On the measurement of expectations, uncertainty, and preferences.

' T ' H E article in this issue by Holden, McBride, and J . Perozek (1997) on expectations of nursing home use reflects an increasingly important phenomenon in the economics literature: the use of direct measures of expectations, uncertainty, and preferences to model economic behavior. Examination of the mainstream economics literature over the past several decades shows two important but essentially inconsistent lines of thought, and one relatively recent body of work designed to remedy the inconsistency. One line of thought reflects the recognition that forward looking phenomena are increasingly important to understanding economic behavior, in that many of the recent models of saving, labor supply, schooling, occupational choices, insurance purchases, etc. are strongly affected by phenomena that are explicitly expectational. Yet until very recently expectations, which are critically important to the models, were not measured directly but typically were inferred from past behavior. For example, in the well-known and widely used permanent income model of saving behavior originating with Milton Friedman (1957), permanent income is defined in explicitly expectational and forward looking terms, but permanent income is operationalized in backward looking terms — estimates of expected future income are based solely on the trajectory of past income or other variables that can be observed both by the individual and by the analyst. The explanation for this apparently inconsistent treatment lies in the deep suspicion with which economists have always viewed statements made about the future by economic actors. The fear is that expectational statements are either frivolous or self-serving, hence unreliable, while actual behavior reflects real constraints and opportunities. Thus, the economics tradition has been that the best way to model future behavior is to infer it from past behavior. There are a great many difficulties with this view of expectations data. Perhaps the most serious problems were pointed out by Manski (1990 and 1993), who notes that modeling income expectations by past income realizations essentially assumes that the process of forming expectations is identical for everyone in the population, and that the analyst knows the complete shape of the function relating information to expectations. The traditional view has seemed to many an excessively restrictive view of how expectations are formed. As a result, there has been an emerging interest in modeling outcomes with the use of direct measures of expectations, as well as with direct measures of uncertainty and preferences. The interest in modeling with these "soft" subjective variables is due in part to the growing realization that these variables are needed for appropriately specified models, and also to the recognition that many such measures seem to yield unbiased (if noisy) estimates of actual behavior. The Holden, McBride, and Perozek piece is very much in the mode of analysis of expectational phenomena using recently available survey measures of expectations. The findings are characteristic of the emerging literature on the role of subjective phenomena in behavior modeling. Holden and colleagues find that the mean probability of entering a nursing home in the Health and Retirement Study (HRS), based on a subjective probability scale ranging from zero (absolutely no chance) to 10 (absolutely certain), looks very much like the actual nursing home experience of a comparable population. They also find that, while some of the characteristics of those who project higher probabilities of nursing home entry are consistent with actual differences, others are not. For example, Blacks have lower subjective probabilities of nursing home entry than Whites, a difference reflected in the actual nursing home experience of Blacks and Whites. But females have a substantially higher likelihood of nursing home entry than males, while the subjective data show about the same mean expectation for men and women. Still, Holden and colleagues find that many of the characteristics that ought to be associated with higher or lower probabilities of nursing home entry (above average mortality, the likelihood of disability given low mortality, experiences of parents, the presence of children, marital status, etc.) show a strong association between the subjective distribution of probabilities of nursing home entry and the actual pattern of nursing home entry. In short, the subjective probability data look as if they represent largely valid and unbiased estimates of future outcomes. One can find much the same characteristics in other studies that focus on the subjective probabilities of future events, as well as in measures that focus on the uncertainty