Multivariate normal distribution approaches for dependently truncated data

作者：Emura Takeshi; Konno Yoshihiko^*

来源：Statistical Papers, 2012, 53(1): 133-149.

DOI：10.1007/s00362-010-0321-x

• Correlation coefficient
• Truncation
• Maximum likelihood
• Missing data
• Multivariate analysis
• Parametric bootstrap

Many statistical methods for truncated data rely on the independence assumption regarding the truncation variable. In many application studies, however, the dependence between a variable X of interest and its truncation variable L plays a fundamental role in modeling data structure. For truncated data, typical interest is in estimating the marginal distributions of (L, X) and often in examining the degree of the dependence between X and L. To relax the independence assumption, we present a method of fitting a parametric model on (L, X), which can easily incorporate the dependence structure on the truncation mechanisms. Focusing on a specific example for the bivariate normal distribution, the score equations and Fisher information matrix are provided. A robust procedure based on the bivariate t-distribution is also considered. Simulations are performed to examine finite-sample performances of the proposed method. Extension of the proposed method to doubly truncated data is briefly discussed.

• 出版日期2012-2