The Validity of Markov Latent Class Analysis for Evaluating Measurement Errors in Complex Panel Surveys

Research Triangle Institute, Durham NC

Investigators

Abstract

Markov latent class analysis (MLCA) comprises a broad class of models and techniques for analyzing categorical longitudinal data subject to misclassification. An important application area is exploring data quality issues in panel surveys. Because MLCA does not rely on gold standard or replicate measurements, it can be applied to virtually any panel survey. For data quality evaluations, MLCA has been used to compare interview modes and alternative questionnaire designs, estimate measurement bias, investigate the causes of misclassification, and investigate many other measurement error issues. Despite its many potential applications in survey work, MLCA has not enjoyed widespread use among survey methodologists because practical guidance on fitting MLC models to complex survey data is lacking. This project will: (1) evaluate the magnitude of the model bias when one or more MLCA assumptions fail when analyzing complex survey data under a wide range of conditions; (2) identify and evaluate the current strategies for diagnosing and repairing MLC model failure and misspecification; (3) address the limitations of current methods by developing improved strategies for diagnosing and repairing MLC model failure and misspecification, particularly in applications to complex surveys; and (4) apply the most effective diagnostic and remedial approaches to real panel survey data to demonstrate the range of modeling issues that can arise in practical applications as well as how to deal with them effectively. As part of the application of these approaches, at least 10 years of data from several national panel surveys will be analyzed to identify temporal trends in measurement error for key national statistics. This research has important implications for MLCA in all branches of science where classification error is an issue, including social science, epidemiology, clinical research, educational testing, and psychology. The project's impact will be felt in at least four ways. The research has particular relevance for complex survey applications because of the emphasis in this research on modeling cluster-correlated data selected with unequal probabilities and subject to nonresponse and measurement error. It also has important implications for disadvantaged and minority populations whose data may be differentially affected by measurement error. In addition, the evaluation of error trends will provide important information on current and historical levels of measurement error for three important federal statistical programs. Finally, theories regarding the relationship between measurement error and survey participation will be formulated and tested. The project is supported by the Methodology, Measurement, and Statistics Program and a consortium of federal statistical agencies as part of a joint activity to support research on survey and statistical methodology.