Three Models of Retirement: Computational Complexity Versus Predictive Validity
NBER Working Paper No. 3558 (Also Reprint No. r1760)
Empirical analysis often raises questions of approximation to underlying individual behavior. Closer approximation may require more complex statistical specifications, On the other hand, more complex specifications may presume computational facility that is beyond the grasp of most real people and therefore less consistent with the actual rules that govern their behavior, even though economic theory may push analysts to increasingly more complex specifications. Thus the issue is not only whether more complex models are worth the effort, but also whether they are better. We compare the in-sample and out-of-sample predictive performance of three models of retirement -- "option value," dynamic programming, and probit -- to determine which of the retirement rules most closely matches retirement behavior in a large firm. The primary measure of predictive validity is the correspondence between the model predictions and actual retirement under the firm's temporary early retirement window plan. The "option value" and dynamic programming models are considerably more successful than the less complex probit model in approximating the rules individuals use to make retirement decisions, but the more complex dynamic programming rule approximates behavior no better than the simpler option value rule.