A Latent Transition Model for Structured Interviews and Surveys With Screening Questions and Skip Patterns

• In Event: New Twists on Latent Transition Analysis: Unique Adaptations and Applications

Abstract

Structured interviews are often used in educational and behavioral assessments when the measurement strategy calls for a sample survey and when essentially the same information must be obtained from numerous participants. These structured assessments are preferred when the researcher requires greater control than is afforded by semi-structured or unstructured interview. The degree to standardization imposed on the instrument also improves inter-rater and test-retest reliability. With highly structured questionnaires, the questions to be asked and allowable responses are completely predetermined. However, when these questionnaires are embedded in a larger study, the actual set of questions asked may vary from participant to participant, with some questions asked only if there is a specific response to a particular question. These possible skip patterns have the advantage of reducing average assessment time and subject fatigue while improving overall participation rates.
Once the data are collected from a structured survey with skip patterns, they are typically run through a deterministic algorithm that computes a total score based on observed responses and corresponding skip patterns. These algorithms are not stochastic-based models and they generally make assumptions about the values of missing responses in those skip patterns. For example, in a structured interview for depression, a smaller set of screening items might be used at the beginning of the assessment, with only the subset of participant endorsing one or more of those screening items being administered any additional depression questions. The diagnostic algorithm would assume that those individuals “screened out” at the beginning would have reported zero levels of all symptoms and normal levels of functioning had they been ask those other questions. The algorithm does not allow for measurement error on the screening items nor does it allow for a non-perfect relationship between screening-out and missing response values in the skip patterns.
This paper presents a novel application of latent transition analysis (LTA) designed to address the limitations of deterministic algorithms applied to structured survey data with skip patterns. Each “period” or “segment” of the LTA represents a subset of items from the assessment grouped according to planned skip patterns. The latent class variable for each segment includes a latent “screened out” class that is measured imperfectly by the corresponding screening questions for that interview segment, and the transition probabilities from one segment to another are freely estimated rather than fixed. To illustrate this approach, we compare the algorithm-based dichotomous diagnostic categorization of the depression module in the World Mental Health Composite International Diagnostic Interview (WMH CIDI) in the United States with the latent structure of responses to WMH CIDI items. The LTA approach reveals a larger group of individuals with experiences of lifetime depressive states than is estimated by the algorithm-based diagnosis: 26% as opposed to 13%.

Authors

• Pamela Scorza, Columbia University

• Katherine E. Masyn, Harvard University