Latent Variable Modeling for Unipolar Constructs in Health Sciences : Partially Ordered Scaling Procedure for Including Reference Population

Edward Ip, Shyh-Huei Chen

DOI: 10.64028/vxdo529620

Keywords: patient reported outcome, IRT, latent class model, depression, graded response, poset

TL;DR: Exploring partially ordered scaling procedure for unipolar constructs in PROs

Abstract: Patient-reported outcome (PRO) data have gained increasing prominence in both medical research and FDA-related regulatory spaces. Inherited from the traditional measurement paradigm, PROs are structured to measure bipolar traits—i.e., traits that have meaning at both ends of the scale. However, in a medical context, certain constructs, such as depression and alcoholism, manifest as unipolar traits, meaning that the trait is only meaningful at one end of the distribution but not the other. For example, a low score signifies the absence of a quality (e.g., not alcoholic) rather than a relatively lower degree of that quality (i.e., less alcoholic) when compared to others. Methods such as zero inflation may not be well-suited for modeling unipolar traits because nonzero low scores, may simply indicate the absence of the condition. One important implication of unipolarity is whether the general population (e.g., the “normal” or “healthy” population) should be included in the calibration process. This decision may not be obvious for researchers aiming to develop PROs targeting patients with specific conditions. Both inclusion and exclusion of the general population present distinct advantages and limitations. For example, including the general population allows for interpretation of the scale using the general population as a reference. In addition to the issue of general population inclusion, individual items should also be evaluated for their relevance to the general population. In this article, we explore methods for addressing unipolarity using a partially ordered set (poset) item response theory (IRT) model. This model decomposes response categories into two components: (1) the lowest category (e.g., "Never"), which is considered qualitatively distinct from the other categories, versus all other categories and (2) the remaining categories (e.g., "Rarely" to "Always") as ordered categories. Poset calibration is performed using a unidimensional IRT approach, which assumes that a common underlying latent trait drives both components. A real dataset (n=653) of breast cancer survivors containing depression data was used to illustrate the exploratory analysis. Scenarios both including and excluding "non-depressed" patients were examined using the poset IRT and the graded response model (GRM). This study also highlights the feasibility of the approach and explores potential extensions, such as incorporating a "Not Applicable" (NA) option and conducting multidimensional analyses.

Submission Number: 27