| Summary: | The first chapter in this dissertation discusses some of the empirical challenges in the Value of a Statistical Life (VSL) literature, and poses some conceptual concerns that can impact how VSL estimates are applied to policy. Identification concerns include bias due to time-invariant unobserved factors, sorting, and unobserved heterogeneity. Conceptual concerns include the application of the VSL to public risk, extrapolating estimates to the policy-relevant variation in risk when risk preferences are non-linear, and misperceptions of risk. The first chapter concludes with potential strategies to address the aforementioned empirical concerns.
The second chapter investigates the trade-off between bias due to measurement error and bias due to unobserved time-invariant factors in the Value of a Statistical Life (VSL) literature. The literature has tended towards addressing measurement error instead of unobservable factors, and is now starting to tackle the latter. Yet, the tools available to researchers are limited by the natural trade-off between bias due to measurement error and bias due to time-invariant unobservables; OLS can address one or the other, but not both. This chapter proposes a new approach that first tackles bias due to unobservables and then simulates and adjusts for bias due to measurement error because it is more easily understood. Using data on truck drivers, this chapter demonstrates this approach using three methods to adjust for measurement error. Without adjustment, VSL estimates are attenuated by 88% and groups that are sampled more than 30 times per year still do not converge to the true value. The measurement error is non-classical, and falsely treating it is as classical will retain some bias. However, classically adjusted VSL estimates converge to the true value when groups are sampled at least 15 times per year. Two methods of simulating measurement error provide consistent estimates as long as groups are sampled at least twice per year.
The final chapter quantifies the bias due to time-invariant unobservables in truck driver VSL estimates. The VSL literature has traditionally exploited variation across occupations, but this strategy is susceptible to bias from unobserved time-invariant factors. Additionally, many occupational death data are measured with error. This chapter addresses these two shortcomings by using truck driver panel data from 1983 to 2002 to bound VSL estimates. Exploiting within-occupation variation abstracts from unobserved between-occupation variation, including productivity shocks and worker risk tolerance. The panel structure allows this study to control for and estimate the effects of unobserved within- and between-occupation heterogeneity. This chapter bounds the VSL between OLS and Instrumental Variables (IV) estimates, because measurement error will attenuate OLS estimates and anti-attenuate IV estimates when another mis-measured risk estimate is used as an instrument. Risk measures are created from the National Highway Traffic Safety Administration---Fatality Analysis Reporting System (NHTSA-FARS), the census of all traffic fatalities on public U.S. roadways. The OLS-IV estimates bound the VSL between $0.4 million and $4.0 million. Within- and between-occupation heterogeneity bias estimates upwards at least $4.6 million, while aggregating workers across heterogeneous occupations biases estimates upward by at least an additional $1.0 million. Although truck drivers do not represent the U.S. labor force, this chapter illustrates the need to address time-invariant unobservables and measurement error concurrently.
|
|---|
