Exact 95% confidence intervals for differences in binomial proportions

A method is described that allows exact 95% confidence intervals to be computed for differences in binomial proportions. The method itself is not iterative, but it uses the iteratively computed 95% confidence intervals for the individual binomial proportions. The method converges with the usual normal approximation method, as denominators become large. It correlates very closely with the Yates-corrected chi-square test. The method works regardless of the definition of success or failure in the proportions and regardless of the order of the proportions.