Estimation and inference for weighted agreement coefficients (Cohen, Fleiss, Conger, Brennan-Prediger) with arbitrary numbers of raters and arbitrary pairwise loss functions, plus coefficient alpha for scored categorical and continuous item batteries. Supports incomplete ratings under MCAR and MAR. Wraps the standalone C++17 misskappa library.
Each estimator is selected with a single estimator= argument:
-
kappa()— raw rating matrices."ipw"(inverse-probability-weighted) and"cat_fiml"(saturated-multinomial FIML) for categorical ratings;"pairwise"(pairwise-available moment) and"nt_fiml"(robust normal-theory FIML) for the quadratically weighted, scored coefficient. -
alpha()— item batteries."pairwise","cat_fiml", or"nt_fiml". -
kappa_counts()— counts-format input (subjects × categories)."fleiss_cuzick"or"cat_fiml".
The MCAR estimators ("pairwise", "ipw") are distribution-free; the FIML estimators ("cat_fiml", "nt_fiml") are valid under ignorable missingness. Coefficients carry a covariance matrix, so coef(), vcov(), and confint() (with an optional Fisher transform) give estimates and Wald confidence intervals. To test whether a coefficient is equal across fits — two groups, two estimators, two timepoints, or several rater pairs — use kappa_test() / alpha_test() (with paired = TRUE for same-subject fits); both return a standard htest.
See the function reference for estimators and inference helpers, and the mathematical guides for the loss-matrix formulation, missing-data estimators, and validation strategy. The underlying C++ library has its own C++ API reference.