Methods for addressing aberrant responding in survey research

In this talk, I highlight two lines of research relevant for survey data consisting of multi-item scales. Part one focuses on confirmatory use of a multidimensional parameterization of the nominal response model. One primary application involves the modeling of response styles, such as a preference for endpoints or midpoint response options, which are thought to be problematic for the measurement of psychological constructs using Likert-type items. This model can be understood in part by starting from sum score-based approaches to studying response styles as well as generalized linear models such as (multinomial) logistic regression. A collection of findings related to this application will be presented along with extensions. Part two considers random responding behavior that may contaminate data collection efforts. A model-agnostic and unsupervised approach to detection of random responding is presented and evaluated in a series of simulation studies. Provided assumptions are met, this approach can calibrate sensitivity without making strong measurement model assumptions. Simulations leverage data from a well-known item response theory model as well as contamination of 35 publicly available datasets to understand the algorithm’s classification accuracy. Through this work, I identify conditions under which classification accuracy is also at or above 95%, but also conditions under which the algorithm should not be used. Extensions of the algorithm and future research is also discussed.