Unexpected events during survey fieldwork are often used to estimate causal effects by comparing respondents interviewed before and after the event. This thesis argues that the credibility of these designs depends not only on whether the event was unexpected, but also on whether fieldwork rollout shaped the composition of respondents observed before and after the event. When rollout is geographically structured, the event cutoff may divide respondents by place and composition in ways related to potential outcomes. I examine this problem using survey data around the 2015 Garissa University College attack in Kenya. The case shows that standard design repairs, including restricting to common support and adjusting for covariates, improve comparability but do not necessarily eliminate rollout-related confounding. Except where stronger assumptions can be justified, unexpected-event survey designs belong in the world of selection on observables. Their advantage is not automatic as-if random timing, but diagnostic transparency: researchers can observe the fieldwork sequence, assess overlap, restrict claims to supported comparisons, and test whether similar estimates appear at non-event cutoffs.