Rho is used to test the generalization of inter rater reliability (IRR) statistics, in this case Cohen's Kappa.
Rho is a Monte Carlo rejective method of interrater reliability statistics, implemented here for Cohen's Kappa. Rho constructs a collection of data sets in which kappa is below a specified threshold, and computes the empirical distribution on kappa based on the specified sampling procedure. Rho returns the percent of the empirical distribution greater than or equal to an observed kappa. As a result, Rho quantifies the type 1 error in generalizing from an observed test set to a true value of agreement between two raters.
Rho starts with an observed kappa value, calculated on a subset of a
codeSet, known as an observed
testSet, and a kappa threshold which indicates what is considered significant agreement between raters.
It then generates a collection of fully-coded, simulated
codeSets (ScS), further described in
createSimulatedCodeSet, all of which have a kappa value below the kappa
threshold and similar properties as the original
Then, kappa is calculated on a
testSet sampled from each of the ScSs in the
collection to create a null hypothesis distribution. These
testSets mirror the observed
testSet in their size and sampling method. How these
testSets are sampled is futher described in
The null hypothesis is that the observed
testSet, was sampled from a data set, which, if both raters were to code in its entirety, would result in a level of agreement below the kappa threshold.
For example, using an alpha level of 0.05, if the observed kappa is greater than 95 percent of the kappas in the null hypothesis distribution, the null hypothesis is rejected. Then one can conclude that the two raters would have acceptable agreement had they coded the entire data set.