Research

Does launching more drones overwhelm air defenses? Current US defense strategy assumes that affordable mass will saturate adversary interception capability, yet little empirical evidence supports this claim. Using 1,941 observations of Russian drone and missile attacks against Ukraine from September 2022 to November 2025, I estimate beta-binomial regression models to test whether larger attacks increase per-unit weapon survivability. I find that mass does not improve per-unit survival for Shahed drones; launching more drones is associated with a lower proportion surviving interception, suggesting that Ukrainian defenses scale effectively with attack size. Weapon type is the dominant predictor of survivability, with ballistic missiles and SAMs surviving at rates above 85% while loitering munitions are routinely intercepted. Theory-driven phase breaks reveal that Russia's adoption of saturation tactics in September 2024 initially tripled Shahed survival odds, but Ukraine's deployment of interceptor drones in March 2025 has progressively restored interception capability. These findings complicate the airpower coercion literature by demonstrating that the physical delivery of airpower, which is a prerequisite for both punishment and denial strategies, is not guaranteed and does not scale linearly with the number of weapons launched.

Do drone strikes reduce terrorist violence? Early studies of the US drone campaign in Pakistan generally answer yes, but these findings depend on research designs with strong assumptions about targeting decisions and baseline violence levels. This paper reassesses the evidence using Bayesian causal forests, a flexible estimator designed for observational settings with strong confounding and heterogeneous effects. We first reanalyze two influential studies, Johnston and Sarbahi (2016) and Mir and Moore (2019), and show that their original negative estimates are not recovered under BCF specifications that separate baseline violence from treatment effects. We then extend the analysis to a broader Pakistan panel of group-location-months from 2010 through 2014, using ACLED event data and an explicit selection-on-observables design. Across alternative geographic scopes, outcome definitions, and treatment windows, posterior intervals generally include zero and the estimated treatment effects do not consistently point toward reductions in violence. Unit-level treatment-effect estimates likewise provide little evidence that null average effects conceal a large subgroup of strongly negative effects. The results suggest that the optimistic empirical case for drone-strike effectiveness in Pakistan is more fragile than the early literature implies.