We present the first evidence of voter-level responses to a climatic disaster — the catastrophic German flooding of 2021, which serves as a natural experiment. Data on previous voting history reveals non-monotonic treatment effects: flood exposure increased the likelihood of voting for the Green Party by four to five percentage points among previous non-Green voters, but decreased future Green voting for previous Green voters. Tracking migration also reveals heterogeneity. Movers-out of flood zones responded more strongly; classifying them in the control group – as geographic panels do – attenuates the treatment effect. Both factors rationalize past findings of null or small effects, emphasizing the importance of microdata.