Conflict forces millions of individuals from their homes each year. Using a simple structural model and new refugee data, we produce the first set of estimates relating outflows to annual conflict magnitudes. The theory underlying the structural model implies that standard panel data approaches will underestimate the impact of conflict violence, by differencing out the effect of prior and expected levels of violence on the decisions to flee. We estimate that whereas a shock that doubles conflict deaths in one year increases outflows in that year by 40% on average, doubling conflict deaths in all years increases annual outflows by 100%. We further estimate an average of 30 refugees per conflict death (median 18), with higher rates for conflicts closer to an OECD country and possibly for ethnic wars and in lower income countries. The analysis illustrates a broader methodological point: It can be hazardous to try to identify a causal effect using shocks to a presumed causal factor if the outcome variable is the result of decisions based not only on shocks but also on levels.