Title: bigIRT: Fitting Item Response Theory Models to Large Scale Binary
Data

Author: Charles Driver

Abstract:
bigIRT is an R package built for fitting item response theory models to large
scale, binary data. Its development was motivated by the need to calibrate the
Mindsteps online learning system, used in Swiss schools, with approximately
300k students assessed multiple times over years on small subsets of the 50k
total items, resulting in extremely sparse data. Estimation is based around
variations of joint maximum likelihood, with empirical Bayesian and
approximate integration options included, and item, ability, and response
level covariates. In this talk I'll discuss the software, some results based
on the Mindsteps data, as well as algorithm details, comparisons to other
software, and some challenges and open questions.