Modeling and imputation of semicontinuous survey variables

Semicontinuous variables have a proportion of responses at some fixed value and a continuous distribution among the remaining responses. Variables of this type occur in economic surveys of individuals or establishments (e.g. specific types of income or expenditures) where distributions are frequently characterized by a mixture of zeros and continuously distributed positive numbers. In this paper, we review strategies for joint statistical modeling and imputation of semicontinuous survey variables. Algorithms are presented for parameter estimation in the presence of unit and item nonresponse and for the imputation of missing values. Methods and software are demonstrated on data from a variety of sources including the Consumer Expenditures Survey.