The Place of the Bifactor Model in Confirmatory Factor Analysis Investigations Into Construct Dimensionality in Language Testing

摘要

For practical and theoretical purposes, language tests commonly aim to measure one overarching trait, general language ability, while simultaneously measuring multiple sub-traits (e.g., reading, grammar, etc). This tension between measuring uni- and multi-dimensional constructs concurrently can generate vociferous debate about the precise nature of the construct(s) being measured. In language testing, this tension is often addressed through the use of a higher-order factor model wherein multidimensional traits representing subskills load on a general ability latent trait. However, an alternative modelling framework that is currently uncommon in language testing, but gaining tract in other disciplines, is the bifactor model. The bifactor model hypothesizes a general factor, onto which all items load, and a series of orthogonal (uncorrelated) skill-specific factors. The model is particularly valuable for evaluating the empirical plausibility of subscales and the practical impact of dimensionality assumptions on test scores. This paper compares the higher order with the bifactor model in terms of theoretical implications and practical considerations, framed for the language testing audience. The models are illustrated using primary data from the British Council’s Aptis English test. The paper is intended to spearhead the uptake of the bifactor model within the cadre of measurement models used in language testing.