Fiche publication
Date publication
septembre 2024
Journal
Statistics in medicine
Auteurs
Membres identifiés du Cancéropôle Est :
Dr DEVILLIERS Hervé
Tous les auteurs :
Bourredjem A, Cardot H, Devilliers H
Lien Pubmed
Résumé
The agreement intra-class correlation coefficient (ICCa) is a suitable statistical index for inter-rater reliability studies. With balanced Gaussian data, we prove the explicit form of ICCa asymptotic normality (ASN), valid both with analysis of variance (ANOVA), maximum likelihood (ML), or restricted ML (REML) estimates. An asymptotic confidence interval is then derived and its performances are examined by simulation compared to the most commonly used methods, under small, moderate and large sample size designs. Then, we deduce sample size calculation formulas, for the number of subjects and observers needed, to achieve a desired confidence interval width or an acceptable ICCa value test power and give concrete examples of their use. Finally, we propose a likelihood ratio test (LRT) to compare two ICCa's from two distinct subpopulations of patients (or raters) and study by simulation its first order risk and power properties. These methods are illustrated using data from two inter-rater reliability studies, one in physiotherapy with 42 patients and 10 raters and the second in neonatology with 80 subjects and 14 raters. In conclusion, we made recommendations to employ the proposed confidence interval for medium to large samples combined with the quantification of the minimal required sample size at the planning step, or the posterior-power at the analysis step, using simple dedicated formulas. Furthermore, with sufficient sizes, the proposed LRT seems suitable to compare inter-rater reliability between two patient subpopulations. Used wisely, this proposed methods toolbox can remedy common current issues in inter-rater reliability studies.
Mots clés
agreement intra‐class correlation coefficient, asymptotic normality (ASN), balanced two‐way crossed random effect model, confidence interval (CI), likelihood ratio test (LRT), sample size
Référence
Stat Med. 2024 09 16;:
