This paper analyzes the persistence of poverty in Sweden using a hazard rate model based on multiple spells. The model also accounts for unobserved heterogeneity and possibly endogenous initial conditions. We estimate the model on a large representative Swedish panel data set, LINDA, for the years 1991 to 2001. The data contains precise information on household disposable income obtained from individual tax files. Poverty is defined using information on annual minimum needs standards determined by the Swedish National Board of Health and Welfare. The data indicates that poverty rates are highest for immigrants, especially refugee immigrants, and for households with children. Further, poverty rates declined, both for natives and for immigrants, between 1991 and 2001, partly as a result of improved labor market conditions. The empirical results suggest that there is significant negative duration dependence in both exit and entry hazard rates. Moreover, the transition rates are significantly affected by immigrant status, educational attainment, labor market conditions, age, and family status. Accounting for multiple spells shows that for two-parent families with two children who are represented by a male person, 44 percent of native households that falls into poverty at any given point in time remain poor in five or more out of the next ten years. For refugee and non-refugee households, the figures are 62 percent and 50 percent, respectively.