Estimation Methods for Cross-Validation Prediction Accuracy: A Comparison of Proportional Bias

Abstract

Using empirical data, the performance of the predictive effectiveness of four algorithms and a bootstrap method for cross-validation of a multiple regression equation were examined. Results indicated that the Browne algorithm was the most accurate in 8 of the 9 data situations. The Rozeboom algorithm, in a majority of conditions, had the second least amount of proportional bias. The Nicholson and Lord and Stein-Darlington formulas demonstrated a consistent pattern of low relative accuracy in many situations, with the most amount of proportional bias in 4 of 9 and in 6 of 9 data sets, respectively. The bootstrap method showed no discernable pattern of relative accuracy with results ranging from the most accurate in a situation to the least accurate in three different data situations.