In applied scientific research as well as in public policy, interest often centers on the influence of some observed characteristics (for example, an individual's age and education) on the probability of a particular event (for example, that an individual's earnings will exceed the poverty threshold). Moreover, these influences may change over time (for example, college graduates may earn more compared to high school graduates today than they did in the 1970's). This project develops quantitative methods for using existing databases to reliably assess the impact of observed characteristics on events that are important in public and private policymaking as well as applied scientific research.
The project extends recent advances in the fields of econometrics, statistics and computer science leading to assessments of the influence of observed characteristics on outcomes of interest that are more reliable and practical than has heretofore been possible. The project has three main components. First, it establishes new results in econometrics on the approximation of conditional distributions under weak conditions, using a sequence of models demonstrated to be practical and reliable in the recent research of the investigator, the smoothly mixing regression models of the title. Second, it addresses a series of practical issues in order to enhance the utility of these models, including the use of parallel computing environments to handle large data sets that are increasingly common in economics, the evaluation of competing models, and application of these procedures when several outcomes are simultaneously of interest. Third, it applies these developments in two leading contexts: the aforementioned example involving the influence of individual characteristics on earnings, using the Current Population Survey of the U.S. Census, and measuring the risk associated with investment in financial assets.
Broader impacts: One of the major themes in econometrics and statistics is the development of general methods that do not require restrictive assumptions yet give precise estimates and are computationally feasible. This project builds on the many past contributions by the investigator on this theme and should lead to a major advance in computationally feasible yet flexible methods. These methods will benefit research in a wide range of disciplines including economics, statistics, public health, biostatistics and environmental sciences. The project addresses questions that arise regularly in the Federal statistical system. The investigator rapidly disseminates and applies research findings and regularly interacts with a diverse group of students from academia, government and the private sector through intensive one-week courses taught in various locations.
