A one-step extrapolation to estimate the performance of a prediction model was proposed by researchers at National Taiwan University (NTU). The study was published online in April in BMJ Open.
Conducted by Dr. Ling-YiWang and her advisor Dr. Wen-Chung Lee, professor of Institute of Epidemiology and Preventive Medicine in the College of Public Health, this study presents a new method to evaluate risk prediction.
Microarray related studies often involve a very large number of genes, but very small sample size (the large-p-small-n problem). To estimate the prediction performance of a gene signature derived from a small study, researchers often need to conduct cross-validating or bootstrapping. However, this makes the effective sample size even smaller. “This study develops a one-step extrapolation method to estimate the prediction performance of the model trained by all the samples, and it does strike a good balance between bias and variance and has small root mean squared error.” said Dr. Lee.
As shown in this study, the proposed method can be applied to normal/non-normal data, and linear/non-linear models. The method takes advantage of a certain learning curve for extrapolation.However, the exact learning curve used is not crucial, because only a mere one step ahead is to be extrapolated.