|
Abstract:
|
Comparing providers based on the ratio of observed to expected number of events such as mortality is the backbone of several report cards. Typically, 95% intervals are presented, and providers are identified as outliers if their interval excludes the null value of 1.0. While multiple methods of calculating such intervals are available, most rely on asymptotic approximations or computationally intensive bootstrapping, and many calculate variance under the false assumption that the true ratio is 1.0. In this talk, two new methods of calculating frequentist confidence intervals are proposed which do not require asymptotic approximations or bootstrapping and which retain at least their nominal coverage probability regardless of the true value of the ratio being estimated. Methods are compared and illustrated using real data from a national cardiac surgery registry.
|
