When coders are reliable: the application of three measures to assess inter-rater reliability/agreement with doctor-patient communication data coded with the VR-CoDES

Abstract

Objective: To investigate whether different measures of inter-rater reliability will compute similar estimates with nominal data commonly encountered in communication studies. To make recommendations how reliability should be computed and described for communication coding instruments.

Methods: The raw data from an inter-rater study with three coders were analysed with; Cohen's κ, sensitivity and specificity measures, Fleiss's multirater κj, and an intraclass correlation coefficient (ICC).

Results: Minor differences were found between Cohen's κ and an ICC model across paired data (largest margin=0.01). There were negligible differences between the multirater estimates e.g. κj (0.52) and ICC (0.53). Sensitivity analyses were in general agreement with the multirater estimates.

Conclusion: It is more practical to analyse nominal data with >2 raters with an appropriate model ICC for inter-rater studies, and little difference exists between Cohen's κ or an ICC.

Practice implication: Alternatives to Cohen's κ are readily available, but researchers need to be aware of the different ICC definitions. An ICC model should be fully described in reports. Investigators are encouraged to supply confidence limits with inter-rater data, and to revisit guidance regarding the relative strengths of agreement of reliability coefficients.