We examine the determinants of low income transitions using first-order Markov models that control for initial conditions effects (those found to be poor in the base year may be a non-random sample) and for attrition (panel retention may also be non-random). Our econometric model is a form of endogeneous switching regression, and is fitted using simulated maximum likelihood methods. The estimates, derived from British panel data for the 1990s, indicate that there is substantial genuine state dependence in poverty. We also provide estimates of low income transition rates and lengths of poverty and non-poverty spells for persons of different types.