A Multidimensional Zero-Inflated Graded Response Model for Ordinal Symptom Data

Zero responses and their equivalents-for example, never, none, not at all-are commonly observed on measures of psychopathology inquiring about symptom frequencies, particularly when these measures are administered to community samples. Measurement researchers typically accommodate multivariate zero inflation by including a nonpathological class of respondents who endorse zero for all symptoms. While this latent class approach accounts for test-level zero inflation (i.e., a proportion of individuals who do not experience any of the symptoms), it may be overly restrictive on questionnaires comprising items of differing severity. For example, an item about suicidal ideation is likely to exhibit a much higher degree of zero inflation than an item about low energy. Existing models do not account for this variability. We propose a multidimensional zero-inflated graded response model (MZI-GRM) as a more flexible approach for modeling zero inflation on questionnaires. According to the model, two distinct but correlated latent variables underlie ordinal item responses; one represents susceptibility to the construct, whereas the other represents severity. As a motivating example, we show how the MZI-GRM can be fit to data from the PHQ-9, a common depression screener. Results suggest that the MZI-GRM is better able to capture zero inflation across items than existing alternative models. Further, we find support for a multidimensional model that allows distinct but correlated latent variables to underlie each response process. Some items better measure susceptibility to depression (symptom presence), whereas others better capture severity of depression (symptom frequency). Implications for scale development and scoring are discussed. Translational Abstract Zero responses and their equivalents-for example, never, none, not at all-are commonly observed on measures of psychopathology inquiring about symptom frequencies, particularly when these measures are administered to community samples where symptom endorsement is rare. This leads to high levels of zero inflation, in which a large proportion of people report zero equivalents for many or all symptoms. Measurement researchers typically take zero inflation into account by including a nonpathological class of respondents who endorse zero for all symptoms; however, this approach may be overly restrictive on questionnaires comprising items of differing severity. For example, an item about suicide ideation is likely to exhibit a much higher degree of zero inflation than an item about low energy levels. Existing models do not account for this variability. We propose a multidimensional zero-inflated graded response model (MZI-GRM) as a more flexible approach for modeling zero inflation on questionnaires. According to the model, two distinct but correlated latent variables underlie ordinal item responses; one represents susceptibility to the construct (for example, susceptibility to depression, addiction, or anxiety), whereas the other represents severity (for example, given that one is susceptible, the severity of depression, addiction, or anxiety). As a motivating example, we show how the MZI-GRM can be fit to data from the Patient Health Questionnaire-9 (PHQ-9), a common depression screener. Results suggest that the MZI-GRM is better able to capture zero inflation across items than existing alternative models. Further, we find support for a multidimensional model that includes two distinct but correlated latent variables. Some items better measure susceptibility to depression, as indicated by symptom presence, whereas others better capture severity of depression, as indicated by symptom frequency. Implications for scale development and scoring are discussed.

Automated summary

Zero responses and their equivalents are common on measures of psychopathology, especially in community samples, resulting in high levels of zero inflation. Researchers propose a multidimensional zero-inflated graded response model (MZI-GRM) to better capture zero inflation on questionnaires by including correlated latent variables representing susceptibility and severity, showing better fit than existing models on capturing zero inflation across items.