PAUL EVANS (*)
The reform of the U.S. social security system is now attracting much attention. A wide range of proposals for reform have been put forward, ranging from modest modifications to complete privatization. (1) In choosing among these reforms, it is important to quantify how they would affect the aggregate economy. A given reform is presumably more likely to command support, the more favorable its aggregate effects are on net.
In recent years, economists have typically quantified the effects of social security reform in large computational general-equilibrium models based on the lifecycle hypothesis; for example, Kotlikoff, Smetters, and Walliser (1998). Although the calibration of these models does make use of data, the models are not designed to fit the aggregate data in any very well-defined sense. Rather, the lifecycle structure of the model is simply imposed, and the parameters are chosen in order to match a few select moments of the aggregate data. The lifecycle hypothesis is not the only possible model of consumer behavior, however. Indeed, like all models, it abstracts from many empirically important phenomena. (2) Moreover, computational general-equilibrium models have not typically been validated on data other than those used for calibrating them. (3) One might therefore reasonably doubt the empirical validity of the effects calculated in such models.
An older literature attempted to estimate the effects of social security directly. The seminal study in this literature is Feldstein (1974), who reports that ceteris paribus, social security increases consumption and lowers saving substantially. As a result, the balanced growth paths for capital, output, and consumption are lowered substantially.
Many other studies followed. (4) They reported mixed results, with perhaps a presumption that social security raises consumption ceteris paribus.
For the most part, this literature investigated the issue by fitting "consumption functions" of the form
(1) [C.sub.t] = [[beta].sub.0] + [[beta].sub.1][y.sub.t] + [[beta].sub.2][a.sub.t] + [[beta].sub.3][S.sub.t] + [[beta].sup.4]'[x.sub.t] + [e.sub.t];
where [C.sub.t] and [y.sub.t] are per capita consumption and disposable income during period t; [a.sub.t] and [S.sub.t] are per capita private wealth and social security wealth at the beginning of period t; [x.sub.t] is a vector of other variables realized during period t; [[beta].sub.1], [[beta].sub.2] and [[beta].sub.3] are parameters; [[beta].sub.4] is a vector of parameters; and [e.sub.t] is an error term. (5) If [[beta].sub.3] turns out to be positive and statistically significant, social security is inferred to raise consumption and lower saving ceteris paribus and to lower the balanced growth paths for private wealth, output, and consumption. Unfortunately, equation (1) is at best an approximate reduced form of the lifecycle model and at worst bears no relationship to any well-known structural model (see Auerbach and Kotlikoff ). If it can be interpreted as a good approximation to a reduced form, its parameters confound structural and expectational parameters. For example, even if Ricardian equiv alence holds so that social security is completely neutral, [[beta].sub.3] can be positive if social security wealth is positively correlated with future disposable wage incomes, conditional on the other regressors. Conversely, even if [[beta].sub.3] is estimated to be insignificantly different from zero, social security may still increase consumption appreciably since social security wealth may be negatively correlated with future disposable wage incomes, conditional on the other regressors. In principle, these problems could be overcome by estimating a variant of equation (1) jointly with auxiliary forecasting equations for the components of future disposable wage income while imposing the cross-equation restrictions implied by rational expectations. …