This paper investigates formation of expected longevity in an elderly population. We use Italian data from the early (2004) release of the Survey of Health, Ageing and Retirement in Europe (SHARE). The SHARE provides a numerical measure for subjective survival probability (SSP). To assess internal consistency and investigate validity of SSP as a proxy of actual mortality, we compare SSP to lifetables and look at the variation with health, smoking and socio-economic variables. In a multivariate framework, we propose a recursive model for expected longevity, self-assessed health and smoking duration, where health and smoking variables are potentially endogenous. Unobservable individual-specific heterogeneity is considered by estimating a finite mixture model via the EM algorithm, which allows division of the population according to different latent classes and estimation of class membership probabilities. Our mixture model fits the data better than the single class model and provides evidence of individual unobserved heterogeneity in the formulation of survival expectations. Expectations are shown to vary most with health status, socio-economic characteristics, parental mortality and age. Two-types of individuals in the population are identified, that differ in terms of unobservable frailty and rationality in addiction. We also find differences between current and former smokers in the way they discount future consequences of tobacco consumption on health and mortality risk. Our findings suggest caution in the use of SSP as a proxy of actual mortality.