Efficient Estimation of Mortality Rates Using Micro and Macro Data

Dynamic discrete choice models are a well established and widely used methodology to study behavior of older individuals. A key aspect of the analysis is to characterize life expectancy and how it is affected by behavioral choices. Researchers, therefore, need to estimate conditional mortality rates to implement these estimators. However, publicly available data sets which follow older individuals are often not large enough to get reliable estimates of mortality rates. Hence estimates of conditional mortality probabilities may not be informative since they have large estimated standard errors. Imbens and Lancaster (1994) have recently proposed a solution to this problem. The key idea is to obtain a more efficient estimator by combining panel data with aggregate data. Following this approach, we estimate qualitative response models of mortality rates by combining panel data from the Health and Retirement Survey with aggregate data from U.S. life tables. Our empirical results show that the Imbens and Lancaster estimator achieves significant efficiency gains over simpler estimators which ignore macro data. We also find that the estimated coefficients of the mortality model change significantly as we add additional orthogonality conditions based on life tables in estimation. These finding supports our conjecture that estimators based on simple qualitative response models may be subject to small sample bias. Finally, we illustrate that the improvements in estimation of mortality probabilities can have significant consequences for evaluating public policies. We consider simple life-cycle computations of health care expenditures associated with smoking and heavy drinking.