Item Response Modeling of Presence-Severity Items: Application to Measurement of Patient-Reported Outcomes
Ying Liu and Jay Verkuilen
Presence-Severity (P-S) items contain two parts: the Present part which is used to check whether the respondent experiences the particular event in question, and the Severity part which is asked about the frequency, density, severity or impact of the event if the answer to the Present part is yes. One common strategy is to combine the parts by appending the Present part to be the bottom category of the Severity part. Another strategy is to analyzed only the Presence component using a binary IRT model and ignore the Severity part. An IRT analysis that does not ignore the Severity part might use ordinal IRT models such as GRM or GPCM. However, the ordinality is assumed to be held when applying the models.
The P-S format item creates systematically missing responses. Such missingness induces conditional dependence between the Presence and Severity parts (referred as conditional dependence by design in this paper).
In this study, the P-S format is transformed to a compound item with multiple categories (just like item bundle) and analyzed by the Nominal Response Model (NRM). It was argued that if the categories are ordering (for example, the a parameter of the second category in a compound item a1 is larger than the a parameter of the first category in a compound item a0), then treating the Present part as the bottom category of the Severity part is reasonable. To determine whether a1>a0, a Wald test for (a1-a0) is used. in addition, category response functions (CRFs) were plotted.
An empirical example which consisted of 27 physical and 4 psychological items were analyzed with NRM and GPCM. It was found that NRM fit better than GPCM for both subscales. However, it was found that 5 of 27 physical items were have (a1-a0) are significantly smaller than 0, indicating that for these items, it is not appropriate to treat the Present part as the bottom category of the Severity part.
The analysis of P-S format were compared between (1) using the compound item using NRM, (2) jus making use the Present part and treating them as dichotomous items. It was found that the first approach yields larger information and smaller standard error.
It was discussed that the P-S form may not be efficient and informative. Sometimes it is better to use more typical Likert-type items. Furthermore, the multidimensionality in P-S form is discussed.
Questions & Comments
1. Mapping the P-S format to compound item may not be practical when the categories the Present and Severity increase. For example, when there are 5 categories in the Present and Severity, respectively, it may generate a compound item with many categories.
2. Based on the result, it was discussed that P-S format item may not be efficient and informative and thus should consider to abandon them. However, given the rationale of P-S format, such as it can reduce respondent burden, a new class of IRT mode is needed.