Bayesian Two-Part Tobit Models with Left-Censoring, Skewness, and Nonignorable Missingness

In a longitudinal HIV/AIDS study with response data, observations may be missing because of patient dropouts due to drug intolerance or other problems, resulting in nonignorable missing data. In addition to nonignorable missingness, there are also problems of skewness and left-censoring in the response variable because of a lower limit of detection (LOD). There has been relatively little work published simultaneously dealing with these features of longitudinal data. In particular, one of the features may sometimes be the existence of a larger proportion of left-censored data falling below LOD than expected under a usually assumed log-normal distribution. When this happens, an alternative model that can account for a high proportion of censored data should be considered. We present an extension of the random effects Tobit model that incorporates a mixture of true undetectable observations and the values from a skew-normal distribution for an outcome with left-censoring, skewness, and nonignorable missingness. A unifying modeling approach is used to assess the impact of left-censoring, skewness, nonignorable missingness and measurement error in covariates on a Bayesian inference. The proposed methods are illustrated using real data from an AIDS clinical study.