針對FDT論文之意見

已收到許清芳老師及呂文賢老師對FDT練習效應論文之回應，其中許老師問了一個比較特別的問題，"You need to be careful with the CIs you are reporting as they are either multivariate in nature or repeated measures. It seems to me that your CIs do not adjust for the correlations."。


我有再確認許老師此問題的意思，他的回應是"Since you took four repeated measurements from the same sample of subjects, these measurements are correlated. Confidence intervals for correlated measurements are not the same as those for independent  measurements just like paired t-test is not the same as two independent-sample t-test."。

與李晏討論此意見。
在古典測量理論中，信度的假設為每個個體的真實分數應該是穩定且一致，因此個體本身之真實分數的相關為1。在觀察到的測量分數中包括隨機誤差 (observed test score = true score + errors of measurement)，而導致觀察到的重複測量分數之間的相關小於1。
就以此研究，我們所估計的CI of RCIp是從信度係數導算出來的，此CI為個體本身的重覆測驗分數之間有相關的假設之下。的確，在此之CI並未調整相關，此CI的用途為呈現個體本身的分數變化是否為真實的進步或退步(已考量評估誤差和練習效應)，但此CI並不適於用來表示不同個體間的分數差異。