Predicting differential item functioning in cross-lingual testing: The case of a high stakes test in the Kyrgyz Republic.

摘要

Cross-lingual tests are assessment instruments created in one language and adapted for use with another language group. Practitioners and researchers use cross-lingual tests for various descriptive, analytical and selection purposes both in comparative studies across nations and within countries marked by linguistic diversity (Hambleton, 2005). Due to cultural, contextual, psychological and linguistic differences between diverse populations, adapting test items for use across groups is a challenging endeavor. The validity of inferences based on cross-lingual tests can only be assured if the content, meaning, and difficulty of test items are similar in the different language versions of the test items (Ercikan, 2002).;Of paramount importance in the test adaptation process is the proven ability of test developers to adapt test items across groups in meaningful ways. One way investigators seek to understand the level of item equivalence on a cross-lingual assessment is to analyze items for differential item functioning, or DIF. DIF is present when examinees from different language groups do not have the same probability of responding correctly to a given item, after controlling for examinee ability (Camilli & Shephard, 1994). In order to detect and minimize DIF, test developers employ both statistical methods and substantive (judgmental) reviews of cross-lingual items. In the Kyrgyz Republic, item developers rely on substantive review of items by bi-lingual professionals. In situations where statistical DIF detection methods are not typically utilized, the accuracy of such professionals in discerning differences in content, meaning and difficulty between items is especially important.;In this study, the accuracy of bi-linguals' predictions about whether differences between Kyrgyz and Russian language test items would lead to DIF was evaluated. The items came from a cross-lingual university scholarship test in the Kyrgyz Republic. Evaluators' predictions were compared to a statistical test of "no difference" in response patterns by group using the logistic regression (LR) DIF detection method (Swaminathan & Rogers, 1990). A small number of test items were estimated to have "practical statistical DIF." There was a modest, positive correlation between evaluators' predictions and statistical DIF levels. However, with the exception of one item type, sentence completion, evaluators were unable to predict which language group was favored by differences on a consistent basis. Plausible explanations for this finding as well as ways to improve the accuracy of substantive review are offered.;Data was also collected to determine the primary sources of DIF in order to inform the test development and adaptation process in the republic. Most of the causes of DIF were attributed to highly contextual (within item) sources of difference related to overt adaptation problems. However, inherent language differences were also noted: Syntax issues with the sentence completion items made the adaptation of this item type from Russian into Kyrgyz problematic. Statistical and substantive data indicated that the reading comprehension items were less problematic to adapt than analogy and sentence completion items. I analyze these findings and interpret their implications to key stakeholders, provide recommendations for how to improve the process of adapting items from Russian into Kyrgyz and highlight cautions to interpreting the data collected in this study.