intraclass correlation coefficient

說明

Intraclass correlation coefficient (ICC)是一個被廣泛使用的信度指標。它可以被使用在再測信度、評分者間信度以及評分者內信度。它可以反應多個測量(評分)結果的相關(correlation)或同意(agreement)程度。

ICC包含多個不同種類(form),Shrout 與 Fleiss(1979)提出6種,而McGraw與Wong(1996)提出10種。根據Koo與Li(2016)的整理,這些不同種類詳列如#ICC 種類表。仔細看可以發現,某些不同類的ICC計算的公式是一樣的。以McGraw與Wong(1996)的架構為例,常用的 Two-way random effects, absolute agreement, single rater/measurement 和 Two-way mixed effects, absolute agreement, single rater/measurement 計算結果是相同的。

研究者應根據研究目的選擇適合的種類。Koo與Li(2016)依照McGraw與Wong(1996)的分類概念提出#ICC選擇流程圖,按照以下4個問題的答案進行選擇:

  1. 是否所有的被評分樣本都是被同一組的評分者所評分?Do we have the same set of raters for all subjects?
  2. 評分者樣本是從某個大群體中隨機選取的,還是特定的評分者樣本?Do we have a sample of raters randomly selected from a larger population or a specific sample of raters?
  3. 我們關注的是各個評分者的信度,還是多個評分者平均值的信度?Are we interested in the reliability of single rater or the mean value of multiple raters?
  4. 我們關心的是一致性還是同意度?Do we concern about consistency or agreement?

結果解釋

依據Koo與Li(2016)

ICC Value Reliability
< 0.5 Poor reliability
0.5 - 0.75 Moderate reliability
0.75 - 0.9 Good reliability
> 0.9 Excellent reliability

R程式碼

psych package

psych::ICC計算是按照Shrout 與 Fleiss(1979)的架構,結果會得到6種ICC的結果。

library(psych)
ICC(x, missing=TRUE, alpha=.05, lmer=TRUE, 
    check.keys=FALSE)

irr package

irr::icc的計算則是按照McGraw與Wong(1996)的架構,需要設定modeltypeunit

library(irr)
icc(ratings, model = c("oneway", "twoway"), 
    type = c("consistency", "agreement"), 
    unit = c("single", "average"), r0 = 0,
    conf.level = 0.95)

附錄

ICC 種類表

McGraw and Wong (1996) Convention Shrout and Fleiss (1979) Convention Formulas for Calculating ICC
One-way random effects, absolute agreement, single rater/measurement ICC (1,1) MSRMSWMSR+(k+1)MSW
Two-way random effects, consistency, single rater/measurement - MSRMSEMSR+(k1)MSE
Two-way random effects, absolute agreement, single rater/measurement ICC (2,1) MSRMSEMSR+(k1)MSE+kn(MSCMSE)
Two-way mixed effects, consistency, single rater/measurement ICC (3,1) MSRMSEMSR+(k1)MSE
Two-way mixed effects, absolute agreement, single rater/measurement - MSRMSEMSR+(k1)MSE+kn(MSCMSE)
One-way random effects, absolute agreement, multiple raters/measurements ICC(1,k) MSRMSWMSR
Two-way random effects, consistency, multiple raters/measurements - MSRMSEMSR
Two-way random effects, absolute agreement, multiple raters/measurements ICC (2,k) MSRMSEMSR+MSCMSEn
Two-way mixed effects, consistency, multiple raters/measurements ICC (3,k) MSRMSEMSR
Two-way mixed effects, absolute agreement, multiple raters/measurements - MSRMSEMSR+MSCMSEn

說明:
MSR: mean square for rows; MSW: mean square for residual sources of variance; MSE: mean square for error; MSC: mean square for columns; n: number of subjects; k: number of raters/measurements

ICC選擇流程

Pasted image 20240528103949.png|來源:Koo & Li (2016)