Latent Class Analysis of Measurement Error in the Consumer Expenditure Survey

Previous research by Tucker et al. (2010), working with the Consumer Expenditure Survey (CE), explores the factor structure of measurement error indicators such as: interview length, extent and type of records used, the monthly patterns of reporting, reporting of income, attempt history information, and response behavior across multiple interviews in a latent class model. Findings from this research, using approximately 30,000 cases from 2005 to 2009, yielded models with slightly poorer fit and less efficacy in predicting household expenditure than models using data from 1996 to 2001 (Tucker et al., 2008). While a number of revisions have been made to the CE in the years since 2001 that may have resulted in less measurement error overall, the differences in model fit and efficacy of the latent construct are worthy of further investigation. In current research we add the use of the information booklet as a possible indicator of measurement error. In addition, we examine the context of the model in much greater detail, examining subgroups where superior model fit and greater efficacy of the latent construct is observed. The description of the context extends beyond characteristics of the responding household to include the mode of administration of the survey (telephone, in-person, likely cell phone), as well as other process variables and their interactions with each other and the latent class measurement error construct.