Statistics Glossary → Survival confidence interval transformations (log-log)
Survival confidence interval transformations (log-log)concept
Scale transformations used when building confidence intervals for the Kaplan-Meier survival probability S(t), so that the interval stays within 0-1 and achieves better coverage.
The simplest ('plain') interval is S(t) ± z·SE; but when the survival probability is close to 0 or 1, it can extend beyond these limits and must be truncated, and its coverage is poor in small samples. With the log transformation the interval is built on the log S(t) scale and transformed back; the lower limit cannot fall below zero, but the upper limit can exceed 1. With the log-log transformation the interval is built on the log(-log S(t)) scale, and back-transforming gives an asymmetric interval of the form S(t)^exp(±z·σ) that always stays between 0 and 1.
Simulation studies have shown that the log-log transformation usually comes closest to nominal coverage, and many programs use it by default. The log-log scale is undefined when S(t) = 1 or 0, so no interval can be computed at those points. Because the confidence interval for the median survival time is usually read from where this pointwise band crosses 0.5, the chosen transformation also affects the median's interval. The same idea is related to the complementary log-log link in ordinal regression and to transformations in parametric survival models.
An everyday example
After a treatment, the 5-year survival probability of 30 patients is estimated at 92%. The plain formula can push the upper limit to an impossible value such as 102%, which then has to be cut back to 100%. With the log-log transformation the interval runs from about 74% to 98%: it is asymmetric, because a probability close to 100% has far less room to move upwards than downwards.
In MerQur
It is chosen with the 'CI transformation (log-log / log / plain)' parameter in Kaplan-Meier Survival Analysis; Log-minus-log is also among the plot types. In Competing Risks the 'CIF confidence intervals (log-log)' extra output gives the cumulative incidence intervals with the log(-log) transformation.
Analyses where this concept appears
Kaplan-Meier Survival AnalysisCompeting Risks
Related concepts
Kaplan-Meier estimator (survival curve)Confidence level and confidence interval