• The Korean Journal of Applied Statistics
• /
• v.20 no.1
• /
• pp.117-132
• /
• 2007

In a $2\times2$ table showing binary agreement between two raters, it is known that Cohen's $\kappa$, a chance-corrected measure of agreement, has two paradoxes. $\kappa$ is substantially sensitive to raters' classification probabilities(marginal probabilities) and does not satisfy conditions as a chance-corrected measure of agreement. However, $\kappa$ and other established measures have a reasonable and similar value when each marginal distribution is close to 0.5. The objectives of this paper are to present a new measure of agreement, H, which resolves paradoxes of $\kappa$ by adjusting unbalanced marginal distributions and to compare the proposed measure with established measures through some examples.