Factors affecting efficiency of interrater reliability estimates from planned missing data designs

Estimating interrater reliability (IRR) requires each of multiple subjects to be observed by multiple raters. Recruiting subjects and raters may be problematic: There may be few available, it may be costly to compensate subjects or to train raters, and participating in an observational study may be burdensome. Planned missing observational designs, in which raters vary across subjects, may accommodate these problems, but little guidance is available about how to optimize a planned missing observational design when estimating IRR. In this study, we used Monte Carlo simulations to optimize an observational design to estimate intraclass correlation coefficients (ICCs), which are very flexible IRR estimators that allow missing observations. We concluded that, given a fixed total number of ratings, the point and credibility estimates of ICCs can be optimized by means of (approximately) continuous measurement scales and assigning small teams of raters to subgroups of subjects. Also, less substantial differences between raters resulted in more efficient IRR estimates. These results highlight the importance of well-designed observational designs and proper training on an observational protocol to avoid substantial differences between raters.