Unipolar IRT and the Author Recognition Test (ART)

Item response theory (IRT) analyses are often used to evaluate measurement error in educational and psychological test instruments. In such contexts, the latent traits/proficiencies are typically assumed normally distributed and a cumulative normal/logistic measurement link function is applied. Such choices are consistent with constructs that are viewed as bipolar in nature and play a critical role in defining the latent proficiency metric against which the measurement error in the test is evaluated. Recently, alternative models that portray the construct as unipolar have been highlighted as being more appropriate for certain psychopathology and personality constructs. In this paper we extend consideration of unipolar IRT models for a recognition task measure, using several example datasets from various versions of the Author Recognition Test (ART), a measure of print exposure. We show how the decision between unipolar versus bipolar IRT modeling has substantial implications for the quantification and interpretation of measurement error in the ART. In sharp contrast to prior bipolar IRT analyses of the ART, under unipolar IRT measurement error in the ART is minimized at low levels of latent print exposure, and increases as latent print exposure increases. Implications for consideration of unipolar IRT with other constructs and measures (e.g., vocabulary, specialized forms of knowledge) that reflect a similar type of response process are considered in the discussion.