Methodology review: Nonparametric IRT approaches to the analysis of dichotomous item scores

Journal/Book: Appl Psychol Meas. 1998; 22: 2455 Teller Rd, Thousand Oaks, CA 91320. Sage Publications Inc. 3-31.

Abstract: This paper reviews developments in nonparametric item response theory (NIRT). Historically, Mokken's (1971) item response theory (IRT) model and procedure of scale analysis have been the starting point for MRT models; the developments that Mokken's work have inspired and those that originated independently are discussed here. Attention is focused on results pertaining to dichotomously scored items. The usefulness of new theoretical results for practical test construction is explained. NIRT is especially useful when the purpose of test application is to order persons, items, or both. The assumptions inherent in several NIRT models, model-data fit methods, item selection procedures, and reliability methods are discussed. NIRT models often allow more items into a test than stronger models, yielding higher reliability and better coverage of the latent trait. Moreover, the theoretical results from NIRT are often relevant to parametric IRT. Examples include conditional association and manifest monotonicity, as well as recent results for polytomous NIRT models that concern the ordering of persons, which also hold for several parametric IRT models.

Note: Review Sijtsma K, Tilburg Univ, Fsw, Vakgrp Methoden Tech Onderzoek, POB 90153, NL-5000 Le Tilburg, NETHERLANDS

Keyword(s): automated item selection; essential unidimensionality; invariant item ordering; model of double monotonicity; model of monotone homogeneity; Mokken model; nonparametric item response theory; ordinal person measurement; LATENT TRAIT UNIDIMENSIONALITY; MONOTONE LIKELIHOOD RATIO; RESPONSE THEORY; MOKKEN SCALE; CONDITIONAL ASSOCIATION; ESSENTIAL INDEPENDENCE; ABILITY ESTIMATION; MODEL; RELIABILITY; PATTERNS

Zurück | Weiter