Age-specific death rates with tobacco smoking and occupational activity: sensitivity to sample length, functional form, and unobserved frailty.

In this article, we estimate accelerated time-to-failure and proportional-hazard functions with about 100,000 members of the Dorn sample, finding greater hazards associated with smoking and some dependence on occupational variables that measure risk and physical activity. We answer three questions: (1) How sensitive are the estimates to sample length, using monthly data for the periods 1954-1969 and 1954-1980? The results differ somewhat between these sample periods. (2) How sensitive are the estimates to alternative functions for the hazard? Within a given time period, the estimates are fairly robust to specification changes in the distribution of the hazard in the accelerated time-to-failure models. (3) How sensitive are the estimates to alternative controls for unobserved frailty? Within a given sample period, the estimates are fairly robust to the allowance for parametric or nonparametric heterogeneity in the proportional-hazard models.