skip to main content skip to footer

Significant Improvements to LOGIST IRT LOGIST

Author(s):
Wingersky, Marilyn S.
Publication Year:
1992
Report Number:
RR-92-22
Source:
ETS Research Report
Document Type:
Report
Page Count:
82
Subject/Key Words:
Computer Software, Estimation (Mathematics), Item Response Theory (IRT), LOGIST (Computer Program)., Maximum Likelihood Statistics, Statistical Analysis

Abstract

Stocking (1989) discovered a problem with the LOGIST estimation procedure. The automatic procedure in this program produced different item parameter estimates when the true item discriminations were used as starting values for the iteration procedure than when the default starting value of one for the item discriminations was used. When a straight run to convergence was performed, the different initial starting values converged to the same item parameter estimates. This study investigated several methods for improving the automatic procedure. When these methods failed to give the improvement necessary, several methods were tried to obtain the same results as the run to convergence in a shorter amount of computer time. A method was devised that takes much less time and gives nearly as good, and in some cases better, estimates as estimates obtained from running to convergence. This method involves adding a step to get better initial item parameter estimates for the automatic procedure. In this step, the abilities are grouped very coarsely and the item parameters and the grouped abilities are estimated iteratively, alternating back and forth between items and abilities until the maximum difference between the estimated item characteristic curves is less than some criterion. The new procedure gives item parameter estimates that are closer to the true values than the current 4- step method does. However, there is a definite nonlinear relationship between the estimated item parameters for the two methods after the parameters have been linearly transformed to the same scale. Consequently, in an ongoing series of calibrations, switching to this procedure will produce a discontinuity in the parameter estimates in the same manner as would be caused by switching from LOGIST to BILOG. The effects of putting a beta prior on c were also investigated. The results were not conclusive. An option to put a beta prior on c was added to the program. This new method has been incorporated into a new version of LOGIST called LOGIST7. (82pp.)

Read More