A random set approach to confidence regions with applications to the effective dose with combinations of agents

The effective dose (ED) is the pharmaceutical dosage required to produce a therapeutic response in a fixed proportion of the patients. When only one drug is considered, the problem is a univariate one and has been well-studied. However, in the multidimensional setting, that is, in the presence of combinations of agents, estimation of the ED becomes more difficult. This study is focused on the plug-in logistic regression estimator of the multidimensional ED. We discuss consistency of such estimators and focus on the problem of simultaneous confidence regions. We develop a bootstrap algorithm to estimate confidence regions for the multidimensional ED. Through simulation, we show that the proposed method gives 95% confidence regions, which have better empirical coverage than the previous method for moderate to large sample sizes. The novel approach is illustrated on a cytotoxicity study on the effect of two toxins in the leukemia cell line HL-60 and a decompression sickness study of the effects of the duration and depth of the dive.