© RSNA, 2006
Performance Benchmarks for Screening Mammography
Appendix E1
It may be useful to understand and calculate the confidence intervals (CIs) for numbers generated in a mammography practice. These CIs are a way to quantify how much random variation is present in the numbers derived from a mammography audit, and to recognize which differences are probably valid. The numbers generated in the audit process should be thought of as representative samples of a larger pool of "underlying data." The sampling process therefore has a potential for error in its estimates of any measurement. It is useful to think of the "true" value to be within the 90% CI, 90% of the time. Similarly, each separate audit of the data has its own potential for sampling error and may differ only because of these random effects. The method given is an approximation for larger numbers only and may not be appropriate for data involving small numbers. Unfortunately, the statistically proper method for data involving small numbers is too complex to be easily used.
The 90% CI for a proportion P, where P = a/n, is P ± 1.645 * √[P * (1 − P)/n].
Sample Calculations
The recall rate is 10% for 910 screening examinations. The proportion P is 91/910, or 0.1, and the 90% CI calculation is 0.1 ± 1.645 * √(0.1 * (1 - 0.1)/910), or 0.1 ± 0.016 (90% CI: 0.084, 0.116), or 84 to 116 recalls per 1000 screening examinations.
The cancer detection rate is nine cancers per 3,000 screening examinations, or approximately three per 1000 screening examinations. The proportion P is 0.003 and the 90% CI calculation is 0.003 ± 1.645 * √ [0.003 * (1 - 0.003)/1000], or 0.003 ± 0.00164 (90% CI: 0.00136, 0.00464), or 1.36 to 4.64 cancers per 1000 screening examinations.
