: Adjustments to the Madness Measure (Recip Model)
Table A19: Dropping Some MIDs (Recip Model)
Table A20: Interaction (Recip Model)
Table A21: Conflict Selection Model (Recip Model)
Table A22: Interactions with Relative Capabilities
Table A23: Tests Mentioned in Footnotes (Initiation Model)
Table A24: More Tests Mentioned in Footnotes (Initiation Model)
Table A25: Tests Mentioned in Footnotes (Recip Model)
Table A26: More Tests Mentioned in Footnotes (Recip Model)
