Title: Practical Higher-Order Inference on Person Parameters in IRT Models

Authors: Ruggero Bellio

Affiliation: University of Udine

Abstract: 
Biehler, Holling & Doebler (2015, Psychometrika) proposed the usage of the
saddlepoint approximation for inference on person parameters in 2PL models.
The saddlepoint approximation is remarkably accurate even for samples with
limited number of items, overcoming the inaccuracy resulting from large-sample
methods. Both median-unbiased point estimation and accurate confidence intervals
are made available by the theory. The drawback of this approach is that some
computational complications arise, and this may discourage its usage. This
fact is even more pronounced for IRT models other than the 2PL one, where the
required analytical expressions may become rather intricate. 

This talk shows how to extend the saddlepoint approximation to general IRT
models by applying the modern theory of likelihood asymptotics, and illustrates
practical application by means of the R package likelihoodAsy, available on the
CRAN repository. The talk mainly covers the case of fixed item parameters,
mentioning in passing how to relax such assumption.