APPLYING PARTICLE SWARM OPTIMIZATION TO ESTIMATE PSYCHOMETRIC MODELS WITH CATEGORICAL RESPONSES

Abstract

Current psychometrics tend to model response data hypothesized to arise from multiple attributes. As a result, the estimation complexity has been greatly increased so that traditional approaches such as the expected-maximization algorithm would fail to produce accurate results. To improve the estimation quality, high-dimensional models are estimated via a global optimization approach- particle swarm optimization (PSO), which is an efficient stochastic method of handling the complexity difficulties. The PSO has been widely used in machine learning fields but remains less-known in the psychometrics community. Details on the integration of the proposed approach to current psychometric model estimation practices are provided. The algorithm tuning process and the accuracy of the proposed approach are demonstrated with simulations. As an illustration, the proposed approach is applied to log-linear cognitive diagnosis models and multi-dimensional item response theory models. These two model families are fairly popular yet challenging frameworks used in assessment and evaluation research to explain how participants respond to item level stimuli. The aim of this dissertation is to fill the gap between the field of psychometric modeling and machine learning estimation techniques.