Comparing regression versus correction formula predictions of passive recall test scores from yes-no test results

Abstract

A novel form of scoring formula for self-report yesno vocabulary tests was presented in Stubbe and Stewart, based on multiple regression models that use both real-word and pseudoword reports to predict subsequent scores on a test of passive recall knowledge (as measured by L2 to L1 translations). The aim of the present study is to determine how well passive recall test scores can be predicted from yes-no test results adjusted using two methods: (1) regression formulas versus (2) the four established correction for guessing formulas outlined in Huibregtse, Admiraal, and Meara: h-f, cfg, Dm and lsdt. After taking a yes-no test followed by a passive recall test of the same 96 real-words, the sample of Japanese university students (N 431) was split into two groups of comparable proficiency (A and B). The original Stubbe and Stewart regression formula was compared to the four correction formulas by analyzing their application with the Group A. Despite having a lower correlation with passive recall test scores than one of the correction formulas, the predicted scores produced were significantly closer. A new regression formula was then created using the Group A’s test results and this was used to predict translation test scores on Group B, along with the four correction formulas. As the resulting predictions were superior to those of any of the correction formulas, and not significantly different from the actual passive recall test scores, plus the correlation with these translation test scores was the highest (0.845), it was concluded that regression formulas produced the best predictions.