Identifying Mischievous Responders through Latent Class Analysis

Introduction: Self-report bias, particularly to sensitive questions (e.g. violent behavior), can occur due to  over-reporting of desirable behaviors or under-reporting of  undesirable behaviors. A common strategy to identify and control for social desirability bias is to use dedicated scales which add non-substantive questions to a survey. However, extending the length of a survey will boost data collection costs and may increase respondents’ fatigue. Moreover, these scales do not necessarily identify respondents who misreport for the ‘fun of it’. A more promising approach is to use person-centered latent variable models to identify groups of participants whose responses are suspicious. This presentation will illustrate the approach using data on sexual violence (SV) perpetration from a national sample of adolescent respondents in the Growing up with Media study.

 

Methods: The sample consists of 887 individual who participated in the last three waves (2010-2012) when extensive questions about SV were added. We focus on two age groups: 16-17 and 18-22. Respondents were asked about their past-year engagement in six types of SV perpetration: sexual harassment, online sexual harassment, attempted rape, rape, coercive sex, and sexual assault. Latent Class Analysis was used to empirically determine the number of latent profiles, focusing on the observed response pattern with respect to class assignment. Two general misclassification patterns are important: Those with “no perpetration” observed patterns assigned to a perpetration profile (under-reporting) and youth with a “perpetration” oberved pattern assigned to a no-perpetration profile (over-reporting).

Results: A three class solution was supported by the data: “no perpetration”, “sexual harassment-based perpetration”, and “varied types of perpetration”, with  above 0.8 classification accuracy for both age groups.

Results indicate some degree of over- and under- reporting. For example, for ages 16-17, three observed patterns of 781 youths indicated no perpetration. However, LCA estimated a  “no perpetration” prevalence of 756, with a 3% (N=24) misclassficaiton. For ages 18-22, two observed patterns of 711 youths indicated no perpetration. However, LCA estimated a “no perperation” prevalence of 706, with a 0.7% (N=5) misclassification.

 

Conclusions: A small, but potentially meaningful portion of responders potentially under-report their perpertation. The presentation will expand on these results by inspecting over-reporting patterns, and individual scores for those misclassified on salient construct (e.g., rape attitudes) and assessing the utility of incorporating an explicit measurement model,i.e., Factor Mixture Model.