Unemployment durations are generally modelled by specifying the conditional probability of leaving unemployment (the hazard function). Existing studies for Britain all use very restrictive parametric specifications of the hazard function, most commonly Weibull in form. These restrictions potentially bias the estimated effects, particularly those of the time-varying economic variables and the baseline hazard. This paper investigates models for the probability of leaving unemployment with these restrictions removed. We use semi-parametric methods to estimate models with completely unrestricted baseline hazards and a model involving flexible step-function approximations to the baseline hazard. The Weibull is found not to give a satisfactory representation of the baseline hazard and its use is found to distort the pattern of unemployment income effects over the length of the spell. The existing studies for Britain also model the exit probability from unemployment rather than the probability of entering a job, despite generally interpreting the evidence as being about the latter probability. We use a competing risks model to distinguish exit into employment from exit into alternative states. We find that the single risk model of exit understates the effects of income in and out of work on the probability of entering a job.