The Predictive Validity of Subjective Probabilities of Survival

Abstract

Although expectations, or more precisely stubjective probability distribtutions, play a prominent role in models of decision making uinder uincertainty, we have had very little data on them. Based on panel data from the Health and Retirement Stu-dy, we stu-dy the evolution of su-bjective stur-vival probabilities and their ability to predict actual mortality. In panel, respondents modify their survival probabilities in response to new information such as the onset of a new disease condition. Stubjective survival probabilities predict acttual survival: those who surnived in the panel reported sturTival probabilities approximately 50% greater at baseline than those who died. Individual behaviour depends notjust on the current state of the world but also on what individuals expect will happen in the future: individuals choose their desired years of schooling based on expected future income; engage in family planning based on expected fertility; and save for retirement based on their expected length of life. Researchers wishing to understand these behaviours must therefore incorporate expectations into their models.' Despite the importance of expectations, we rarely have data on the expectations of individuals and, therefore, have had to make assumptions about them. Often, individuals are assumed to face probabilities equal to the average population probabilities. For example, when survival probabilities are needed in economic models of savings behaviour, researchers use survival probabilities calculated from life tables. When forecasting future income, average incorne of a similar demographic group is often used. In many cases, however, average values will not be the appropriate measure for most individuals: subjective probabilities are likely to differ across individuals, and individuals will choose their behaviours based on their own probabilities. Studies which assume that individuals base their behaviours on population probabilities may therefore give misleading results. A potential improvement over population probabilities is to base models on an individual's own subjective probability.2 Such probabilities could be constructed