A problem that arises in bilingual testing is that some persons are partial speakers of two languages, but not totally competent in either. In these cases, the respondent might answer in English sometimes and Spanish sometimes. We ran into this situation in preschool readiness testing with Hispanic children in Florida.
The Woodcock/Muñoz-Sandoval approach (RMT 7:4 p. 329) is designed for first language speakers of different languages. It equates parallel tests, then transforms measures on the non-English language form to their English equivalent. But a simple translation of a test into another language ignores cultural bias - for example, a cross has specific cultural significance for Hispanic preschoolers.
We took a different approach with the Rasch model. We pooled responses from Spanish and English forms and calibrated the items together. Then we used the person "between" (Hispanic students vs. Anglo students) fit statistic provided by IPARM (Smith 1991) to compare Hispanic and English cultural bias on each item, without regard to the language that the student used when responding. Mixed language responses were allowed. Credit was given when the answer was correct in either language.
The person between fit statistic proved more sensitive to bias than the usual fit statistics. We were able to identify items for inclusion, exclusion or editing in a bilingual test, so avoiding the need for equated test versions and equivalent measures. Without this approach, the more or less unpredictable use of two languages simultaneously could render some preschoolers unmeasurable.
Steve Lang 1994 RMT 8:1 p. 343
University of South Florida at St. Petersburg
Bilingual testing. Lang S. … Rasch Measurement Transactions, 1994, 8:1 p.343
The URL of this page is www.rasch.org/rmt/rmt81k.htm
Website: www.rasch.org/rmt/contents.htm