ECOCIDE: Quantifying War-Time Environmental Damage: A Difference-in-Differences Analysis of the Kakhovka Dam Destruction

SAKSHI D. MASKE 

Abstract:
International law organisations are actively looking into declaring "ecocide" -- mass environmental destruction -- as a prosecutable international crime and in September 2024, Vanuatu, Fiji and Samoa submitted a proposed amendment to Rome Statute. No consistent, statistically sound method yet has been developed to document such a claim of harm, however, because prior environmental trends are not easily differentiated from harm caused by conflict, existing satellite-based assessments of environmental war damage are qualitative and based on visual interpretations of images. In an attempt to fill this gap, I estimate the impact of the Kakhovka Dam's destruction on Ukraine and employ a causal-inference framework of Difference-in-Differences to those data. The design compares monthly NDVI in Kherson Oblast, a region generally affected by the conflict, against a "control" panel of four counties in Romania which are comparable on the pre-conflict ecology of the river deltas, steppe regions, and coastal areas along the Danube/Black Sea corridor but which are indeed not affected by the conflict. It is statistically significant with respect to the decrease in vegetation caused by the event (coefficient = −0.0703, 95% CI [−0.130, −0.010]; p = 0.022 [HAC-robust]). The result has been confirmed by a clean placebo-control test based on a counterfactual date preceding the event. As a robustness check, I run the same analysis at the level of the 4-county panel as a whole, obtaining a similar result (coefficient = −0.0600, HAC p = 0.029), and three of the four controls individually reproduce it, while the fourth, Constanța, does not — which I report as an open question rather than resolve away. In addition, there's this very real methodological challenge that I didn't anticipate, where I find after running the event study on a quarterly basis that there is a significant effect in the pre-treatment quarter, but this relates to a banally acknowledged 'conflict status' of conflict prior to the dam's destruction, where in fact lots of territory was already involved in conflict when the dam went up, which I report openly as a specification I know I'm dealing with, rather than hide and ignore, and a similar challenge shows up in a different variant the moment I run the event study for the four-county panel, where I find that a cluster-robust inference is problematic at five clusters. An independent verified flood extent for a flood rise-peak-recession cycle is obtained from the multi-sensor validated flood extent data (UNOSAT). I demonstrate that for this event, which has never before been the subject of a satellite-based environmental war-damage assessment beyond qualitative measures, causal-inference methods can supplement the existing qualitative approach with quantitative measures based on statistical significance — directly answering the call the existing literature itself makes for future research.