This paper revisits the latest statistical evidence for the nuclear emboldenment thesis—nuclear-armed states are more likely to initiate military aggression than non-nuclear states—from (Bell and Miller 2015). If correct, their findings have important theoretical and policy implications regarding the effect of nuclear proliferation on international conflict. This paper shows, however, that Bell and Miller’s findings heavily rely on two important components of their statistical analysis: (1) using all state dyad observations, and (2) employing pooled regression models to analyze time-series-cross-sectional (TSCS) data. I argue that those components are based on questionable assumptions on heterogeneity in their dataset. Based on alternative strategies dealing with heterogeneity in dyadic data, my reanalysis shows that the emboldening effect of nuclear weapons is not as robust as originally claimed. Instead, I find the robust deterrent effect of nuclear weapons: nuclear-armed states are less likely to be targeted in military disputes. These findings highlight the need for careful application of quantitative methods to produce a more robust understanding of nuclear issues.