Inter Rater Reliability is one of those statistics I seem to need just seldom enough that I forget all the details and have to look it up every time.
Luckily, there are a few really great web sites by experts that explain it (and related concepts) really well, in language that is accessible to non-statisticians.
So rather than reinvent the wheel and write about it, I’m going to refer you to these really great sites:
- Computing Intraclass Correlations (ICC) as Estimates of Interrater Reliability in SPSS by Richard Landers
- Statistical Methods for Rater and Diagnostic Agreement by John Uebersax
If you know of any others, please share in the comments. I’ll be happy to add to the list.
Paul van Haard says
Dear Mrs. Grace-Martin,
Please advocate Krippendorff’s alpha and the 95%CI for the population, which can be easily calculated in R (2019 CRAN) via scripts available.
Solid arguments in favour for this broadly applicable IRR test can be found in the papers from Krippendorff.
BioStatistician/ Clinical Biochemist
daniel klein says
Before advocating Krippendorff’s alpha, note that solid arguments can be made against it (e.g., Zhao et al. 2018). Actually, all arguments for using Krippendorff’s alpha that I have come across are made by its author. I am not aware of a single paper by another author that empirically shows how alpha is superior to other inter-rater-reliability measures.
Gwet (2014) shows that Krippendorff’s alpha is mathematically almost identical to Fleiss version of the kappa coefficient (especially when there are no missing ratings). Therefore, it shares some of the shortcomings of kappa: most notably, Krippendorff’s alpha (re-)produces the so-called high agreement low kappa paradox (cf. Feinstein and Cicchetti 1990).
Concerning the confidence interval estimation proposed by Krippendorff has also been criticized (Zapf et al. 2016). The authors propose using the standard bootstrap method while Gwet (2014) suggests yet another variance estimator.
Moreover, Gwet (2014) also shows how various other coefficients can be extended to multiple raters, any level of measurement, and handling missing values just like Krippendorff’s alpha. Thus, contrary to common claims, the latter is all but unique in these respects.
Gwet, K. L. 2014. Handbook of Inter-Rater Reliability: The Definitive Guide to Measuring
the Extent of Agreement Among Raters. 4th ed. Gaithersburg, MD: Advanced
Feinstein, A. R., and D. V. Cicchetti. 1990. High agreement but low kappa: I. The
problems of two paradoxes. Journal of Clinical Epidemiology 43: 543–549.
Zapf, A., S. Castell, L. Morawietz, and A. Karch. 2016. Measuring inter-rater reliability
for nominal data—Which coefficients and confidence intervals are appropriate? BMC
Medical Research Methodology 16: 93
Zhao, X., Feng, G., Liu, J., and Ke Deng. 2018. We agreed to measure agreement – Redefining reliability de-justifies Krippendorff’s
alpha. China Media Research, 14: 1-15. (Authors’ final version:
Sean Murphy says
Conceptual, not computation focused overview…
Good overview of reliability in general and how IRR fits into that picture…
Computational resources and explanations using Excel.
Not free, but the book here is worth getting if you need to think through reliability in greater depth. There are a few free R functions available on the site for computing various IRR statistics and importantly generalizations that handle missing data. The R functions work well, I have not used the software. The book is quite comprehensive and well organized with worked examples in online spreadsheets.
Also a couple of classic papers…
Shrout and Fleiss (1979). Intraclass correlations: Uses in assessing rater reliability.
McGraw and Wong (1996). Forming inferences about some intraclass correlation coefficients. (Unfortunately this is behind a pay wall, but a good article if you have library access.)
Joshua Cruz says
Thanks for the great resources. I found another one that was really helpful (also by a Karen)