Bayesian Cross-Sectional Analysis of the Conditional Distribution of Earnings of Men in the United States, 1967-1996

This study develops practical methods for Bayesian nonparametric inference in regression models. The emphasis is on extending a nonparametric treatment of the regression function to the full conditional distribution. It applies these methods to the relationship of earnings of men in the United States to their age and education over the period 1967 through 1996. Principal findings include increasing returns to both education and experience over this period, rising variance of earnings conditional on age and education, a negatively skewed and leptokurtic conditional distribution of log earnings, and steadily increasing inequality with asymmetric and changing impacts on highand low-wage earners. These results are insensitive to several alternative nonparametric specifications of distribution of earnings conditional on age and education. Acknowledgement 1 Grant R01-HD37060-01 from the National Institutes of Health provided financial support for this work. Much of applied statistics and econometrics is concerned with the measurement and interpretation of conditional distributions. In statistics the core curriculum devotes substantial time to regression, a topic that practically defines elementary econometrics and is the main point of departure in advanced treatments. Typically the conditional distribution is that of a univariate random variable y conditional on a random vector x. Regression, narrowly defined, is concerned only with E (y | x), yet this topic alone is the basis of a huge literature in mathematical statistics and theoretical econometrics. The strong assumption that regression is linear in x, made in