Keywords: subgroups analysis; companion diagnostic test; cut-points; multiple comparisons
Many modern medicines are targeted therapies, targeting specific pathways. A biomarker (such as baseline HbA1c for diabetic patients) that is informative on how sick a patient is might allow such medicine to be personalized if it is sufficiently predictive of the effect patients will receive. If a candidate biomarker is continuously valued, it is typically dichotomized to classify patients into target (marker-positive) and non-target (marker-negative) subgroups.
For each candidate biomarker cut-point value, to answer the question of which patients to target, we provide simultaneous confidence intervals for the effect of the drug on the marker-positive patients, on the marker-negative patients, and on the entire patients. To counter the situation that a subgroup has a wide confidence interval due to small sample size (which makes its interpretation challenging), we can allocate larger alpha to that subgroup to provide equal width simultaneous confidence intervals.
Often time, the cut-point value is uncertain and scientists search through many cut-point values for the most meaningful one. The simultaneous confidence intervals we provide are multiplicity adjusted for searching through all possible cut-point values. Our method is surprisingly neat, with a convenient app implementation.