Climate change is driving extreme weather, heat and flooding, which increasingly require evacuations. Recent studies have found inconclusive results on the determinants of evacuation-related decisions and have reported widely varying evacuation rates, especially in high-density areas. We use a large dataset of geotagged evacuation choices in Paris, France (n = 2976) during a flood, to show that while they are rarely addressed, location, social vulnerability, length of residence, and hazard exposure are critical predictors. They can be used to infer the impact of previous experience on evacuation intentions, which is usually difficult to collect at scales relevant to decision-making. We address multiple evacuation choices over time that have previously been overlooked, including gradually self-fueling spontaneous evacuations after observations of peers evacuating or flooding proximity and post-impact evacuation due to infrastructure disruption. Our findings reveal that many people wait until their home is flooded to evacuate. The gap between the initial share of people leaving immediately and the minority refusing to leave in any case even after their housing is flooded gradually fills when considering different evacuation behaviors over time. Such chronology might explain the wide range of evacuation rates reported in the literature. However, people in vulnerable situations and with increased hazard exposure are more likely to declare that they would disregard evacuation instructions. One key implication is that policies aimed at reducing social vulnerability might be more effective than communication campaigns to increase preparedness and support evacuation.